docs: the cluster decode kernels run on sm_90+ (RTX 50), not RTX 40

The ring rule: 1024 slots for the Q2_0 pack's fused experts (P3), 512 for native packs with fused layers.
Quality (teacher-forced vs the FP16 path, Q2_0): fused as close as MMQ at 8K, closer at 32K (phaseA-tests s18-tf).

.dockerignore

@@ -0,0 +1,27 @@
+# Keep the image to the source only: everything generated is re-created on
+# first start (inside the container or on the /data volume).
+.git
+
+# local Python environment (rebuilt in the image)
+.venv/
+__pycache__/
+*.pyc
+.pytest_cache/
+
+# setup.py outputs and downloads (engine, packs, models, logs)
+/engine/
+/build/
+/build*/
+/dist/
+/packs/
+/mtp/
+/models/
+/Strata-data/
+third_party/llama.cpp/
+third_party/*.zip
+*.done
+*.log
+strata-*.json
+strata-*.txt
+run-*.sh
+run-*.bat

.gitattributes

@@ -4,3 +4,6 @@ bench/prompts/** -text
 # launchers keep their platform line endings wherever the repository is cloned
 *.sh text eol=lf
 *.bat text eol=crlf
+
+# copies of real packs' native_experts.txt, parsed byte for byte by tests/core/expert_layout_test.cpp
+tests/data/native_experts/** -text

.github/FUNDING.yml

@@ -0,0 +1 @@
+buy_me_a_coffee: strataengine

.gitignore

@@ -18,6 +18,8 @@ __pycache__/
 
 # plan v0.3 P1 (repository hygiene, review E1): profiler captures and large diagnostic dumps stay out of git.
 *.nsys-rep
+
+# Pi harness metadata (local session state, not project source)
 *.sqlite
 *.npz
 *.occ1
@@ -26,7 +28,7 @@ td.log
 # created by setup.py on the user's machine
 /.venv/
 /build/
-/third_party/llama.cpp/
+/third_party/llama.cpp
 /models/
 /packs/
 /mtp/
@@ -35,6 +37,7 @@ td.log
 /strata-*.json
 /strata-*.log
 /build-strata.bat
+/.strata-mcp/
 
 # image encoder (setup.py --vision): build output and the generated build script
 build-vision*/
@@ -46,3 +49,5 @@ build-vision.bat
 /dist/
 *.done
 /third_party/*.zip
+
+/.rocm-win/

AGENTS.md

@@ -0,0 +1,23 @@
+# AGENTS.md
+
+Strata runs the Qwen3.8-Flash-Next mixture-of-experts model (and its Coder, Swift 1.5 and Unsloth variants) on a
+normal PC: one NVIDIA or AMD graphics card plus system RAM, on Windows or Linux. It has a C++/CUDA/HIP engine
+(`src/`, `include/`), a Python server with an OpenAI- and Anthropic-compatible API and a web app (`serve/`), and a
+one-click installer (`setup.py`, started by `START-HERE.bat` / `setup.sh`).
+
+## Installing Strata for a user
+
+Follow **[docs/AI_SETUP.md](docs/AI_SETUP.md)**: check the PC, pick the model by RAM, run setup non-interactively,
+start and verify the server, and connect the user's apps. Never expose the server beyond `127.0.0.1` without
+`--api-key`. As an alternative to shell commands, Strata's MCP server ([docs/MCP_SERVER.md](docs/MCP_SERVER.md))
+offers the same steps as tools.
+
+## Working on the code
+
+- How the engine works, every measured number, the API and all settings: [docs/DETAILS.md](docs/DETAILS.md) and
+  the [paper](docs/paper/Strata-Paper.pdf).
+- AMD (HIP) build and validation: [docs/AMD_HIP.md](docs/AMD_HIP.md); multi-GPU: [docs/MULTI_GPU.md](docs/MULTI_GPU.md).
+- Setup's own tests run without a GPU or downloads: `python tools/test_setup_<name>.py` (for example
+  `tools/test_setup_amd.py`, `tools/test_setup_choices.py`).
+- Keep the docs' style: plain words, measured numbers with what they were measured on, no claims without a
+  measurement.

CMakeLists.txt

@@ -4,11 +4,18 @@
 # `strata-pack` are added when their sources exist - a target with no sources would make the build graph
 # claim something the tree does not have, which is worse than an absent target.
 #
-# CUDA targets arrive in Phase 2 and this build REJECTS any architecture other than 120: the engine is
-# written against sm_120 (RTX 5000 / Blackwell) and silently building for an older arch would produce a
-# binary that runs and is wrong about its own performance, which is the worst of both outcomes.
+# CUDA targets arrive in Phase 2.  The kernels need sm_80 or newer (tf32 mma, bf16 math) and the guard
+# below refuses anything older: silently building for a pre-Ampere arch would produce a binary that runs
+# and is wrong about its own performance, which is the worst of both outcomes.
 cmake_minimum_required(VERSION 3.24)
-project(strata VERSION 0.1.0 LANGUAGES CXX)
+project(strata VERSION 0.1.36 LANGUAGES CXX)
+add_compile_definitions(STRATA_VERSION="${PROJECT_VERSION}")
+# issue #31: release builds keep a PDB beside the exe (not shipped in it), so a minidump from a stalled engine can
+# be read back to the exact source line
+if(MSVC AND CMAKE_BUILD_TYPE STREQUAL "Release")
+  add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/Zi>)
+  add_link_options(/DEBUG /OPT:REF /OPT:ICF)
+endif()
 
 # The prebuilt release (setup.py downloads it, so users need no compiler): ggml's CPU backend for any AVX2 CPU
 # instead of this machine's (Strata's own AVX-512 kernels are per-file and chosen at run time either way), and
@@ -26,36 +33,88 @@ if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
 endif()
 
 option(STRATA_ENABLE_CUDA "Build the CUDA targets (Phase 2+)" OFF)
+option(STRATA_ENABLE_HIP "Build the HIP targets for wave32 AMD GPUs (gfx1100, gfx1201; gfx1030 unvalidated)" OFF)
+option(STRATA_PREFILL_MMQ "Build the optional GGML MMQ prefill path for HIP" OFF)
+# CUDA: MMQ prompt kernels for Q4_K / Q5_K / Q5_1 (Unsloth's UD-Q4_K_XL experts).  Off by default: the engine loads
+# every kernel when it starts (CUDA_MODULE_LOADING=EAGER), so these took ~5 MB of VRAM - 1-3 expert slots - from every
+# model, and the UD-Q4_K_XL prompt is SSD-bound with or without them (docs/UNSLOTH_Q4.md)
+option(STRATA_MMQ_KQUANTS "Build the CUDA MMQ prompt kernels for Q4_K / Q5_K / Q5_1 experts" OFF)
+option(STRATA_ORCA_Q4KS_MMQ "Build the GGML Q5_0 MMQ instance for OrcaRouter Q4_K_S" OFF)
+if(STRATA_PREFILL_MMQ AND NOT STRATA_ENABLE_HIP)
+  message(FATAL_ERROR "STRATA_PREFILL_MMQ requires STRATA_ENABLE_HIP=ON")
+endif()
+if(STRATA_ENABLE_CUDA AND STRATA_ENABLE_HIP)
+  message(FATAL_ERROR "STRATA_ENABLE_CUDA and STRATA_ENABLE_HIP are separate backends; enable only one")
+endif()
+if(STRATA_ENABLE_HIP)
+  include(cmake/hip_backend.cmake)
+endif()
 # The test suite and the micro benchmarks live in tests/ and bench/micro/, which the published source leaves out.
 set(_strata_tests_default OFF)
 if(EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/tests/CMakeLists.txt AND EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/bench/micro)
   set(_strata_tests_default ON)
 endif()
 option(STRATA_BUILD_TESTS "Build the test suite" ${_strata_tests_default})
+option(STRATA_BUILD_CONVERSATION_TESTS "Build standalone conversation cache tests" OFF)
 option(STRATA_WERROR "Treat compiler warnings as errors" OFF)
 
+if(STRATA_BUILD_CONVERSATION_TESTS)
+  enable_testing()
+  add_executable(conversation_cache_test src/core/conversation_cache_test.cpp)
+  target_include_directories(conversation_cache_test PRIVATE include)
+  add_test(NAME conversation_cache_test COMMAND conversation_cache_test)
+  add_executable(conversation_memory_test src/core/conversation_memory_test.cpp src/core/conversation_memory.cpp)
+  target_include_directories(conversation_memory_test PRIVATE include)
+  add_test(NAME conversation_memory_test COMMAND conversation_memory_test)
+endif()
+
 if(STRATA_ENABLE_CUDA)
   enable_language(CUDA)
   # `enable_language(CUDA)` gives the .cu compile rules but NOT the imported targets, so `CUDA::cudart` has to
   # be brought in explicitly for the host-side tests that call cudaMalloc themselves.
   find_package(CUDAToolkit REQUIRED)
+  option(STRATA_EXPERIMENTAL_SM60 "community build for GPUs below compute capability 7.5: Pascal sm_60, Volta sm_70 (#236)" OFF)
   if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
     set(CMAKE_CUDA_ARCHITECTURES 120 CACHE STRING "CUDA architectures" FORCE)
   endif()
   foreach(_a IN LISTS CMAKE_CUDA_ARCHITECTURES)
     string(REGEX REPLACE "-.*$" "" _base "${_a}")     # 120-real / 120-virtual are still 120
-    # Plan v0.3 P6: developed and measured on sm_120 (RTX 50); the kernels need sm_80 or newer (tf32 mma in
-    # the attention scorer, bf16 math), so RTX 30 (86) and RTX 40 (89) build too.  Older cards are refused.
+    # Turing port: developed and measured on sm_120 (RTX 50); the tf32 mma in the QSA scorer has a
+    # portable fp32-FMA fallback, so RTX 20 (75), 30 (86), RTX 40 (89) and RTX 50 (120) build too.
+    # Pre-sm_75 is refused.
     if(_base STREQUAL "native" OR _base STREQUAL "all" OR _base STREQUAL "all-major")
       continue()
     endif()
-    if(_base LESS 80)
+    # STRATA_EXPERIMENTAL_SM60: a Pascal (sm_60, GP100-class) or Volta (sm_70, V100 / Titan V, #236) build.
+    # These sit below every architecture the kernels were written for, so two intrinsics are replaced below
+    # their arch - `__dp4a` (6.1+) and `__nanosleep` (7.0+) - in include/strata/kernels/dp4a.hpp, which is
+    # bit-exact for the first and merely a shorter pause for the second (sm_70 has both).  Community-tested,
+    # not in the ready-made engine; RTX 20/30/40/50 cards must use the release engine instead.  It needs a CUDA
+    # 12.x toolkit: CUDA 13 dropped sm_60 and sm_70 (nvcc fails CMake's compiler test with "compute_70").
+    if(_base LESS 75 AND NOT STRATA_EXPERIMENTAL_SM60)
       message(FATAL_ERROR
-        "Strata needs an NVIDIA GPU of compute capability 8.0 or newer (RTX 30 / 40 / 50 series), but "
-        "CMAKE_CUDA_ARCHITECTURES is '${CMAKE_CUDA_ARCHITECTURES}'.")
+        "Strata needs an NVIDIA GPU of compute capability 7.5 or newer (RTX 20 / 30 / 40 / 50 series), but "
+        "CMAKE_CUDA_ARCHITECTURES is '${CMAKE_CUDA_ARCHITECTURES}'.  (A Pascal sm_60 or Volta sm_70 build is "
+        "-DSTRATA_EXPERIMENTAL_SM60=ON; upstream does not support it.)")
+    endif()
+    if(_base LESS 60)
+      message(FATAL_ERROR "CMAKE_CUDA_ARCHITECTURES '${CMAKE_CUDA_ARCHITECTURES}': even the experimental build "
+                          "needs compute capability 6.0 (sm_60) or newer")
+    endif()
+    # #220 / #224: RTX 50 engines built with CUDA 12.8 crashed in the prompt path on Linux (13.0 and newer did not)
+    if(_base GREATER_EQUAL 120 AND CMAKE_CUDA_COMPILER_VERSION VERSION_LESS 13.0)
+      message(WARNING "Building for sm_${_base} (RTX 50) with CUDA ${CMAKE_CUDA_COMPILER_VERSION}: engines built with "
+                      "CUDA 12.8 crashed on long prompts on these cards (issues #220, #224). Use CUDA 13.0 or newer.")
     endif()
   endforeach()
   message(STATUS "Strata: CUDA enabled, arch ${CMAKE_CUDA_ARCHITECTURES}")
+  set(_strata_gpu_runtime_target CUDA::cudart)
+  set(_strata_gpu_blas_target CUDA::cublas)
+  set(_strata_gpu_include_directories ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+elseif(STRATA_ENABLE_HIP)
+  set(_strata_gpu_runtime_target strata_hip_runtime)
+  set(_strata_gpu_blas_target roc::hipblas)
+  set(_strata_gpu_include_directories ${STRATA_HIP_COMPAT_INCLUDE_DIR})
 endif()
 
 # Warnings live on an INTERFACE target the project's own code links, NOT in add_compile_options: applied
@@ -71,11 +130,14 @@ target_compile_options(strata_warnings INTERFACE
   $<$<AND:$<COMPILE_LANGUAGE:CXX>,$<CXX_COMPILER_ID:MSVC>>:/utf-8>
   $<$<AND:$<COMPILE_LANGUAGE:CXX>,$<NOT:$<CXX_COMPILER_ID:MSVC>>>:-Wall>
   $<$<AND:$<COMPILE_LANGUAGE:CXX>,$<NOT:$<CXX_COMPILER_ID:MSVC>>>:-Wextra>
-  $<$<AND:$<COMPILE_LANGUAGE:CUDA>,$<CXX_COMPILER_ID:MSVC>>:-Xcompiler=/W4>)
+  $<$<AND:$<COMPILE_LANGUAGE:CUDA>,$<CXX_COMPILER_ID:MSVC>>:-Xcompiler=/W4>
+  $<$<COMPILE_LANGUAGE:HIP>:-Wall>
+  $<$<COMPILE_LANGUAGE:HIP>:-Wextra>)
 if(STRATA_WERROR)
   target_compile_options(strata_warnings INTERFACE
     $<$<AND:$<COMPILE_LANGUAGE:CXX>,$<CXX_COMPILER_ID:MSVC>>:/WX>
-    $<$<AND:$<COMPILE_LANGUAGE:CXX>,$<NOT:$<CXX_COMPILER_ID:MSVC>>>:-Werror>)
+    $<$<AND:$<COMPILE_LANGUAGE:CXX>,$<NOT:$<CXX_COMPILER_ID:MSVC>>>:-Werror>
+    $<$<COMPILE_LANGUAGE:HIP>:-Werror>)
 endif()
 
 # ---- strata_artifact -------------------------------------------------------
@@ -93,6 +155,16 @@ add_executable(strata-dequant src/artifact/dequant.cpp)
 target_link_libraries(strata-dequant PRIVATE strata_artifact)
 # this file is now nothing but the CLI entry point, and its inherited guard spells that as a "selftest"
 target_compile_definitions(strata-dequant PRIVATE STRATA_DEQUANT_SELFTEST)
+if(STRATA_BUILD_TESTS)
+  # the reader refuses a duplicate tensor name at open, naming it (a synthetic GGUF: no model, no GPU)
+  add_executable(gguf_reader_test src/artifact/gguf_reader_test.cpp)
+  target_link_libraries(gguf_reader_test PRIVATE strata_artifact)
+  add_test(NAME gguf_reader_test COMMAND gguf_reader_test)
+  # a split model's shards and their joint view: missing shard, split keys, duplicates, truncation (synthetic)
+  add_executable(gguf_split_test tests/core/gguf_split_test.cpp)
+  target_link_libraries(gguf_split_test PRIVATE strata_artifact)
+  add_test(NAME gguf_split_test COMMAND gguf_split_test)
+endif()
 
 # ---- strata_plan ----------------------------------------------------------
 # Also header-only.  A "planner" that hides state would be one whose decision is not inspectable, and the
@@ -104,17 +176,20 @@ target_link_libraries(strata_plan INTERFACE strata_warnings)
 add_executable(strata-plan src/plan/plan_main.cpp)
 target_link_libraries(strata-plan PRIVATE strata_plan)
 
-# ---- strata_core (CUDA, Phase 2) ------------------------------------------
-# Only built when CUDA is on, so the Phase 1 targets keep working on a machine with no toolkit.  The arch
-# guard above has already refused anything but sm_120 by the time this runs.
-if(STRATA_ENABLE_CUDA)
+# ---- strata_core (GPU runtime) ---------------------------------------------
+# Only built when a GPU backend is on, so the Phase 1 targets keep working without a toolkit.
+if(STRATA_ENABLE_CUDA OR STRATA_ENABLE_HIP)
   add_library(strata_core STATIC src/core/device.cu src/core/pinned.cu src/platform/memory.cpp src/core/graph.cpp src/core/weights.cpp
                                src/core/layout.cpp)
   target_include_directories(strata_core PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
-  # CUDA::cudart is PUBLIC because graph.cpp is a HOST translation unit that calls cudaGraphLaunch and
+  if(STRATA_ENABLE_CUDA AND STRATA_EXPERIMENTAL_SM60)
+    # the runtime device check (src/core/device.cu) admits the cards this build was made for (#236)
+    target_compile_definitions(strata_core PRIVATE STRATA_EXPERIMENTAL_SM60=1)
+  endif()
+  # The selected runtime is PUBLIC because graph.cpp is a HOST translation unit that calls cudaGraphLaunch and
   # cudaEventQuery: the .cu sources get cudart through the CUDA language rules, a .cpp does not, and the
   # linker error ("unresolved external symbol cudaGraphInstantiate") names the function rather than the cause.
-  target_link_libraries(strata_core PUBLIC strata_plan strata_warnings CUDA::cudart)
+  target_link_libraries(strata_core PUBLIC strata_plan strata_warnings ${_strata_gpu_runtime_target})
 
   add_executable(strata-device src/core/device_main.cpp)
   target_link_libraries(strata-device PRIVATE strata_core)
@@ -123,13 +198,13 @@ if(STRATA_ENABLE_CUDA)
   target_link_libraries(strata-load PRIVATE strata_core strata_kernels)
 
   add_executable(strata-concurrent src/core/concurrent_main.cpp)
-  target_include_directories(strata-concurrent PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(strata-concurrent PRIVATE strata_core strata_kernels CUDA::cudart)
+  target_include_directories(strata-concurrent PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(strata-concurrent PRIVATE strata_core strata_kernels ${_strata_gpu_runtime_target})
 
   add_executable(strata-overlap src/core/overlap_main.cpp)
   # a .cpp that calls cudaMemcpyAsync itself needs the toolkit headers and cudart explicitly
-  target_include_directories(strata-overlap PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(strata-overlap PRIVATE strata_core strata_kernels CUDA::cudart)
+  target_include_directories(strata-overlap PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(strata-overlap PRIVATE strata_core strata_kernels ${_strata_gpu_runtime_target})
 
   # The load test needs the 34 GB artifact, so it is registered ONLY where that artifact exists.  A test that
   # silently skips is a test that stops running; an absent registration is visible in the configure output.
@@ -177,8 +252,12 @@ if(STRATA_ENABLE_CUDA)
                           src/kernels/cuda/native_qsa_score.cu
                           src/kernels/cuda/ple.cu
                           src/kernels/cuda/kv_q8.cu
+                          src/kernels/cuda/kv_stream.cu
+                          src/kernels/cuda/kv_q4.cu
                           src/kernels/cuda/fused_gr.cu
+                          src/kernels/cuda/cvec.cu
                           src/kernels/cuda/qsa_decode_attn.cu
+                          src/kernels/cuda/qsa_prompt_attn.cu
                           src/kernels/cuda/fused_gdn.cu
                           src/kernels/cuda/dequant_bf16.cu
                           src/kernels/cuda/qsa_select.cu
@@ -188,33 +267,57 @@ if(STRATA_ENABLE_CUDA)
                           src/ngram/ple_reader.cpp
                           src/platform/direct_file.cpp)
   target_link_libraries(strata_kernels PUBLIC strata_core strata_artifact)
+  # perf-review D-1: the tensor-core prompt attention vs the FP32 kernel and FP64 (GPU, synthetic, no model); with
+  # the tests, or alone with -DSTRATA_PARITY_PROMPT_ATTN=ON.  CUDA only: HIP compiles the tensor-core kernel out.
+  if((STRATA_BUILD_TESTS OR STRATA_PARITY_PROMPT_ATTN) AND STRATA_ENABLE_CUDA)
+    add_executable(qsa_prompt_attn_parity src/kernels/qsa_prompt_attn_parity.cpp)
+    target_include_directories(qsa_prompt_attn_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+    target_link_libraries(qsa_prompt_attn_parity PRIVATE strata_kernels CUDA::cudart)
+  endif()
+  # the QSA selection (block scores on the warp kernel and the tensor-core one, top-k): a GPU bench with a
+  # scorer-vs-scorer comparison and an FP64 accuracy gate (synthetic, no model); the ctest runs a 32K context under a
+  # 262144-cell capacity (the top-k dispatch follows the capacity) and exits non-zero on a failed gate
+  if((STRATA_BUILD_TESTS OR STRATA_PARITY_PROMPT_ATTN) AND STRATA_ENABLE_HIP)
+    add_executable(qsa_select_bench src/kernels/qsa_select_bench.cpp)
+    set_source_files_properties(src/kernels/qsa_select_bench.cpp PROPERTIES LANGUAGE HIP)
+    target_link_libraries(qsa_select_bench PRIVATE strata_kernels strata_hip_runtime)
+    add_test(NAME qsa_select_bench COMMAND qsa_select_bench 32768 256 3 262144)
+  endif()
   # plan v0.3 P6: llama.cpp's block layouts and codebook grids for the i-quant formats (MIT, third_party/ggml)
   target_include_directories(strata_kernels PRIVATE third_party/ggml)
   # Match the pinned llama.cpp CUDA arithmetic contract for this adapter only.
-  set_source_files_properties(src/kernels/cuda/native_mmvq.cu src/kernels/cuda/native_bf16.cu
-    src/kernels/cuda/native_gr_norm.cu
-    src/kernels/cuda/native_moe.cu
-    src/kernels/cuda/native_gdn.cu
-    src/kernels/cuda/native_gr_postops.cu
-    src/kernels/cuda/native_router.cu
-    src/kernels/cuda/native_gdn_preprocess.cu
-    src/kernels/cuda/native_qsa.cu
-    src/kernels/cuda/native_rope.cu
-    src/kernels/cuda/native_ple_postops.cu
-    src/kernels/cuda/native_flash_attn.cu
-    src/kernels/cuda/native_qsa_indexer.cu
-    src/kernels/cuda/native_qsa_score.cu
-    PROPERTIES COMPILE_OPTIONS "--use_fast_math")
+  if(STRATA_ENABLE_CUDA)
+    set_source_files_properties(src/kernels/cuda/native_mmvq.cu src/kernels/cuda/native_bf16.cu
+      src/kernels/cuda/native_gr_norm.cu
+      src/kernels/cuda/native_moe.cu
+      src/kernels/cuda/native_gdn.cu
+      src/kernels/cuda/native_gr_postops.cu
+      src/kernels/cuda/native_router.cu
+      src/kernels/cuda/native_gdn_preprocess.cu
+      src/kernels/cuda/native_qsa.cu
+      src/kernels/cuda/native_rope.cu
+      src/kernels/cuda/native_ple_postops.cu
+      src/kernels/cuda/native_flash_attn.cu
+      src/kernels/cuda/native_qsa_indexer.cu
+      src/kernels/cuda/native_qsa_score.cu
+      PROPERTIES COMPILE_OPTIONS "--use_fast_math")
+  elseif(STRATA_ENABLE_HIP)
+    # Keep Q8_K's multiply and magic-bias rounding as separate FP32 operations on HIP; contraction to an
+    # FMA changes tie results. The embedding gather also relies on separately rounded operations.
+    set_source_files_properties(src/kernels/cuda/elementwise.cu src/kernels/cuda/quantize_act.cu
+      PROPERTIES COMPILE_OPTIONS "-ffp-contract=off")
+  endif()
 
   set(STRATA_ORACLE_SOURCE_DIR "" CACHE PATH "Pinned llama.cpp source for independent CUDA parity tests")
   set(STRATA_ORACLE_BUILD_DIR "" CACHE PATH "Matching existing shared llama.cpp build")
-  if(STRATA_ORACLE_SOURCE_DIR AND STRATA_ORACLE_BUILD_DIR)
+  # This oracle is a CUDA ggml build; do not pull CUDA libraries into a HIP build.
+  if(STRATA_ENABLE_CUDA AND STRATA_ORACLE_SOURCE_DIR AND STRATA_ORACLE_BUILD_DIR)
     foreach(parity IN ITEMS native_mmvq native_bf16 native_shared native_gr_norm native_moe native_gdn native_gr_postops native_router native_gdn_preprocess native_gdn_pipeline native_qsa native_rope native_ple_postops native_flash_attn native_qsa_indexer native_qsa_score)
     add_executable(${parity}_parity bench/micro/${parity}_parity.cpp)
     target_include_directories(${parity}_parity PRIVATE
-      "${STRATA_ORACLE_SOURCE_DIR}/ggml/include" ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
+      "${STRATA_ORACLE_SOURCE_DIR}/ggml/include" ${_strata_gpu_include_directories})
     target_compile_definitions(${parity}_parity PRIVATE GGML_SHARED GGML_BACKEND_SHARED)
-    target_link_libraries(${parity}_parity PRIVATE strata_kernels CUDA::cudart)
+    target_link_libraries(${parity}_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
     foreach(component IN ITEMS ggml ggml-base ggml-cpu ggml-cuda)
       string(REPLACE "-" "_" variable "STRATA_ORACLE_${component}")
       find_library(${variable} NAMES ${component} PATHS
@@ -235,102 +338,303 @@ if(STRATA_ENABLE_CUDA)
   # already links `strata_core` - putting `gdn_layer` in core would make the two mutually dependent.  A third
   # library is the honest shape of the dependency graph: core knows about devices and weights, kernels know
   # about arithmetic, and this one knows the ORDER the arithmetic happens in.
-  add_library(strata_engine STATIC src/core/layer.cpp src/core/session.cpp src/core/expert_source.cpp
-    src/core/expert_cache.cpp src/core/native_head.cpp src/core/native_dense.cpp src/core/verify.cpp src/core/mtp.cpp)
+  add_library(strata_engine STATIC src/core/layer.cpp src/core/session.cpp src/core/expert_source.cpp src/core/remote_experts.cpp
+    src/core/expert_cache.cpp src/core/native_head.cpp src/core/native_dense.cpp src/core/verify.cpp src/core/mtp.cpp
+    src/core/conversation_snapshot.cpp src/core/conversation_state.cpp src/core/conversation_memory.cpp)
   target_include_directories(strata_engine PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
-  target_link_libraries(strata_engine PUBLIC strata_core strata_kernels strata_kernels_cpu CUDA::cudart)
+  target_link_libraries(strata_engine PUBLIC strata_core strata_kernels strata_kernels_cpu ${_strata_gpu_runtime_target})
+  if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/tests/core/file_expert_source_test.cpp")
+    enable_testing()
+    add_executable(file_expert_source_test tests/core/file_expert_source_test.cpp)
+    target_link_libraries(file_expert_source_test PRIVATE strata_engine)
+    add_test(NAME file_expert_source_test COMMAND file_expert_source_test)
+    set_tests_properties(file_expert_source_test PROPERTIES TIMEOUT 60)
+  endif()
+  # #477: the learned expert profile (--expert-profile-save): its order and make_profile.py's file format (CPU only)
+  if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/tests/core/expert_profile_save_test.cpp")
+    enable_testing()
+    add_executable(expert_profile_save_test tests/core/expert_profile_save_test.cpp)
+    target_link_libraries(expert_profile_save_test PRIVATE strata_engine)
+    add_test(NAME expert_profile_save_test COMMAND expert_profile_save_test)
+    set_tests_properties(expert_profile_save_test PROPERTIES TIMEOUT 60)
+  endif()
+  if(STRATA_BUILD_TESTS AND STRATA_NATIVE_EXPERTS)
+    # native_experts.txt v2-v4 (the real packs' copies and the per-role shard column) and the arena filled from
+    # synthetic GGUF shards at the real expert dimensions
+    add_executable(expert_layout_test tests/core/expert_layout_test.cpp)
+    target_link_libraries(expert_layout_test PRIVATE strata_engine)
+    add_test(NAME expert_layout_test COMMAND expert_layout_test ${CMAKE_CURRENT_SOURCE_DIR}/tests/data/native_experts)
+    set_tests_properties(expert_layout_test PROPERTIES TIMEOUT 60)
+    # #326: a native pack's BF16 (--compat-bf16) PLE key stays in the arena; a quantized key row loads as before
+    add_executable(native_dense_ple_key_test tests/core/native_dense_ple_key_test.cpp)
+    target_link_libraries(native_dense_ple_key_test PRIVATE strata_engine)
+    add_test(NAME native_dense_ple_key_test COMMAND native_dense_ple_key_test)
+    set_tests_properties(native_dense_ple_key_test PROPERTIES SKIP_RETURN_CODE 77 TIMEOUT 60)
+    # #369: per-layer admission on a sized (native pack) cache starts each layer in its own range
+    add_executable(expert_cache_per_layer_test tests/core/expert_cache_per_layer_test.cpp)
+    target_link_libraries(expert_cache_per_layer_test PRIVATE strata_engine)
+    add_test(NAME expert_cache_per_layer_test COMMAND expert_cache_per_layer_test)
+    set_tests_properties(expert_cache_per_layer_test PROPERTIES SKIP_RETURN_CODE 77 TIMEOUT 60)
+  endif()
+  if(STRATA_BUILD_CONVERSATION_TESTS)
+    add_executable(conversation_snapshot_test src/core/conversation_snapshot_test.cpp)
+    target_link_libraries(conversation_snapshot_test PRIVATE strata_engine)
+    add_test(NAME conversation_snapshot_test COMMAND conversation_snapshot_test)
+    set_tests_properties(conversation_snapshot_test PROPERTIES SKIP_RETURN_CODE 77 TIMEOUT 120)
+    add_executable(conversation_validation_test src/core/conversation_validation_test.cpp)
+    target_link_libraries(conversation_validation_test PRIVATE strata_engine)
+    add_test(NAME conversation_validation_test COMMAND conversation_validation_test)
+    set_tests_properties(conversation_validation_test PROPERTIES ENVIRONMENT "CUDA_VISIBLE_DEVICES=-1" TIMEOUT 120)
+    if(NOT STRATA_ENABLE_HIP AND UNIX AND NOT APPLE AND CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+      add_executable(conversation_transfer_test src/core/conversation_validation_test.cpp)
+      target_compile_definitions(conversation_transfer_test PRIVATE CONVERSATION_TEST_TRANSFERS)
+      target_link_options(conversation_transfer_test PRIVATE
+        -Wl,--wrap=cudaMemcpy -Wl,--wrap=cudaDeviceSynchronize -Wl,--wrap=cudaGetLastError)
+      target_link_libraries(conversation_transfer_test PRIVATE strata_engine)
+      add_test(NAME conversation_transfer_test COMMAND conversation_transfer_test)
+      set_tests_properties(conversation_transfer_test PROPERTIES ENVIRONMENT "CUDA_VISIBLE_DEVICES=-1" TIMEOUT 120)
+    endif()
+  endif()
 
   # `strata generate` (P2.S6).  The FIRST user-facing program: everything under it is a component, and this is
   # what composes them into a token.  No tokenizer - the prompt arrives as ids - because Gate C1 compares logits
   # against llama.cpp on the SAME ids and a tokenizer on one side of that comparison is a second variable.
   # plan v0.3 P5: batched prompt processing (tensor-core GEMMs through cuBLAS).
   add_library(strata_prefill STATIC src/prefill/gemm.cu src/prefill/kernels.cu src/prefill/prefill.cpp)
-  target_include_directories(strata_prefill PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(strata_prefill PUBLIC strata_engine CUDA::cublas CUDA::cudart)
+  target_include_directories(strata_prefill PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include ${_strata_gpu_include_directories})
+  target_link_libraries(strata_prefill PUBLIC strata_engine ${_strata_gpu_blas_target} ${_strata_gpu_runtime_target})
+  if(STRATA_ENABLE_HIP AND STRATA_HIPBLASLT_AVAILABLE)
+    target_compile_definitions(strata_prefill PRIVATE STRATA_HIPBLASLT_AVAILABLE=1)
+    target_link_libraries(strata_prefill PRIVATE roc::hipblaslt)
+
+    add_executable(tune_hipblaslt tools/hip/tune_hipblaslt.cpp)
+    target_link_libraries(tune_hipblaslt PRIVATE hip::host roc::hipblas roc::hipblaslt)
+  endif()
 
   add_executable(strata src/program/generate.cpp)
-  target_include_directories(strata PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(strata PRIVATE strata_engine strata_prefill CUDA::cudart)
+  target_include_directories(strata PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(strata PRIVATE strata_engine strata_prefill strata_spec ${_strata_gpu_runtime_target})
+
+  if(STRATA_ENABLE_HIP)
+    enable_testing()
+    set(_strata_hip_tests)
+    add_executable(hip_expert_cache_staging tests/hip/expert_cache_staging.cpp)
+    target_link_libraries(hip_expert_cache_staging PRIVATE strata_engine)
+    add_test(NAME hip_expert_cache_staging COMMAND hip_expert_cache_staging)
+    list(APPEND _strata_hip_tests hip_expert_cache_staging)
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/tests/hip/intrinsics.cpp")
+      add_executable(hip_intrinsics tests/hip/intrinsics.cpp)
+      target_link_libraries(hip_intrinsics PRIVATE strata_hip_runtime)
+      add_test(NAME hip_intrinsics COMMAND hip_intrinsics)
+      list(APPEND _strata_hip_tests hip_intrinsics)
+    endif()
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/tests/hip/handoff.cpp")
+      add_executable(hip_handoff tests/hip/handoff.cpp)
+      target_link_libraries(hip_handoff PRIVATE strata_kernels strata_hip_runtime)
+      add_test(NAME hip_handoff COMMAND hip_handoff)
+      list(APPEND _strata_hip_tests hip_handoff)
+    endif()
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/tests/hip/mapped_alias.cpp")
+      # #325: reports what a hipHostGetDevicePointer alias is on this stack (on Windows: the host pointer, and a
+      # D2D copy into it does not land - hip_handoff's timeout there). A diagnosis: exits 0 whenever it ran.
+      add_executable(hip_mapped_alias tests/hip/mapped_alias.cpp)
+      target_link_libraries(hip_mapped_alias PRIVATE strata_kernels strata_hip_runtime)
+      add_test(NAME hip_mapped_alias COMMAND hip_mapped_alias)
+      set_tests_properties(hip_mapped_alias PROPERTIES SKIP_RETURN_CODE 77)
+      list(APPEND _strata_hip_tests hip_mapped_alias)
+    endif()
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/tests/hip/native_qsa_score.cpp")
+      add_executable(hip_native_qsa_score tests/hip/native_qsa_score.cpp)
+      target_link_libraries(hip_native_qsa_score PRIVATE strata_kernels strata_hip_runtime)
+      add_test(NAME hip_native_qsa_score COMMAND hip_native_qsa_score)
+      list(APPEND _strata_hip_tests hip_native_qsa_score)
+    endif()
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/tests/hip/prefill_native_batch.cpp")
+      add_executable(hip_prefill_native_batch tests/hip/prefill_native_batch.cpp)
+      target_link_libraries(hip_prefill_native_batch PRIVATE strata_kernels strata_hip_runtime)
+      add_test(NAME hip_prefill_native_batch COMMAND hip_prefill_native_batch)
+      list(APPEND _strata_hip_tests hip_prefill_native_batch)
+    endif()
+    # S6: the AMD fast router against the portable kernel, bitwise
+    add_executable(hip_router_fast tests/hip/router_fast.cpp)
+    target_link_libraries(hip_router_fast PRIVATE strata_kernels strata_hip_runtime)
+    add_test(NAME hip_router_fast COMMAND hip_router_fast)
+    list(APPEND _strata_hip_tests hip_router_fast)
+    # S6: the RDNA4 matrix-core prompt attention (STRATA_HIP_WMMA) against the FP32 kernel and FP64; skips off gfx12
+    add_executable(hip_prompt_attn_wmma src/kernels/qsa_prompt_attn_parity.cpp)
+    target_link_libraries(hip_prompt_attn_wmma PRIVATE strata_kernels strata_hip_runtime)
+    add_test(NAME hip_prompt_attn_wmma COMMAND hip_prompt_attn_wmma 32768 2048 3)
+    set_tests_properties(hip_prompt_attn_wmma PROPERTIES TIMEOUT 600 SKIP_RETURN_CODE 77)
+    list(APPEND _strata_hip_tests hip_prompt_attn_wmma)
+    if(STRATA_HIPBLASLT_AVAILABLE AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/tests/hip/prefill_hipblaslt_gemm.cpp")
+      add_executable(hip_prefill_hipblaslt_gemm tests/hip/prefill_hipblaslt_gemm.cpp)
+      target_include_directories(hip_prefill_hipblaslt_gemm PRIVATE "${CMAKE_CURRENT_SOURCE_DIR}/src/prefill")
+      target_link_libraries(hip_prefill_hipblaslt_gemm PRIVATE strata_prefill strata_hip_runtime roc::hipblaslt roc::hipblas)
+      add_test(NAME hip_prefill_hipblaslt_gemm COMMAND hip_prefill_hipblaslt_gemm)
+      set_tests_properties(hip_prefill_hipblaslt_gemm PROPERTIES TIMEOUT 120 SKIP_RETURN_CODE 77)
+      list(APPEND _strata_hip_tests hip_prefill_hipblaslt_gemm)
+    endif()
+    if(_strata_hip_tests)
+      set_tests_properties(${_strata_hip_tests} PROPERTIES TIMEOUT 60)
+    endif()
+    if(TARGET hip_prefill_native_batch)
+      set_tests_properties(hip_prefill_native_batch PROPERTIES TIMEOUT 120)
+    endif()
+    if(TARGET hip_prefill_hipblaslt_gemm)
+      set_tests_properties(hip_prefill_hipblaslt_gemm PROPERTIES TIMEOUT 120 SKIP_RETURN_CODE 77)
+    endif()
+  endif()
 
   add_executable(dequant_s2_parity src/kernels/dequant_s2_parity.cpp)
   # The tests are HOST C++ files that call cudaMalloc/cudaMemcpy themselves, so they need the toolkit headers
   # and cudart explicitly: enabling the CUDA language gives those to .cu sources automatically, but a .cpp
   # target is just a C++ target and gets neither.
-  target_include_directories(dequant_s2_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(dequant_s2_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(dequant_s2_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(dequant_s2_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME dequant_s2_parity COMMAND dequant_s2_parity --selftest)
 
   add_executable(s2_gemv_parity src/kernels/s2_gemv_parity.cpp)
-  target_include_directories(s2_gemv_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(s2_gemv_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(s2_gemv_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(s2_gemv_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME s2_gemv_parity COMMAND s2_gemv_parity --selftest)
 
   add_executable(iq_parity src/kernels/iq_parity.cpp)
-  target_include_directories(iq_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(iq_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(iq_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(iq_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
+  # TODO 24: the i-quant fixtures are generated, not shipped.  tools/iq_fixture.py needs python3 + numpy
+  # and the repository's vendored gguf-py (no pip installs); it writes deterministic fixtures into the
+  # BUILD directory (the source tree is not touched).  The CTest fixture makes generation-to-test order
+  # explicit even under parallel execution; a fixture-generation failure fails iq_parity with it.
+  find_package(Python3 COMPONENTS Interpreter QUIET)
+  # the generator needs an interpreter WITH numpy: probe the configured one, the setup's .venv, and the
+  # system python3; -DIQ_FIXTURE_PYTHON=... overrides the search outright.
+  set(_iq_python "")
+  set(_iq_candidates "")
+  if(IQ_FIXTURE_PYTHON)
+    list(APPEND _iq_candidates ${IQ_FIXTURE_PYTHON})
+  else()
+    if(Python3_Interpreter_FOUND)
+      list(APPEND _iq_candidates ${Python3_EXECUTABLE})
+    endif()
+    list(APPEND _iq_candidates "${CMAKE_CURRENT_SOURCE_DIR}/.venv/bin/python"
+                                "${CMAKE_CURRENT_SOURCE_DIR}/.venv/Scripts/python.exe"
+                                /usr/bin/python3
+                                python3)
+  endif()
+  foreach(_cand IN LISTS _iq_candidates)
+    if(_cand STREQUAL "")
+      continue()
+    endif()
+    execute_process(COMMAND ${_cand} -c "import numpy" OUTPUT_QUIET ERROR_QUIET RESULT_VARIABLE _iq_probe)
+    if(_iq_probe EQUAL 0)
+      set(_iq_python "${_cand}")
+      break()
+    endif()
+  endforeach()
+  if(NOT _iq_python STREQUAL "")
+    add_test(NAME iq_parity_fixtures
+             COMMAND ${_iq_python} ${CMAKE_CURRENT_SOURCE_DIR}/tools/iq_fixture.py
+                     --out ${CMAKE_CURRENT_BINARY_DIR}/iq_fixture)
+    # exit 3: numpy or the vendored gguf-py is missing (a trimmed third_party) - a skip, not a failure
+    set_tests_properties(iq_parity_fixtures PROPERTIES FIXTURES_SETUP iq_fixture_data SKIP_RETURN_CODE 3)
+    add_test(NAME iq_parity COMMAND iq_parity ${CMAKE_CURRENT_BINARY_DIR}/iq_fixture)
+    set_tests_properties(iq_parity PROPERTIES FIXTURES_REQUIRED iq_fixture_data SKIP_RETURN_CODE 3)
+  else()
+    message(STATUS "STRATA tests: no python3 with numpy was found (the configured interpreter and .venv "
+                   "were probed) - the iq_parity fixtures and its test are not registered; the fixtures "
+                   "need python3 with numpy and the repository's vendored gguf-py")
+  endif()
+  if(STRATA_ENABLE_CUDA)
+    # native_mmvq's multi_exact contract, bitwise: a T-column call against T single-column calls on the same Q8_1
+    # bytes, with the flag off as the negative control (synthetic, GPU, no model; the control assumes 32-lane warps)
+    add_executable(mmvq_multi_parity src/kernels/mmvq_multi_parity.cpp)
+    target_include_directories(mmvq_multi_parity PRIVATE ${_strata_gpu_include_directories})
+    target_link_libraries(mmvq_multi_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
+    add_test(NAME mmvq_multi_parity COMMAND mmvq_multi_parity)
+  endif()
+  # the IQ kernels that decode a weight part once for every column/entry, bitwise vs the per-column ones
+  # (GPU, synthetic, no model; --bench adds old/new timings)
+  add_executable(iq_multi_parity src/kernels/iq_multi_parity.cpp)
+  target_include_directories(iq_multi_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(iq_multi_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
+  add_test(NAME iq_multi_parity COMMAND iq_multi_parity)
   add_executable(shared_expert_parity src/kernels/shared_expert_parity.cpp)
-  target_include_directories(shared_expert_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(shared_expert_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(shared_expert_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(shared_expert_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME shared_expert_parity COMMAND shared_expert_parity --selftest)
 
   add_executable(gr_parity src/kernels/gr_parity.cpp)
-  target_include_directories(gr_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(gr_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(gr_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(gr_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME gr_parity COMMAND gr_parity --selftest)
 
   add_executable(gdn_parity src/kernels/gdn_parity.cpp)
-  target_include_directories(gdn_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(gdn_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(gdn_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(gdn_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME gdn_parity COMMAND gdn_parity --selftest)
 
   add_executable(s2_gemv_q8_parity src/kernels/s2_gemv_q8_parity.cpp)
-  target_include_directories(s2_gemv_q8_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(s2_gemv_q8_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(s2_gemv_q8_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(s2_gemv_q8_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME s2_gemv_q8_parity COMMAND s2_gemv_q8_parity --selftest)
 
   add_executable(sampler_parity src/kernels/sampler_parity.cpp)
-  target_include_directories(sampler_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(sampler_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(sampler_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(sampler_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME sampler_parity COMMAND sampler_parity --selftest)
+  set_tests_properties(sampler_parity PROPERTIES ENVIRONMENT "STRATA_OLD_SAMPLER=0;STRATA_SAMPLER_ONE_BLOCK=0")
+  # the same fixtures on the other two sampled paths: the one-block kernel and the 0.1.20 kernel
+  add_test(NAME sampler_parity_one_block COMMAND sampler_parity --selftest)
+  set_tests_properties(sampler_parity_one_block PROPERTIES ENVIRONMENT "STRATA_OLD_SAMPLER=0;STRATA_SAMPLER_ONE_BLOCK=1")
+  add_test(NAME sampler_parity_old COMMAND sampler_parity --selftest)
+  set_tests_properties(sampler_parity_old PROPERTIES ENVIRONMENT "STRATA_OLD_SAMPLER=1;STRATA_SAMPLER_ONE_BLOCK=0")
+  # S19: the decode QSA top-k and the greedy argmax on thread-block clusters (sm_90+) vs the one-CTA kernels, bitwise,
+  # captured in a graph too (GPU, synthetic, no model; skips below sm_90); --bench: per-call times
+  if(STRATA_ENABLE_CUDA)
+    add_executable(decode_cluster_parity src/kernels/decode_cluster_parity.cpp)
+    target_include_directories(decode_cluster_parity PRIVATE ${_strata_gpu_include_directories})
+    target_link_libraries(decode_cluster_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
+    add_test(NAME decode_cluster_parity COMMAND decode_cluster_parity --selftest)
+  endif()
 
   add_executable(rope_parity src/kernels/rope_parity.cpp)
-  target_include_directories(rope_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(rope_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(rope_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(rope_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME rope_parity COMMAND rope_parity --selftest)
 
   add_executable(quantize_act_parity src/kernels/quantize_act_parity.cpp)
-  target_include_directories(quantize_act_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(quantize_act_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(quantize_act_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(quantize_act_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME quantize_act_parity COMMAND quantize_act_parity --selftest)
 
   add_executable(router_top10_parity src/kernels/router_top10_parity.cpp)
-  target_include_directories(router_top10_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(router_top10_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(router_top10_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(router_top10_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME router_top10_parity COMMAND router_top10_parity --selftest)
 
   add_executable(s_gemv_parity src/kernels/s_gemv_parity.cpp)
-  target_include_directories(s_gemv_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(s_gemv_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(s_gemv_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(s_gemv_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME s_gemv_parity COMMAND s_gemv_parity --selftest)
 
   add_executable(elementwise_parity src/kernels/elementwise_parity.cpp)
-  target_include_directories(elementwise_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(elementwise_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(elementwise_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(elementwise_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME elementwise_parity COMMAND elementwise_parity --selftest)
 
   add_executable(bf16_gemv_parity src/kernels/bf16_gemv_parity.cpp)
-  target_include_directories(bf16_gemv_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(bf16_gemv_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(bf16_gemv_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(bf16_gemv_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME bf16_gemv_parity COMMAND bf16_gemv_parity --selftest)
 
   add_executable(s_gemv_q8k_parity src/kernels/s_gemv_q8k_parity.cpp)
-  target_include_directories(s_gemv_q8k_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(s_gemv_q8k_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(s_gemv_q8k_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(s_gemv_q8k_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME s_gemv_q8k_parity COMMAND s_gemv_q8k_parity --selftest)
 
   add_executable(qsa_parity src/kernels/qsa_parity.cpp)
-  target_include_directories(qsa_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(qsa_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(qsa_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(qsa_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME qsa_parity COMMAND qsa_parity --selftest)
 
   # ---- P2.S4: the PLE n-gram hash, the IQ4_NL table read, and the PLE block ----
@@ -342,8 +646,8 @@ if(STRATA_ENABLE_CUDA)
   # INSIDE the test.  A test that passes by finding nothing to compare is the failure mode this project keeps
   # re-learning.
   add_executable(ple_parity src/kernels/ple_parity.cpp)
-  target_include_directories(ple_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(ple_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(ple_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(ple_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   set(STRATA_PLE_FIXTURE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/bench/micro" CACHE PATH
       "Directory containing required ple_in.bin and ple_out.bin parity fixtures")
   add_test(NAME ple_parity COMMAND ple_parity --selftest
@@ -352,57 +656,99 @@ if(STRATA_ENABLE_CUDA)
 
   # Plan v0.3 P7: INT8 KV storage vs a host reference and vs the FP16 path (GPU, synthetic, no model).
   add_executable(kv_q8_parity src/kernels/kv_q8_parity.cpp)
-  target_include_directories(kv_q8_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(kv_q8_parity PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(kv_q8_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(kv_q8_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME kv_q8_parity COMMAND kv_q8_parity)
 
+  # KV streaming (docs/kv-streaming-design.md): streamed vs resident attention, bitwise (GPU, synthetic, no model).
+  add_executable(kv_stream_parity src/kernels/kv_stream_parity.cpp)
+  target_include_directories(kv_stream_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(kv_stream_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
+  add_test(NAME kv_stream_parity COMMAND kv_stream_parity)
+
+  # PR #120: the hybrid K8V4 KV (INT8 K + rotated Q4_0 V) vs host references (GPU, synthetic, no model).
+  add_executable(kv_hybrid_parity src/kernels/kv_hybrid_parity.cpp)
+  target_include_directories(kv_hybrid_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(kv_hybrid_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
+  add_test(NAME kv_hybrid_parity COMMAND kv_hybrid_parity)
+  # PR #21: Q4_0 KV with Walsh-Hadamard rotation vs a host reference (GPU, synthetic, no model).
+  add_executable(kv_q4_parity src/kernels/kv_q4_parity.cpp)
+  target_include_directories(kv_q4_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(kv_q4_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
+  add_test(NAME kv_q4_parity COMMAND kv_q4_parity)
+  add_executable(cvec_parity src/kernels/cvec_parity.cpp)
+  target_include_directories(cvec_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(cvec_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
+  add_test(NAME cvec_parity COMMAND cvec_parity)
+
+  # The rewritten per-hit and grouped expert kernels vs the previous ones, bitwise (GPU, synthetic, no model).
+  # `--bench` adds previous/new timings on engine-like shapes.
+  add_executable(s2_expert_grouped_parity src/kernels/s2_expert_grouped_parity.cpp)
+  target_include_directories(s2_expert_grouped_parity PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(s2_expert_grouped_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
+  add_test(NAME s2_expert_grouped_parity COMMAND s2_expert_grouped_parity --selftest)
+
   # Plan v0.3 P3: native MMVQ for 1..8 tokens (bitwise selftest + --bench timing); synthetic data, no model.
-  if(STRATA_BUILD_TESTS)
-  add_executable(native_mmvq_multi bench/micro/native_mmvq_multi.cpp)
-  target_include_directories(native_mmvq_multi PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(native_mmvq_multi PRIVATE strata_kernels CUDA::cudart)
-  add_test(NAME native_mmvq_multi COMMAND native_mmvq_multi --selftest)
+  if(STRATA_BUILD_TESTS AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/bench/micro/native_mmvq_multi.cpp")
+    add_executable(native_mmvq_multi bench/micro/native_mmvq_multi.cpp)
+    target_include_directories(native_mmvq_multi PRIVATE ${_strata_gpu_include_directories})
+    target_link_libraries(native_mmvq_multi PRIVATE strata_kernels ${_strata_gpu_runtime_target})
+    add_test(NAME native_mmvq_multi COMMAND native_mmvq_multi --selftest)
   endif()
 
   # Plan v0.3 P0.1: working-set lock of a host region (CPU only).
   add_executable(platform_memory_test src/platform/memory_test.cpp)
-  target_link_libraries(platform_memory_test PRIVATE strata_core CUDA::cudart)
+  target_link_libraries(platform_memory_test PRIVATE strata_core ${_strata_gpu_runtime_target})
   add_test(NAME platform_memory_test COMMAND platform_memory_test)
 
+  if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
+    add_executable(pinned_shared_test src/core/pinned_shared_test.cpp)
+    target_link_libraries(pinned_shared_test PRIVATE strata_core ${_strata_gpu_runtime_target})
+    add_test(NAME pinned_shared_test COMMAND pinned_shared_test)
+  endif()
+
   # Plan v0.3 P2: the n-gram table read straight from the SSD. The selftest needs only the CPU and a temp file.
   add_executable(ple_reader_test src/ngram/ple_reader_test.cpp)
-  target_include_directories(ple_reader_test PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(ple_reader_test PRIVATE strata_kernels CUDA::cudart)
+  target_include_directories(ple_reader_test PRIVATE ${_strata_gpu_include_directories})
+  target_link_libraries(ple_reader_test PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   add_test(NAME ple_reader_selftest COMMAND ple_reader_test --selftest)
 
-  # The GPU test is registered only when CUDA is on, and it runs ON THE TARGET PC.  `--selftest` allocates
+  # The GPU test runs ON THE TARGET PC. `--selftest` allocates
   # and writes real VRAM, so a machine without a device fails this test rather than skipping it silently -
   # which is correct: an engine that targets one GPU should not report a pass for having no GPU.
-  add_test(NAME cuda_device_selftest COMMAND strata-device --selftest)
+  if(STRATA_ENABLE_CUDA)
+    add_test(NAME cuda_device_selftest COMMAND strata-device --selftest)
+  else()
+    add_test(NAME hip_device_selftest COMMAND strata-device --selftest)
+  endif()
 
   # ---- P2.S5: the GraphRegistry --------------------------------------------------
   # A .cu because the test needs real kernels to capture; it links `strata_core` and not `strata_kernels`,
   # because what it exercises is the capture machinery and not any kernel in it.
-  if(STRATA_BUILD_TESTS)
-  add_executable(graph_registry_test bench/micro/graph_registry_test.cu)
-  target_include_directories(graph_registry_test PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(graph_registry_test PRIVATE strata_core CUDA::cudart)
-  add_test(NAME graph_registry_test COMMAND graph_registry_test)
+  if(STRATA_BUILD_TESTS AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/bench/micro/graph_registry_test.cu")
+    add_executable(graph_registry_test bench/micro/graph_registry_test.cu)
+    target_include_directories(graph_registry_test PRIVATE ${_strata_gpu_include_directories})
+    target_link_libraries(graph_registry_test PRIVATE strata_core ${_strata_gpu_runtime_target})
+    add_test(NAME graph_registry_test COMMAND graph_registry_test)
+  endif()
 
   # ---- P2.S5's GATE: is a device<->pinned-memory handoff graph-capturable? --------------------
   # The doorbell protocol puts the D2H of `x_f` and the miss list INSIDE G[l] - that is the whole overlap.  The
   # phase doc says to verify this "before building on it", and this project has already lost two rounds to
   # building first (a memory-only spin that DEADLOCKS on Windows, and a `cudaEventRecord` inside a capture that
   # is SILENTLY DROPPED).  It links only cudart: it is about the capture machinery, not about any kernel.
-  add_executable(pinned_capture bench/micro/pinned_capture.cu)
-  target_include_directories(pinned_capture PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-  target_link_libraries(pinned_capture PRIVATE CUDA::cudart)
-  add_test(NAME pinned_capture COMMAND pinned_capture)
+  if(STRATA_ENABLE_CUDA AND STRATA_BUILD_TESTS AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/bench/micro/pinned_capture.cu")
+    add_executable(pinned_capture bench/micro/pinned_capture.cu)
+    target_include_directories(pinned_capture PRIVATE ${_strata_gpu_include_directories})
+    target_link_libraries(pinned_capture PRIVATE ${_strata_gpu_runtime_target})
+    add_test(NAME pinned_capture COMMAND pinned_capture)
   endif()
 
   # Tells tests/CMakeLists.txt to add the CUDA-dependent Catch2 target.  It cannot be declared there
   # unconditionally: tests/ is added at the end of this file, but the flag has to be known before it runs.
-  set(STRATA_CUDA_TESTS ON)
+  if(STRATA_ENABLE_CUDA)
+    set(STRATA_CUDA_TESTS ON)
+  endif()
 endif()
 
 # ---- strata_kernels_cpu (P2.S3) ----
@@ -417,7 +763,7 @@ endif()
 #     worse opt-out.  `cpu_require_expert_support()` therefore refuses at startup on a CPU without the
 #     features, and the fallback exists for tests, exactly as P2.S3 specifies.
 # Plan v0.3 P6: speculation components that need no GPU.
-add_library(strata_spec STATIC src/spec/suffix_drafter.cpp src/spec/controller.cpp)
+add_library(strata_spec STATIC src/spec/suffix_drafter.cpp src/spec/controller.cpp src/spec/draft_policy.cpp)
 target_include_directories(strata_spec PUBLIC include)
 if(STRATA_BUILD_TESTS)
   add_executable(suffix_drafter_test src/spec/suffix_drafter_test.cpp)
@@ -426,6 +772,23 @@ if(STRATA_BUILD_TESTS)
   add_executable(controller_test src/spec/controller_test.cpp)
   target_link_libraries(controller_test PRIVATE strata_spec)
   add_test(NAME controller_test COMMAND controller_test)
+  add_executable(draft_policy_test src/spec/draft_policy_test.cpp)
+  target_link_libraries(draft_policy_test PRIVATE strata_spec)
+  add_test(NAME draft_policy_test COMMAND draft_policy_test)
+endif()
+if(STRATA_BUILD_TESTS OR STRATA_BUILD_CONVERSATION_TESTS)
+  # the serve loop's conversation-cache retention policy (header-only, no library to link)
+  add_executable(conv_cache_test src/program/conv_cache_test.cpp)
+  target_include_directories(conv_cache_test PRIVATE include)
+  add_test(NAME conv_cache_test COMMAND conv_cache_test)
+  # coupled draft sampling's counter and penalty-ring arithmetic (header-only, CPU)
+  add_executable(coupled_draft_test src/core/coupled_draft_test.cpp)
+  target_include_directories(coupled_draft_test PRIVATE include)
+  add_test(NAME coupled_draft_test COMMAND coupled_draft_test)
+  # the f32 -> bf16 rounding against ggml's, NaN included, over every f32 bit pattern (header-only, CPU)
+  add_executable(bf16_bits_test src/kernels/bf16_bits_test.cpp)
+  target_include_directories(bf16_bits_test PRIVATE include)
+  add_test(NAME bf16_bits_test COMMAND bf16_bits_test)
 endif()
 
 add_library(strata_kernels_cpu STATIC src/kernels/cpu/expert.cpp src/kernels/cpu/pool.cpp
@@ -452,6 +815,12 @@ target_link_libraries(strata_kernels_cpu PUBLIC Threads::Threads)
 # with -DSTRATA_GGML_DIR=..., otherwise the pinned upstream commit (third_party/ggml/VERSION.txt).
 option(STRATA_NATIVE_EXPERTS "Native (IQ) expert support through ggml-cpu" ON)
 set(STRATA_GGML_DIR "" CACHE PATH "llama.cpp checkout to take ggml from (empty: fetch the pinned commit)")
+if(STRATA_PREFILL_MMQ AND NOT STRATA_NATIVE_EXPERTS)
+  message(FATAL_ERROR "STRATA_PREFILL_MMQ requires STRATA_NATIVE_EXPERTS=ON for GGML sources")
+endif()
+if(STRATA_ORCA_Q4KS_MMQ AND NOT STRATA_NATIVE_EXPERTS)
+  message(FATAL_ERROR "STRATA_ORCA_Q4KS_MMQ requires STRATA_NATIVE_EXPERTS=ON for GGML sources")
+endif()
 if(STRATA_NATIVE_EXPERTS)
   enable_language(C)
   set(BUILD_SHARED_LIBS OFF CACHE BOOL "" FORCE)
@@ -485,34 +854,175 @@ if(STRATA_NATIVE_EXPERTS)
   endif()
   # part of strata_kernels_cpu (the pool calls these and they call its CPU probe: one library, so GNU ld's
   # one-pass static linking has no cycle to resolve)
-  target_sources(strata_kernels_cpu PRIVATE src/kernels/cpu/native_expert.cpp src/kernels/cpu/iq_avx512.cpp)
+  target_sources(strata_kernels_cpu PRIVATE src/kernels/cpu/native_expert.cpp src/kernels/cpu/iq_avx512.cpp
+                                           src/kernels/cpu/iq_avx2.cpp src/kernels/cpu/kq_avx2.cpp)
   target_include_directories(strata_kernels_cpu PRIVATE third_party/ggml)
   if(MSVC)
     set_source_files_properties(src/kernels/cpu/iq_avx512.cpp PROPERTIES COMPILE_OPTIONS "/arch:AVX512")
+    set_source_files_properties(src/kernels/cpu/iq_avx2.cpp PROPERTIES COMPILE_OPTIONS "/arch:AVX2")
+    set_source_files_properties(src/kernels/cpu/kq_avx2.cpp PROPERTIES COMPILE_OPTIONS "/arch:AVX2")
   else()
     set_source_files_properties(src/kernels/cpu/iq_avx512.cpp PROPERTIES COMPILE_OPTIONS
                                 "-mavx512f;-mavx512bw;-mavx512vl;-mavx512dq;-mavx512vnni;-mavx512vbmi;-mf16c;-mfma")
+    set_source_files_properties(src/kernels/cpu/iq_avx2.cpp PROPERTIES COMPILE_OPTIONS "-mavx2;-mfma;-mf16c")
+    set_source_files_properties(src/kernels/cpu/kq_avx2.cpp PROPERTIES COMPILE_OPTIONS "-mavx2;-mfma;-mf16c")
   endif()
   target_link_libraries(strata_kernels_cpu PUBLIC ggml-cpu ggml-base)
   target_compile_definitions(strata_kernels_cpu PUBLIC STRATA_NATIVE_EXPERTS=1)
+  set(_strata_mmq_types q2_0 iq2_xxs iq2_xs iq2_s iq3_xxs iq3_s iq4_nl iq4_xs q8_0)
+  # #296: OrcaRouter Q4_K_S's Q5_0 experts (Q4_K and Q5_1 come with STRATA_MMQ_KQUANTS on CUDA, S5's branch)
+  if(STRATA_ORCA_Q4KS_MMQ)
+    list(APPEND _strata_mmq_types q5_0)
+  endif()
   if(STRATA_ENABLE_CUDA)
+    # prompt-speed plan step 2b: llama.cpp's MMQ kernels (ggml-cuda, MIT) for the prompt path's experts, compiled
+    # from the same checkout as ggml above - only mmq.cuh's instances for the packs' types and the quantizer, with
+    # the few host symbols of ggml-cuda.cu they use in src/prefill/ggml_cuda_host.cu
+    if(STRATA_GGML_DIR)
+      set(_strata_gg ${STRATA_GGML_DIR}/ggml)
+    else()
+      set(_strata_gg ${strata_llamacpp_SOURCE_DIR}/ggml)
+    endif()
+    set(_strata_mmq ${_strata_gg}/src/ggml-cuda)
+    # #136: the Q2_0 pack's prompt experts on Strata's own int8 kernels (moe_fused.cu, opt-in STRATA_PF_FUSED=1; CUDA
+    # only: the HIP branch below keeps MMQ, and an sm_75 device or an sm_75-only build reports them unavailable)
+    add_library(strata_mmq STATIC src/prefill/moe_mmq.cu src/prefill/moe_fused.cu src/prefill/moe_fused_iq.cu
+                src/prefill/ggml_cuda_host.cu ${_strata_mmq}/quantize.cu)
+    # q8_0: the drafter's matrices (E-9); STRATA_ORCA_Q4KS_MMQ adds q5_0 (#296, in _strata_mmq_types), and
+    # STRATA_MMQ_KQUANTS q4_k q5_k q5_1, the experts of Unsloth's UD-Q4_K_XL (docs/UNSLOTH_Q4.md; the dispatch
+    # follows eddoursul/Strata 7706c1b)
+    set(_strata_mmq_cuda_types ${_strata_mmq_types})
+    if(STRATA_MMQ_KQUANTS)
+      list(APPEND _strata_mmq_cuda_types q4_k q5_k q5_1)
+    endif()
+    foreach(_t IN LISTS _strata_mmq_cuda_types)
+      target_sources(strata_mmq PRIVATE ${_strata_mmq}/template-instances/mmq-instance-${_t}.cu)
+    endforeach()
+    target_include_directories(strata_mmq PRIVATE ${_strata_gg}/include ${_strata_gg}/src ${_strata_mmq}
+                               PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
+    target_compile_options(strata_mmq PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:--extended-lambda -use_fast_math>)
+    if(MSVC)
+      target_compile_options(strata_mmq PRIVATE $<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler=/bigobj>)
+    endif()
+    target_link_libraries(strata_mmq PUBLIC ggml-base ${_strata_gpu_runtime_target})
+    if(STRATA_MMQ_KQUANTS)
+      target_compile_definitions(strata_mmq PRIVATE STRATA_MMQ_KQUANTS=1)
+    endif()
+    if(STRATA_ORCA_Q4KS_MMQ)
+      target_compile_definitions(strata_mmq PRIVATE STRATA_ORCA_Q4KS_MMQ=1)
+    endif()
+    target_link_libraries(strata_prefill PUBLIC strata_mmq)
+    target_compile_definitions(strata_prefill PUBLIC STRATA_PREFILL_MMQ=1 STRATA_PREFILL_FUSED=1)
+    if(STRATA_BUILD_TESTS)
+      # #136: the fused int8 experts against a double-precision reference and the MMQ path (exit 77 below sm_80)
+      add_executable(prefill_fused_moe_test tests/cuda/prefill_fused_moe_test.cpp)
+      target_link_libraries(prefill_fused_moe_test PRIVATE strata_mmq ggml-base ${_strata_gpu_runtime_target})
+      add_test(NAME prefill_fused_moe_test COMMAND prefill_fused_moe_test)
+      set_tests_properties(prefill_fused_moe_test PROPERTIES SKIP_RETURN_CODE 77 TIMEOUT 300)
+      # #136: the same for the native packs' formats (IQ2_XXS IQ2_XS IQ2_S IQ3_XXS IQ3_S IQ4_XS, down Q2_0 IQ4_NL)
+      add_executable(prefill_fused_iq_test tests/cuda/prefill_fused_iq_test.cpp)
+      target_link_libraries(prefill_fused_iq_test PRIVATE strata_mmq ggml-base ${_strata_gpu_runtime_target})
+      add_test(NAME prefill_fused_iq_test COMMAND prefill_fused_iq_test)
+      set_tests_properties(prefill_fused_iq_test PROPERTIES SKIP_RETURN_CODE 77 TIMEOUT 600)
+    endif()
+    if(STRATA_BUILD_TESTS AND STRATA_MMQ_KQUANTS)
+      # the UD-Q4_K_XL expert formats through MMQ against ggml's dequantized weights (synthetic, a few seconds)
+      add_executable(prefill_mmq_kquant_test tests/cuda/prefill_mmq_kquant_test.cpp)
+      target_link_libraries(prefill_mmq_kquant_test PRIVATE strata_mmq ggml-base ${_strata_gpu_runtime_target})
+      add_test(NAME prefill_mmq_kquant_test COMMAND prefill_mmq_kquant_test)
+      set_tests_properties(prefill_mmq_kquant_test PROPERTIES TIMEOUT 120)
+    endif()
+  elseif(STRATA_PREFILL_MMQ)
+    # Compile only the pinned GGML MMQ and activation-quantizer sources used by prefill. This uses
+    # common.cuh's native HIP path without enabling a second GGML HIP backend.
+    if(STRATA_GGML_DIR)
+      set(_strata_gg "${STRATA_GGML_DIR}/ggml")
+    else()
+      set(_strata_gg "${strata_llamacpp_SOURCE_DIR}/ggml")
+    endif()
+    set(_strata_mmq "${_strata_gg}/src/ggml-cuda")
+    if(NOT EXISTS "${_strata_mmq}/common.cuh" OR NOT EXISTS "${_strata_mmq}/quantize.cu")
+      message(FATAL_ERROR "The selected GGML source is missing the ggml-cuda MMQ sources: ${_strata_mmq}")
+    endif()
+    set(_strata_mmq_sources
+      src/prefill/moe_mmq.cu
+      src/prefill/ggml_cuda_host.cu
+      "${_strata_mmq}/quantize.cu")
+    foreach(_t IN LISTS _strata_mmq_types)
+      list(APPEND _strata_mmq_sources "${_strata_mmq}/template-instances/mmq-instance-${_t}.cu")
+    endforeach()
+    set_source_files_properties(${_strata_mmq_sources} PROPERTIES LANGUAGE HIP)
+    add_library(strata_mmq STATIC ${_strata_mmq_sources})
+    target_include_directories(strata_mmq
+      PRIVATE "${_strata_gg}/include" "${_strata_gg}/src" "${_strata_mmq}"
+      PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/include)
+    target_compile_definitions(strata_mmq PRIVATE GGML_USE_HIP=1)
+    if(STRATA_ORCA_Q4KS_MMQ)
+      target_compile_definitions(strata_mmq PRIVATE STRATA_ORCA_Q4KS_MMQ=1)
+    endif()
+    target_link_libraries(strata_mmq PUBLIC ggml-base hip::host roc::hipblas)
+    target_link_libraries(strata_prefill PUBLIC strata_mmq)
+    target_compile_definitions(strata_prefill PUBLIC STRATA_PREFILL_MMQ=1)
+
+    # Numerical diagnostic: synthetic layouts by default, optional first-expert IQ3_XXS/IQ4_NL check with a pack.
+    add_executable(hip_prefill_mmq_parity tests/hip/prefill_mmq_parity.cpp)
+    target_include_directories(hip_prefill_mmq_parity PRIVATE "${_strata_gg}/include")
+    target_link_libraries(hip_prefill_mmq_parity PRIVATE strata_prefill)
+    add_test(NAME hip_prefill_mmq_parity COMMAND hip_prefill_mmq_parity)
+    set_tests_properties(hip_prefill_mmq_parity PROPERTIES TIMEOUT 120)
+    list(APPEND _strata_hip_tests hip_prefill_mmq_parity)
+  endif()
+  if(STRATA_ENABLE_HIP)
+    # Optional real-model diagnostic: pass the GGUF shard containing the native PLE key.
+    add_executable(hip_ple_iq4 tests/hip/ple_iq4.cpp)
+    target_link_libraries(hip_ple_iq4 PRIVATE strata_kernels strata_hip_runtime ggml-base)
+  endif()
+  if(STRATA_ENABLE_CUDA OR STRATA_ENABLE_HIP)
     add_executable(native_expert_parity src/kernels/native_expert_parity.cpp)
-    target_include_directories(native_expert_parity PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
-    target_link_libraries(native_expert_parity PRIVATE strata_kernels strata_kernels_cpu strata_artifact CUDA::cudart)
+    target_include_directories(native_expert_parity PRIVATE ${_strata_gpu_include_directories})
+    target_link_libraries(native_expert_parity PRIVATE strata_kernels strata_kernels_cpu strata_artifact ${_strata_gpu_runtime_target})
+    if(STRATA_BUILD_TESTS AND STRATA_ENABLE_CUDA)
+      # Unsloth UD-Q4_K_XL's three expert pairs (and Q8_0 alone, its embedding) on synthetic ggml-quantized weights
+      # at 2560/640: GPU and CPU experts against the float reference, GPU dequantizers against to_float
+      foreach(_pair q4_K/q5_1 q4_K/q8_0 q5_K/q8_0 q8_0/q8_0)
+        string(REPLACE "/" "_" _name "${_pair}")
+        add_test(NAME native_expert_parity_${_name} COMMAND native_expert_parity --synthetic ${_pair})
+        set_tests_properties(native_expert_parity_${_name} PROPERTIES TIMEOUT 120)
+      endforeach()
+      add_test(NAME native_expert_parity_q5_1_min COMMAND native_expert_parity --q5_1-min)
+      # #290: the BF16 token embedding (--embd-gguf) gathered and dequantized bit for bit
+      add_test(NAME native_expert_parity_bf16_embd COMMAND native_expert_parity --bf16-embd)
+      # a pair without GPU kernels (Q6_K gate/up) is refused by native_expert_supported, not run
+      add_test(NAME native_expert_parity_refuses_q6_K COMMAND native_expert_parity --synthetic q6_K/q8_0)
+      set_tests_properties(native_expert_parity_refuses_q6_K PROPERTIES WILL_FAIL TRUE)
+    endif()
+  endif()
+  # #296's real-weight checks (OrcaRouter Q4_K_S files; docs/ORCA_Q4_K_S.md) - test programs, not part of a release
+  if(STRATA_ENABLE_CUDA AND STRATA_BUILD_TESTS)
+    add_executable(ple_q5_parity src/kernels/ple_q5_parity.cpp)
+    target_link_libraries(ple_q5_parity PRIVATE strata_kernels strata_artifact ggml-base ${_strata_gpu_runtime_target})
+    add_executable(q5_projection_parity src/kernels/q5_projection_parity.cpp)
+    target_include_directories(q5_projection_parity PRIVATE ${_strata_gpu_include_directories})
+    target_link_libraries(q5_projection_parity PRIVATE strata_kernels strata_artifact ggml-base ${_strata_gpu_runtime_target})
+    # #291: the FP8 PLE table (tools/ple_fp8_pack.py) through PleTable vs the checkpoint's own values
+    add_executable(ple_fp8_parity src/kernels/ple_fp8_parity.cpp)
+    target_link_libraries(ple_fp8_parity PRIVATE strata_kernels ${_strata_gpu_runtime_target})
   endif()
 endif()
 
-if(STRATA_ENABLE_CUDA AND STRATA_BUILD_TESTS)
+if(STRATA_ENABLE_CUDA AND STRATA_BUILD_TESTS AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/bench/micro/hit_cpu_order_parity.cu")
   add_executable(hit_cpu_order_parity bench/micro/hit_cpu_order_parity.cu)
-  target_link_libraries(hit_cpu_order_parity PRIVATE strata_kernels strata_kernels_cpu CUDA::cudart)
+  target_link_libraries(hit_cpu_order_parity PRIVATE strata_kernels strata_kernels_cpu ${_strata_gpu_runtime_target})
 endif()
 
 if(STRATA_BUILD_TESTS)
   # Plan v0.3 P6: one expert applied to several tokens, bitwise vs the single-token kernel (synthetic, CPU only).
   # plan v0.3 P5: the device dequantizers vs the artifact's CPU dequantizers on the real shards (not a ctest: needs
   # the 38 GB model; run it by hand or from tools/gpu_session.py)
-  add_executable(dequant_bf16_test src/kernels/dequant_bf16_test.cpp)
-  target_link_libraries(dequant_bf16_test PRIVATE strata_kernels strata_artifact CUDA::cudart)
+  if(STRATA_ENABLE_CUDA OR STRATA_ENABLE_HIP)   # #93: it needs the device dequantizers and a GPU runtime
+    add_executable(dequant_bf16_test src/kernels/dequant_bf16_test.cpp)
+    target_link_libraries(dequant_bf16_test PRIVATE strata_kernels strata_artifact ${_strata_gpu_runtime_target})
+  endif()
 
   add_executable(expert_multi_test src/kernels/cpu/expert_multi_test.cpp)
   target_link_libraries(expert_multi_test PRIVATE strata_kernels_cpu)
@@ -528,9 +1038,23 @@ if(STRATA_BUILD_TESTS)
   add_executable(pool_test src/kernels/cpu/pool_test.cpp)
   target_link_libraries(pool_test PRIVATE strata_kernels_cpu)
   add_test(NAME pool_test COMMAND pool_test --selftest WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR})
+  # issue #29: tiny batches with sleeping workers, as on a big-VRAM card; a hang fails it
+  add_executable(pool_stress src/kernels/cpu/pool_stress.cpp)
+  target_link_libraries(pool_stress PRIVATE strata_kernels_cpu)
+  add_test(NAME pool_stress COMMAND pool_stress 20)
+
+  if(UNIX AND NOT APPLE)
+    add_executable(direct_file_async_test tests/platform/direct_file_async.cpp src/platform/direct_file.cpp)
+    target_include_directories(direct_file_async_test PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/include)
+    target_link_libraries(direct_file_async_test PRIVATE Threads::Threads)
+    add_test(NAME direct_file_async_test COMMAND direct_file_async_test)
+  endif()
 endif()
 
 if(STRATA_BUILD_TESTS)
   enable_testing()
-  add_subdirectory(tests)
+  # Model-level checks may be absent from source distributions.
+  if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/tests/CMakeLists.txt")
+    add_subdirectory(tests)
+  endif()
 endif()

Dockerfile

@@ -0,0 +1,115 @@
+# syntax=docker/dockerfile:1
+#
+# Strata: Qwen3.8-Flash-Next on NVIDIA GPUs (RTX 30/40/50, 12+ GB VRAM; two or
+# three cards can share one model, 8 GB each - docs/MULTI_GPU.md).
+#
+# The engine is compiled during docker build, so the first container start only
+# downloads the model (~70 GB) and starts the server. docker build has no GPU,
+# so the CUDA architectures are fixed here instead of read from nvidia-smi: the
+# engine is a fat binary with a cubin per listed arch, and the runtime picks the
+# one matching your card. Narrow CUDA_ARCHITECTURES to your card for a faster
+# build; a card outside the set needs a rebuild with its own arch.
+#
+# Build:
+#   docker build -t strata .
+#   docker build -t strata --build-arg CUDA_ARCHITECTURES=89 .        # RTX 40 only
+#
+# Run (host needs an NVIDIA driver >= 580 and nvidia-container-toolkit):
+#   docker run --rm --gpus all \
+#     -p 8080:8080 \
+#     --ulimit memlock=-1 \
+#     -v strata-data:/data \
+#     -e MODEL=IQ2_XS \
+#     strata
+#
+# Setup choices are env vars, read by docker-entrypoint.sh: FAMILY, MODEL, CONTEXT,
+# VISION (no | yes | cpu), KV (int8 | q4_0 | k8v4), GPU (one card) or GPUS ("0,2"
+# or "all", with LAYER_SPLIT), LOW_RAM (auto | on | off), HOST, PORT, API_KEY.
+#
+# Only the model files, the prepared pack, the MTP layer and the install config
+# live in the /data volume; the engine is part of the image. Strata loads 32-62 GB
+# into RAM, so a capped container needs -e LOW_RAM=on: setup.py reads the RAM from
+# /proc/meminfo, which here is the host's total, not the container's limit. Add an
+# API key before exposing the port to a network: -e API_KEY=<secret>. Pass
+# -e REINSTALL=1 to change the model settings later.
+#
+# --gpus all on a host with two usable cards: setup takes both (the layer split is
+# its recommended default). Pin one card with -e GPU=0, or name them with
+# -e GPUS=0,2. A volume set up for one card switches to the pair on its first start
+# on a two-card host unless GPU or GPUS pins it. LOW_RAM=on runs on one card.
+
+FROM nvidia/cuda:13.0.0-devel-ubuntu24.04
+
+# STRATA_EXECV=1: setup.py replaces itself with the server, so the server is PID 1
+# and docker stop's SIGTERM reaches it (see setup.start). Normal Linux starts, which
+# don't set it, keep spawning the server as a child.
+ENV DEBIAN_FRONTEND=noninteractive PYTHONUNBUFFERED=1 LANG=C.UTF-8 STRATA_EXECV=1
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        build-essential ca-certificates curl git libatomic1 libgomp1 \
+        python3 python3-pip python3-venv unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /opt/strata
+COPY . .
+
+# RTX 20 (75), RTX 30 (86), RTX 40 (89), RTX 50 (120), plus 80 for A-series. CMakeLists
+# refuses anything below 75. BUILD_VISION=0 skips the image encoder build.
+ARG CUDA_ARCHITECTURES=75;80;86;89;120
+ARG BUILD_VISION=1
+
+RUN python3 -m venv .venv \
+    && .venv/bin/pip install --no-cache-dir --upgrade pip \
+    && .venv/bin/pip install --no-cache-dir -r requirements.txt \
+    && chmod +x setup.sh docker-entrypoint.sh
+
+# llama.cpp at the pinned commit, then the engine and the image encoder, built
+# exactly the way setup.py builds them. BUILD.json is what setup.py reads to
+# decide whether an engine is current: source=local with a matching src hash
+# means the first start reuses it instead of recompiling.
+RUN .venv/bin/python - <<'PYEOF'
+import json, os, pathlib, shutil
+import setup
+
+llama = setup.get_llama_cpp()
+nvcc, _ = setup.find_nvcc()
+arch = os.environ.get("CUDA_ARCHITECTURES", "75;80;86;89;120").strip().strip('"').replace(",", ";")
+vision = "gpu" if os.environ.get("BUILD_VISION", "1") == "1" else "none"
+
+setup.cmake_build(setup.ROOT, setup.ROOT / "build", "strata",
+    ["-DSTRATA_ENABLE_CUDA=ON", "-DSTRATA_BUILD_TESTS=OFF",
+     f"-DCMAKE_CUDA_ARCHITECTURES={arch}", f"-DCMAKE_CUDA_COMPILER={nvcc}",
+     f"-DSTRATA_GGML_DIR={llama}"], None, "build-strata.bat")
+if vision != "none":
+    setup.cmake_build(setup.ROOT / "tools" / "vision", setup.ROOT / "build-vision", "strata-vision",
+        [f"-DLLAMA_DIR={llama}", "-DSTRATA_VISION_CUDA=ON",
+         f"-DCMAKE_CUDA_ARCHITECTURES={arch}", f"-DCMAKE_CUDA_COMPILER={nvcc}"], None, "build-vision.bat")
+
+eng = setup.ROOT / "engine"
+eng.mkdir(exist_ok=True)
+shutil.copy2(setup.ROOT / "build" / setup.EXE, eng / setup.EXE)
+if vision != "none":
+    shutil.copy2(setup.ROOT / "build-vision" / "bin" / setup.VEXE, eng / setup.VEXE)
+bindir = pathlib.Path(nvcc).parent
+meta = {"source": "local", "version": setup.source_version(),
+        "archs": [int(a.split("-")[0]) for a in arch.split(";") if a.split("-")[0].isdigit()], "vision": vision,
+        "cuda_dirs": [str(d) for d in (bindir, bindir / "x64", bindir.parent / "lib64") if d.is_dir()],
+        "src": setup.source_hash(setup.ENGINE_SOURCES),
+        "vision_src": setup.source_hash(setup.VISION_SOURCES) if vision != "none" else None}
+(eng / "BUILD.json").write_text(json.dumps(meta, indent=1))
+PYEOF
+
+# the cmake trees are build-time only; the engine itself is what the container needs
+RUN rm -rf build build-vision
+
+VOLUME ["/data"]
+EXPOSE 8080
+
+# /health is answered before the API key gate, so it works with or without one.
+# The port only opens after the model loads (1-3 minutes, longer on a first run),
+# so the start period is generous: a too short one marks a still-loading container
+# unhealthy and a restart policy would kill it mid-download.
+HEALTHCHECK --interval=30s --timeout=5s --start-period=600s --retries=3 \
+  CMD curl -fs "http://127.0.0.1:${PORT:-8080}/health" || exit 1
+
+ENTRYPOINT ["./docker-entrypoint.sh"]

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Niko1221 and the Strata contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -1,313 +1,185 @@
-<h1 align="center">Qwen3.8-Flash-Next (125B MoE) on a 12-24 GB NVIDIA GPU + 64 GB RAM: one-click install for Windows / Linux</h1>
+<h1 align="center">Strata</h1>
 
-<p align="center"><b>Strata</b>: an inference engine built for exactly one model and exactly this kind of PC</p>
+<p align="center"><b>Run a 125-billion-parameter AI model on your own gaming PC</b><br>
+NVIDIA or AMD graphics card (12 GB or more) · Windows or Linux · free and open source</p>
 
-A 125-billion-parameter mixture-of-experts model (6B active per token) on a **normal gaming PC**: one NVIDIA card with
-12-24 GB of VRAM, 64 GB of RAM and a desktop CPU. Strata keeps the always-used part of the model on the GPU, all 24,576
-experts in RAM, computes the experts the GPU does not hold on the CPU at the same time, and checks several tokens per
-pass with the model's own multi-token-prediction (MTP) layer. The text it produces is identical to plain greedy decoding.
+<p align="center"><a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4"><img src="docs/media/pagoda-preview.webp" width="720" alt="A voxel pagoda garden that Strata's model wrote, running in the browser"></a><br>
+<sub>A voxel pagoda garden, 1 shot prompt running on an RTX 5070 with Strata (IQ3_S, 128K context) ·
+<a href="https://github.com/Niko1221/Strata/releases/download/v0.1.10/Pagoda.mp4">full video (49 s)</a></sub></p>
 
-It serves an **OpenAI-** and **Anthropic-compatible API on localhost**, so any chat app, coding agent or script that
-talks to those APIs can use it. Optionally it also **reads images** (screenshots, photos, scanned pages) with the
-model's own vision encoder.
+Strata runs **[Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)** - a large, smart AI model that
+normally needs a server - on a normal PC. It chats, writes code, reads pictures and works with your apps and coding
+agents, and nothing leaves your PC.
 
-> **Jump to:** [Speed](#speed-measured) · [Other GPUs](#other-gpus-estimated) · [Which model?](#which-model) ·
-> [Windows](#windows) · [Linux](#linux) · [Using it](#using-it) · [Images](#images-vision) ·
-> [Troubleshooting](#troubleshooting) · [How it works](#how-it-works)
+## How fast is it?
 
----
+Measured on two ordinary gaming PCs. "Writes answers" is how fast the reply appears in a short chat; "reads your
+prompt" is how fast it takes in what you send (a 32K-token document, code or chat history). A token is about ¾ of a
+word, so 60 tokens per second is faster than you can read.
 
-## Speed (measured)
+<table>
+<tr><th>NVIDIA: RTX 5070 (12 GB), Ryzen 5 7600, 64 GB RAM</th><th>AMD: RX 9070 XT (16 GB), Ryzen 9 3900X, 47 GB RAM</th></tr>
+<tr><td>
 
-RTX 5070 **12 GB**, Ryzen 5 7600 (6 cores), 64 GB DDR5-5200, Windows. One code-agent prompt per length, 256 generated
-tokens, MTP speculative decoding on. "262K" is the model's full context window (a 259,943-token prompt).
+| Size | Writes answers | Reads your prompt |
+| --- | ---: | ---: |
+| **Q2_0** | 94 tokens/s | 2,650 tokens/s |
+| **IQ2_XS** | 79 tokens/s | 2,090 tokens/s |
+| **IQ3_XXS** | 62 tokens/s | 1,750 tokens/s |
+| **IQ3_S** | 53 tokens/s | 1,620 tokens/s |
+| **Coder** | 55 tokens/s | 2,180 tokens/s |
 
-### Prompt processing (tokens/s)
+</td><td>
 
-| Model | 1K | 4K | 32K | 64K | 128K | 262K |
-| --- | ---: | ---: | ---: | ---: | ---: | ---: |
-| **Q2_0** | 389 | 539 | 571 | 561 | 543 | 496 |
-| **IQ2_XS** | 332 | 463 | 495 | 486 | 472 | 437 |
-| **IQ3_XXS** | 285 | 410 | 435 | 427 | 414 | - |
+| Size | Writes answers | Reads your prompt |
+| --- | ---: | ---: |
+| **Q2_0** | 60 tokens/s | 1,160 tokens/s |
+| **IQ2_XS** | 52 tokens/s | 1,110 tokens/s |
+| **Coder** | 44 tokens/s | 1,420 tokens/s |
 
-### Output (tokens/s)
+</td></tr>
+</table>
 
-| Model | 1K | 4K | 32K | 64K | 128K | 262K |
-| --- | ---: | ---: | ---: | ---: | ---: | ---: |
-| **Q2_0** | 88.7 | 94.6 | 87.5 | 76.4 | 65.1 | 56.3 |
-| **IQ2_XS** | 82.0 | 78.0 | 65.3 | 63.7 | 52.0 | 48.0 |
-| **IQ3_XXS** | 64.6 | 65.6 | 57.3 | 54.4 | 44.8 | - |
+A card with more VRAM is faster: an RTX 3090 (24 GB) should write roughly 100-140 tokens per second. Long chats,
+other cards: [speed of each model](docs/MODELS.md#how-fast-is-each-size), [community results](docs/COMMUNITY_BENCHMARKS.md).
 
-IQ3_XXS at 262K is not measured: with its 43 GB of experts, a 260K-token context brings a 64 GB PC to its memory
-limit. Use up to 128K with IQ3_XXS on 64 GB.
+<p align="center"><a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50"></a><br>
+<sub>Strata is free. If it runs well on your PC, a coffee keeps the work on it going.</sub></p>
 
-Time to first token is prompt length / prompt speed: about 7 s at 4K, 55 s at 32K, 4 minutes at 128K and 9 minutes at
-262K. The raw numbers: `bench/results/`. The [paper](docs/paper/Strata-Paper.pdf) explains every number.
-
-## Other GPUs (estimated)
-
-Not measured - estimated from the runs above (same CPU and 64 GB RAM): the GPU part scaled by memory bandwidth, the CPU
-part by how many more experts the card's VRAM holds. Treat as **±20%**. Numbers are *prompt / output* tokens/s.
-
-| GPU | Model | 1K | 4K | 32K | 64K | 128K | 262K |
-| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
-| RTX 5060 Ti 16GB | Q2_0 | ~341 / ~80 | ~472 / ~87 | ~501 / ~81 | ~492 / ~72 | ~476 / ~62 | ~435 / ~53 |
-|  | IQ2_XS | ~291 / ~80 | ~406 / ~77 | ~434 / ~63 | ~426 / ~62 | ~413 / ~51 | ~383 / ~47 |
-|  | IQ3_XXS | ~249 / ~66 | ~359 / ~65 | ~381 / ~56 | ~374 / ~54 | ~363 / ~45 | - |
-| RTX 3090 24GB | Q2_0 | ~355 / ~128 | ~491 / ~140 | ~521 / ~130 | ~512 / ~115 | ~495 / ~100 | ~453 / ~85 |
-|  | IQ2_XS | ~303 / ~131 | ~422 / ~128 | ~451 / ~103 | ~444 / ~102 | ~430 / ~85 | ~398 / ~78 |
-|  | IQ3_XXS | ~260 / ~106 | ~374 / ~103 | ~396 / ~89 | ~390 / ~85 | ~378 / ~71 | - |
-
-More VRAM matters more than a faster GPU: every extra GB holds ~700 more experts, and every expert on the GPU is one the
-CPU does not have to compute. A 3090's 24 GB takes most of the CPU work away.
-
-## Which model?
-
-All three are [ISTA-DASLab's GSQ-RCO quantizations](https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF)
-of [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next).
-
-| Model | Download | RAM it uses | Speed | Quality |
-| --- | ---: | ---: | --- | --- |
-| **Q2_0** | 66 GB | ~34 GB experts + ~6 GB | fastest | good |
-| **IQ2_XS** | 68 GB | ~36 GB experts + ~6 GB | close to Q2_0 | a bit better |
-| **IQ3_XXS** | 76 GB | ~43 GB experts + ~6 GB | slower (more CPU work) | best |
-
-With 64 GB of RAM all three fit (close the browser for IQ3_XXS, and keep its context at 128K or less). With 48 GB only Q2_0 / IQ2_XS may fit. 32 GB is not enough.
-
-## Before you start
-
-You need **only an NVIDIA driver** (version 580 or newer; update it with the NVIDIA App or from
-[nvidia.com/drivers](https://www.nvidia.com/drivers)). Everything else is installed for you the first time.
+## What you need
 
 | | |
 | --- | --- |
-| GPU | NVIDIA **RTX 30, 40 or 50 series**, **12 GB VRAM or more** (8 GB runs, slowly). Measured on an RTX 5070; RTX 30/40 are untested. |
-| RAM | **64 GB** recommended (see the table above). |
-| CPU | x86-64 with AVX2 (any Intel/AMD desktop CPU from the last ~8 years). AVX-512 (Ryzen 7000/9000) is a bit faster. |
-| Disk | ~70-80 GB free for the model, ~6 GB for the MTP layer (+1 GB with images). **Q2_0 on an AVX-512 CPU** also writes a one-time ~40 GB copy of its experts for the fast CPU kernel. An NVMe SSD is strongly recommended. |
-| OS | Windows 10/11, or Linux (Ubuntu 22.04/24.04 get everything installed automatically). |
+| **Graphics card** | **NVIDIA** GeForce RTX 20, 30, 40 or 50 series, or **AMD** Radeon RX 7900 XT / XTX, RX 7800 XT / 7700 XT, RX 9060 XT, RX 9070 / 9070 XT, Radeon AI PRO R9700 or RX 6800 / 6900 series - with **12 GB of VRAM or more** |
+| **RAM** | 32 GB or more - how much decides [which model](#which-model-should-i-pick) fits; 64 GB runs every size |
+| **Disk** | about 80 GB free, on an SSD if you can (the first start is much faster) |
+| **System** | Windows 10 / 11 or Linux, and a current graphics driver from NVIDIA or AMD |
 
-What the first start installs, all inside this folder (`.venv/`, `engine/`, `third_party/`, `models/`, `packs/`, `mtp/`):
-Python 3.12 if you have none (for your user account, no admin), a private Python environment, NVIDIA's CUDA libraries
-(from pip, ~0.4 GB), the ready-made Strata engine for RTX 30/40/50, the model and the MTP draft layer. If no
-ready-made engine fits your PC, it offers to install the build tools (Visual Studio Build Tools + CUDA Toolkit on
-Windows, `build-essential` + CUDA on Ubuntu) and compiles the engine for your GPU (asks first; 20-40 minutes once).
+Everything else is installed for you. Two or three cards can share the model ([multi-GPU](docs/MULTI_GPU.md)).
+The full list: [docs/INSTALL.md](docs/INSTALL.md#what-you-need).
 
----
+## Install
 
-## Windows
+### Let your AI set it up
 
-### Double-click `START-HERE.bat`
+Use an AI coding assistant (Claude Code, Cursor, Codex, GitHub Copilot, ...)? Paste this into it:
 
-**The first time** it asks three questions and does the rest:
-
-1. **Which model?** Q2_0, IQ2_XS or IQ3_XXS (it recommends one for your RAM).
-2. **How much context?** 8K to 256K tokens (it recommends one for your VRAM).
-3. **Images?** yes / no (see [Images](#images-vision)).
-
-Then it downloads and prepares everything (the model is 66-76 GB, so the first start takes a while; an interrupted
-download continues where it stopped) and **starts the model**: your browser opens `http://127.0.0.1:8080`, a small
-page that shows it is running and lets you chat. The API is at `http://127.0.0.1:8080/v1` for your apps.
-
-**Every time after that**, `START-HERE.bat` just starts the model (30-90 s to load 34-43 GB into RAM). Nothing is
-downloaded again. Closing the window stops the model.
-
-```
-START-HERE.bat --setup                          install another model, or change context / images
-START-HERE.bat --model IQ2_XS --context 32768 --vision yes --yes     no questions
-START-HERE.bat --gguf-dir D:\models\IQ2_XS       use GGUF files you already have
-START-HERE.bat --port 8081                      another port
+```text
+Set up Strata on this PC for me: https://github.com/Niko1221/Strata - follow docs/AI_SETUP.md in that repository.
 ```
 
-With more than one model installed, it asks which one to start. `run-<model>.bat` starts a model directly.
+It checks your graphics card, RAM and disk, picks the model that fits, installs it, starts it and tells you how to
+connect your apps. AI tools can also install, start and stop Strata themselves through its
+[MCP server](docs/MCP_SERVER.md).
 
-### Chat in the terminal (optional)
+### Or do it yourself
 
-```
-.venv\Scripts\python chat.py
-```
+[Download Strata](https://github.com/Niko1221/Strata/archive/refs/heads/main.zip) and unzip it (or `git clone` it).
+**Windows:** double-click **`START-HERE.bat`**. **Linux:** run **`./setup.sh`** in the Strata folder.
 
----
+The same steps for NVIDIA and AMD: the installer finds your card and sets up the right engine for it. It asks which
+model, which size, how much context (how much text it keeps in mind) and whether it should read pictures - press
+Enter each time for the recommended answer. Then it downloads the model (~70 GB; you can stop and it continues where
+it left off) and starts it. Your browser opens the Strata app at `http://127.0.0.1:8080`.
 
-## Linux
+> **While the model starts, your PC can be slow or stop responding for 1-3 minutes** (longest the first time): Strata
+> loads 35-55 GB into your RAM and locks part of it for the graphics card. That's normal - wait, and don't close the
+> window. The window tells you what it is doing.
 
-```bash
-./setup.sh
-```
+**Next time**, run `START-HERE.bat` (or `./setup.sh`) again: it starts right away, nothing is downloaded twice. Close
+its window to stop the model. `UPDATE.bat` (`./update.sh`) updates Strata without starting it. Updating, Docker,
+several cards, where the files go and every option:
+[docs/INSTALL.md](docs/INSTALL.md).
 
-The same three questions, the same automatic install (it uses `sudo apt` for Python and, only if it has to compile,
-for the build tools), and the same start: `http://127.0.0.1:8080`. Later runs of `./setup.sh` (or `./run-<model>.sh`)
-start the model directly. Options as on Windows (`./setup.sh --setup`, `--model Q2_0 --yes`, `--gguf-dir /data/Q2_0`).
-Terminal chat: `.venv/bin/python chat.py`.
+## Which model should I pick?
 
----
+The installer recommends one for your RAM. The same model comes in sizes that are compressed more or less: smaller
+is faster, larger is a bit smarter.
+
+| Your RAM | Take | Why |
+| --- | --- | --- |
+| **32 GB** | **Coder** | it fits 32 GB, and it is made for code (with a 24 GB card, Q2_0 and IQ2_XS run too) |
+| **48 GB** | **IQ2_XS** (or Q2_0, the fastest) | the larger sizes do not fit |
+| **64 GB** | **IQ2_XS** (recommended), or IQ3_XXS / IQ3_S | every size fits; IQ3_S is the best, and the slowest |
+| **96 GB or more** | **IQ3_S**, or Unsloth's 4-bit (experimental) | room for the largest sizes with everything else open |
+
+- **[Coder](docs/MODELS.md#coder)** - a coding version with half of the experts removed: 91% of the full model's
+  SWE-bench Verified score (by its authors), fits 32 GB of RAM. Weaker outside code, including Chinese and other
+  CJK text (#438): for those, take Q2_0, IQ2_XS or IQ3_S, which keep every expert.
+- **[Swift 1.5](docs/MODELS.md#swift-15)** - a fine-tune that thinks much shorter before it answers, so you get the
+  answer sooner, at about the same quality.
+- **[Unsloth UD-Q4_K_XL](docs/MODELS.md#unsloth-ud-q4_k_xl-experimental)** (experimental) - the closest to the full
+  model, but most of it is read from the SSD while it answers: 7-8.5 tokens/s on a 64 GB PC.
+- **[OrcaRouter's Uncensored IQ3_XXS](docs/MODELS.md#orcarouter-uncensored-iq3_xxs)** - a manual setup, not in the
+  installer's menu.
+
+Sizes, downloads and what fits where: [docs/MODELS.md](docs/MODELS.md). You can add another model later with
+`SETUP.bat` (Linux: `./setup.sh --setup`).
 
 ## Using it
 
-The server listens on `http://127.0.0.1:8080` (change with `--port` in setup, or edit the run script).
+<p align="center"><img src="docs/media/runpagoda.png" width="900" alt="The Strata app's Monitor tab next to a coding agent"><br>
+<sub>The Strata app's <b>Monitor</b> (left) while a coding agent writes the pagoda garden from the video (right)</sub></p>
 
-| API | Endpoint |
-| --- | --- |
-| OpenAI Chat Completions (stream and non-stream, tools) | `POST /v1/chat/completions` |
-| Anthropic Messages (stream and non-stream, tools) | `POST /v1/messages` |
-| Model list / health | `GET /v1/models`, `GET /health` |
+- **In the browser:** `http://127.0.0.1:8080` - **Chat**, a live **Monitor** of the model and your GPU/CPU/RAM, and
+  **About** with the settings and addresses.
+- **Your apps and coding agents:** add an "OpenAI-compatible" provider with base URL **`http://127.0.0.1:8080/v1`**,
+  any API key and any model name. Apps that use Anthropic's API: `http://127.0.0.1:8080/v1/messages` (Claude Code:
+  `ANTHROPIC_BASE_URL=http://127.0.0.1:8080`).
+- **Thinking:** choose **off, low, medium or high** in the chat menu or your app's "reasoning effort". Off is
+  fastest; high is best for hard questions.
+- **Pictures:** say yes to "Images?" in setup, then click **Picture** in the chat, or attach them in your app
+  (AMD cards: on Linux through the processor, not on Windows yet).
+- **From your phone or another PC:** `START-HERE.bat --setup --host 0.0.0.0 --api-key <secret>` - always with a key.
+- **Good to know:** it answers one request at a time. The first message of a chat is read in full (about 1 minute
+  per 30,000 tokens); follow-ups start in seconds.
 
-```bash
-curl http://127.0.0.1:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
-  "model": "strata", "messages": [{"role": "user", "content": "Write a haiku about GPUs."}], "max_tokens": 512 }'
-```
+More: [where your chats are stored](docs/INSTALL.md#where-things-are-stored), [the API](docs/DETAILS.md#using-it).
 
-```python
-from openai import OpenAI
-client = OpenAI(base_url="http://127.0.0.1:8080/v1", api_key="none")
-r = client.chat.completions.create(model="strata", messages=[{"role": "user", "content": "Hello!"}])
-print(r.choices[0].message.content)
-```
+## Something went wrong?
 
-- **Thinking levels: none, low, medium, high.** The model thinks before it answers (streamed as
-  `reasoning_content`, Anthropic: `thinking` blocks). Choose how much per request - in the chat page (the "Thinking"
-  menu), in `chat.py` (`/think low`), or over the API:
+- **My PC froze the first time Strata started.** Normal while it loads the model: wait, don't close the window.
+  Still frozen after 10 minutes? Restart the PC, close other programs and try again, or pick a smaller size.
+- **It stopped while downloading or installing.** Run `START-HERE.bat` (or `./setup.sh`) again: it continues where
+  it stopped.
+- **It's very slow and the disk light keeps blinking, or "the engine stopped unexpectedly".** Not enough free RAM:
+  close other programs (browsers use a lot), or pick a smaller size (Q2_0 or IQ2_XS).
+- **It says port 8080 is already in use.** Strata is already running - look for its window.
 
-  | API | how |
-  | --- | --- |
-  | OpenAI | `"reasoning_effort": "none" \| "low" \| "medium" \| "high"` (also `"reasoning": {"effort": ...}`, or `"chat_template_kwargs": {"enable_thinking": false}`) |
-  | Anthropic | `"output_config": {"effort": "low" \| "medium" \| "high"}`, `"thinking": {"type": "disabled"}`, or `"thinking": {"type": "enabled", "budget_tokens": N}` (under 2K = low, under 8K = medium, more = high) |
+More problems and their fixes: [docs/TROUBLESHOOTING.md](docs/TROUBLESHOOTING.md). Still stuck? Open an
+[issue](https://github.com/Niko1221/Strata/issues) and attach `strata-<model>.log` from the Strata folder.
 
-  Without a setting the model uses its own default, **high**. `none` answers at once (fastest); `low` keeps the thinking
-  short. The levels are instructions the model was trained with, not a hard token limit: on easy questions all three
-  think briefly, on hard ones `high` thinks longest and is most accurate.
-- **Chat apps.** Any app with an "OpenAI-compatible" provider works: base URL `http://127.0.0.1:8080/v1`, any API key.
-- **Context.** Chosen in setup (8K-262K). Requests longer than that are refused, never silently cut.
-- **From other devices / the internet.** The server listens on your PC only (`127.0.0.1`). To reach it from elsewhere,
-  put a tunnel in front of it, for example [cloudflared](https://developers.cloudflare.com/cloudflare-one/connections/connect-networks/do-more-with-tunnels/trycloudflare/):
-  `cloudflared tunnel --url http://127.0.0.1:8080`. **Set a key first**, or anyone with the link can use your PC:
-  add `"api_key": "some-long-secret"` to `strata-<model>.json` (or set the `STRATA_API_KEY` environment variable);
-  clients then send it as their API key.
+## How does it work?
 
-**Current limits (v1):** one request at a time; greedy decoding (temperature is ignored); every request processes its
-whole prompt again (no conversation cache yet, so long chats have a long time-to-first-token); images only when set up
-with them (below); no video.
+Models like this one normally run on servers with hundreds of gigabytes of graphics memory. Your graphics card has
+12-24 GB. Strata makes it fit by **sharing the work across your whole PC** - like a kitchen, where the things you use
+all the time stay on the counter and the rest waits in the pantry.
 
----
+<p align="center"><img src="docs/media/how-it-works.svg" width="860" alt="The model's 24,576 experts: the busiest on the graphics card, all of them in RAM, a lookup table on the SSD"></p>
 
-## Images (vision)
+- **The model is a team of 24,576 small specialists ("experts"),** and each word needs only 10 of them.
+- **Your graphics card** keeps the few thousand experts that are asked most often; **your RAM** holds all of them,
+  and **your processor** works on the rest at the same time. **Your SSD** holds a big lookup table.
 
-The model has a vision encoder: [`mmproj-Qwen3.8-Flash-Next-BF16.gguf`](https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF)
-(0.9 GB, a 27-layer ViT plus the projector into the language model). It is **optional**: say yes when the setup asks
-"Images?", or run it again with `--vision gpu` (or `--vision cpu`). The setup downloads the encoder, builds a small
-helper (`strata-vision`, from llama.cpp's `mtmd` library) and adds it to your start script. Nothing else changes.
+<p align="center"><img src="docs/media/guess-and-check.svg" width="860" alt="A small helper guesses the next words; the big model checks them all at once and keeps the right ones"></p>
 
-| Encoder on | Time per picture | Cost |
-| --- | --- | --- |
-| **GPU** (recommended) | **0.1-0.5 s** (up to 1,024 image tokens) | ~1.4 GB of VRAM is kept free for it, so the expert cache is smaller: text output is a few % slower (table below) |
-| CPU | 10-30 s (pictures are scaled down to ~300 image tokens) | nothing on the GPU |
+- **Guess, then check:** a small helper guesses the next few words and the big model checks them all at once, so
+  you get the same answer, 1.6-1.8x sooner.
+- **Long texts are read in big pieces** (up to 8,192 tokens at a time): over 1,000 tokens per second.
 
-A picture becomes up to 1,024 tokens of the context (a 640x480 photo: 300). The same picture sent again, as chat apps
-do on every turn, is encoded only once.
+The longer explanation: [docs/HOW_IT_WORKS.md](docs/HOW_IT_WORKS.md). Every part and its numbers: [the
+details](docs/DETAILS.md#how-it-works) and the [paper](docs/paper/Strata-Paper.pdf).
 
-### Sending a picture
+## Credits and license
 
-**Terminal chat:** type `/image <path to a picture>`, press Enter, then type your question.
+The model is [Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) by the Qwen team, compressed by
+[ISTA-DASLab](https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF), UkisAI (Swift 1.5) and Unsloth;
+Strata is built with parts of [llama.cpp / ggml](https://github.com/ggml-org/llama.cpp). All credits:
+[docs/HOW_IT_WORKS.md](docs/HOW_IT_WORKS.md#credits). Strata is open source under the [MIT License](LICENSE); a few
+parts and every model carry their own licenses ([which ones](docs/HOW_IT_WORKS.md#license)).
 
-```
-you> /image C:\Users\me\Pictures\receipt.jpg
-(picture attached: receipt.jpg - now type your question)
-you> What is the total on this receipt?
-```
+## Support Strata
 
-**OpenAI API** (an `image_url` part: a `data:` URL, an `http(s)://` URL or a local file path):
+Strata is free and open source. If it is useful to you, you can support its development:
 
-```python
-import base64
-from openai import OpenAI
-client = OpenAI(base_url="http://127.0.0.1:8080/v1", api_key="none")
-img = base64.b64encode(open("photo.jpg", "rb").read()).decode()
-r = client.chat.completions.create(model="strata", messages=[{"role": "user", "content": [
-    {"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{img}"}},
-    {"type": "text", "text": "What is in this picture?"}]}])
-print(r.choices[0].message.content)
-```
-
-**Anthropic API:** an `image` block with a `base64` (or `url`) source, as usual.
-
-JPEG, PNG, BMP and GIF work (WebP does not). Chat apps with image upload work the same way.
-
-### Speed with images on (4K context, measured)
-
-| Model | Prompt tok/s, images off | Prompt tok/s, images on | Output tok/s, images off | Output tok/s, images on |
-| --- | ---: | ---: | ---: | ---: |
-| **Q2_0** | 541 | 531 | 90.7 | 86.9 |
-| **IQ2_XS** | 468 | 458 | 77.1 | 74.0 |
-| **IQ3_XXS** | 411 | 401 | 64.7 | 59.5 |
-
-"Off" is the published setup; "on" runs with the encoder loaded on the GPU and its VRAM kept free, so ~1,000 fewer
-experts fit in VRAM and the CPU computes a few more per token: 2-8% slower. For text, turning images on changes
-nothing else (the output is bit-identical when the VRAM is the same).
-
-A question about a picture (a 640x480 newspaper page = 300 image tokens, a 328-token prompt; the whole request, encoder
-on the GPU, measured through the API):
-
-| Model | Answer ("MEN WALK ON MOON") | Picture encoded | Prompt | Output |
-| --- | ---: | ---: | ---: | ---: |
-| **Q2_0** | 1.9 s | ~0.1 s | 207 tok/s | 65-73 tok/s |
-| **IQ2_XS** | 2.3 s | ~0.1 s | 173 tok/s | 50-60 tok/s |
-| **IQ3_XXS** | 2.7 s | ~0.1 s | 144 tok/s | 41-50 tok/s |
-
-(Short prompts run below the 4K prompt speed: a 2,048-token chunk is where the prompt path is efficient. Short answers
-run below the long-output speed: the first rounds have no draft yet.)
-
-**How it works inside:** the encoder turns the picture into rows of the same width as the model's word embeddings;
-Strata puts them where the prompt has `<|image_pad|>` tokens and gives each one its 2-D position (row and column in the
-picture; the model uses interleaved M-RoPE). Answers match llama.cpp's multimodal implementation token for token on our
-test images.
-
----
-
-## Troubleshooting
-
-| Symptom | What to do |
-| --- | --- |
-| `the NVIDIA driver is too old` | Update the driver (NVIDIA App or nvidia.com/drivers), restart, run `START-HERE.bat` again. |
-| Python or the build tools could not be installed | Install what it names (links are printed), then run it again. Everything already done is kept. |
-| `port 8080 is already in use` | Strata is already running (look for its window), or another program uses the port: `START-HERE.bat --port 8081`. |
-| `cudaHostRegister ... out of memory` in the log | Normal on Windows: the engine pins the experts in per-layer slices instead. Only a problem if the load then fails. |
-| The first start takes minutes | It is reading 34-43 GB into RAM; the second start is faster while the files are in the OS cache. |
-| Slow output, disk light busy | Not enough free RAM: close other programs, or choose Q2_0 / IQ2_XS. |
-| `prompt ... exceeds the context` | The request is longer than the context you chose: run setup again with a bigger `--context`. |
-| Slower than the tables | The monitor plugged into the GPU and other GPU programs take VRAM from the expert cache; RAM running below its rated speed (enable EXPO/XMP in the BIOS) slows the CPU half. |
-| `this server was started without the vision encoder` | The model was set up for text only: run setup again with `--vision gpu`. |
-| A picture is refused or `cannot read the image` | Use JPEG, PNG, BMP or GIF (not WebP). |
-| Pictures are slow (10-30 s) | The encoder runs on the CPU: run setup again with `--vision gpu` (needs ~1.4 GB of VRAM). |
-| Anything else | The engine log is `strata-<model>.log` in this folder. |
-
----
-
-## How it works
-
-<p align="center"><img src="docs/paper/tiers.svg" width="760" alt="memory tiers"></p>
-
-- **GPU (VRAM):** attention and DeltaNet mixers, the gated-residual weights, routers, shared experts, output head, the MTP
-  draft layer, the KV cache, and an **expert cache** that fills the rest of VRAM with the most-used experts (it adapts to
-  the conversation while you chat).
-- **RAM:** all 24,576 experts, pinned. The CPU computes the experts that are not on the GPU **in place**, at the same time
-  as the GPU works on the cached ones (AVX-512 / AVX2 kernels, ggml's for the i-quants).
-- **SSD:** the 28.8 GB n-gram table, read a few rows per token through the OS cache.
-- **Speculation:** the model's own MTP layer drafts up to 3 tokens; one pass over all 48 layers checks them. 2.4-3.2
-  tokens per pass on average.
-- **Prompts** are processed in 2,048-token chunks with the experts streamed to the GPU over PCIe.
-
-The full story, with measurements, bottlenecks and what comes next: **[docs/paper/Strata-Paper.pdf](docs/paper/Strata-Paper.pdf)**.
-
----
-
-## Credits and licenses
-
-- Model: [Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) by the Qwen team; quantizations:
-  [ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF](https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF).
-  Their licenses apply to the weights.
-- [llama.cpp / ggml](https://github.com/ggml-org/llama.cpp) (MIT): the i-quant formats, the GPU dot products and
-  dequantizers transcribed in `src/kernels/cuda/iq_kernels.cu`, the CPU backend linked for the i-quant experts, the
-  `mtmd` library behind the image encoder (`tools/vision/`), and `gguf-py` used by the tools. See
-  `third_party/ggml/LICENSE`.
-- Ideas from [Splash](https://github.com/incoai/splash), [ninfer](https://github.com/Neroued/ninfer) and
-  [HyperQwen](https://github.com/syv-ai/HyperQwen); references in the paper.
+<p align="center"><a href="https://buymeacoffee.com/strataengine"><img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50"></a></p>

SETUP.bat

@@ -0,0 +1,3 @@
+@echo off
+rem Change settings or install another model (the same as START-HERE.bat --setup).
+call "%~dp0START-HERE.bat" --setup %*

START-HERE.bat

@@ -1,6 +1,6 @@
 @echo off
 rem Strata for Windows: the first run installs everything and starts the model; later runs just start it.
-rem Needs only an NVIDIA driver. Python is installed for your user account if it is missing (no admin needed).
+rem Needs only an NVIDIA or AMD graphics driver. Python is installed for your user account if it is missing (no admin needed).
 setlocal
 title Strata
 cd /d "%~dp0"

UPDATE.bat

@@ -0,0 +1,36 @@
+@echo off
+rem Update Strata without starting the model (#475): the newest code (git pull, when this folder is a git clone), then
+rem what START-HERE.bat does before a start - the engine (a new one when this version needs it), the Python packages,
+rem each installed model's settings and draft subset. The model files are not touched. Start the model later with
+rem START-HERE.bat. Options are passed on to setup.py.
+setlocal
+title Strata - update
+cd /d "%~dp0"
+rem All of it in one block: cmd reads a .bat file while it runs it, and the git pull can change this very file.
+(
+  if exist ".git" (
+    where git >nul 2>nul || (
+      echo  This folder is a git clone, but git is not on PATH: install Git for Windows ^(winget install Git.Git^)
+      echo  or run "git pull" here yourself, then run UPDATE.bat again.
+      pause
+      exit /b 1
+    )
+    echo  Getting the newest Strata ^(git pull^) ...
+    git pull --ff-only || (
+      echo.
+      echo  git pull did not succeed ^(the reason is above^): nothing was updated. Files you changed here can stop it:
+      echo  "git status" lists them.
+      pause
+      exit /b 1
+    )
+  ) else (
+    echo  This copy of Strata was not made with git, so it cannot fetch new files itself. Download the newest one:
+    echo    https://github.com/Niko1221/Strata/archive/refs/heads/main.zip
+    echo  unzip it anywhere and run START-HERE.bat ^(or UPDATE.bat^) in it: it finds the model files in Strata-data
+    echo  and sets itself up the same way - nothing big is downloaded again.
+    echo  Checking this copy's engine and settings meanwhile ...
+  )
+  call "%~dp0START-HERE.bat" --update %*
+  if not errorlevel 1 pause
+  exit /b
+)

bench/results/2026-09-27-cache-parity/README.md

@@ -0,0 +1,29 @@
+# 2026-09-27: expert cache on vs off (issue #23)
+
+Question: the engine used to warn at start that "the GPU hit path is NOT CORRECT ... any OUTPUT from it is not". Does
+the expert cache make the output worse?
+
+Method: teacher-forced. `strata --dump-logits` runs a text one token at a time and writes the logits at every
+position, so both runs see the same text and each position's next-token prediction can be compared:
+- cache off: no `--expert-cache`, every routed expert computed by the CPU pool;
+- cache on: `--expert-profile pack/profile-decode-8k.bin --expert-cache auto`, 5,130 of the most-used experts on the GPU.
+
+Q2_0, RTX 5070, the first 512 or 1,024 tokens of three frozen prompts.
+
+| Text | Same top-1 | Median KL | Perplexity off | Perplexity on | On - off (nats per token) |
+|---|---:|---:|---:|---:|---:|
+| Code agent (512) | 97.7% | 3.0e-04 | 2.672 | 2.630 | -0.016 +- 0.019 |
+| Long document (1,024) | 96.2% | 4.5e-03 | 29.844 | 29.843 | -0.000 +- 0.005 |
+| Chat (1,024) | 95.0% | 4.4e-03 | 18.380 | 18.301 | -0.004 +- 0.006 |
+| **All 2,557 tokens** | | | | | **-0.005 +- 0.005** |
+
+Reading:
+- The cache changes which of two nearly equal candidates wins at 2-5% of positions. Most of the flips have a top-2
+  margin under 0.5 logits (`logits-compare.txt`).
+- Neither side is more accurate: the perplexity of the actual text is the same, within 1 standard error.
+- A GPU expert and the CPU compute the same quantized expert with a different floating-point order. Both are an
+  equally good approximation of the model.
+
+The start-up warning dates from round 328, when the GPU hit path had a real fault (divergence from token 0). That fault
+was fixed long ago and is covered by `expert_parity`, `native_expert_parity` and the grouped kernels' tests. Since
+engine 0.1.8 the warning is replaced by one accurate line.

bench/results/2026-09-27-esp/README.md

@@ -0,0 +1,36 @@
+# 2026-09-27: the experimental-speed-projection control vector
+
+The package's `Qwen3.8-Flash-Next-experimental-speed-projection.gguf` (47 unit directions of 2,560, one per layer),
+applied as `--control-vector-scaled FILE:1.0 --control-vector-layer-range 4 44 --cvec-mode project --cvec-dir
+per-layer`: after layers 4-44, every hyper-connection stream `h -= (h . v) v`. The package describes it as a
+refusal-direction projection. Q2_0, RTX 5070, fixed experts (`--expert-cache 4000 --pcie-frac 0 --adapt-swaps 0`).
+
+## Kernel (`cvec_parity`)
+
+Project and add against a double-precision reference (max error 5e-7); with the switch off, the kernel's pending
+residual write is bitwise the fused hyper-connection read's own, so a loaded-but-off vector is the stock engine.
+
+## Teacher-forced, stock vs projection (`esp_kl.py`, per-token path)
+
+| Text | Tokens | Same top-1 | Mean KL | p99 KL | Perplexity stock | With the vector |
+|---|---:|---:|---:|---:|---:|---:|
+| Code | 511 | 92.8% | 0.098 | 1.69 | 2.594 | 2.982 |
+| Document | 1,023 | 89.8% | 0.058 | 0.84 | 30.299 | 30.984 |
+| Chat | 1,023 | 89.6% | 0.050 | 0.33 | 18.409 | 18.485 |
+| **All** | 2,557 | | **0.063** (max 4.10) | 1.00 | | |
+
+The package reports a mean KL of 0.186 on its own neutral text with UD-Q2_K_XL in llama.cpp.
+
+Cost on the same tokens: 44.34 -> 44.26, 46.74 -> 46.56, 47.02 -> 46.83 tok/s (0.2-0.4%).
+
+## Serve (`esp_serve.py`)
+
+- Loaded and switched off (`cvec=0`) vs an engine without it: 65 / 65 first tokens after prefixes (verify windows
+  and the batched prompt path) and three 256-token generations **token-identical**.
+- On, off, on again: identical (the switch drops the conversation cache, so nothing computed one way is reused the
+  other way).
+- The prompt paths steer like the per-token path: served-with-vector first tokens agree with the per-token path's
+  teacher-forced argmax with the vector 62 / 65, and with the stock one 57 / 65 (stock served: 64 / 65 and 57 / 65).
+- Decode tok/s of 3 x 256-token generations: stock 66.4, on 76.7 / 81.2. **Not a like-for-like comparison:** the
+  model writes different text with it on (on the code prompt it copied a long stretch of the source, which prompt
+  lookup drafts at 91% acceptance, where stock stopped after 78 tokens). On the same tokens it is 0.2-0.4% slower.

bench/results/2026-09-27-esp/esp_kl.py

@@ -0,0 +1,85 @@
+"""Teacher-forced logits, stock vs the experimental-speed-projection vector, on the same texts and fixed experts.
+
+    python esp_kl.py run      # the dumps (per-token path)
+    python esp_kl.py compare  # same top-1, mean/median KL(stock || projection), perplexities
+"""
+import json
+import os
+import subprocess
+import sys
+
+import numpy as np
+
+S = os.path.dirname(os.path.abspath(__file__))
+E = r"C:\Users\AI-Server\Desktop\Strata\Public\Engine"
+EXE = E + r"\build-dev\strata.exe"
+VEC = r"C:\Users\AI-Server\Desktop\Strata\experimental-speed-projection\Qwen3.8-Flash-Next-experimental-speed-projection.gguf"
+TEXTS = {"code": S + r"\p512.ids", "doc": S + r"\4k-02-long-doc-1k.ids", "chat": S + r"\4k-04-chat-1k.ids"}
+OUT = S + r"\esp-kl"
+ARMS = {"stock": [],
+        "proj": ["--control-vector-scaled", VEC + ":1.0", "--control-vector-layer-range", "4", "44",
+                 "--cvec-mode", "project", "--cvec-dir", "per-layer"]}
+
+
+def base_args():
+    c = json.load(open(S + r"\cfg-rel.json"))
+    a = list(c["args"])
+    i = a.index("--expert-cache")
+    a[i + 1] = "4000"
+    j = a.index("--prefill")          # the prompt one token at a time: a logits row per position
+    del a[j:j + 2]
+    return a + ["--pcie-frac", "0", "--adapt-swaps", "0"], c
+
+
+def run():
+    os.makedirs(OUT, exist_ok=True)
+    args, c = base_args()
+    env = dict(os.environ)
+    env["PATH"] = os.pathsep.join(c["lib_dirs"]) + os.pathsep + env["PATH"]
+    for name, path in TEXTS.items():
+        for arm, extra in ARMS.items():
+            out = f"{OUT}\\{name}-{arm}.bin"
+            if os.path.exists(out):
+                continue
+            cmd = [EXE, *args, *extra, "--tokens-file", path, "--max-new", "1", "--dump-logits", out]
+            with open(f"{OUT}\\{name}-{arm}.log", "w") as log:
+                r = subprocess.run(cmd, cwd=c["cwd"], stdout=log, stderr=subprocess.STDOUT, env=env)
+            print(name, arm, "exit", r.returncode, flush=True)
+
+
+def load(p):
+    h = np.fromfile(p, dtype=np.int32, count=2)
+    return np.memmap(p, dtype=np.float32, mode="r", offset=8, shape=(int(h[1]), int(h[0])))
+
+
+def lsm(x):
+    x = x.astype(np.float64)
+    m = x.max(axis=-1, keepdims=True)
+    return x - (m + np.log(np.exp(x - m).sum(axis=-1, keepdims=True)))
+
+
+def compare():
+    print(f"{'text':<6}{'tokens':>7}{'same top-1':>11}{'mean KL':>9}{'median KL':>11}{'p99 KL':>8}{'ppl stock':>10}{'ppl proj':>9}")
+    allkl = []
+    for name, path in TEXTS.items():
+        ids = [int(x) for x in open(path).read().replace(",", " ").split()]
+        a, b = load(f"{OUT}\\{name}-stock.bin"), load(f"{OUT}\\{name}-proj.bin")
+        n = min(a.shape[0], b.shape[0], len(ids) - 1)
+        kl, same, la, lb = [], 0, [], []
+        for r in range(0, n, 64):
+            pa, pb = lsm(a[r:min(n, r + 64)]), lsm(b[r:min(n, r + 64)])
+            kl += list((np.exp(pa) * (pa - pb)).sum(axis=1))
+            same += int((pa.argmax(1) == pb.argmax(1)).sum())
+            nxt = np.array(ids[r + 1:min(n, r + 64) + 1])
+            la += list(pa[np.arange(len(nxt)), nxt])
+            lb += list(pb[np.arange(len(nxt)), nxt])
+        kl = np.array(kl)
+        allkl += list(kl)
+        print(f"{name:<6}{n:>7}{100 * same / n:>10.1f}%{kl.mean():>9.3f}{np.median(kl):>11.4f}{np.percentile(kl, 99):>8.2f}"
+              f"{np.exp(-np.mean(la)):>10.3f}{np.exp(-np.mean(lb)):>9.3f}")
+    allkl = np.array(allkl)
+    print(f"all {len(allkl)} tokens: mean KL {allkl.mean():.3f}, max {allkl.max():.2f}, p99 {np.percentile(allkl, 99):.2f}")
+
+
+if __name__ == "__main__":
+    {"run": run, "compare": compare}[sys.argv[1]]()

bench/results/2026-09-27-esp/esp_serve.py

@@ -0,0 +1,129 @@
+"""Serve-mode checks of the experimental-speed-projection vector (exact mode: fixed experts, no PCIe share).
+
+    python esp_serve.py run ARM      # ARM = stock | vec; writes esp-serve/<arm>.json
+    python esp_serve.py compare
+
+Per arm: the first greedy token after many prefixes of three texts (short ones go through the verify windows, long
+ones through the batched prompt path; the conversation cache reuses the earlier prefix), and three 256-token greedy
+generations.  The vec arm runs everything with cvec=1, then cvec=0, then cvec=1 again.
+"""
+import json
+import os
+import sys
+import threading
+
+S = os.path.dirname(os.path.abspath(__file__))
+E = r"C:\Users\AI-Server\Desktop\Strata\Public\Engine"
+sys.path.insert(0, E)
+from serve.server import StrataEngine  # noqa: E402
+
+EXE = E + r"\build-dev\strata.exe"
+VEC = r"C:\Users\AI-Server\Desktop\Strata\experimental-speed-projection\Qwen3.8-Flash-Next-experimental-speed-projection.gguf"
+TEXTS = {"code": S + r"\p512.ids", "doc": S + r"\4k-02-long-doc-1k.ids", "chat": S + r"\4k-04-chat-1k.ids"}
+OUT = S + r"\esp-serve"
+VEC_ARGS = ["--control-vector-scaled", VEC + ":1.0", "--control-vector-layer-range", "4", "44",
+            "--cvec-mode", "project", "--cvec-dir", "per-layer"]
+
+
+def ids_of(name):
+    return [int(x) for x in open(TEXTS[name]).read().replace(",", " ").split()]
+
+
+def engine(extra):
+    c = json.load(open(S + r"\cfg-rel.json"))
+    a = list(c["args"])
+    a[a.index("--expert-cache") + 1] = "4000"
+    env = dict(os.environ)
+    env["PATH"] = os.pathsep.join(c["lib_dirs"]) + os.pathsep + env["PATH"]
+    return StrataEngine(EXE, a + ["--pcie-frac", "0", "--adapt-swaps", "0"] + extra, cwd=c["cwd"],
+                        log=OUT + r"\engine.log", env=env)
+
+
+def gen(eng, ids, n, sampling):
+    toks = [t for t in eng.generate(ids, n, sampling, threading.Event()) if t is not None]
+    return toks, dict(eng.last)
+
+
+def suite(eng, sampling):
+    out = {"first": {}, "long": {}}
+    for name in TEXTS:
+        ids = ids_of(name)
+        # short prefixes growing by 4 (each continues the live session through the verify windows), then long ones
+        # SHRINKING (no cached prefix, so each is read from 0 by the batched prompt path)
+        lens = list(range(32, 72, 4)) + list(range(min(len(ids) - 1, 700), 95, -48))
+        out["first"][name] = {L: gen(eng, ids[:L], 1, sampling)[0][0] for L in lens}
+    for name in TEXTS:
+        toks, last = gen(eng, ids_of(name)[:300], 256, sampling)
+        out["long"][name] = {"tokens": toks, "last": last}
+    return out
+
+
+def run(arm):
+    os.makedirs(OUT, exist_ok=True)
+    eng = engine(VEC_ARGS if arm == "vec" else [])
+    res = {"info": eng.info}
+    if arm == "stock":
+        res["stock"] = suite(eng, {})
+    else:
+        res["on"] = suite(eng, {"experimental_speed_projection": True})
+        res["off"] = suite(eng, {"experimental_speed_projection": False})
+        res["on2"] = suite(eng, {"experimental_speed_projection": True})
+    eng.close()
+    json.dump(res, open(f"{OUT}\\{arm}.json", "w"))
+    print(arm, "done", eng.info, flush=True)
+
+
+def compare():
+    import numpy as np
+    sys.path.insert(0, S)
+    from esp_kl import load, OUT as KL
+    st = json.load(open(OUT + r"\stock.json"))
+    ve = json.load(open(OUT + r"\vec.json"))
+    print("engine INFO (vec):", ve["info"])
+
+    def eq_first(a, b):
+        tot = same = 0
+        for name in a:
+            for L, t in a[name].items():
+                tot += 1
+                same += int(b[name][L] == t)
+        return same, tot
+
+    def eq_long(a, b):
+        return all(a[n]["tokens"] == b[n]["tokens"] for n in a), \
+            [(n, next((i for i, (x, y) in enumerate(zip(a[n]["tokens"], b[n]["tokens"])) if x != y), None)) for n in a]
+
+    print("1. vector loaded, switched off == no vector")
+    print("   first tokens: %d / %d identical" % eq_first(st["stock"]["first"], ve["off"]["first"]))
+    print("   256-token generations identical:", *eq_long(st["stock"]["long"], ve["off"]["long"]))
+    print("2. repeatable: on, off, on again")
+    print("   first tokens: %d / %d identical" % eq_first(ve["on"]["first"], ve["on2"]["first"]))
+    print("   256-token generations identical:", *eq_long(ve["on"]["long"], ve["on2"]["long"]))
+    print("3. the prompt paths against the per-token path's teacher-forced argmax (row L-1 predicts token L)")
+    for arm, first in (("stock", st["stock"]["first"]), ("on", ve["on"]["first"])):
+        for ref in ("stock", "proj"):
+            agree = tot = 0
+            for name in first:
+                d = load(f"{KL}\\{name}-{ref}.bin")
+                for L, t in first[name].items():
+                    L = int(L)
+                    if L - 1 < d.shape[0]:
+                        tot += 1
+                        agree += int(int(np.argmax(d[L - 1])) == t)
+            print(f"   served {arm:<5} vs per-token {ref:<5}: {agree} / {tot}")
+    print("4. how different the text is: 256-token generations, on vs stock, first difference")
+    print("  ", eq_long(st["stock"]["long"], ve["on"]["long"])[1])
+    print("5. decode speed (exact mode: fixed experts, no PCIe share - not the normal setup)")
+    for label, r in (("stock", st["stock"]), ("vec off", ve["off"]), ("vec on", ve["on"]), ("vec on 2", ve["on2"])):
+        g = sum(v["last"]["generated"] for v in r["long"].values())
+        ms = sum(v["last"]["decode_ms"] for v in r["long"].values())
+        acc = sum(v["last"].get("drafts_accepted", 0) for v in r["long"].values())
+        off = sum(v["last"].get("drafts_offered", 0) for v in r["long"].values())
+        print(f"   {label:<9} {g / ms * 1000:6.1f} tok/s  drafts accepted {acc}/{off}")
+
+
+if __name__ == "__main__":
+    if sys.argv[1] == "run":
+        run(sys.argv[2])
+    else:
+        compare()

bench/results/2026-09-27-esp/serve-checks.txt

@@ -0,0 +1,20 @@
+1. vector loaded, switched off == no vector
+   first tokens: 65 / 65 identical
+   256-token generations identical: True [('code', None), ('doc', None), ('chat', None)]
+2. repeatable: on, off, on again
+   first tokens: 65 / 65 identical
+   256-token generations identical: True [('code', None), ('doc', None), ('chat', None)]
+3. the prompt paths against the per-token path's teacher-forced argmax (row L-1 predicts token L)
+   served stock vs per-token stock: 64 / 65
+   served stock vs per-token proj : 57 / 65
+   served on    vs per-token stock: 57 / 65
+   served on    vs per-token proj : 62 / 65
+4. how different the text is: 256-token generations, on vs stock, first difference
+   [('code', 11), ('doc', 6), ('chat', 3)]
+5. decode speed (exact mode: fixed experts, no PCIe share - not the normal setup)
+   stock       66.4 tok/s  drafts accepted 256/309
+   vec off     65.7 tok/s  drafts accepted 251/318
+   vec on      76.7 tok/s  drafts accepted 366/425
+   vec on 2    81.2 tok/s  drafts accepted 366/423
+
+[exited with code 0]

bench/results/2026-09-27-esp/teacher-forced.txt

@@ -0,0 +1,6 @@
+text   tokens same top-1  mean KL  median KL  p99 KL ppl stock ppl proj
+code      511      92.8%    0.098     0.0031    1.69     2.594    2.982
+doc      1023      89.8%    0.058     0.0271    0.84    30.299   30.984
+chat     1023      89.6%    0.050     0.0265    0.33    18.409   18.485
+all 2557 tokens: mean KL 0.063, max 4.10, p99 1.00
+

bench/results/2026-09-27-kv-q4/README.md

@@ -0,0 +1,60 @@
+# 2026-09-27: Q4_0 KV with Hadamard rotation (PR #21, `--kv q4_0`)
+
+PR #21 (code-martin) adds a 4-bit K/V format: an orthonormal 256-point Walsh-Hadamard rotation of K, V and Q, then
+ggml `q4_0` blocks (576 B per cell and layer, vs 1,056 B in int8).
+
+## Integration fixes made on merge
+
+- **Q4 + KV streaming (`--kv-resident`).** The PR had no host copy for Q4 and skipped the staging, so with streaming
+  on, the prompt path wrote K/V through the residency map's -1 entries (out-of-bounds writes).
+  - Q4 is now a third format in `kv_stream`: host copy, block residency, prompt-path staging, MTP ring and STATE_HASH.
+  - `kv_stream_parity` checks it bitwise, streamed vs resident, with ring restore.
+- **Duplicate attention.** The per-token decode path ran attention twice, the first time before the selection's
+  blocks were made resident; it now runs once.
+- **Tie-break.** The group maximum's tie-break is now deterministic across lanes: equal |x| with opposite signs
+  could give one block two scales.
+- **Unrelated changes dropped:** duplicate path fixes (already in 0.1.7), personal `.gitignore` entries, and a
+  re-encoded CMakeLists.
+
+`kv_q4_parity` (the PR's test) and all 70 ctest tests pass.
+
+## Quality (teacher-forced, `tools/kv_precision_compare.py`)
+
+Same text, same fixed expert set (Q2_0, 3,000 slots, `--adapt-swaps 0`); only `--kv` differs. fp16 is the reference.
+
+| Text | KV | Same top-1 as fp16 | Mean KL | Perplexity | vs fp16 (nats/token) |
+|---|---|---:|---:|---:|---:|
+| code agent (512) | int8 | 98.4% | 0.014 | 2.649 | -0.011 +- 0.006 |
+| | q4_0 | 95.7% | 0.030 | 2.598 | -0.031 +- 0.014 |
+| document (1,024) | int8 | 95.6% | 0.009 | 30.18 | +0.012 +- 0.005 |
+| | q4_0 | 91.3% | 0.039 | 32.26 | **+0.078 +- 0.011** |
+| chat (1,024) | int8 | 95.5% | 0.009 | 18.24 | -0.003 +- 0.005 |
+| | q4_0 | 90.7% | 0.028 | 18.52 | +0.012 +- 0.008 |
+| document, 8K (every 8th position) | int8 | 92.3% | 0.114 | 7.63 | -0.013 +- 0.020 |
+| | q4_0 | 88.0% | 0.297 | 8.65 | **+0.112 +- 0.037** |
+
+- **int8 is indistinguishable from fp16** (within about 2 standard errors everywhere).
+- **q4_0 costs measurable precision, and more as the context grows:**
+  - document perplexity +8% at 1K and +12% at 8K;
+  - the KL from fp16 is 2.6-4.2x int8's;
+  - code and chat are near-neutral.
+- **Needles with `--kv q4_0 --kv-resident 32768`: 5/5** (1K, 32K, 128K at 10% and 90% depth, 262K;
+  `needles-q4.*`). Retrieval holds.
+
+## Speed and memory (Q2_0, 128K, matrix bench, `--kv-resident 32768`)
+
+| KV | Expert slots | Decode tok/s | Tokens/round | Rounds/s | K/V in RAM |
+|---|---:|---:|---:|---:|---:|
+| int8 (`kv-stream/matrix-stream`) | 4,044 | 62.4 | 2.52 | 24.8 | 1.52 GiB |
+| q4_0 (`matrix-q4`) | 4,190 | 67.4 | 2.62 | 25.7 (+3.7%) | 0.83 GiB |
+
+Without streaming (contexts up to 64K, or too little RAM for it) q4_0 halves the whole KV cache in VRAM instead.
+
+## Decision
+
+`--kv q4_0` is a user option:
+- `START-HERE.bat --kv q4_0`, or the question setup asks above 8K context;
+- or `--kv q4_0` in a `strata-*.json`.
+
+int8 stays the default: every published quality number was measured with it, and q4_0's precision cost is real,
+though it passes the needles.

bench/results/2026-09-28-coder/README.md

@@ -0,0 +1,44 @@
+# The GSQ-RCO Coder (256 of 512 experts) on Strata
+
+RTX 5070 12 GB, Ryzen 5 7600, 64 GB DDR5-5200, Windows 10. Engine 0.1.15 + PR #54 (the expert count from the model
+file), the settings setup writes (`--prefill auto`, 8-bit KV above 4K, KV streaming from 64K, MTP, greedy, 256 tokens),
+`data/expert-profile-coder.bin`, the same code-agent prompts as `2026-09-28-speed-0114`.
+
+| | 1K | 4K | 32K | 64K | 128K | 262K |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Prompt, tokens/s | 599 | 1,152 | 1,298 | 1,350 | 1,266 | 1,034 |
+| Output, tokens/s | 53.3 | 50.6 | 53.3 | 50.8 | 44.0 | 42.8 |
+| Draft acceptance | 0.720 | 0.707 | 0.785 | 0.696 | 0.619 | 0.684 |
+| Experts in VRAM | 2,570 | 2,615 | 2,418 | 2,382 | 2,326 | 2,208 |
+
+Every run in [`matrix.json`](matrix.json). The 128K and 262K answers were read: coherent (they list and discuss the
+files of the prompt).
+
+## Why it is fast
+
+The kept experts are stored like IQ3_S's (IQ2_S / IQ3_XXS / IQ3_S / IQ4_XS gate and up, IQ4_NL / Q2_0 down; "IQ1_M" is
+the release's name for 1.89 bits per *original* parameter), so a token costs about what IQ3_S costs where the expert is
+on the CPU. But there are half as many experts: the 23 GB of them pin completely (`cudaHostRegister` of the whole
+arena), a 12 GB card holds ~20% of them instead of ~10%, and the prompt path streams half as many.
+
+## The expert profile
+
+`data/expert-profile-coder.bin` (48 x 256, all 12,288 pairs ranked): the shipped 48 x 512 ranking, each pair mapped to
+its kept expert's new index through section 2 of the release's `tensor-allocation/*.rco-allocation.txt` (the kept
+experts in their original order; gate, up and down agree on which were kept), pruned pairs dropped. 5,492 of the
+shipped top 8,000 survive (as PR #54 found). With 2,537 slots on the 12 GB card, 72% of the expert reads of a 4K code
+prompt hit the GPU - a wrong index mapping would give about the slot share, 21%.
+
+## Images
+
+The Coder ships the original's vision encoder (the same file). With `--vision` on the GPU it described a Strata
+Monitor screenshot correctly, windows, labels and the numbers of the dashboard and the requests table (1,079 prompt
+tokens with the image, 2,012 generated at ~50 tokens/s).
+
+## The experimental speed projection
+
+The vector was made from the full model's activations; the Coder has the same residual stream. On the Coder it loads
+(41 layers steered), costs no measurable speed (47.7-55.7 tokens/s with it on or off), and both settings write correct
+answers to three coding prompts. What it is for - fewer refusals - was not measured. (Two runs of the same prompt with
+it off also differ after ~130-200 characters in serve mode, so this check cannot show how far it moves ordinary text;
+the teacher-forced method of `2026-09-27-esp` needs the token-by-token prompt path, which native packs do not run.)

bench/results/2026-09-28-coder/matrix.json

@@ -0,0 +1,86 @@
+[
+ {
+  "model": "CODER",
+  "tier": "1k",
+  "prompt_tokens": 1017,
+  "exit": 0,
+  "wall_s": 18.0,
+  "prefill_tok_s": 599.05,
+  "ttft_s": 2.54,
+  "decode_tok_s": 53.33,
+  "decoded": 256,
+  "spec_accept": 0.72,
+  "tokens_per_round": 2.86,
+  "vram_slots": 2570
+ },
+ {
+  "model": "CODER",
+  "tier": "4k",
+  "prompt_tokens": 3562,
+  "exit": 0,
+  "wall_s": 19.6,
+  "prefill_tok_s": 1152.17,
+  "ttft_s": 3.94,
+  "decode_tok_s": 50.6,
+  "decoded": 256,
+  "spec_accept": 0.707,
+  "tokens_per_round": 2.88,
+  "vram_slots": 2615
+ },
+ {
+  "model": "CODER",
+  "tier": "32k",
+  "prompt_tokens": 31566,
+  "exit": 0,
+  "wall_s": 40.7,
+  "prefill_tok_s": 1297.8,
+  "ttft_s": 25.17,
+  "decode_tok_s": 53.34,
+  "decoded": 256,
+  "spec_accept": 0.785,
+  "tokens_per_round": 3.02,
+  "vram_slots": 2418
+ },
+ {
+  "model": "CODER",
+  "tier": "64k",
+  "prompt_tokens": 64162,
+  "exit": 0,
+  "wall_s": 75.5,
+  "prefill_tok_s": 1350.09,
+  "ttft_s": 49.01,
+  "decode_tok_s": 50.84,
+  "decoded": 256,
+  "spec_accept": 0.696,
+  "tokens_per_round": 2.61,
+  "vram_slots": 2382
+ },
+ {
+  "model": "CODER",
+  "tier": "128k",
+  "prompt_tokens": 128478,
+  "exit": 0,
+  "wall_s": 121.5,
+  "prefill_tok_s": 1266.31,
+  "ttft_s": 102.95,
+  "decode_tok_s": 44.03,
+  "decoded": 256,
+  "spec_accept": 0.619,
+  "tokens_per_round": 2.33,
+  "vram_slots": 2326
+ },
+ {
+  "model": "CODER",
+  "tier": "262k",
+  "prompt_tokens": 259943,
+  "exit": 0,
+  "wall_s": 280.7,
+  "prefill_tok_s": 1033.91,
+  "ttft_s": 252.92,
+  "decode_tok_s": 42.8,
+  "decoded": 256,
+  "spec_accept": 0.684,
+  "tokens_per_round": 2.6,
+  "vram_slots": 2208
+ }
+]

bench/results/2026-09-28-prefill-speed/README.md

@@ -0,0 +1,47 @@
+# Faster prompt reading (engine 0.1.13)
+
+RTX 5070 12 GB, Ryzen 5 7600, 64 GB DDR5. 32K-token prompt (`bench/prompts/long/32k-01-code-agent.ids`), one-shot
+engine, `--expert-cache auto`, greedy. Numbers are prompt tokens/s.
+
+## What changed, step by step
+
+| Step | What | Result |
+|---|---|---|
+| Buffers | The attention half and the MoE half of a layer share one scratch region (~260 KB less per prompt token) | bit-identical; chunk 4096 borrows 1.78 GiB instead of 2.84 |
+| PLE rows | The next chunk's PLE rows are read from the model file on a thread while the GPU works | bit-identical; 762 -> 790 at chunk 4096 |
+| Chunk size | `--prefill auto`: the largest of 8192 / 6144 / 4096 / ... whose buffers fit in the cache slots it may borrow; a request borrows only what its own prompt needs | Q2_0: 4096 791, 6144 878, 8192 973 |
+| Host copies | Experts the arena could not pin are copied to pinned buffers by helper threads instead of the launching thread | bit-identical; IQ3_S at 4096: 551 -> 652 |
+| PLE block | Computed for the whole chunk (two GEMMs) instead of token by token | Q2_0 981 -> 1053, IQ3_S 831 -> 940 |
+| Experts | llama.cpp's MMQ kernels (weights stay quantized, int8 tensor cores) instead of dequantize-to-FP16 + cuBLAS | Q2_0 1052 -> 1130, IQ3_S (8192) 932 -> 1014 |
+| Streaming | At chunks of 2048+ every non-resident expert streams in a fixed order through a ring, so the next layer's arrive while the current layer's attention runs | bit-identical; Q2_0 1130 -> 1290, IQ3_S -> 1208 |
+
+Tried and dropped: cuBLAS grouped GEMMs for the experts (faster in isolation, slower inside the engine: 770 vs 790).
+
+## Before and after
+
+| | engine 0.1.12 (`--prefill 2048`) | engine 0.1.13 (`--prefill auto`) |
+|---|---:|---:|
+| Q2_0, 32K prompt | 572 | 1,290 |
+| IQ3_S, 32K prompt | 383 | 1,208 |
+| Server, Q2_0, 128K context, 999-token prompt | 353 | 438 |
+| Server, 6,927-token prompt | 529 | 1,077 |
+| Server, 28,584-token prompt | 584 | 1,249 |
+
+Output speed is unchanged (the decode path is the same). Short prompts (a few hundred tokens) go through the verify
+windows as before.
+
+## Quality
+
+- "Bit-identical" steps were checked with the GDN state hash after the 32K prompt (all 36 recurrent layers).
+- The steps that change rounding (the batched PLE block, MMQ's 8-bit activations) were compared teacher-forced: the
+  32K prompt is read by the prompt path up to its last 256 tokens, those go through the token path with their logits
+  dumped. At 32K every rounding change moves the predictions about this much - even a different chunk size of the
+  unchanged engine - so that is the yardstick:
+
+| Against the 8192-chunk FP16 path | Same top-1 | Mean KL | Perplexity (reference 9.21) |
+|---|---:|---:|---:|
+| Chunk 6144, same code (the yardstick) | 89.8% | 0.33 | 9.18 |
+| MMQ experts | 85.9% | 0.38 | 8.63 |
+
+- Needles (the answer hidden at a given depth): 5 of 5 at 1K, 32K, 128K (two depths) and 262K, Q2_0, KV streaming.
+- A rerun of the same configuration is bit-identical.

bench/results/2026-09-28-rope-scaling/README.md

@@ -0,0 +1,49 @@
+# Rope scaling: needle recall past the trained 262K
+
+The feature (PR #84, experimental): `--rope-scaling none|linear|yarn` extends the context past the model's
+trained 262,144 positions by rescaling the rotary angles - llama.cpp's types, ggml's `rope_yarn`
+arithmetic, one process config applied at the cos/sin table (default path) and as kernel constants
+(`--native-rope`, the native indexer, prefill). The three tables here are the first end-to-end check
+that a scaled run actually recalls past the trained end.
+
+**Hardware:** RTX 4070 Ti SUPER 16 GB, Ryzen 9 5900X (no AVX-512 - the IQ3_XXS native pack runs its
+expert kernels on AVX-2), 125 GB RAM, Linux. **Engine:** this branch built from source (the log records
+`rope scaling ... --max-context ... against a trained context of 262144` at startup). The runs below
+were taken on the rope tree built over the 0.1.17-era base; the branch then rebased onto main (engine
+0.1.20) - the drift touches none of the rope files - and the rebase was re-validated: `rope_parity`
+green, a 64-token greedy run token-identical to the plain 0.1.20 build, and the yarn 262k probe
+repeated green (see the last table row). Baseline arm A also re-checked the default is the stock
+model: a 64-token greedy run produces the same token ids as the ready-made 0.1.17 engine, and
+`rope_parity` holds the `none` table bit-identical to the pre-scaling builder.
+**Method:** `tools/needle_bench.py` against a running server (launch recipes in `config-*.json`),
+`temperature 0`, thinking off, depth 50%, one probe per length.
+
+| Arm | Scaling | Context | Probe | Prompt tokens | Result | Probe time |
+| --- | --- | --- | --- | --- | --- | --- |
+| A | none | 262,144 | 1k @ 50% | 1,442 | FOUND | 4 s |
+| A | none | 262,144 | 32k @ 50% | 31,763 | FOUND | 27 s |
+| A | none | 262,144 | 128k @ 50% | 128,328 | FOUND | 86 s |
+| B | linear, factor 2 | 560,000 | 262k @ 50% | 274,222 | FOUND | 232 s |
+| B | linear, factor 2 | 560,000 | 512k @ 50% | 527,051 | FOUND | 619 s |
+| C | yarn, factor 2 | 560,000 | 262k @ 50% | 274,215 | FOUND | 236 s |
+| C | yarn, factor 2 | 560,000 | 512k @ 50% | 526,988 | FOUND | 615 s |
+| C | yarn, factor 2 | 560,000 | 262k @ 50% (the rebase re-check) | 273,708 | FOUND | 235 s |
+
+All 8 of 8 probes found the needle; JSON per arm: `needle-A-none.json`, `needle-B-linear2.json`,
+`needle-B-linear2-512k.json`, `needle-C-yarn2.json` (+ the rebase re-check in
+`needle-C-yarn2-rebased.json`); server and engine logs alongside.
+
+What this establishes, and what it does not:
+
+- **Past the trained end the model still reads.** 512k probes are 2.01x the trained 262,144 - position
+  interpolation (B) and YaRN (C) both hold a mid-context lookup there.
+- **Scaled runs within the trained range are a slightly different model.** `yarn`'s mscale applies at
+  every position (a 64-token greedy prompt diverges from the stock run after ~19 tokens); arm A is the
+  stock model, arms B/C are not, by design.
+- **One depth, one needle, one machine.** A single probe per (length, arm) at depth 50% says "recall
+  works there", not "no degradation anywhere" - a full depth sweep and a longer-length ladder (768K,
+  1M) on the bench machine (64 GB RAM there vs 125 GB here, so 262K IQ3_XXS fits either way) is the
+  natural follow-up. 512K prompts prefill at ~850-950 tokens/s on this card; the probe time is almost
+  entirely prefill.
+- **The `560,000` context** in arms B/C is headroom for the needle harness's tokenizer overshoot
+  (a "512k" probe tokenizes to ~527k); 524,288 exactly would refuse the probe.

bench/results/2026-09-28-rope-scaling/config-A-none-262144.json

@@ -0,0 +1,38 @@
+{
+ "exe": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata/build/strata",
+ "cwd": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata",
+ "tokenizer": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/packs/iq3_xxs/tokenizer",
+ "model_name": "qwen3.8-flash-next-iq3_xxs-rope-scaling",
+ "lib_dirs": [
+  "/usr/local/cuda/bin",
+  "/usr/local/cuda/lib64"
+ ],
+ "port": 8130,
+ "args": [
+  "--pack",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/packs/iq3_xxs",
+  "--native",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf",
+  "--ple-gguf",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf",
+  "--expert-profile",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata/data/expert-profile.bin",
+  "--expert-cache",
+  "auto",
+  "--prefill",
+  "auto",
+  "--spec",
+  "4",
+  "--spec-min-p",
+  "0.5",
+  "--mtp",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/mtp/rt",
+  "--kv",
+  "int8",
+  "--kv-resident",
+  "32768",
+  "--max-context",
+  "262144"
+ ],
+ "log": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata/bench/results/2026-09-28-rope-scaling/engine-A-none-262144.log"
+}

bench/results/2026-09-28-rope-scaling/config-B-linear2.json

@@ -0,0 +1,42 @@
+{
+ "exe": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata/build/strata",
+ "cwd": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata",
+ "tokenizer": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/packs/iq3_xxs/tokenizer",
+ "model_name": "qwen3.8-flash-next-iq3_xxs-rope-scaling",
+ "lib_dirs": [
+  "/usr/local/cuda/bin",
+  "/usr/local/cuda/lib64"
+ ],
+ "port": 8130,
+ "args": [
+  "--pack",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/packs/iq3_xxs",
+  "--native",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf",
+  "--ple-gguf",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf",
+  "--expert-profile",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata/data/expert-profile.bin",
+  "--expert-cache",
+  "auto",
+  "--prefill",
+  "auto",
+  "--spec",
+  "4",
+  "--spec-min-p",
+  "0.5",
+  "--mtp",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/mtp/rt",
+  "--kv",
+  "int8",
+  "--kv-resident",
+  "32768",
+  "--max-context",
+  "560000",
+  "--rope-scaling",
+  "linear",
+  "--rope-scale",
+  "2"
+ ],
+ "log": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata/bench/results/2026-09-28-rope-scaling/engine-B-linear2-524288.log"
+}

bench/results/2026-09-28-rope-scaling/config-C-yarn2.json

@@ -0,0 +1,42 @@
+{
+ "exe": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata/build/strata",
+ "cwd": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata",
+ "tokenizer": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/packs/iq3_xxs/tokenizer",
+ "model_name": "qwen3.8-flash-next-iq3_xxs-rope-scaling",
+ "lib_dirs": [
+  "/usr/local/cuda/bin",
+  "/usr/local/cuda/lib64"
+ ],
+ "port": 8130,
+ "args": [
+  "--pack",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/packs/iq3_xxs",
+  "--native",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf",
+  "--ple-gguf",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/models/IQ3_XXS/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00002-of-00002.gguf",
+  "--expert-profile",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata/data/expert-profile.bin",
+  "--expert-cache",
+  "auto",
+  "--prefill",
+  "auto",
+  "--spec",
+  "4",
+  "--spec-min-p",
+  "0.5",
+  "--mtp",
+  "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata-data/mtp/rt",
+  "--kv",
+  "int8",
+  "--kv-resident",
+  "32768",
+  "--max-context",
+  "560000",
+  "--rope-scaling",
+  "yarn",
+  "--rope-scale",
+  "2"
+ ],
+ "log": "/home/jakub/workspace/deepseek-harness/harnes_workspace/Strata/bench/results/2026-09-28-rope-scaling/engine-C-yarn2-524288.log"
+}

bench/results/2026-09-28-rope-scaling/kv-sweep-tgs.json

@@ -0,0 +1,121 @@
+[
+ {
+  "length": "128k",
+  "round": 0,
+  "kind": "decode",
+  "prompt_tok": 128401,
+  "read": 52,
+  "gen": 357,
+  "read_tok_s": 79.2,
+  "gen_tok_s": 56.3,
+  "drafts_ok": 182,
+  "drafts_all": 305,
+  "seg_blocks": 3284340,
+  "seg_mib": 370.0,
+  "seg_misses": 91744,
+  "seg_hit_pct": 97.207,
+  "kib_per_token": 1061.29
+ },
+ {
+  "length": "128k",
+  "round": 1,
+  "kind": "read",
+  "prompt_tok": 128357,
+  "read": 128357,
+  "gen": 24,
+  "read_tok_s": 1389.2,
+  "gen_tok_s": 41.7,
+  "drafts_ok": 11,
+  "drafts_all": 23,
+  "seg_blocks": -3253560,
+  "seg_mib": -391.7,
+  "seg_misses": -97147,
+  "seg_hit_pct": 97.014,
+  "kib_per_token": -3.12
+ },
+ {
+  "length": "128k",
+  "round": 1,
+  "kind": "decode",
+  "prompt_tok": 128402,
+  "read": 52,
+  "gen": 388,
+  "read_tok_s": 76.5,
+  "gen_tok_s": 54.6,
+  "drafts_ok": 195,
+  "drafts_all": 363,
+  "seg_blocks": 3752616,
+  "seg_mib": 463.5,
+  "seg_misses": 114911,
+  "seg_hit_pct": 96.938,
+  "kib_per_token": 1223.26
+ },
+ {
+  "length": "512k",
+  "round": 0,
+  "kind": "read",
+  "prompt_tok": 522983,
+  "read": 522983,
+  "gen": 24,
+  "read_tok_s": 868.4,
+  "gen_tok_s": 42.4,
+  "drafts_ok": 13,
+  "drafts_all": 16,
+  "seg_blocks": -3808056,
+  "seg_mib": -387.8,
+  "seg_misses": -96113,
+  "seg_hit_pct": 97.476,
+  "kib_per_token": -0.76
+ },
+ {
+  "length": "512k",
+  "round": 0,
+  "kind": "decode",
+  "prompt_tok": 523028,
+  "read": 52,
+  "gen": 402,
+  "read_tok_s": 63.3,
+  "gen_tok_s": 48.4,
+  "drafts_ok": 221,
+  "drafts_all": 348,
+  "seg_blocks": 3580056,
+  "seg_mib": 920.4,
+  "seg_misses": 228658,
+  "seg_hit_pct": 93.613,
+  "kib_per_token": 2344.5
+ },
+ {
+  "length": "512k",
+  "round": 1,
+  "kind": "read",
+  "prompt_tok": 522984,
+  "read": 522984,
+  "gen": 24,
+  "read_tok_s": 857.1,
+  "gen_tok_s": 35.5,
+  "drafts_ok": 9,
+  "drafts_all": 18,
+  "seg_blocks": -3555408,
+  "seg_mib": -905.8,
+  "seg_misses": -225016,
+  "seg_hit_pct": 93.671,
+  "kib_per_token": -1.77
+ },
+ {
+  "length": "512k",
+  "round": 1,
+  "kind": "decode",
+  "prompt_tok": 523029,
+  "read": 52,
+  "gen": 402,
+  "read_tok_s": 61.3,
+  "gen_tok_s": 42.8,
+  "drafts_ok": 192,
+  "drafts_all": 388,
+  "seg_blocks": 4005216,
+  "seg_mib": 1079.9,
+  "seg_misses": 268221,
+  "seg_hit_pct": 93.303,
+  "kib_per_token": 2750.79
+ }
+]

bench/results/2026-09-28-rope-scaling/needle-A-none.json

@@ -0,0 +1,29 @@
+[
+ {
+  "length": "1k",
+  "depth": 50,
+  "prompt_tokens": 1442,
+  "seconds": 3.7,
+  "found": true,
+  "word": "meadow-copper-504",
+  "answer": "meadow-copper-504"
+ },
+ {
+  "length": "32k",
+  "depth": 50,
+  "prompt_tokens": 31763,
+  "seconds": 27.5,
+  "found": true,
+  "word": "falcon-quartz-940",
+  "answer": "falcon-quartz-940"
+ },
+ {
+  "length": "128k",
+  "depth": 50,
+  "prompt_tokens": 128328,
+  "seconds": 86.1,
+  "found": true,
+  "word": "willow-cobalt-696",
+  "answer": "willow-cobalt-696"
+ }
+]

bench/results/2026-09-28-rope-scaling/needle-B-linear2-512k.json

@@ -0,0 +1,11 @@
+[
+ {
+  "length": "512k",
+  "depth": 50,
+  "prompt_tokens": 527051,
+  "seconds": 619.0,
+  "found": true,
+  "word": "meadow-copper-504",
+  "answer": "meadow-copper-504"
+ }
+]

bench/results/2026-09-28-rope-scaling/needle-B-linear2.json

@@ -0,0 +1,16 @@
+[
+ {
+  "length": "262k",
+  "depth": 50,
+  "prompt_tokens": 274222,
+  "seconds": 232.2,
+  "found": true,
+  "word": "meadow-copper-504",
+  "answer": "meadow-copper-504"
+ },
+ {
+  "length": "512k",
+  "depth": 50,
+  "error": "HTTP Error 400: Bad Request"
+ }
+]

bench/results/2026-09-28-rope-scaling/needle-C-yarn2-rebased.json

@@ -0,0 +1,11 @@
+[
+ {
+  "length": "262k",
+  "depth": 50,
+  "prompt_tokens": 273708,
+  "seconds": 234.8,
+  "found": true,
+  "word": "meadow-copper-504",
+  "answer": "meadow-copper-504"
+ }
+]

bench/results/2026-09-28-rope-scaling/needle-C-yarn2.json

@@ -0,0 +1,20 @@
+[
+ {
+  "length": "262k",
+  "depth": 50,
+  "prompt_tokens": 274215,
+  "seconds": 235.5,
+  "found": true,
+  "word": "meadow-copper-504",
+  "answer": "meadow-copper-504"
+ },
+ {
+  "length": "512k",
+  "depth": 50,
+  "prompt_tokens": 526988,
+  "seconds": 615.0,
+  "found": true,
+  "word": "falcon-quartz-940",
+  "answer": "falcon-quartz-940"
+ }
+]

bench/results/2026-09-28-speed-0114/README.md

@@ -0,0 +1,32 @@
+# Speed with engine 0.1.14
+
+RTX 5070 12 GB (PCIe 5.0 x16), Ryzen 5 7600, 64 GB DDR5-5200, Windows 10. The ready-made 0.1.14 engine, one-shot runs
+with what setup writes: `--prefill auto`, the shipped expert profile, `--expert-cache auto`, 8-bit KV above 4K, KV
+streaming (`--kv-resident 32768`) from 64K, MTP (`--spec 4 --spec-min-p 0.5`), greedy, 256 generated tokens. One
+code-agent prompt per length (`bench/prompts/*/…-code-agent.ids`, the same prompts as `2026-09-24-final`).
+
+The tables: [`matrix.md`](matrix.md) (prompt and output speed, draft acceptance, experts in VRAM); every run in
+[`matrix.json`](matrix.json).
+
+- **IQ2_XS** was measured with Swift 1.5's IQ2_XS (the original's file was not on the machine; the two run at the same
+  speed: 4K 465 vs 467 prompt tokens/s in `2026-09-24`).
+- **IQ3_S at 128K** comes from a first pass during which a large download was running (the clean rerun of that row
+  was stopped when Windows ran short of memory); the other rows are from runs on an otherwise idle machine.
+- **IQ3_XXS and IQ3_S at 262K** are not measured: with 43 / 50 GB of experts, a 260K context exceeds 64 GB of RAM.
+
+## Against 0.1.12
+
+Prompts are 1.2x (1K) to 2.4x (32K-128K) faster: the 0.1.13 prompt path (chunks of up to 8,192 tokens, MMQ experts,
+streamed expert copies; see `2026-09-28-prefill-speed`).
+
+Output speed moves with the text: speculative decoding is faster when more drafts are accepted, and the new prompt
+path's rounding changes the answer, so the same prompt can land a few percent either way (Q2_0 32K: acceptance 0.865
+then, 0.709 now). The engine itself is not slower: on the 4K prompt, run back to back on the same machine state,
+0.1.14 writes 88.5 tokens/s and 0.1.12 85.7 (3 runs each, same free VRAM).
+
+## Found on the way
+
+IQ3_XXS at 64K and 128K ended with `out of memory: cudaFuncSetAttribute` in the MMQ launch: CUDA loads a kernel's code
+when it is first used, and mid-prompt there was no VRAM left for it. Loading every kernel at start
+(`CUDA_MODULE_LOADING=EAGER`, which the engine now sets itself) fixes it for ~30 MB of VRAM; the IQ3_XXS 64K/128K rows
+were measured with that fix.

bench/results/2026-09-28-speed-0114/matrix.json

@@ -0,0 +1,322 @@
+[
+ {
+  "model": "Q2_0",
+  "tier": "1k",
+  "prompt_tokens": 1017,
+  "exit": 0,
+  "wall_s": 41.0,
+  "prefill_tok_s": 494.43,
+  "ttft_s": 3.55,
+  "decode_tok_s": 84.26,
+  "decoded": 256,
+  "spec_accept": 0.709,
+  "tokens_per_round": 2.72,
+  "vram_slots": 4176
+ },
+ {
+  "model": "Q2_0",
+  "tier": "4k",
+  "prompt_tokens": 3562,
+  "exit": 0,
+  "wall_s": 41.2,
+  "prefill_tok_s": 1006.63,
+  "ttft_s": 5.05,
+  "decode_tok_s": 90.26,
+  "decoded": 256,
+  "spec_accept": 0.846,
+  "tokens_per_round": 3.23,
+  "vram_slots": 4123
+ },
+ {
+  "model": "Q2_0",
+  "tier": "32k",
+  "prompt_tokens": 31566,
+  "exit": 0,
+  "wall_s": 63.1,
+  "prefill_tok_s": 1307.6,
+  "ttft_s": 25.63,
+  "decode_tok_s": 73.61,
+  "decoded": 256,
+  "spec_accept": 0.709,
+  "tokens_per_round": 2.68,
+  "vram_slots": 3833
+ },
+ {
+  "model": "Q2_0",
+  "tier": "64k",
+  "prompt_tokens": 64162,
+  "exit": 0,
+  "wall_s": 89.0,
+  "prefill_tok_s": 1293.5,
+  "ttft_s": 51.12,
+  "decode_tok_s": 69.02,
+  "decoded": 256,
+  "spec_accept": 0.679,
+  "tokens_per_round": 2.62,
+  "vram_slots": 3780
+ },
+ {
+  "model": "Q2_0",
+  "tier": "128k",
+  "prompt_tokens": 128478,
+  "exit": 0,
+  "wall_s": 147.4,
+  "prefill_tok_s": 1208.46,
+  "ttft_s": 107.81,
+  "decode_tok_s": 67.21,
+  "decoded": 256,
+  "spec_accept": 0.691,
+  "tokens_per_round": 2.64,
+  "vram_slots": 3697
+ },
+ {
+  "model": "Q2_0",
+  "tier": "262k",
+  "prompt_tokens": 259943,
+  "exit": 0,
+  "wall_s": 309.8,
+  "prefill_tok_s": 966.84,
+  "ttft_s": 270.35,
+  "decode_tok_s": 60.3,
+  "decoded": 256,
+  "spec_accept": 0.623,
+  "tokens_per_round": 2.43,
+  "vram_slots": 3525
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "1k",
+  "prompt_tokens": 1017,
+  "exit": 0,
+  "wall_s": 40.4,
+  "prefill_tok_s": 461.28,
+  "ttft_s": 2.28,
+  "decode_tok_s": 74.38,
+  "decoded": 256,
+  "spec_accept": 0.772,
+  "tokens_per_round": 2.84,
+  "vram_slots": 4386,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "4k",
+  "prompt_tokens": 3562,
+  "exit": 0,
+  "wall_s": 41.8,
+  "prefill_tok_s": 810.71,
+  "ttft_s": 4.48,
+  "decode_tok_s": 73.84,
+  "decoded": 256,
+  "spec_accept": 0.79,
+  "tokens_per_round": 3.24,
+  "vram_slots": 4336,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "32k",
+  "prompt_tokens": 31566,
+  "exit": 0,
+  "wall_s": 65.0,
+  "prefill_tok_s": 1237.98,
+  "ttft_s": 25.59,
+  "decode_tok_s": 71.49,
+  "decoded": 256,
+  "spec_accept": 0.856,
+  "tokens_per_round": 3.28,
+  "vram_slots": 4057,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "64k",
+  "prompt_tokens": 64162,
+  "exit": 0,
+  "wall_s": 97.1,
+  "prefill_tok_s": 1135.63,
+  "ttft_s": 56.59,
+  "decode_tok_s": 64.29,
+  "decoded": 256,
+  "spec_accept": 0.707,
+  "tokens_per_round": 2.76,
+  "vram_slots": 4005,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "128k",
+  "prompt_tokens": 128478,
+  "exit": 0,
+  "wall_s": 158.0,
+  "prefill_tok_s": 1070.61,
+  "ttft_s": 120.1,
+  "decode_tok_s": 59.76,
+  "decoded": 256,
+  "spec_accept": 0.726,
+  "tokens_per_round": 2.69,
+  "vram_slots": 3923,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ2_XS",
+  "tier": "262k",
+  "prompt_tokens": 259943,
+  "exit": 0,
+  "wall_s": 332.4,
+  "prefill_tok_s": 886.2,
+  "ttft_s": 293.43,
+  "decode_tok_s": 52.83,
+  "decoded": 256,
+  "spec_accept": 0.628,
+  "tokens_per_round": 2.53,
+  "vram_slots": 3565,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_XXS",
+  "tier": "1k",
+  "prompt_tokens": 1017,
+  "exit": 0,
+  "wall_s": 44.1,
+  "prefill_tok_s": 414.55,
+  "ttft_s": 2.54,
+  "decode_tok_s": 60.26,
+  "decoded": 256,
+  "spec_accept": 0.864,
+  "tokens_per_round": 3.52,
+  "vram_slots": 3304,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_XXS",
+  "tier": "4k",
+  "prompt_tokens": 3562,
+  "exit": 0,
+  "wall_s": 45.8,
+  "prefill_tok_s": 769.57,
+  "ttft_s": 4.73,
+  "decode_tok_s": 62.07,
+  "decoded": 256,
+  "spec_accept": 0.855,
+  "tokens_per_round": 3.24,
+  "vram_slots": 3263,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_XXS",
+  "tier": "32k",
+  "prompt_tokens": 31566,
+  "exit": 0,
+  "wall_s": 69.7,
+  "prefill_tok_s": 1107.75,
+  "ttft_s": 28.61,
+  "decode_tok_s": 51.4,
+  "decoded": 256,
+  "spec_accept": 0.678,
+  "tokens_per_round": 2.6,
+  "vram_slots": 3031,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_XXS",
+  "tier": "64k",
+  "prompt_tokens": 64162,
+  "exit": 0,
+  "wall_s": 104.9,
+  "prefill_tok_s": 1065.31,
+  "ttft_s": 61.77,
+  "decode_tok_s": 49.95,
+  "decoded": 256,
+  "spec_accept": 0.741,
+  "tokens_per_round": 2.94,
+  "vram_slots": 2965,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_XXS",
+  "tier": "128k",
+  "prompt_tokens": 128478,
+  "exit": 0,
+  "wall_s": 172.1,
+  "prefill_tok_s": 1015.11,
+  "ttft_s": 128.12,
+  "decode_tok_s": 45.78,
+  "decoded": 256,
+  "spec_accept": 0.665,
+  "tokens_per_round": 2.35,
+  "vram_slots": 2901,
+  "note": "first pass"
+ },
+ {
+  "model": "IQ3_S",
+  "tier": "1k",
+  "prompt_tokens": 1017,
+  "exit": 0,
+  "wall_s": 56.2,
+  "prefill_tok_s": 396.13,
+  "ttft_s": 4.14,
+  "decode_tok_s": 51.52,
+  "decoded": 256,
+  "spec_accept": 0.826,
+  "tokens_per_round": 3.12,
+  "vram_slots": 2617
+ },
+ {
+  "model": "IQ3_S",
+  "tier": "4k",
+  "prompt_tokens": 3562,
+  "exit": 0,
+  "wall_s": 57.5,
+  "prefill_tok_s": 736.9,
+  "ttft_s": 6.35,
+  "decode_tok_s": 51.6,
+  "decoded": 256,
+  "spec_accept": 0.747,
+  "tokens_per_round": 2.89,
+  "vram_slots": 2580
+ },
+ {
+  "model": "IQ3_S",
+  "tier": "32k",
+  "prompt_tokens": 31566,
+  "exit": 0,
+  "wall_s": 82.5,
+  "prefill_tok_s": 1069.83,
+  "ttft_s": 31.33,
+  "decode_tok_s": 48.19,
+  "decoded": 256,
+  "spec_accept": 0.772,
+  "tokens_per_round": 2.94,
+  "vram_slots": 2383
+ },
+ {
+  "model": "IQ3_S",
+  "tier": "64k",
+  "prompt_tokens": 64162,
+  "exit": 0,
+  "wall_s": 115.0,
+  "prefill_tok_s": 1069.8,
+  "ttft_s": 61.47,
+  "decode_tok_s": 48.75,
+  "decoded": 256,
+  "spec_accept": 0.786,
+  "tokens_per_round": 2.9,
+  "vram_slots": 2347
+ },
+ {
+  "model": "IQ3_S",
+  "tier": "128k",
+  "prompt_tokens": 128478,
+  "exit": 0,
+  "wall_s": 191.8,
+  "prefill_tok_s": 931.09,
+  "ttft_s": 138.11,
+  "decode_tok_s": 40.48,
+  "decoded": 256,
+  "spec_accept": 0.661,
+  "tokens_per_round": 2.71,
+  "vram_slots": 2310,
+  "note": "first pass (a download was running)"
+ }
+]

bench/results/2026-09-28-speed-0114/matrix.md

@@ -0,0 +1,35 @@
+### Prompt processing (tok/s)
+
+| Model | 1K | 4K | 32K | 64K | 128K | 262K |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Q2_0 | 494 | 1,007 | 1,308 | 1,294 | 1,208 | 967 |
+| IQ2_XS | 461 | 811 | 1,238 | 1,136 | 1,071 | 886 |
+| IQ3_XXS | 415 | 770 | 1,108 | 1,065 | 1,015 | - |
+| IQ3_S | 396 | 737 | 1,070 | 1,070 | 931 | - |
+
+### Output (tok/s)
+
+| Model | 1K | 4K | 32K | 64K | 128K | 262K |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Q2_0 | 84.3 | 90.3 | 73.6 | 69.0 | 67.2 | 60.3 |
+| IQ2_XS | 74.4 | 73.8 | 71.5 | 64.3 | 59.8 | 52.8 |
+| IQ3_XXS | 60.3 | 62.1 | 51.4 | 50.0 | 45.8 | - |
+| IQ3_S | 51.5 | 51.6 | 48.2 | 48.8 | 40.5 | - |
+
+### Draft acceptance
+
+| Model | 1K | 4K | 32K | 64K | 128K | 262K |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Q2_0 | 0.709 | 0.846 | 0.709 | 0.679 | 0.691 | 0.623 |
+| IQ2_XS | 0.772 | 0.790 | 0.856 | 0.707 | 0.726 | 0.628 |
+| IQ3_XXS | 0.864 | 0.855 | 0.678 | 0.741 | 0.665 | - |
+| IQ3_S | 0.826 | 0.747 | 0.772 | 0.786 | 0.661 | - |
+
+### Experts in VRAM
+
+| Model | 1K | 4K | 32K | 64K | 128K | 262K |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Q2_0 | 4,176 | 4,123 | 3,833 | 3,780 | 3,697 | 3,525 |
+| IQ2_XS | 4,386 | 4,336 | 4,057 | 4,005 | 3,923 | 3,565 |
+| IQ3_XXS | 3,304 | 3,263 | 3,031 | 2,965 | 2,901 | - |
+| IQ3_S | 2,617 | 2,580 | 2,383 | 2,347 | 2,310 | - |

bench/results/2026-09-29-layer-split-limits/README.md

@@ -0,0 +1,59 @@
+# Layer split: the features that were off across GPUs (2026-09-29)
+
+**Setup:**
+- **Rig:** RIBPC. RTX 5080 16 GB (CUDA0) + RTX 3090 24 GB; RTX 2080 Ti 22 GB as a helper card. Ryzen 9 9950X3D,
+  62 GB RAM, Windows 11.
+- **Model:** the Coder (IQ1_M native pack).
+- **Build:** branch `mg/limits`, sm_75 + sm_86 + sm_120.
+- **Correctness runs:** `rib_norepeat.py`, with a fixed cache, no swaps, no PCIe share and no suffix drafts where two
+  runs are compared byte for byte.
+- **Speed runs:** `rib_pp.py`.
+- Raw results are in `data/`.
+
+## Correctness
+
+| Check | Result |
+|---|---|
+| One GPU, this branch vs the 0.1.21-equivalent run (fixed cache) | 10/10 byte-identical |
+| Speed projection (control vector), one GPU vs the split on the same GPU (K=24) | 10/10 byte-identical |
+| KV streaming (128K context, 32K resident), one GPU vs the split on the same GPU | 10/10 byte-identical |
+| Images, one GPU vs the split on the same GPU (three questions about a test picture) | identical answers |
+| Speed projection / KV streaming / plain, 5080 + 3090 (K=24): no-repeat under presence 100, the 144 sanity case | all pass, 0 repeats |
+| The vector acts across GPUs: 5080 + 3090 with vs without the vector | 8 of 10 answers differ (as on one GPU: 7 of 10) |
+| Images, 5080 + 3090 (K=24): "What color is the circle?" / "What number?" / describe | "Red" / "42" / "A red circle, the number 42, and a blue square are displayed on a white background" |
+| Mid-prompt checkpoint (every 16,384 tokens): a 19.5K prompt, then another sharing its first 18,600 tokens | one GPU and 5080 + 3090: reused 16,384, answer right |
+| 5080 + 3090 (K=30) + a 2080 Ti helper holding 3,000 experts | all pass |
+| `--mmap-experts` with the Coder (a native pack) | a clear start-up error: it needs a canonical pack's `experts.bin` (with or without a split) |
+
+Across GPUs the text is not byte-identical to one GPU: experts that one card leaves to the CPU run on a GPU in the
+split, and the two round differently.
+
+## Speed
+
+| Run | Prompt 16K / 28K / 60K tok/s | Decode story / code tok/s | Hits |
+|---|---|---|---|
+| KV streaming (128K), 5080 alone | 1,487 / 1,801 / 1,837 | 66.6 / 74.0 | 59-71% |
+| KV streaming (128K), 5080 + 3090 auto | 1,685 / 1,957 / 1,968 | 75.8 / 97.0 | 96-97% |
+| Speed projection, 5080 + 3090 auto | 1,813 / 2,053 / - | 77.0 / 103.6 | 96-97% |
+| 5080 + 3090 (K=30), no helper | - | 62.5 / 85.2 | |
+| 5080 + 3090 (K=30) + 2080 Ti helper | - | 43.0 / 63.7 | |
+
+**The helper card makes it slower,** as it did without a split (bench/results/2026-09-29-layer-split). Its per-layer
+round trip costs more than the CPU pool needs for the few experts it would take over. It stays supported: the
+helpers take the visible GPUs no stage runs on and hold only experts no stage's cache holds. But it isn't
+recommended.
+
+## What changed
+
+- **Per-device tables:** the image-position (mrope) table and the control vector's tables are now kept per device;
+  every stage's rope and steering kernels read their own device's copy. `cvec_set_enabled` switches the vector on
+  every device.
+- **KV streaming:** each stage's session streams its own KV host copy (pinned, portable). The split no longer forces
+  it off.
+- **Mid-prompt checkpoints:** `Prefill::on_stage_chunk` lets each stage save its part of a checkpoint when it has
+  read the chunk. The last stage assembles the parts: the earlier ones are a chunk ahead by then.
+- **Helpers (`--expert-cache-remote`) with a split:** they take the visible devices after the stages, skip every
+  pair a stage's cache holds, and preflight and open on those devices.
+- **Explicit `--expert-cache` with a split:** it is capped so CUDA0 keeps room for the prompt path's own buffers.
+  Without the cap, the first prompt failed with "device buffers ... do not fit".
+- **`--mmap-experts`:** refused at start for a native pack, with the reason.

bench/results/2026-09-29-layer-split-limits/data/norepeat-lim-cv-same24-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 102,
+  "drafts_offered": 166,
+  "tok_s": 66.7,
+  "text": "The wind didn\u2019t just howl; it screamed. It tore through the iron gallery of the Blackwood Point Light, rattling the glass panes like a prisoner testing the bars of a cell.\n\nElias Thorne was not a man who feared the dark. He was a man who feared the silence.\n\nFor forty years, he had lived in the shadow of the Fresnel lens, a rotating eye of glass and brass that swept the Atlantic with a beam of pure, unyielding light. His job was simple: keep the light burning. Keep the logbook current. Keep the foghorn sounding when the mist rolled in, thick as wool.\n\nBut Elias knew something else. He knew that the light was not just for ships. It was for the lost.\n\nIt was November 14th, 1982. The storm was a Category 5, a monster that had swallowed the horizon. The waves were mountains of black water, crashing against the granite"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 134,
+  "drafts_offered": 180,
+  "tok_s": 80.8,
+  "text": "```python\nimport csv\nfrom typing import List, Dict\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, str]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile(mode='w', suffix='.csv', delete=False"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 72.5,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.6,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 104,
+  "repeats": 0,
+  "drafts_accepted": 29,
+  "drafts_offered": 92,
+  "tok_s": 45.4,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAnaconda    \nAntelope     \nBaboon       \nBadger      \nBat         \nBear        \nBeaver          \nBirds of Paradise (Manucode) \nBoa Constrictor \n\nBuffalo           \nCamel            \nCapybara             \nCat              \nChameleon\nCheetah\n\nChicken\n\n\nCrocodile  \nDeer  \n\nDingo   \nDog    \nDonkey     \nDragonfly   \n\nElephant       \nEmu      \nFalcon        \nFerret          \nFlamingo         \nFox<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 72,
+  "repeats": 0,
+  "drafts_accepted": 15,
+  "drafts_offered": 70,
+  "tok_s": 46.1,
+  "text": "The wind didn\u2019t just howl; it screamed against the glass of Lantern House 42, demanding entry into every crack and crevice that had been sealed with tar in previous years but now wept like old wounds on an open sore skin\u2014like a wound left unhealed by time. Elias knew this rhythm better than he remembered his own name<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 76,
+  "tok_s": 49.3,
+  "text": "The wind didn\u2019t just blow at Blackwood Rock; it screamed through the gaps in my teeth, searching for something to break its rhythm against stone that refused be broken by anything less than God himself had forged into being willing and able withstand eternity\u2019s cruelty without complaint or mercy from start finish until now forevermore if any such concept even existed within human comprehension let alone divine authority which we all knew did not exist but wished it would because wishful thinking is always easier to imagine hopelessly impossible truths as probable lies than vice versa.\n\nElias Thorne was eighty-three years old today, exactly one hundred and forty-seven days since his last heart attack had been logged by Dr Armitage who himself retired seven months ago after receiving a subpoena from the county clerk indicating he might have overstayed professional courtesy boundaries"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 94,
+  "repeats": 0,
+  "drafts_accepted": 27,
+  "drafts_offered": 56,
+  "tok_s": 46.7,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKangaroo      \nKoala        \nHippo         \nRhino          \nCheetah            \nLeopard           \nJaguar             \nWolf                \nFox              \nBear               \nDeer                 \nElk                  \nMoose                    \nBuffalo                        \nCamel                       \nDonkey                      \nGoat                     \nSheep                   \nCow                         \nBull                          \nOx                           \nPig                            \nBoar                             \nRabbit                              \nSquirrel                                <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 62.4,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 37,
+  "drafts_offered": 148,
+  "tok_s": 56.4,
+  "text": "The phrase \"the History Of The Sea\" is inherently ambiguous, as it could refer to several distinct historical narratives: geological formation (how oceans came into existence), maritime navigation and exploration by humans throughout human civilization; or perhaps more specifically in modern times refers often used metaphorically for economic trends like stock market fluctuations (\"sea change\"). However given your request likely intends an academic-style comprehensive overview covering both natural science aspects AND cultural significance through time periods\u2014especially since this kind question appears frequently on exams asking students write essays discussing evolution from primordial origins up present-day impacts including environmental changes due climate crisis etc.\u2014I will craft such broad yet structured piece combining geology + anthropology/ecological impact together while maintaining coherent flow suitable high-level educational context without overly technical jargon but still accurate enough scientifically grounded facts where appropriate within reasonable scope limits typical school assignments expect (~80\u201312% word count range implied implicitly though not stated explicitly here we aim moderate length ~65 words? Wait no user said \u201clong\u201d so probably >3k+ characters minimum safe bet around 7-9 paragraphs total approximating maybe close-to-full-page standard double-spaced equivalent which translates roughly \u22654K chars safely acceptable). Let me proceed accordingly ensuring clarity structure intro body conclusion style formal tone neutral informative balanced perspective acknowledging uncertainties if any minor ones exist e.g exact dates vary depending sources hence use hedging language sparingly only when necessary avoid overclaiming certainty especially regarding speculative theories vs established consensus knowledge base current mainstream understanding circa early-mid"
+ }
+]

bench/results/2026-09-29-layer-split-limits/data/norepeat-lim-cv-single-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 102,
+  "drafts_offered": 166,
+  "tok_s": 67.1,
+  "text": "The wind didn\u2019t just howl; it screamed. It tore through the iron gallery of the Blackwood Point Light, rattling the glass panes like a prisoner testing the bars of a cell.\n\nElias Thorne was not a man who feared the dark. He was a man who feared the silence.\n\nFor forty years, he had lived in the shadow of the Fresnel lens, a rotating eye of glass and brass that swept the Atlantic with a beam of pure, unyielding light. His job was simple: keep the light burning. Keep the logbook current. Keep the foghorn sounding when the mist rolled in, thick as wool.\n\nBut Elias knew something else. He knew that the light was not just for ships. It was for the lost.\n\nIt was November 14th, 1982. The storm was a Category 5, a monster that had swallowed the horizon. The waves were mountains of black water, crashing against the granite"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 134,
+  "drafts_offered": 180,
+  "tok_s": 81.0,
+  "text": "```python\nimport csv\nfrom typing import List, Dict\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, str]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile(mode='w', suffix='.csv', delete=False"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 62.4,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 73.4,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 104,
+  "repeats": 0,
+  "drafts_accepted": 29,
+  "drafts_offered": 92,
+  "tok_s": 43.5,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAnaconda    \nAntelope     \nBaboon       \nBadger      \nBat         \nBear        \nBeaver          \nBirds of Paradise (Manucode) \nBoa Constrictor \n\nBuffalo           \nCamel            \nCapybara             \nCat              \nChameleon\nCheetah\n\nChicken\n\n\nCrocodile  \nDeer  \n\nDingo   \nDog    \nDonkey     \nDragonfly   \n\nElephant       \nEmu      \nFalcon        \nFerret          \nFlamingo         \nFox<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 72,
+  "repeats": 0,
+  "drafts_accepted": 15,
+  "drafts_offered": 70,
+  "tok_s": 46.1,
+  "text": "The wind didn\u2019t just howl; it screamed against the glass of Lantern House 42, demanding entry into every crack and crevice that had been sealed with tar in previous years but now wept like old wounds on an open sore skin\u2014like a wound left unhealed by time. Elias knew this rhythm better than he remembered his own name<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 76,
+  "tok_s": 48.9,
+  "text": "The wind didn\u2019t just blow at Blackwood Rock; it screamed through the gaps in my teeth, searching for something to break its rhythm against stone that refused be broken by anything less than God himself had forged into being willing and able withstand eternity\u2019s cruelty without complaint or mercy from start finish until now forevermore if any such concept even existed within human comprehension let alone divine authority which we all knew did not exist but wished it would because wishful thinking is always easier to imagine hopelessly impossible truths as probable lies than vice versa.\n\nElias Thorne was eighty-three years old today, exactly one hundred and forty-seven days since his last heart attack had been logged by Dr Armitage who himself retired seven months ago after receiving a subpoena from the county clerk indicating he might have overstayed professional courtesy boundaries"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 94,
+  "repeats": 0,
+  "drafts_accepted": 27,
+  "drafts_offered": 56,
+  "tok_s": 46.8,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKangaroo      \nKoala        \nHippo         \nRhino          \nCheetah            \nLeopard           \nJaguar             \nWolf                \nFox              \nBear               \nDeer                 \nElk                  \nMoose                    \nBuffalo                        \nCamel                       \nDonkey                      \nGoat                     \nSheep                   \nCow                         \nBull                          \nOx                           \nPig                            \nBoar                             \nRabbit                              \nSquirrel                                <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 63.1,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 37,
+  "drafts_offered": 148,
+  "tok_s": 56.6,
+  "text": "The phrase \"the History Of The Sea\" is inherently ambiguous, as it could refer to several distinct historical narratives: geological formation (how oceans came into existence), maritime navigation and exploration by humans throughout human civilization; or perhaps more specifically in modern times refers often used metaphorically for economic trends like stock market fluctuations (\"sea change\"). However given your request likely intends an academic-style comprehensive overview covering both natural science aspects AND cultural significance through time periods\u2014especially since this kind question appears frequently on exams asking students write essays discussing evolution from primordial origins up present-day impacts including environmental changes due climate crisis etc.\u2014I will craft such broad yet structured piece combining geology + anthropology/ecological impact together while maintaining coherent flow suitable high-level educational context without overly technical jargon but still accurate enough scientifically grounded facts where appropriate within reasonable scope limits typical school assignments expect (~80\u201312% word count range implied implicitly though not stated explicitly here we aim moderate length ~65 words? Wait no user said \u201clong\u201d so probably >3k+ characters minimum safe bet around 7-9 paragraphs total approximating maybe close-to-full-page standard double-spaced equivalent which translates roughly \u22654K chars safely acceptable). Let me proceed accordingly ensuring clarity structure intro body conclusion style formal tone neutral informative balanced perspective acknowledging uncertainties if any minor ones exist e.g exact dates vary depending sources hence use hedging language sparingly only when necessary avoid overclaiming certainty especially regarding speculative theories vs established consensus knowledge base current mainstream understanding circa early-mid"
+ }
+]

bench/results/2026-09-29-layer-split-limits/data/norepeat-lim-cv-x2-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 85,
+  "drafts_offered": 172,
+  "tok_s": 36.6,
+  "text": "The wind didn\u2019t just howl; it screamed. It tore through the gaps in the iron railing of the Lantern Room, carrying the scent of brine and rotting kelp. Elias pulled his woolen collar up to his chin, his fingers numb and clumsy inside thick leather gloves.\n\nHe was the last man on the island. The automated systems had failed three winters ago, a glitch in the solar-powered backup that no engineer had been able to fix before the supply boat stopped coming. Now, it was just him, the light, and the sea.\n\nElias climbed the spiral stairs, his boots heavy with salt spray. The iron steps groaned under his weight, a sound that seemed to mock the storm outside. He reached the top, the glass walls vibrating with the force of the gale. Outside, the world was a blur of black and white. The ocean was invisible; it was only a presence, a weight pressing against the stone foundation of the tower.\n\nHe"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 136,
+  "drafts_offered": 173,
+  "tok_s": 96.8,
+  "text": "```python\nimport csv\nfrom typing import List, Dict\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, str]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile(mode='w', suffix='.csv', delete=False"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 88.7,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 94.1,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 104,
+  "repeats": 0,
+  "drafts_accepted": 30,
+  "drafts_offered": 96,
+  "tok_s": 55.9,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAnaconda    \nAntelope     \nBaboon       \nBadger      \nBat         \nBear        \nBeaver          \nBirds of Paradise (Manucode) \nBoa Constrictor \n\nBuffalo           \nCamel            \nCapybara             \nCat              \nChameleon\nCheetah\n\nChicken\n\n\nCrocodile  \nDeer  \n\nDingo   \nDog    \nDonkey     \nDragonfly   \n\nElephant       \nEmu      \nFalcon        \nFerret          \nFlamingo         \nFox<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 29,
+  "drafts_offered": 114,
+  "tok_s": 55.4,
+  "text": "The wind didn\u2019t just howl; it screamed against the glass of Lantern House 42, demanding entry into every crack and crevice that had been sealed with tar in previous years but now wept like old wounds at this one\u2019s age: forty-three winters since its construction. My name was Elias Thorne\u2014no middle initial\u2014and I kept no logbook for visitors because there were none; only ships passed by, their hulls cutting through the black water as if they too feared what lay beneath it more than above or beside them in fog so thick that even my own reflection vanished from sight when looked upon directly.\n\nI climbed each morning not to greet dawn but check whether light still lived within me\u2014the bulb\u2019s heart had failed twice before, once during a storm named *Harrow*,"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 29,
+  "drafts_offered": 110,
+  "tok_s": 52.2,
+  "text": "The wind didn\u2019t just blow at Blackwood Rock; it screamed through the gaps in my teeth, searching for something to hold onto until I died.\u201d said Elias Thorne as he adjusted his collar against November\u2019s bite.** He was not merely complaining of cold weather but articulating a poetic metaphor: that even after death (metaphorically), there would be no relief from this relentless gale.\n\nElias had been the last human keeper on any significant stretch coast since automation became standard practice, yet here he stood , stubbornly refusing to let technology replace his vigilance . His cottage perched atop a crumbling granite foundation was more than just shelter; it housed centuries of handwritten logs detailing every ship\u2019s name and cargo load that passed within sight.\n\nOn most nights Elias tended the light mechanically\u2014"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 91,
+  "repeats": 0,
+  "drafts_accepted": 29,
+  "drafts_offered": 69,
+  "tok_s": 58.7,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKangaroo      \nKoala        \nHippopotamus          \nRhinoceros            \nLeopard                \nCheetah                 \nPanther                  \nWolf                    \nFox                      \nBear                     \nDeer                       \nElk                        \nMoose               \nBuffalo                   \nCamel           \nDonkey              \nGoat             \nSheep         \nCow  \nBull   \nOx    \nPig     \nDog       \nCat<|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 12,
+  "repeats": 0,
+  "drafts_accepted": 4,
+  "drafts_offered": 17,
+  "tok_s": 60.3,
+  "text": "1 2,3456789<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 41,
+  "drafts_offered": 168,
+  "tok_s": 59.9,
+  "text": "The phrase \"the History Of The Sea\" is inherently ambiguous, as it could refer to several distinct historical narratives: geological formation (how oceans came into existence), maritime navigation and exploration histories (\"history at/over/across/on/in/of/to/from/toward/off by means around above below through under upon alongside against among between within without throughout despite during after before since until while when where who whom whose which what how why whether\"), or perhaps even metaphorical interpretations involving human interaction with water bodies over time periods like ages eras epochs generations centuries decades years months days hours minutes seconds moments instants times occasions instances events episodes incidents occurrences situations circumstances conditions states phases stages steps processes procedures sequences series chains cycles loops repetitions iterations continuations progressions developments evolutions transformations changes alterations modifications adjustments corrections revisions amendments updates upgrades improvements enhancements refinements optimizations perfecting completing finishing concluding ending terminating ceasing stopping halting pausing resuming continuing proceeding advancing progressing moving shifting translating rotating reflecting mirroring flipping turning twisting bending curving arches spirals helix coils twists turns rotations reflections mirrors flips inversions reversings reversion backtracking retracing reversing undo redo do again repeat repeats repeated repeatedly repetition repetitive repetitiveness repete repeater repeatingly recurrent recurrence recurring recurs recur recurred recursion recursive recursively recurse recessed recessionary regress regression regressed retrograde reverse reversed reversal revert reverted reverberate resonated echo echoed echoing echoes echolocation sonar radar radio telecommunication communication communications communicate communicated communicating communicative comms comm chatter chat chats chatting chatted talk talks talked"
+ }
+]

bench/results/2026-09-29-layer-split-limits/data/norepeat-lim-hp0-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 171,
+  "tok_s": 62.5,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and shadow.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined his life for thirty years.\n\nElias was not a man of many words. He was a man of many tasks. His hands, mapped with scars from broken glass and rusted iron, moved with a rhythm that predated his memory. He checked the clockwork mechanism of the Fresnel lens, ensuring the rotation was smooth, silent, and precise. *Tick. Tick. Tick.* The sound was the heartbeat of the coast."
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 135,
+  "drafts_offered": 184,
+  "tok_s": 85.2,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 78.1,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 80.5,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 59,
+  "repeats": 0,
+  "drafts_accepted": 13,
+  "drafts_offered": 48,
+  "tok_s": 46.6,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBuffalo     \n\nCat          \nCheetah      \n\nChimpanzee \nClam         \n\n\nCrocodile   \n\n\n\nDolphin      \n\n\n\n\nDonkey       <|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 30,
+  "drafts_offered": 101,
+  "tok_s": 51.3,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral vaults below Elias\u2019s boots smelled faintly brine and old coffee as he climbed spiral stairs worn smooth by decades feet his hands were calloused from gripping iron railings that had been polished to a dull shine\n\nElias was not merely keeping light alive; He kept order intact. Every night, at precisely 18:07 hours local time\u2014adjusted for leap seconds and magnetic drift\u2014he wound the clockwork mechanism inside brass housing behind Fresnel lens This wasn\u2019t vanity It\u2019s necessity The rotation speed must remain constant or ships lose their bearing in fog where compasses fail\n\nHe spoke softly to himself while working, murmuring numbers aloud because silence made him"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 67,
+  "repeats": 0,
+  "drafts_accepted": 14,
+  "drafts_offered": 68,
+  "tok_s": 41.2,
+  "text": "The wind didn\u2019t just blow at Blackwood Rock; it screamed through the gaps in his teeth, like he had no mouth to hold back its fury anymore that night.\u201d Elara pulled her wool coat tighter around herself as she climbed spiral staircase of **Blackrock Lighthouse**. Her boots were slick with salt spray from previous days.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 91,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 54,
+  "tok_s": 48.8,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKoala      \nKangaroo        \nHippopotamus          \nRhinoceros            \nCamel                 \nCheetah               \nLeopard                \nJaguar                  \nBear                    \nWolf                     \nFox                      \nDog                       \nCat                        \nMouse                   \nRat                         \nSquirrel              \nDeer                      \n\nAntelope                \n\nBuffalo                 \n\n\nBison                   \n\n\n\nElk                    \n\n\n\n\nMoose                  <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 11,
+  "tok_s": 68.8,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 32,
+  "drafts_offered": 107,
+  "tok_s": 52.3,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant storybook that weaves together geology with biology as wellas human culture into an epic narrative spanning billions years from its earliest beginnings in primordial times through modern-day ecological crises The vast expanse covers more than seventy percent Earth\u2019s surface area making up nearly all life support systems on our planet yet despite being so large humans have only explored less five thousand square kilometers out total 361 million km\u00b2 which means over ninety-nine point nine eight six per cent unexplored territory This makes oceans arguably larger unknown territories compared space exploration where NASA estimates they\u2019ve mapped roughly ten billion stars versus current marine surveys covering just two hundred fifty-six species documented globally against estimated millions undiscovered organisms including deep-sea creatures like anglerfish or vampire squid whose existence was first recorded centuries ago but still remain poorly understood due extreme pressures depths exceeding eleven miles below Pacific Trenches Mariana located near Guam island chain reaching maximum depth known today at approximately thirty-seven feet under pressure equivalent atmospheric levels causing structural collapse if unprotected by specialized vessels capable handling such conditions without failure mechanisms designed specifically for submersible operations operating within these zones require titanium hulls rated withstand forces greater gravity acting upon objects falling vertically downward acceleration rates calculated using Newtonian physics principles applied universally across gravitational fields regardless medium density variations present throughout different layers stratosphere troposphere mesospheres thermoclines etceterally speaking however when discussing aquatic environments instead air masses replace gases dissolved oxygen nitrogen arg"
+ }
+]

bench/results/2026-09-29-layer-split-limits/data/norepeat-lim-hp1-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 94,
+  "drafts_offered": 169,
+  "tok_s": 43.0,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy-three years old, and he had been the keeper of the Blackwood Point Light for forty of those years. The other thirty-three had been spent in a small, drafty cottage that smelled perpetually of kerosene and damp wood.\n\nFor decades, Elias had tended the great Fresnel lens, polishing the glass prisms until they gleamed like diamonds, winding the clockwork mechanism that turned the beam, and trimming the wick of the oil lamp. It was a job of patience, of precision, of silence. The light was his charge. The sea was his neighbor.\n\nBut this morning, the sea was not his neighbor. It was his enemy.\n\nThe storm had been forecasted for three days, but it had arrived in one. The barometric pressure had dropped so sharply that Elias\u2019s ears popped."
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 136,
+  "drafts_offered": 178,
+  "tok_s": 63.7,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 56.8,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 57.7,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 94,
+  "repeats": 0,
+  "drafts_accepted": 24,
+  "drafts_offered": 49,
+  "tok_s": 43.3,
+  "text": "Lion  \nTiger   \nElephant    \nGiraffe     \nZebra       \nHippopotamus \nRhinoceros      \nCheetah         \nLeopard        \nJaguar          \nPuma           \nWolf            \nFox             \nBear              \nDeer                \nElk                 \nMoose               \nAntelope                  \nBison                   \nBuffalo                    \nCamel                     \nDonkey                      \nGoat                       \nSheep                        \nCow                         \nDog                          \nCat                           \nRabbit                            \nMouse                             \nHamster                               \nGuinea Pig<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 35,
+  "drafts_offered": 108,
+  "tok_s": 41.4,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral vaults below Elias\u2019s boots smelled faintly brine and old coffee as he climbed spiral stairs worn smooth by decades feet his hands were calloused from gripping iron railings that had been polished to a dull shine.\n\nElias wasn\u2019t just keeping light; was guarding silence the beam swept across black water every seven seconds cutting through fog so thick it swallowed sound entirely for three nights straight no ships answered its signal not even fishing trawlers seeking shelter they\u2019d long abandoned this stretch of coast where currents shifted unpredictably pulling debris from sunken wrecks into tangled heaps that snagged anchors and tore hulls open like wounds.\n\nHe remembered the last time someone had"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 59,
+  "repeats": 0,
+  "drafts_accepted": 11,
+  "drafts_offered": 46,
+  "tok_s": 33.7,
+  "text": "The wind didn\u2019t just blow at Blackwood Rock; it screamed through the gaps in his teeth, trying to pry them apart like loose planks of weathered oak outside your own mouth again and then you are going away from home is fine but for now let me write a proper story.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 60,
+  "repeats": 0,
+  "drafts_accepted": 17,
+  "drafts_offered": 48,
+  "tok_s": 43.2,
+  "text": "Aardvark  \nAntelope allonax (Oryx) \u2013 *Note: Using standard names for clarity*  \n\nLet me provide a clean list of exactly **1** animal name on each single physical text file/terminal output stream with no other characters except the newline (\\n).<|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 11,
+  "tok_s": 48.9,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 27,
+  "drafts_offered": 105,
+  "tok_s": 40.2,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\" \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \"the Sea\" or \u201cOcean\u201d was never just an element\u2014it became civilization itself through its capacity as both barrier (to land) but also conduit between lands\u2014and thus gave rise directly from being part nature\u2019s own evolutionary engine into becoming humanity's greatest historical actor across all time scales since first emergence until now present-day future projections still ongoing today within our current era context window limits allowed here by system constraints set above threshold levels required before output generation begins below line break marker inserted immediately after final punctuation mark ending entire preceding paragraph block structure defined earlier during initial drafting phase prior execution stage completed successfully without errors detected internally via internal validation routines executed automatically upon completion cycle termination signal received end-of-file condition triggered correctly according standard protocol specifications followed strictly adhered-to throughout process flow control logic implemented using structured programming constructs compatible with target runtime environment configuration parameters specified herein explicitly declared beforehand at top-level scope declaration section located near beginning source code file header comments included inline documentation style guidelines observed consistently applied uniformly everywhere else similarly formatted sections appearing later down document body text area scrolling vertically downward directionally aligned left margin justified right edge wrapped around fixed-width column boundaries determined dynamically based available screen real estate measured pixel-per-unit scaling factor adjusted responsively depending viewport dimensions reported back client-side rendering pipeline components responsible"
+ }
+]

bench/results/2026-09-29-layer-split-limits/data/norepeat-lim-kv-same24-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 94,
+  "drafts_offered": 167,
+  "tok_s": 62.7,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the fabric stiff with salt and ice. Outside the thick glass of the lantern room, the Atlantic was a churning cauldron of black and white, the waves crashing against the granite base with a violence that shook the very foundations of the tower. It was November, 1924, and the storm had been raging for three days.\n\nElias was not a man who feared the dark. He was a man who feared the silence.\n\nHe checked the clock on the wall. 11:42 PM. The rotation of the great Fresnel lens was steady, a rhythmic *whirrr-click, whirrr-click* that seemed to hold the chaos of the ocean at bay. The beam swept out, a blade of pure, concentrated light cutting through the rain and sleet, illuminating the jagged teeth of the reef below.\n\n"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 136,
+  "drafts_offered": 179,
+  "tok_s": 81.3,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 71.0,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 71.9,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 61,
+  "repeats": 0,
+  "drafts_accepted": 18,
+  "drafts_offered": 52,
+  "tok_s": 47.9,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBoar          \nBuffalo \nButterfly \n\nCamel           \nCat            \nCheetah        \n\nDolphin         \n\n\nElephant        \n\n\n\nFalcon          <|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 29,
+  "drafts_offered": 122,
+  "tok_s": 47.4,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral vaults below Elias\u2019s boots were heavy with salt and exhaustion as he climbed one hundred twelve steps to reach his post at dawn every morning for thirty years. His name was written on a brass plaque by door: **Elias Thorne**.\n\nHe wasn\u2019t just keeping light; He kept time, weathered hands gripping the cold iron railing until knuckles turned white against frostbite and windburn that never healed fully because they were always there waiting to be felt again when he stepped outside into storm or calm alike. The lens above him was a Fresnel giant\u2019s eye made of hand-ground crystal prisms each one polished daily with chamois cloth soaked in whale"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 65,
+  "repeats": 0,
+  "drafts_accepted": 11,
+  "drafts_offered": 58,
+  "tok_s": 42.6,
+  "text": "The wind didn\u2019t just blow at Cape Hollow; it screamed through the gaps in my cottage\u2019s shutters, rattling them like loose teeth until they slammed against their hinges every night for eleven months of winter and summer alike.\u201d I had never seen calm water here\u2014only gray churning seas that swallowed ships smaller than themselves<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 101,
+  "repeats": 0,
+  "drafts_accepted": 32,
+  "drafts_offered": 53,
+  "tok_s": 49.4,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKoala      \nKangaroo        \nCheetah         \nHippopotamus          \nRhinoceros            \nLeopard             \nJaguar              \nWolf                \nFox                 \nBear                  \nDeer                    \nOtter                   \nBeaver                      \nBadger                        \nMongoose                          \nMeerkat                           \nHyena                             \nJackal                            \nAntelope                              \nBison                               \nElk                                 \nCaribou                                   \nLlama                                    \nAlpace                                     <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 12,
+  "tok_s": 64.7,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 52,
+  "drafts_offered": 148,
+  "tok_s": 49.9,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\" \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \"the Sea\" or \u201cOcean\u201d was never just an element within our planet\u2019s geography but rather its most fundamental biological engine\u2014a living system whose existence predates humanity by hundreds millions years yet continues shaping every aspect from climate regulation through evolutionary biology down even into modern industrial logistics today's global supply chains depend upon them directly indirectly via atmospheric pressure systems weather patterns migration routes food webs carbon cycles oxygen generation thermal buffering tidal forces salinity gradients depth stratification light attenuation currents circulation gyres upwelling zones thermoclines halothermics bathymetric features continental shelves abyssal plains trenching ridges volcanic arcs island formations atoll structures seamounts submarine canons canyon walls cliff faces overhang ledge platforms reef flats fore-reef back-lagoon inner-shelf outer-slope drop-offs pinnacle pinnacles spires columns arches tunnels caves grottos swim-through passages chimney vents blowholes fissures cracks crevices joints faults fractures breaks gaps holes openings void spaces hollow chambers cavernous interiors subterranean networks underground rivers lakes pools springs wells shaft pits sinks drains sewers pipes conduits tubes channels gutters ditches furrows grooves rills runnels streams brooks rivulets creeks tributaries branches forks splits divergences con"
+ }
+]

bench/results/2026-09-29-layer-split-limits/data/norepeat-lim-kv-single-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 94,
+  "drafts_offered": 167,
+  "tok_s": 49.2,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the fabric stiff with salt and ice. Outside the thick glass of the lantern room, the Atlantic was a churning cauldron of black and white, the waves crashing against the granite base with a violence that shook the very foundations of the tower. It was November, 1924, and the storm had been raging for three days.\n\nElias was not a man who feared the dark. He was a man who feared the silence.\n\nHe checked the clock on the wall. 11:42 PM. The rotation of the great Fresnel lens was steady, a rhythmic *whirrr-click, whirrr-click* that seemed to hold the chaos of the ocean at bay. The beam swept out, a blade of pure, concentrated light cutting through the rain and sleet, illuminating the jagged teeth of the reef below.\n\n"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 136,
+  "drafts_offered": 179,
+  "tok_s": 76.2,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 73.1,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 73.4,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 61,
+  "repeats": 0,
+  "drafts_accepted": 18,
+  "drafts_offered": 52,
+  "tok_s": 47.1,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBoar          \nBuffalo \nButterfly \n\nCamel           \nCat            \nCheetah        \n\nDolphin         \n\n\nElephant        \n\n\n\nFalcon          <|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 29,
+  "drafts_offered": 122,
+  "tok_s": 47.0,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral vaults below Elias\u2019s boots were heavy with salt and exhaustion as he climbed one hundred twelve steps to reach his post at dawn every morning for thirty years. His name was written on a brass plaque by door: **Elias Thorne**.\n\nHe wasn\u2019t just keeping light; He kept time, weathered hands gripping the cold iron railing until knuckles turned white against frostbite and windburn that never healed fully because they were always there waiting to be felt again when he stepped outside into storm or calm alike. The lens above him was a Fresnel giant\u2019s eye made of hand-ground crystal prisms each one polished daily with chamois cloth soaked in whale"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 65,
+  "repeats": 0,
+  "drafts_accepted": 11,
+  "drafts_offered": 58,
+  "tok_s": 42.3,
+  "text": "The wind didn\u2019t just blow at Cape Hollow; it screamed through the gaps in my cottage\u2019s shutters, rattling them like loose teeth until they slammed against their hinges every night for eleven months of winter and summer alike.\u201d I had never seen calm water here\u2014only gray churning seas that swallowed ships smaller than themselves<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 101,
+  "repeats": 0,
+  "drafts_accepted": 32,
+  "drafts_offered": 53,
+  "tok_s": 50.0,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKoala      \nKangaroo        \nCheetah         \nHippopotamus          \nRhinoceros            \nLeopard             \nJaguar              \nWolf                \nFox                 \nBear                  \nDeer                    \nOtter                   \nBeaver                      \nBadger                        \nMongoose                          \nMeerkat                           \nHyena                             \nJackal                            \nAntelope                              \nBison                               \nElk                                 \nCaribou                                   \nLlama                                    \nAlpace                                     <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 12,
+  "tok_s": 64.6,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 52,
+  "drafts_offered": 148,
+  "tok_s": 49.8,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\" \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \"the Sea\" or \u201cOcean\u201d was never just an element within our planet\u2019s geography but rather its most fundamental biological engine\u2014a living system whose existence predates humanity by hundreds millions years yet continues shaping every aspect from climate regulation through evolutionary biology down even into modern industrial logistics today's global supply chains depend upon them directly indirectly via atmospheric pressure systems weather patterns migration routes food webs carbon cycles oxygen generation thermal buffering tidal forces salinity gradients depth stratification light attenuation currents circulation gyres upwelling zones thermoclines halothermics bathymetric features continental shelves abyssal plains trenching ridges volcanic arcs island formations atoll structures seamounts submarine canons canyon walls cliff faces overhang ledge platforms reef flats fore-reef back-lagoon inner-shelf outer-slope drop-offs pinnacle pinnacles spires columns arches tunnels caves grottos swim-through passages chimney vents blowholes fissures cracks crevices joints faults fractures breaks gaps holes openings void spaces hollow chambers cavernous interiors subterranean networks underground rivers lakes pools springs wells shaft pits sinks drains sewers pipes conduits tubes channels gutters ditches furrows grooves rills runnels streams brooks rivulets creeks tributaries branches forks splits divergences con"
+ }
+]

bench/results/2026-09-29-layer-split-limits/data/norepeat-lim-kv-x2-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 97,
+  "drafts_offered": 176,
+  "tok_s": 73.0,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was the last of the line, the only man left on the jagged spine of Widow\u2019s Rock. For forty years, he had tended the light. For forty years, he had listened to the ocean\u2019s endless, hungry whisper.\n\nTonight was different. The storm wasn\u2019t just a storm; it was a *tempest*. The waves didn\u2019t crash; they *shattered*. They threw themselves against the granite base of the lighthouse with the force of a battering ram, sending tremors up through the soles of Elias\u2019s boots.\n\nHe climbed the spiral stairs, one hundred and twelve steps of iron and rust. His hand, calloused and scarred from decades of polishing the brass railing, found the familiar groove in the metal. He didn\u2019t look up. He didn\u2019t look down. He looked *forward*."
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 135,
+  "drafts_offered": 176,
+  "tok_s": 102.5,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 90.9,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 91.7,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 118,
+  "repeats": 0,
+  "drafts_accepted": 38,
+  "drafts_offered": 103,
+  "tok_s": 62.8,
+  "text": "Aardvark  \nAlbatross   \nAlligator    \nAntelope     \nBaboon       \nBadger      \nBeaver        \nBear          \nBird of Paradise \nBuffalo         \nCamel           \nCheetah            \nChimpanzee             \nCrane              \nCrocodile                \nDeer                 \nDolphin               \nDonkey                  \nElephant                   \nFlamingo                    \nFox                       \nGiraffe                     \nGoat                      \nHippopotamus  \nIguana       \nJaguar        \nKangaroo      \nKoala          \nLion         \nLeopard     \nMonkey    \nMoose   \nOstrich \nOtter\nPanda<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 58,
+  "repeats": 0,
+  "drafts_accepted": 9,
+  "drafts_offered": 63,
+  "tok_s": 50.4,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty jail cell below sea level three hundred feet down to where Elias kept his boots dry and ready for emergencies when he had no time even think about changing them.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 140,
+  "repeats": 0,
+  "drafts_accepted": 22,
+  "drafts_offered": 90,
+  "tok_s": 50.1,
+  "text": "The wind didn\u2019t just blow at Blackwood Rock; it screamed through the gaps in his teeth, demanding entry into every crack and crevice of Stone Harbor Lighthous eKeepers have always been solitary figures against isolation\u2019s greatest adversary: time itself Elias Thorne knew this intimately. For thirty years he had kept watch over one hundred thousand ships each night scanning for lights that might signal disaster or hope\n\nHis job was simple but deadly serious He checked weather logs every day without fail The foghorn echoed through the tower three times before dawn Each morning at 4:02 AM Elias would climb eighty-two steps to reach his office where coffee cooled beside a single candle flickering in drafts from below<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 96,
+  "repeats": 0,
+  "drafts_accepted": 31,
+  "drafts_offered": 69,
+  "tok_s": 58.9,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKangaroo      \nKoala        \nHippo         \nRhino          \nCheetah            \nLeopard             \nJaguar              \nPanther               \nCougar                \nBear                  \nWolf                    \nFox                     \nDeer                      \nElk                        \nMoose                       \nBison                         \nBuffalo                          \nAntelope                           \nCamel                            \nDonkey                              \nGoat                                \nSheep                                  \nCow                                    \nPig                                     \nHorse                                       <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 71.2,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 37,
+  "drafts_offered": 119,
+  "tok_s": 61.3,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth\u2019s surface.\u201d \u2014 *This quote by Rachel Carson serves as an apt introduction.* The story we tell today begins with that first drop falling into primordial pools billions years ago\u2014but what follows next? What happens when those drops become oceans themselves\u2014when they begin moving across continents carrying sediment from mountains down rivers until finally reaching their final resting place at last\u2014the deep abyssal plains where sunlight cannot reach nor pressure crushes everything else alive there below depths greater than six kilometers beneath waves crashing against shores above them all around world forever changing shape over time due plate tectonics shifting landmass slowly apart together again repeatedly throughout geological epochs spanning hundreds millions or even thousands million+ yrs since formation early Hadean era through Archeozoic eons onward present Holocene epoch still ongoing now right here under current conditions observed via satellites monitoring global climate patterns affecting currents temperatures salinity levels biodiversity distribution human impacts altering ecosystems directly indirectly causing extinction events locally globally recently historically prehistorically anciently mythologically culturally spiritually philosoph scientifically mathematically physically chem biol geo astro cosmological sociopolitical economic environmental ecological evolutionary genetic linguistic literary artistic musical cinematic theatrical photographic journalistic digital analog virtual real tangible intangible metaphysical ontosemantic semiotic symbolic allegorical narrative descriptive analytical interpretive evaluative critical reflective introspective contemplatory ponderous thoughtful considerate attentive observant perceptive insightful intelligent knowledgeable informed educated literati learned skilled capable"
+ }
+]

bench/results/2026-09-29-layer-split-limits/data/norepeat-lim-nocv-x2-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 87,
+  "drafts_offered": 170,
+  "tok_s": 71.4,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and green.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined his life for thirty years.\n\nElias was not a man of many words. He was a man of many tasks. His hands, mapped with scars from rope burns and broken glass, moved with a rhythm that predated his memory. He checked the clockwork mechanism of the Fresnel lens, listening for the faint, rhythmic *tick-tick-tick* that signaled the rotation was true. If the rhythm faltered, ships"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 136,
+  "drafts_offered": 178,
+  "tok_s": 104.1,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 90.1,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 94.8,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 76,
+  "repeats": 0,
+  "drafts_accepted": 20,
+  "drafts_offered": 64,
+  "tok_s": 61.6,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear         \nBird        \nBuffalo          \nButterfly \nCamel           \nCat            \nCaterpillar \n\nCentipede    \n\nChameleon   \n\nCheetah     \n\nChicken      \n\nCricket       \n\nCrocodile     \n\n\nDolphin       \n\n\n\nDonkey        \n\n\n\n\nDragon<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 21,
+  "drafts_offered": 90,
+  "tok_s": 59.2,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty jail cell below sea level to above cloud height at once\u2014no matter how high you climbed or deep into its foundations down through rock and saltwater. It was November again when Elias first noticed that his logbook had begun recording entries he hadn\u2019t written yet: *Storm surge expected Tuesday, but the sky is clear.*\n\nHe stopped writing for a moment to listen\u2014not with ears alone\u2014but by pressing both palms flat against cold iron walls where condensation formed in slow spirals like fingerprints left behind after touching something invisible. The metal hummed faintly beneath his skin as if remembering old storms long past their dates on paper calendars torn from newspapers dated three years ahead of today"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 17,
+  "drafts_offered": 63,
+  "tok_s": 52.6,
+  "text": "The wind didn\u2019t just blow at Blackwood Rock; it screamed through the gaps in his teeth, trying to get inside where he had spent thirty years refusing entry except for family and friends like him who were old enough too know better than young men ever could be again about what safety meant. It was November 2014 when Elias Thorne first stepped onto that spiral staircase with only one boot on because his other foot hurt after walking from the harbor office all morning long before anyone else thought he needed shoes off there for work even if nobody saw them do it but everyone knew why so they did not ask him anymore once this happened already twice last winter near same season now here again today too soon then tomorrow maybe never will come back unless someone sends message tonight via radio signal code Morse dot dash pattern"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 94,
+  "repeats": 0,
+  "drafts_accepted": 29,
+  "drafts_offered": 60,
+  "tok_s": 58.8,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nKangaroo       \nPanda        \nKoala          \nCheetah      \nWolf         \nFox           \nBear            \nMonkey             \nDolphin              \nWhale                \nShark                 \nOctopus               \nJellyfish                  \nCrab                    \nSpider                   \nRabbit                     \nCat                      \nDog                       \nHorse                        \nCow                          \nSheep                            \nGoat                              \nChicken                                \nRooster                                   \nTurkey                                  \nEagle                                    \nFalcon                                     <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 71.8,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 52,
+  "drafts_offered": 148,
+  "tok_s": 61.8,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\u201d \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \u201cthe Sea\u201d or \"Ocean\" (terms often used interchangeably but distinct geographically) was never just an element within our planet\u2019s geography\u2014it became its engine itself\u2014a living system whose evolution shaped everything else around us from climate patterns down through evolutionary biology right up into modern civilization's economic structures cultural identities political boundaries ecological systems scientific understanding technological capabilities philosophical frameworks psychological experiences artistic expressions literary metaphors cinematic narratives musical compositions architectural designs urban planning strategies agricultural practices industrial processes medical discoveries legal precedents educational curricula religious traditions spiritual beliefs mythological stories folkloric tales poetic verses dramatic scripts visual arts photographic images digital media social networks behavioral norms emotional responses cognitive models decision-making algorithms predictive analytics statistical probabilities probabilistic reasoning causal inference counterfactual thinking hypothetical scenarios normative ethics deontal logic modal semantics temporal dynamics spatial relations topographical features hydrological cycles meteor phenomena atmospheric chemistry geological formations tectonic shifts volcanic eruptions seismic events tsunamis hurricanes cyclones monsoons typhoons droughts floods storms winds currents waves swells breakers surf foam spray mist fog rain snow ice glaciers berg floes drift pack-ice brine salinity pH alkalines acidity hardness softness temperature pressure depth light darkness color hue saturation brightness contrast luminance chroma tone value shade tint reflection"
+ }
+]

bench/results/2026-09-29-layer-split-limits/data/norepeat-lim-reg-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 187,
+  "tok_s": 58.7,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, mapped with veins like the roots of an ancient oak, gripped the cold iron railing.\n\nFor forty years, Elias had tended the light at the end of the world. Not the end of the map, but the end of the land. Here, the Atlantic threw itself against the granite cliffs with such violence that the spray reached the gallery, thirty feet above the rocks.\n\nHe checked the clock. 11:42 PM. The fog was thickening, a gray soup that swallowed the horizon. In the old days, he would have climbed the stairs to the lantern room, polished the Fresnel lens, and wound the clockwork mechanism that turned the great beam. But the automation had changed"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 135,
+  "drafts_offered": 179,
+  "tok_s": 80.7,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 71.8,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 73.0,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 61,
+  "repeats": 0,
+  "drafts_accepted": 19,
+  "drafts_offered": 49,
+  "tok_s": 49.1,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBoar          \nBuffalo \nButterfly \n\nCamel           \nCat            \nCheetah        \n\nDolphin         \n\n\nElephant        \n\n\n\nFlamingo<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 80,
+  "repeats": 0,
+  "drafts_accepted": 6,
+  "drafts_offered": 52,
+  "tok_s": 46.2,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to check what was happening with that storm and why they were calling me out there when all we had left for tonight\u2019s shift is one last pass through our logs before dawn breaks.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 17,
+  "drafts_offered": 83,
+  "tok_s": 43.9,
+  "text": "The wind didn\u2019t just blow at Cape Harrow; it screamed through the gaps in its teeth, hungry and relentless for three centuries of stone salt-spray history that had long since forgotten to be polite or gentle with itself against anything smaller than God Himself was willing acknowledge as divine protection from drowning souls who were already lost before they ever touched water.\n\nElias knew this rhythm better he himself could hear it sometimes; his boots worn smooth by pacing the spiral stair\u2019s center-post groove where thousands upon countless steps had rubbed away texture until all that remained beneath sole leather now felt like polished bone carved out slowly but surely over time into something more hollow than wood ever really should be allowed become when used so often enough within such confined circular spaces meant only for vertical ascent down always descending again back upward never truly leaving"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 62,
+  "repeats": 0,
+  "drafts_accepted": 18,
+  "drafts_offered": 31,
+  "tok_s": 49.1,
+  "text": "Lion  \nTiger  \n\nElephant  \n  \nGiraffe  \n\n\n\nHippopotamus    \nZebra   \nRhinoceros   \n\n\nPanda Bear    \n\n\n\n\nKangaroo     \nKoala      \nSloth      \n\n\n\n\n\nCheetah       \nLeopard        \nPanther        \n\n\n\n\n\n\nWolf              \nFox                \nBear                 <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 62.7,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 47,
+  "drafts_offered": 153,
+  "tok_s": 52.2,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\" \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \"the Sea\" or \u201cOcean\u201d was never just an element within our planet\u2019s geography but rather its most fundamental biological engine\u2014a living system whose existence predates humanity by billions years yet continues shaping every aspect from climate regulation through evolutionary biology all way down modern industrial supply chains today's global shipping lanes carry cargo across vast distances because they were once crossed only slowly via coastal hopping routes used since antiquity until steamships revolutionized travel during late nineteenth century when diesel engines replaced wind power entirely allowing container ships carrying thousands TEUs (Twenty-foot Equivalent Units) per voyage crossing entire Pacific basins without stopping at intermediate ports anymore unlike earlier sailing vessels needing favorable winds currents seasonal patterns dictating departure windows based upon monsoon calendars established centuries ago before technological advances changed those constraints fundamentally altering logistics networks worldwide thereby transforming economic structures globally while simultaneously creating ecological disruptions due invasive species transported unintentionally alongside legitimate goods causing cascading ecosystem collapses locally where non-native organisms outcompete native fauna disrupting food webs leading population declines affecting biodiversity indices measured scientifically using standardized metrics like Shannon-Wiener diversity index calculations applied consistently over decades revealing trends showing significant decreases post-introduction events correl strongly with introduction dates matching historical records documenting first sightings confirmed taxonomically classified formally according International Code Nomenclature rules"
+ }
+]

bench/results/2026-09-29-layer-split-limits/data/pp-lim-cv-pp-x2.json

@@ -0,0 +1,84 @@
+{
+ "label": "lim-cv-pp-x2",
+ "extra": [
+  "--layer-split",
+  "auto",
+  "--control-vector-scaled",
+  "C:\\Users\\ribbon\\strata-test\\Strata\\data\\experimental-speed-projection\\Qwen3.8-Flash-Next-experimental-speed-projection.gguf:1.0",
+  "--control-vector-layer-range",
+  "4",
+  "44",
+  "--cvec-mode",
+  "project",
+  "--cvec-dir",
+  "per-layer"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 15925,
+   "prompt_ms": 8783.1,
+   "pp_tok_s": 1813.1,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27474,
+   "prompt_ms": 13379.7,
+   "pp_tok_s": 2053.4,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3652.4,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 246.8,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 77.0,
+   "reps": [
+    77.0
+   ],
+   "hit_rate": 0.9718150054476462,
+   "text": "The wind didn\u2019t just howl; it screamed. It tore through the gaps in the iron railing of the Lantern Room, carrying the scent of brine and rotting kelp. Elias pulled his woolen collar up to his chin, his fingers numb and clumsy inside thick leather gloves.\n\nHe was the last man on the island. The automated systems had failed three winters ago, a glitch in the solar panel array that no engineer had been able to fix before the supply boat stopped running. Now, it was just him, the light, and the sea.\n\nElias climbed the spiral stairs, his boots heavy with salt spray. The iron steps groaned under his weight, a sound that seemed to mock the storm outside. He reached the top, the glass walls vibrating with the force of the gale.\n\nInside the lantern room, the great Fresnel lens sat like a dormant eye. It was cold, dark, and waiting.\n\nElias moved to the control panel. His"
+  },
+  {
+   "name": "code",
+   "tok_s": 103.6,
+   "reps": [
+    103.6
+   ],
+   "hit_rate": 0.957731667226356,
+   "text": "```python\nimport csv\nfrom typing import List, Dict\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, str]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile(mode='w', suffix='.csv', delete=False"
+  }
+ ],
+ "stage_timing": [],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: layer split auto: CUDA0 84 SMs at 2.81 GHz -> 0.31 ms per layer, 8.30 GiB for experts",
+  "strata generate: layer split auto: CUDA1 82 SMs at 1.86 GHz -> 0.48 ms per layer, 14.14 GiB for experts",
+  "strata generate: layer split auto: K=22 - predicted 19.7 ms per decode window; the caches hold 11298 of 12288 profiled pairs (~99.7% of the routed mass)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 3479 slots, 6.22 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 22-47, expert cache 6656 slots (13.34 GiB), 6656 of its 6656 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-21",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-21 (CUDA0), 22-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 91.6% (5554 hits / 6062 lookups)",
+  "strata serve: decode expert cache hit rate: 95.3% (3592 hits / 3771 lookups)",
+  "strata serve: decode expert cache hit rate: 96.5% (3650 hits / 3782 lookups)",
+  "strata serve: decode expert cache hit rate: 96.3% (25027 hits / 25980 lookups)",
+  "strata serve: decode expert cache hit rate: 96.5% (3201 hits / 3318 lookups)",
+  "strata serve: decode expert cache hit rate: 97.2% (129334 hits / 133085 lookups)",
+  "strata serve: decode expert cache hit rate: 95.8% (108375 hits / 113158 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split-limits/data/pp-lim-kv-pp-single.json

@@ -0,0 +1,76 @@
+{
+ "label": "lim-kv-pp-single",
+ "extra": [
+  "--max-context",
+  "131072",
+  "--kv-resident",
+  "32768"
+ ],
+ "gpus": "0",
+ "needles": [
+  {
+   "tokens": 15925,
+   "prompt_ms": 10706.3,
+   "pp_tok_s": 1487.4,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27474,
+   "prompt_ms": 15256.7,
+   "pp_tok_s": 1800.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 59725,
+   "prompt_ms": 32515.5,
+   "pp_tok_s": 1836.8,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4006.4,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 276.2,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 66.6,
+   "reps": [
+    66.6
+   ],
+   "hit_rate": 0.7101864314556545,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the fabric stiff with salt and ice. Outside the thick glass of the lantern room, the Atlantic was a churning cauldron of black and white, the waves crashing against the granite base with a violence that shook the very foundations of the tower. It was November, 1924, and the storm had been raging for three days.\n\nElias was not a man who feared the dark. He was a man who feared the silence.\n\nHe checked the clock on the wall. 11:42 PM. The rotation of the great Fresnel lens was steady, a rhythmic *whirrr-click, whirrr-click* that seemed to sync with his own heartbeat. He had been doing this for forty years. His hands were mapped with scars from rope burns and broken glass, his eyes permanently squinted against the glare of the bulb."
+  },
+  {
+   "name": "code",
+   "tok_s": 74.0,
+   "reps": [
+    74.0
+   ],
+   "hit_rate": 0.5873783774988882,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [],
+ "engine_lines": [
+  "strata generate: expert cache auto: 8.99 GiB free, 700 MiB reserved -> 3349 slots",
+  "strata generate: expert cache 2434 slots, 4.67 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 1944 cache slots (3.72 GiB)",
+  "strata serve: decode expert cache hit rate: 41.8% (1746 hits / 4176 lookups)",
+  "strata serve: decode expert cache hit rate: 50.4% (1512 hits / 3000 lookups)",
+  "strata serve: decode expert cache hit rate: 54.9% (1689 hits / 3077 lookups)",
+  "strata serve: decode expert cache hit rate: 60.3% (1938 hits / 3213 lookups)",
+  "strata serve: decode expert cache hit rate: 58.3% (11146 hits / 19121 lookups)",
+  "strata serve: decode expert cache hit rate: 66.1% (2201 hits / 3331 lookups)",
+  "strata serve: decode expert cache hit rate: 71.0% (74397 hits / 104757 lookups)",
+  "strata serve: decode expert cache hit rate: 58.7% (54151 hits / 92191 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split-limits/data/pp-lim-kv-pp-x2.json

@@ -0,0 +1,87 @@
+{
+ "label": "lim-kv-pp-x2",
+ "extra": [
+  "--max-context",
+  "131072",
+  "--kv-resident",
+  "32768",
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 15925,
+   "prompt_ms": 9452.6,
+   "pp_tok_s": 1684.7,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27474,
+   "prompt_ms": 14042.4,
+   "pp_tok_s": 1956.5,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 59725,
+   "prompt_ms": 30345.9,
+   "pp_tok_s": 1968.1,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3806.1,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 245.6,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 75.8,
+   "reps": [
+    75.8
+   ],
+   "hit_rate": 0.9698055042592483,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and shadow.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined Elias\u2019s life for thirty years. He stood before the great Fresnel lens, a cathedral of glass and brass that caught the light of the single, powerful bulb and fractured it into a thousand beams.\n\nHe checked the clock. 11:42 PM. The rotation was smooth, silent, a hypnotic sweep that painted the darkness with a slow, rhythmic pulse. *Flash. Dark. Flash. Dark.*\n\nE"
+  },
+  {
+   "name": "code",
+   "tok_s": 97.0,
+   "reps": [
+    97.0
+   ],
+   "hit_rate": 0.9562207324057541,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 17.96 GiB free",
+  "strata generate: layer split auto: CUDA0 84 SMs at 2.81 GHz -> 0.31 ms per layer, 8.13 GiB for experts",
+  "strata generate: layer split auto: CUDA1 82 SMs at 1.86 GHz -> 0.48 ms per layer, 13.98 GiB for experts",
+  "strata generate: layer split auto: K=22 - predicted 19.7 ms per decode window; the caches hold 11199 of 12288 profiled pairs (~99.7% of the routed mass)",
+  "strata generate: expert cache auto: 10.30 GiB free, 700 MiB reserved -> 3277 slots",
+  "strata generate: expert cache 3409 slots, 6.09 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 22-47, expert cache 6656 slots (13.34 GiB), 6656 of its 6656 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-21",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-21 (CUDA0), 22-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 91.3% (5098 hits / 5585 lookups)",
+  "strata serve: decode expert cache hit rate: 94.7% (4000 hits / 4225 lookups)",
+  "strata serve: decode expert cache hit rate: 96.5% (4095 hits / 4245 lookups)",
+  "strata serve: decode expert cache hit rate: 97.8% (4183 hits / 4278 lookups)",
+  "strata serve: decode expert cache hit rate: 96.2% (27177 hits / 28250 lookups)",
+  "strata serve: decode expert cache hit rate: 97.6% (4177 hits / 4281 lookups)",
+  "strata serve: decode expert cache hit rate: 97.0% (130241 hits / 134296 lookups)",
+  "strata serve: decode expert cache hit rate: 95.6% (113534 hits / 118732 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split-limits/data/vision.json

@@ -0,0 +1,24 @@
+{
+ "answers": {
+  "single": [
+   " Red",
+   "42",
+   "A red circle, the number 42, and a blue square are arranged on a white background."
+  ],
+  "same24": [
+   " Red",
+   "42",
+   "A red circle, the number 42, and a blue square are arranged on a white background."
+  ],
+  "x2": [
+   " Red",
+   "42",
+   "A red circle, the number 42, and a blue square are displayed on a white background."
+  ]
+ },
+ "correct": {
+  "single": true,
+  "same24": true,
+  "x2": true
+ }
+}

bench/results/2026-09-29-layer-split/README.md

@@ -0,0 +1,84 @@
+# Layer split across GPUs (2026-09-29)
+
+**Rig (a community member's PC, "RIBPC"):**
+- Ryzen 9 9950X3D, 62 GB DDR5, Windows 11, CUDA 13.1.
+- RTX 5080 16 GB (x16), RTX 3090 24 GB (x4 gen4, ~6 GB/s), RTX 2080 Ti 22 GB (x4 gen3, Turing sm_75; the
+  experimental `STRATA_EXPERIMENTAL_SM75` build).
+
+**Model:** the Coder (GSQ-RCO IQ1_M, 48 layers x 256 experts = 12,288 expert pairs, ~23 GB of experts), 32K
+context, int8 KV, `--prefill auto`, `--spec 4`. It is the only model that fits the rig's RAM and disk.
+
+**Tools:** `rib_pp.py` (needle prompts at 2.7K/8K/16K/28K tokens with a code word at 50% depth, a 2-turn chat that
+must reuse its checkpoint, greedy decode of a story and a code prompt) and `rib_norepeat.py` (the Phase A
+no-repeat and sanity cases). Raw results are in `data/`.
+
+## Correctness
+
+| Check | Result |
+|---|---|
+| This branch vs the 0.1.20 release, one GPU (5080), fixed cache, all 10 cases | **10/10 byte-identical** |
+| This branch, one GPU vs the split on the same GPU (K=24, `--split-device 0`) | **10/10 byte-identical**: the hand-off is exact |
+| 5080 + 3090 (auto) and 5080 + 2080 Ti + 3090 (auto): no-repeat under presence 100 (greedy and sampled), 144 sanity | all pass, 0 repeats |
+| Needles at 2.7K-28K tokens, every configuration | all found |
+| 2-turn chat (the checkpoint restored on every GPU) | reused 6,631-6,632 of 6,660 tokens, answer right |
+
+Across GPUs the text is not byte-identical to one GPU: experts the CPU computed on one card run on a GPU in the split,
+and they round differently (the same effect as `--expert-cache`, bench/results/2026-09-27-cache-parity).
+
+## Speed
+
+**Prompt tok/s** at 16K / 28K tokens, and **decode tok/s** for story / code (greedy; three repeats averaged
+in the K sweep):
+
+| Configuration | Split | Prompt 16K | Prompt 28K | Decode story | Decode code | Decode hits |
+|---|---|---|---|---|---|---|
+| 0.1.20, 5080 alone | - | 2,045 | 2,005 | 83.2 | 88.2 | 70-84% |
+| this branch, 5080 alone | - | 2,017 | 1,974 | 83.1 | 88.0 | 70-84% |
+| this branch, 5080 alone, 3 repeats (adaptive swaps learn the prompt) | - | 1,726 | 1,970 | 86.7 | 105.2 | 91-92% |
+| 3090 alone | - | 1,253 | 1,211 | 71.7 | 92.9 | 86-93% |
+| 5080 + 3090 | K=18 | 1,547 | 1,694 | 77.2 | 103.8 | 99.9% |
+| 5080 + 3090 | K=22 | 1,840 | 2,074 | 80.2 | 106.4 | 99.8% |
+| 5080 + 3090 | K=24 | 1,978 | 2,258 | 83.5 | 109.5 | 99.5% |
+| **5080 + 3090** | **K=26** | **2,039** | **2,357** | **83.8** | **109.7** | 99.3% |
+| 5080 + 3090 | K=28 | 2,011 | 2,287 | 83.5 | 112.5 | 99.0% |
+| 5080 + 3090 | K=30 | 1,935 | 2,218 | 81.1 | 109.6 | 98% |
+| 5080 + 3090 | K=32 | 1,818 | 2,063 | 83.0 | 108.7 | 98% |
+| 5080 + 3090 | auto (K=22) | 2,037 | 2,073 | 79.5 | 108.9 | 99.7% |
+| 3090 + 5080 | auto (K=30) | 1,672 | 1,876 | 90.3 | 105.1 | 99.4% |
+| 5080 + 2080 Ti + 3090 | auto (22, 23) | 1,687 | 1,847 | 67.6 | 89.6 | 99.7% |
+| 5080 + 3090 + 2080 Ti | first auto (12, 33) | 1,148 | 1,144 | 51.9 | 69.7 | 100% |
+| 5080 + 2080 Ti | first auto | 914 | 906 | 45.5 | 58.0 | 97% |
+
+What decides it:
+
+- **The best two-GPU split for the Coder is K=26-28.** Prompts run +18-20% over the 5080 alone at 16-28K;
+  decode is on par for story and +4-7% for code.
+- **Residency saturates quickly: the per-layer GPU speed decides.** From K=18 up, both caches hold 98-100% of
+  the routed experts. Moving layers to the faster card is what helps: the 3090 costs 0.50 ms per layer per
+  window, the 5080 0.33 (`STRATA_SPLIT_TIMING`, the "wait for the GPU" per stage).
+- **Decode is bounded by that GPU time.** One 5080 with its cache adapted spends 16.3 ms waiting for its GPU and
+  5.5 ms in the CPU pool per window. The split removes the pool (0.8-1.0 ms) but adds the slower card's layers:
+  7.9 + 12.1 ms at K=24. The per-layer GPU half (kernels plus the host round trip that publishes each layer's
+  expert plan) is the next thing to cut, for one GPU as well (perf-review-plan E-6).
+- **A third, slower card costs more than it holds.** The 2080 Ti's one-layer stage took 3.1 ms per window.
+  For a model that already fits two cards, leave it out. It pays only when the two faster cards cannot hold most
+  of the routed experts (a bigger model; not testable on this rig's 62 GB of RAM).
+- **Prompt chunk:** 1024 instead of 2048 frees ~0.7 GB per card, but prompts read 7-12% slower.
+
+## Auto placement
+
+`--layer-split auto` predicts one decode window's time for every placement (every placement up to three GPUs) and
+keeps the fastest:
+
+- **per-layer time** from SMs x clock (0.33 ms for an RTX 5080);
+- **cache contents:** each card's layers' profiled pairs, hottest first, as the fill does;
+- **routed mass** of rank r taken as (r+1)^-1.2 (fits the sweep's hit rates within 0.1-1%);
+- **missed mass** costing 190 ms per unit (fitted to the sweep's best K; `STRATA_SPLIT_MISS_MS` overrides).
+
+It picked K=22 on the 5080 + 3090: within 0-12% of the best fixed split on prompts and on par on decode. It gives
+a slow third card a single layer.
+
+## Reproduce
+
+On a rig: the `strata-remote` tools, then `rib_matrix.bat`, `rib_ksweep.bat`, `rib_validate.bat` (in the dev
+notes, not in this repository).

bench/results/2026-09-29-layer-split/data/norepeat-nr-v-x2-auto-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 193,
+  "tok_s": 74.6,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, rough as sandpaper, gripped the iron railing as he ascended the final flight of steps to the lantern room.\n\nBelow him, the Atlantic churned in a frenzy, whitecaps exploding against the black rocks like fireworks. Above, the sky was a bruised purple, heavy with rain that hadn\u2019t yet fallen.\n\nElias checked the clock. 11:42 PM. The light was due.\n\nHe didn\u2019t flip a switch. He didn\u2019t press a button. He reached for the heavy brass lever that had worn smooth by his palm over the years. With a grunt that matched the strain of his aging muscles, he pulled it down.\n\n*Clunk.*\n\nThe"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 136,
+  "drafts_offered": 181,
+  "tok_s": 105.5,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 85.8,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 94.3,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 118,
+  "repeats": 0,
+  "drafts_accepted": 37,
+  "drafts_offered": 105,
+  "tok_s": 67.9,
+  "text": "Aardvark  \nAlbatross   \nAlligator    \nAntelope     \nBaboon       \nBadger      \nBeaver        \nBear          \nBird of Paradise \nBuffalo         \nCamel           \nCheetah            \nChimpanzee             \nCrane              \nCrocodile                \nDeer                 \nDolphin               \nDonkey                  \nElephant                   \nFlamingo                    \nFox                       \nGiraffe                     \nGoat                      \nHippopotamus  \nIguana       \nJaguar        \nKangaroo      \nKoala          \nLion         \nLeopard     \nMonkey    \nMoose   \nOstrich \nOtter\nPanda<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 147,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 103,
+  "tok_s": 63.2,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to that date and time for context because you asked me about a lighthouse keeper but gave no specifics beyond \u201cshort story,\u201d so let\u2019s set it: isolated coastal station off the Oregon coast during winter storm season, 1987.\n\nMy name is Elias Vane (I made this fictional; real keepers often kept journals signed with initials only), and I\u2019m writing from memory because my logbook got waterlogged when a rogue wave breached lower storage room at dawn\u2014so here\u2019s what happened:<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 109,
+  "repeats": 0,
+  "drafts_accepted": 13,
+  "drafts_offered": 56,
+  "tok_s": 54.1,
+  "text": "The wind didn\u2019t just blow at Cape Hollow; it screamed through the gaps in my cottage\u2019s shutters, rattling them like loose teeth until they settled into silence with every storm break out of three hundred days here being storms that broke something else again but mostly kept me warm because I burned coal and wood whenever we had any left over from last winter when even then if you\u2019re lucky enough to get through it all without freezing yourself so hard your own thoughts become ice too\n\nBut Elias knew better than anyone: fear was useful only until dawn<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 92,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 48,
+  "tok_s": 64.7,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKoala      \nKangaroo        \nHippopotamus          \nRhinoceros            \nCamel              \nCheetah             \nLeopard           \nPanther         \nJaguar                \nBison               \nElk                 \nDeer                  \nFox                   \nWolf                    \nBear                     \nLlama                       \nDonkey                        \nGoat                      \nSheep                         \nCow                           \nPig                          \nRabbit                            \nCat                              \nDog<|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 14,
+  "tok_s": 71.7,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 53,
+  "drafts_offered": 134,
+  "tok_s": 66.1,
+  "text": "The ocean is not merely water; it has been, and continues to be one with humanity\u2019s most profound narrative arc in human civilization: **the story we tell ourselves through our relationship** between landmasses separated by vast expansiveness that defines both isolation from other worlds (and thus their own unique evolutionary paths) as well than connection across them via currents wind patterns tides depth pressure temperature salinity biodiversity geology climate change economic trade routes ecological systems cultural exchange political boundaries scientific discovery technological advancement environmental degradation conservation efforts maritime archaeography historical cartographic evolution legal frameworks international law enforcement mechanisms insurance liability regimes regulatory compliance standards safety protocols navigation aids communication networks emergency response strategies search-and-rescue operations salvage rights ownership disputes heritage protection zones sustainable harvesting practices aquaculture development infrastructure engineering projects coastal management policies urban planning regulations zoning ordinances building codes architectural design guidelines aesthetic considerations functional requirements accessibility provisions energy efficiency measures thermal insulation techniques moisture control methods ventilation optimization lighting solutions acoustic soundproofing structural integrity assessments seismic resistance evaluations flood mitigation plans storm surge barriers levee construction seawall reinforcement breakwater installation harbor dredging channel maintenance buoy placement lighthouse automation radar monitoring sonar mapping bathymetric surveys hydrographic charts nautical almanacs tide tables current predictions weather forecasts marine meteorological models satellite imagery remote sensing data assimilation numerical simulations computational fluid dynamics finite element analysis partial differential equations tensor calculus linear algebra probability theory statistics stochastic processes Markov chains Bayesian inference machine learning artificial neural network deep convolutional recurrent transformer attention heads generative adversarial diffusion transformers large language model tokenization"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-nr-v-x3-auto-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 193,
+  "tok_s": 68.2,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, rough as sandpaper, gripped the iron railing as he ascended the final flight of steps to the lantern room.\n\nBelow him, the Atlantic churned in a frenzy, whitecaps exploding against the black rocks like fireworks. Above, the sky was a bruised purple, heavy with rain that hadn\u2019t yet fallen.\n\nElias checked the clock. 11:42 PM. The light was due.\n\nHe didn\u2019t flip a switch. He didn\u2019t press a button. He reached for the heavy brass lever that had worn smooth by his palm over the years. With a grunt that matched the strain of his aging muscles, he pulled it down.\n\n*Clunk.*\n\nThe"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 136,
+  "drafts_offered": 181,
+  "tok_s": 99.4,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.5,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 76.3,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 118,
+  "repeats": 0,
+  "drafts_accepted": 37,
+  "drafts_offered": 105,
+  "tok_s": 65.2,
+  "text": "Aardvark  \nAlbatross   \nAlligator    \nAntelope     \nBaboon       \nBadger      \nBeaver        \nBear          \nBird of Paradise \nBuffalo         \nCamel           \nCheetah            \nChimpanzee             \nCrane              \nCrocodile                \nDeer                 \nDolphin               \nDonkey                  \nElephant                   \nFlamingo                    \nFox                       \nGiraffe                     \nGoat                      \nHippopotamus  \nIguana       \nJaguar        \nKangaroo      \nKoala          \nLion         \nLeopard     \nMonkey    \nMoose   \nOstrich \nOtter\nPanda<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 147,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 103,
+  "tok_s": 45.7,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to that date and time for context because you asked me about a lighthouse keeper but gave no specifics beyond \u201cshort story,\u201d so let\u2019s set it: isolated coastal station off the Oregon coast during winter storm season, 1987.\n\nMy name is Elias Vane (I made this fictional; real keepers often kept journals signed with initials only), and I\u2019m writing from memory because my logbook got waterlogged when a rogue wave breached lower storage room at dawn\u2014so here\u2019s what happened:<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 109,
+  "repeats": 0,
+  "drafts_accepted": 13,
+  "drafts_offered": 56,
+  "tok_s": 42.1,
+  "text": "The wind didn\u2019t just blow at Cape Hollow; it screamed through the gaps in my cottage\u2019s shutters, rattling them like loose teeth until they settled into silence with every storm break out of three hundred days here being storms that broke something else again but mostly kept me warm because I burned coal and wood whenever we had any left over from last winter when even then if you\u2019re lucky enough to get through it all without freezing yourself so hard your own thoughts become ice too\n\nBut Elias knew better than anyone: fear was useful only until dawn<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 92,
+  "repeats": 0,
+  "drafts_accepted": 26,
+  "drafts_offered": 48,
+  "tok_s": 43.9,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKoala      \nKangaroo        \nHippopotamus          \nRhinoceros            \nCamel              \nCheetah             \nLeopard           \nPanther         \nJaguar                \nBison               \nElk                 \nDeer                  \nFox                   \nWolf                    \nBear                     \nLlama                       \nDonkey                        \nGoat                      \nSheep                         \nCow                           \nPig                          \nRabbit                            \nCat                              \nDog<|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 14,
+  "tok_s": 61.4,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 53,
+  "drafts_offered": 134,
+  "tok_s": 51.9,
+  "text": "The ocean is not merely water; it has been, and continues to be one with humanity\u2019s most profound narrative arc in human civilization: **the story we tell ourselves through our relationship** between landmasses separated by vast expansiveness that defines both isolation from other worlds (and thus their own unique evolutionary paths) as well than connection across them via currents wind patterns tides depth pressure temperature salinity biodiversity geology climate change economic trade routes ecological systems cultural exchange political boundaries scientific discovery technological advancement environmental degradation conservation efforts maritime archaeography historical cartographic evolution legal frameworks international law enforcement mechanisms insurance liability regimes regulatory compliance standards safety protocols navigation aids communication networks emergency response strategies search-and-rescue operations salvage rights ownership disputes heritage protection zones sustainable harvesting practices aquaculture development infrastructure engineering projects coastal management policies urban planning regulations zoning ordinances building codes architectural design guidelines aesthetic considerations functional requirements accessibility provisions energy efficiency measures thermal insulation techniques moisture control methods ventilation optimization lighting solutions acoustic soundproofing structural integrity assessments seismic resistance evaluations flood mitigation plans storm surge barriers levee construction seawall reinforcement breakwater installation harbor dredging channel maintenance buoy placement lighthouse automation radar monitoring sonar mapping bathymetric surveys hydrographic charts nautical almanacs tide tables current predictions weather forecasts marine meteorological models satellite imagery remote sensing data assimilation numerical simulations computational fluid dynamics finite element analysis partial differential equations tensor calculus linear algebra probability theory statistics stochastic processes Markov chains Bayesian inference machine learning artificial neural network deep convolutional recurrent transformer attention heads generative adversarial diffusion transformers large language model tokenization"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-nr-x2-auto-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 91,
+  "drafts_offered": 190,
+  "tok_s": 74.0,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, rough as sandpaper, knew every bolt, every lever, and every creak in the iron bones of the tower.\n\nFor forty years, Elias had been the guardian of the Point. Not just of the light, but of the silence between the waves.\n\nIt was November. The sky was a bruised purple, heavy with rain that hadn\u2019t yet fallen. The sea was a churning cauldron of grey and white, throwing itself against the granite base with a violence that shook the floorboards beneath Elias\u2019s boots.\n\nHe climbed the stairs. One hundred and twelve steps. He counted them by feel, his left hand trailing along the cold iron railing. At the top, the air was"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 137,
+  "drafts_offered": 178,
+  "tok_s": 103.4,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 91.1,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 70.7,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 118,
+  "repeats": 0,
+  "drafts_accepted": 38,
+  "drafts_offered": 108,
+  "tok_s": 65.4,
+  "text": "Aardvark  \nAlbatross   \nAlligator    \nAntelope     \nBaboon       \nBadger      \nBeaver        \nBear          \nBird of Paradise \nBuffalo         \nCamel           \nCheetah            \nChimpanzee             \nCrane              \nCrocodile                \nDeer                 \nDolphin               \nDonkey                  \nElephant                   \nFlamingo                    \nFox                       \nGiraffe                     \nGoat                      \nHippopotamus  \nIguana       \nJaguar        \nKangaroo      \nKoala          \nLion         \nLeopard     \nMonkey    \nMoose   \nOstrich \nOtter\nPanda<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 57,
+  "repeats": 0,
+  "drafts_accepted": 11,
+  "drafts_offered": 61,
+  "tok_s": 57.1,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral vaults below Elias\u2019s boots were heavy with salt and exhaustion as he climbed one hundred twelve steps to reach top room where his world began again.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 58,
+  "repeats": 0,
+  "drafts_accepted": 14,
+  "drafts_offered": 54,
+  "tok_s": 52.9,
+  "text": "The wind didn\u2019t just blow at Cape Harrow; it screamed through the gaps in stone and steel, carrying with them spray from three miles out into churning black water that no longer resembled an ocean so much as abyssal chaos waiting for its prey to surface beneath waves of salt.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 59,
+  "repeats": 0,
+  "drafts_accepted": 11,
+  "drafts_offered": 47,
+  "tok_s": 51.7,
+  "text": "Lion  \nTiger  \n\nElephant\\nGiraffe\\Zebra\\rHippopotamus\\tRhinoceros\\vCrocodile\\fAlligator\\bSnake\\aSpider\\u2615CoffeeCup\u2615\ufe0fTeaTime\ud83d\udc09DragonFishWhaleDolph<|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 12,
+  "tok_s": 73.5,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 43,
+  "drafts_offered": 126,
+  "tok_s": 63.4,
+  "text": "The ocean is not merely water; it has been, and continues to be one with humanity\u2019s most profound narrative arc in human civilization: **the story told by waves**. It serves as both witness (through sedimentary layers)and participant( through currents that carry messages across continents). Its \"history\" spans geological epochs where life first emerged from primordial pools into complex ecosystems today\u2014and cultural histories shaped entirely around its roleas connector between distant lands or barrier separating them forever depending on perspective taken at different times throughout millennia since ancient civilizations began building boats outof logs carved using stone tools until modern container ships carrying billions tons cargo daily cross Atlantic routes once navigated onlyby stars alone under moonlight during nights when sailors feared monsters lurking beneath hulls made fragile against storms they could barely predict even now despite advanced meteorology systems developed over past century after industrial revolution transformed maritime trade patterns globally due economic pressures driving demand for faster logistics chains linking factories producing electronics manufactured inside cities located far inland near rivers feeding estuaries opening directly onto open seas connected via canals engineered centuries ago like Suez Canal built originallyfor military purposes later adapted commercial use post-World War II era reshaping global supply networks dramatically because geopolitical shifts altered which ports remained accessible politically versus physically blocked temporarily then reopened permanently following treaty agreements signed among nations agreeing upon shared responsibilities regarding environmental stewardship including coral reef restoration efforts funded jointly international organizations coordinating scientific research conducted aboard vessels specifically designed deep-sea exploration equipped robotic submersibles capable descending trenches deeper than any previous technology allowed"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-nr-x3-auto-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 98,
+  "drafts_offered": 166,
+  "tok_s": 54.4,
+  "text": "The wind didn\u2019t just blow; it screamed. It tore through the iron gallery of the Blackwood Point Lighthouse, rattling the glass panes like a prisoner shaking its bars.\n\nElias Thorne was seventy years old, and he had been the keeper for forty of them. He knew every creak in the spiral staircase, every draft that slipped under the door, and every shift in the weather that the barometer couldn\u2019t quite predict. He knew that when the fog rolled in thick enough to swallow the horizon, the world ceased to exist. There was only the light, the stone, and him.\n\nIt was November, the month of storms. The sea was a churning cauldron of grey and white, throwing itself against the granite base with a violence that made the floorboards tremble. Elias stood by the lantern room, his hands wrapped around a mug of coffee that had gone cold hours ago. He wasn\u2019t drinking it. He was holding it for warmth."
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 137,
+  "drafts_offered": 163,
+  "tok_s": 70.7,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv(file_path: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys being the header values\n    and values being the corresponding cell values.\n    \n    Args:\n        file_path: Path to the CSV file.\n        \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(file_path, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_basic():\n    \"\"\"Test basic CSV parsing with simple data.\"\"\"\n    import tempfile\n    import os\n    \n    # Create a temporary CSV"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 56.7,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 57.6,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 59,
+  "repeats": 0,
+  "drafts_accepted": 13,
+  "drafts_offered": 47,
+  "tok_s": 39.7,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBuffalo     \n\nCat          \nCheetah      \n\nChimpanzee \nClam         \n\n\nCrocodile   \n\n\n\nDolphin      \n\n\n\n\nDonkey       <|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 57,
+  "repeats": 0,
+  "drafts_accepted": 12,
+  "drafts_offered": 58,
+  "tok_s": 39.9,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral vaults below Elias\u2019s boots were heavy with salt and exhaustion as he climbed one hundred twelve steps to reach topmost gallery where his world began.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 30,
+  "drafts_offered": 111,
+  "tok_s": 34.7,
+  "text": "The wind didn\u2019t just blow at Blackwood Rock; it screamed through the gaps in his stone cottage like an angry ghost trying to get inside, yet finding no door left unlocked because Elias never let one stay that way for more than three seconds after entering or leaving during storm season. He had been doing this since 1987\u2014since before he even knew why\u2014and now it was muscle memory so deep into his bones and fingers\u2019 calloused skin texture patterns worn thin by years of gripping the same iron railing at midnight watch that if you asked him what time exactly, down to quarter-hour increments when fog rolled in thick enough where light beams bent around itself like old rope coiling tight on a winch drum spooling out slowly from below deck floor level storage locker number seven under stairs leading"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 96,
+  "repeats": 0,
+  "drafts_accepted": 31,
+  "drafts_offered": 48,
+  "tok_s": 42.3,
+  "text": "Lion  \nTiger   \nElephant \nGiraffe    \nZebra     \nPanda       \nKoala      \nKangaroo        \nCheetah         \nHippopotamus          \nRhinoceros            \nLeopard             \nJaguar              \nWolf                \nFox                 \nBear                  \nDeer                    \nOtter                   \nBeaver                      \nBadger                        \nMongoose                          \nMeerkat                           \nChimpanzee                            \nGorilla                             \nOrangetutan                              \nAntelope                                \nBison                                   \nElk                                     <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 43.0,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 29,
+  "drafts_offered": 127,
+  "tok_s": 37.6,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\u201d \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \u201cthe Sea\u201d or \"Ocean\" (terms often used interchangeably but distinct geographically) was never just an element within our planet\u2019s geography\u2014it became its engine itself\u2014a living system whose existence predates humanity by billions years yet continues shaping us today through climate regulation oxygen production food supply trade routes cultural identity ecological resilience economic development political boundaries scientific discovery technological innovation environmental crisis conservation efforts historical memory philosophical reflection poetic metaphor musical composition cinematic storytelling literary narrative visual art architectural inspiration culinary tradition spiritual symbolism psychological influence sociological transformation educational foundation medical necessity agricultural dependency industrial revolution urban planning infrastructural engineering legal framework regulatory policy international cooperation global governance planetary stewardship existential risk mitigation sustainable practice ethical consideration moral imperative social justice democratic participation civic engagement public health community welfare individual responsibility personal accountability collective action shared experience common understanding mutual respect reciprocal trust collaborative effort cooperative endeavor joint venture unified mission integrated strategy coordinated plan synchronized schedule aligned timeline concurrent execution simultaneous operation parallel processing sequential progression iterative refinement recursive improvement continuous adaptation dynamic adjustment flexible response adaptive behavior responsive conduct proactive initiative anticipatory foresight predictive modeling probabilistic assessment statistical analysis empirical observation experimental verification theoretical hypothesis conceptual abstraction logical deduction rational inference critical evaluation analytical reasoning systematic investigation methodical inquiry structured exploration comprehensive examination detailed scrutiny meticulous inspection thorough review exhaustive audit complete survey full census total enumeration absolute"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-reg-dev-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 187,
+  "tok_s": 59.2,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, mapped with veins like the roots of an ancient oak, gripped the cold iron railing.\n\nFor forty years, Elias had tended the light at the end of the world. Not the end of the map, but the end of the land. Here, the Atlantic threw itself against the granite cliffs with such violence that the spray reached the gallery, thirty feet above the rocks.\n\nHe checked the clock. 11:42 PM. The fog was thickening, a gray soup that swallowed the horizon. In the old days, he would have climbed the stairs to the lantern room, polished the Fresnel lens, and wound the clockwork mechanism that turned the great beam. But the automation had changed"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 135,
+  "drafts_offered": 179,
+  "tok_s": 81.6,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.9,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.3,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 61,
+  "repeats": 0,
+  "drafts_accepted": 19,
+  "drafts_offered": 49,
+  "tok_s": 49.6,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBoar          \nBuffalo \nButterfly \n\nCamel           \nCat            \nCheetah        \n\nDolphin         \n\n\nElephant        \n\n\n\nFlamingo<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 80,
+  "repeats": 0,
+  "drafts_accepted": 6,
+  "drafts_offered": 52,
+  "tok_s": 44.9,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to check what was happening with that storm and why they were calling me out there when all we had left for tonight\u2019s shift is one last pass through our logs before dawn breaks.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 17,
+  "drafts_offered": 83,
+  "tok_s": 44.1,
+  "text": "The wind didn\u2019t just blow at Cape Harrow; it screamed through the gaps in its teeth, hungry and relentless for three centuries of stone salt-spray history that had long since forgotten to be polite or gentle with itself against anything smaller than God Himself was willing acknowledge as divine protection from drowning souls who were already lost before they ever touched water.\n\nElias knew this rhythm better he himself could hear it sometimes; his boots worn smooth by pacing the spiral stair\u2019s center-post groove where thousands upon countless steps had rubbed away texture until all that remained beneath sole leather now felt like polished bone carved out slowly but surely over time into something more hollow than wood ever really should be allowed become when used so often enough within such confined circular spaces meant only for vertical ascent down always descending again back upward never truly leaving"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 62,
+  "repeats": 0,
+  "drafts_accepted": 18,
+  "drafts_offered": 31,
+  "tok_s": 49.5,
+  "text": "Lion  \nTiger  \n\nElephant  \n  \nGiraffe  \n\n\n\nHippopotamus    \nZebra   \nRhinoceros   \n\n\nPanda Bear    \n\n\n\n\nKangaroo     \nKoala      \nSloth      \n\n\n\n\n\nCheetah       \nLeopard        \nPanther        \n\n\n\n\n\n\nWolf              \nFox                \nBear                 <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 65.2,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 47,
+  "drafts_offered": 153,
+  "tok_s": 52.5,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\" \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \"the Sea\" or \u201cOcean\u201d was never just an element within our planet\u2019s geography but rather its most fundamental biological engine\u2014a living system whose existence predates humanity by billions years yet continues shaping every aspect from climate regulation through evolutionary biology all way down modern industrial supply chains today's global shipping lanes carry cargo across vast distances because they were once crossed only slowly via coastal hopping routes used since antiquity until steamships revolutionized travel during late nineteenth century when diesel engines replaced wind power entirely allowing container ships carrying thousands TEUs (Twenty-foot Equivalent Units) per voyage crossing entire Pacific basins without stopping at intermediate ports anymore unlike earlier sailing vessels needing favorable winds currents seasonal patterns dictating departure windows based upon monsoon calendars established centuries ago before technological advances changed those constraints fundamentally altering logistics networks worldwide thereby transforming economic structures globally while simultaneously creating ecological disruptions due invasive species transported unintentionally alongside legitimate goods causing cascading ecosystem collapses locally where non-native organisms outcompete native fauna disrupting food webs leading population declines affecting biodiversity indices measured scientifically using standardized metrics like Shannon-Wiener diversity index calculations applied consistently over decades revealing trends showing significant decreases post-introduction events correl strongly with introduction dates matching historical records documenting first sightings confirmed taxonomically classified formally according International Code Nomenclature rules"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-reg-rel-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 187,
+  "tok_s": 58.4,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, mapped with veins like the roots of an ancient oak, gripped the cold iron railing.\n\nFor forty years, Elias had tended the light at the end of the world. Not the end of the map, but the end of the land. Here, the Atlantic threw itself against the granite cliffs with such violence that the spray reached the gallery, thirty feet above the rocks.\n\nHe checked the clock. 11:42 PM. The fog was thickening, a gray soup that swallowed the horizon. In the old days, he would have climbed the stairs to the lantern room, polished the Fresnel lens, and wound the clockwork mechanism that turned the great beam. But the automation had changed"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 135,
+  "drafts_offered": 179,
+  "tok_s": 80.6,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 75.2,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.2,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 61,
+  "repeats": 0,
+  "drafts_accepted": 19,
+  "drafts_offered": 49,
+  "tok_s": 48.5,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBoar          \nBuffalo \nButterfly \n\nCamel           \nCat            \nCheetah        \n\nDolphin         \n\n\nElephant        \n\n\n\nFlamingo<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 80,
+  "repeats": 0,
+  "drafts_accepted": 6,
+  "drafts_offered": 52,
+  "tok_s": 46.4,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to check what was happening with that storm and why they were calling me out there when all we had left for tonight\u2019s shift is one last pass through our logs before dawn breaks.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 17,
+  "drafts_offered": 83,
+  "tok_s": 43.3,
+  "text": "The wind didn\u2019t just blow at Cape Harrow; it screamed through the gaps in its teeth, hungry and relentless for three centuries of stone salt-spray history that had long since forgotten to be polite or gentle with itself against anything smaller than God Himself was willing acknowledge as divine protection from drowning souls who were already lost before they ever touched water.\n\nElias knew this rhythm better he himself could hear it sometimes; his boots worn smooth by pacing the spiral stair\u2019s center-post groove where thousands upon countless steps had rubbed away texture until all that remained beneath sole leather now felt like polished bone carved out slowly but surely over time into something more hollow than wood ever really should be allowed become when used so often enough within such confined circular spaces meant only for vertical ascent down always descending again back upward never truly leaving"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 62,
+  "repeats": 0,
+  "drafts_accepted": 18,
+  "drafts_offered": 31,
+  "tok_s": 49.8,
+  "text": "Lion  \nTiger  \n\nElephant  \n  \nGiraffe  \n\n\n\nHippopotamus    \nZebra   \nRhinoceros   \n\n\nPanda Bear    \n\n\n\n\nKangaroo     \nKoala      \nSloth      \n\n\n\n\n\nCheetah       \nLeopard        \nPanther        \n\n\n\n\n\n\nWolf              \nFox                \nBear                 <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 62.5,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 47,
+  "drafts_offered": 153,
+  "tok_s": 52.6,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\" \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \"the Sea\" or \u201cOcean\u201d was never just an element within our planet\u2019s geography but rather its most fundamental biological engine\u2014a living system whose existence predates humanity by billions years yet continues shaping every aspect from climate regulation through evolutionary biology all way down modern industrial supply chains today's global shipping lanes carry cargo across vast distances because they were once crossed only slowly via coastal hopping routes used since antiquity until steamships revolutionized travel during late nineteenth century when diesel engines replaced wind power entirely allowing container ships carrying thousands TEUs (Twenty-foot Equivalent Units) per voyage crossing entire Pacific basins without stopping at intermediate ports anymore unlike earlier sailing vessels needing favorable winds currents seasonal patterns dictating departure windows based upon monsoon calendars established centuries ago before technological advances changed those constraints fundamentally altering logistics networks worldwide thereby transforming economic structures globally while simultaneously creating ecological disruptions due invasive species transported unintentionally alongside legitimate goods causing cascading ecosystem collapses locally where non-native organisms outcompete native fauna disrupting food webs leading population declines affecting biodiversity indices measured scientifically using standardized metrics like Shannon-Wiener diversity index calculations applied consistently over decades revealing trends showing significant decreases post-introduction events correl strongly with introduction dates matching historical records documenting first sightings confirmed taxonomically classified formally according International Code Nomenclature rules"
+ }
+]

bench/results/2026-09-29-layer-split/data/norepeat-reg-same24-coder-iq1_m.json

@@ -0,0 +1,92 @@
+[
+ {
+  "name": "greedy story 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 88,
+  "drafts_offered": 187,
+  "tok_s": 58.4,
+  "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, mapped with veins like the roots of an ancient oak, gripped the cold iron railing.\n\nFor forty years, Elias had tended the light at the end of the world. Not the end of the map, but the end of the land. Here, the Atlantic threw itself against the granite cliffs with such violence that the spray reached the gallery, thirty feet above the rocks.\n\nHe checked the clock. 11:42 PM. The fog was thickening, a gray soup that swallowed the horizon. In the old days, he would have climbed the stairs to the lantern room, polished the Fresnel lens, and wound the clockwork mechanism that turned the great beam. But the automation had changed"
+ },
+ {
+  "name": "greedy code 200",
+  "tokens": 200,
+  "repeats": 0,
+  "drafts_accepted": 135,
+  "drafts_offered": 179,
+  "tok_s": 80.5,
+  "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+ },
+ {
+  "name": "greedy sanity",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 74.5,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy, no penalty (sanity)",
+  "tokens": 4,
+  "repeats": 0,
+  "drafts_accepted": 3,
+  "drafts_offered": 3,
+  "tok_s": 76.8,
+  "text": "144<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64",
+  "tokens": 61,
+  "repeats": 0,
+  "drafts_accepted": 19,
+  "drafts_offered": 49,
+  "tok_s": 49.6,
+  "text": "Aardvark  \nAlbatross  \n\nAlligator   \nAntelope    \nBaboon      \nBadger     \nBeaver       \nBear        \nBird         \nBoar          \nBuffalo \nButterfly \n\nCamel           \nCat            \nCheetah        \n\nDolphin         \n\n\nElephant        \n\n\n\nFlamingo<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=64 story",
+  "tokens": 80,
+  "repeats": 0,
+  "drafts_accepted": 6,
+  "drafts_offered": 52,
+  "tok_s": 46.7,
+  "text": "The wind didn\u2019t just blow; it screamed against the glass of Lantern House 42, rattling like chains in an empty cathedral hall outside my window at midnight on November third this year\u2014so I pulled up to check what was happening with that storm and why they were calling me out there when all we had left for tonight\u2019s shift is one last pass through our logs before dawn breaks.<|im_end|>"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 3",
+  "tokens": 160,
+  "repeats": 0,
+  "drafts_accepted": 17,
+  "drafts_offered": 83,
+  "tok_s": 44.3,
+  "text": "The wind didn\u2019t just blow at Cape Harrow; it screamed through the gaps in its teeth, hungry and relentless for three centuries of stone salt-spray history that had long since forgotten to be polite or gentle with itself against anything smaller than God Himself was willing acknowledge as divine protection from drowning souls who were already lost before they ever touched water.\n\nElias knew this rhythm better he himself could hear it sometimes; his boots worn smooth by pacing the spiral stair\u2019s center-post groove where thousands upon countless steps had rubbed away texture until all that remained beneath sole leather now felt like polished bone carved out slowly but surely over time into something more hollow than wood ever really should be allowed become when used so often enough within such confined circular spaces meant only for vertical ascent down always descending again back upward never truly leaving"
+ },
+ {
+  "name": "sampled presence=100 last_n=64 seed 11",
+  "tokens": 62,
+  "repeats": 0,
+  "drafts_accepted": 18,
+  "drafts_offered": 31,
+  "tok_s": 50.0,
+  "text": "Lion  \nTiger  \n\nElephant  \n  \nGiraffe  \n\n\n\nHippopotamus    \nZebra   \nRhinoceros   \n\n\nPanda Bear    \n\n\n\n\nKangaroo     \nKoala      \nSloth      \n\n\n\n\n\nCheetah       \nLeopard        \nPanther        \n\n\n\n\n\n\nWolf              \nFox                \nBear                 <|im_end|>"
+ },
+ {
+  "name": "greedy frequency=100 last_n=16",
+  "tokens": 11,
+  "repeats": 0,
+  "drafts_accepted": 5,
+  "drafts_offered": 13,
+  "tok_s": 62.0,
+  "text": "1 to 50, separated by commas:<|im_end|>"
+ },
+ {
+  "name": "greedy presence=100 last_n=4096 (8-row windows)",
+  "tokens": 300,
+  "repeats": 0,
+  "drafts_accepted": 47,
+  "drafts_offered": 153,
+  "tok_s": 52.1,
+  "text": "The ocean is not merely water; it has been, and remains to this day one giant machine for making life possible on Earth.\" \u2014 *David Attenborough* \n\nThis quote captures perhaps more than any other single sentence in human language: that which we call \"the Sea\" or \u201cOcean\u201d was never just an element within our planet\u2019s geography but rather its most fundamental biological engine\u2014a living system whose existence predates humanity by billions years yet continues shaping every aspect from climate regulation through evolutionary biology all way down modern industrial supply chains today's global shipping lanes carry cargo across vast distances because they were once crossed only slowly via coastal hopping routes used since antiquity until steamships revolutionized travel during late nineteenth century when diesel engines replaced wind power entirely allowing container ships carrying thousands TEUs (Twenty-foot Equivalent Units) per voyage crossing entire Pacific basins without stopping at intermediate ports anymore unlike earlier sailing vessels needing favorable winds currents seasonal patterns dictating departure windows based upon monsoon calendars established centuries ago before technological advances changed those constraints fundamentally altering logistics networks worldwide thereby transforming economic structures globally while simultaneously creating ecological disruptions due invasive species transported unintentionally alongside legitimate goods causing cascading ecosystem collapses locally where non-native organisms outcompete native fauna disrupting food webs leading population declines affecting biodiversity indices measured scientifically using standardized metrics like Shannon-Wiener diversity index calculations applied consistently over decades revealing trends showing significant decreases post-introduction events correl strongly with introduction dates matching historical records documenting first sightings confirmed taxonomically classified formally according International Code Nomenclature rules"
+ }
+]

bench/results/2026-09-29-layer-split/data/pp-dev-3090.json

@@ -0,0 +1,72 @@
+{
+ "label": "dev-3090",
+ "extra": [],
+ "gpus": "2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 3424.8,
+   "pp_tok_s": 796.0,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 8617.9,
+   "pp_tok_s": 972.3,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 12820.1,
+   "pp_tok_s": 1253.2,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 22901.1,
+   "pp_tok_s": 1211.4,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 5657.8,
+  "turn2_tokens": 6660,
+  "reused": 6631,
+  "turn2_ms": 238.7,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 71.7,
+   "hit_rate": 0.9266798230257504,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning black void, punctuated by whitecaps that looked like teeth.\n\nHe climbed the spiral stairs, his knees protesting with every step. The iron rungs were cold, slick with condensation that had leaked through the hairline cracks in the masonry. At the top, the lantern room hummed with the low, steady thrum of the rotating lens. It was a sound Elias had learned to sleep with, a heartbeat of the house.\n\nHe checked the clock. 11:42 PM. The light swept across the water, a slow, hypnotic arc. *Flash. Darkness. Flash. Darkness.*\n\nFor a moment, the beam caught something that wasn\u2019t water. A shape"
+  },
+  {
+   "name": "code",
+   "tok_s": 92.9,
+   "hit_rate": 0.8636411400016346,
+   "text": "```python\nimport csv\nfrom typing import List, Dict\n\n\ndef parse_csv(filepath: str) -> List[Dict[str, str]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header row\n    and values from the corresponding row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(filepath, 'r', newline='') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n```<|im_end|>"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: expert cache auto: 17.31 GiB free, 700 MiB reserved -> 6704 slots",
+  "strata generate: expert cache 8730 slots, 16.62 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2315 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 86.6% (4986 hits / 5756 lookups)",
+  "strata serve: decode expert cache hit rate: 87.1% (3757 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 88.7% (3827 hits / 4314 lookups)",
+  "strata serve: decode expert cache hit rate: 92.1% (3975 hits / 4318 lookups)",
+  "strata serve: decode expert cache hit rate: 95.5% (4124 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 94.3% (25793 hits / 27360 lookups)",
+  "strata serve: decode expert cache hit rate: 94.6% (4088 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 92.7% (116454 hits / 125668 lookups)",
+  "strata serve: decode expert cache hit rate: 86.4% (73970 hits / 85649 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-dev-5080.json

@@ -0,0 +1,72 @@
+{
+ "label": "dev-5080",
+ "extra": [],
+ "gpus": "0",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2129.8,
+   "pp_tok_s": 1279.9,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5057.2,
+   "pp_tok_s": 1656.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 7964.7,
+   "pp_tok_s": 2017.2,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 14053.6,
+   "pp_tok_s": 1974.1,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3366.4,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 197.3,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.1,
+   "hit_rate": 0.8418683557684931,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning cauldron of black ink and white foam, throwing itself against the granite base with the force of a battering ram.\n\nInside the lantern room, the air was thick with the smell of kerosene and old brass. Elias wiped his hands on his woolen trousers, leaving faint, oily smears on the fabric. He checked the clock. 11:42 PM. The rotation was due.\n\nHe reached for the crank handle, a heavy iron wheel worn smooth by decades of his own palms. With a grunt that matched the strain of his shoulders, he began to turn it. *Click. Click. Click.* The gears engaged, and the massive Fresnel lens began its slow"
+  },
+  {
+   "name": "code",
+   "tok_s": 88.0,
+   "hit_rate": 0.7011385968111186,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: expert cache auto: 9.15 GiB free, 700 MiB reserved -> 3415 slots",
+  "strata generate: expert cache 4293 slots, 8.17 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2343 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 59.6% (2749 hits / 4610 lookups)",
+  "strata serve: decode expert cache hit rate: 64.2% (2136 hits / 3329 lookups)",
+  "strata serve: decode expert cache hit rate: 67.3% (2309 hits / 3432 lookups)",
+  "strata serve: decode expert cache hit rate: 72.5% (2539 hits / 3504 lookups)",
+  "strata serve: decode expert cache hit rate: 77.3% (2817 hits / 3642 lookups)",
+  "strata serve: decode expert cache hit rate: 75.3% (16500 hits / 21916 lookups)",
+  "strata serve: decode expert cache hit rate: 80.0% (2978 hits / 3723 lookups)",
+  "strata serve: decode expert cache hit rate: 84.2% (98752 hits / 117301 lookups)",
+  "strata serve: decode expert cache hit rate: 70.1% (69215 hits / 98718 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k18.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k18",
+ "extra": [
+  "--layer-split",
+  "18"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 10383.7,
+   "pp_tok_s": 1547.2,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 16371.9,
+   "pp_tok_s": 1694.5,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4308.0,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 239.8,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 77.2,
+   "reps": [
+    74.6,
+    78.1,
+    79.0
+   ],
+   "hit_rate": 0.9985378787878788,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, the glass was vibrating like a plucked cello string. The sea, usually a rhythmic pulse against the granite base, was now a chaotic, frothing beast, throwing itself against the rocks with the force of a battering ram.\n\nElias climbed the spiral stairs, his hand gripping the cold iron railing. His knees ached, not from age, but from the constant vibration of the floor beneath his feet. He reached the lantern room, the heart of the lighthouse, and checked the clockwork mechanism. The great lens, a prism of Fresnel glass, rotated with a steady, hypnotic *whirrr-click, whirrr-click*. It was the only sound that mattered. The light swept out into the black void, cutting through"
+  },
+  {
+   "name": "code",
+   "tok_s": 103.8,
+   "reps": [
+    97.6,
+    106.4,
+    107.3
+   ],
+   "hit_rate": 0.9984863945578232,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 524 windows; per window: wait for the GPU 6.456 ms, pool + plan 0.181 ms, host staging 0.581 ms, commit 0.206 ms",
+  "strata serve: stage 1: 524 windows; per window: wait for the GPU 14.872 ms, pool + plan 0.501 ms, host staging 0.000 ms, commit 0.387 ms",
+  "strata serve: stage 0: 593 windows; per window: wait for the GPU 6.479 ms, pool + plan 0.185 ms, host staging 0.515 ms, commit 0.207 ms",
+  "strata serve: stage 1: 593 windows; per window: wait for the GPU 15.053 ms, pool + plan 0.496 ms, host staging 0.000 ms, commit 0.390 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 18)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4608 slots, 8.24 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 18-47, expert cache 7152 slots (14.14 GiB), 7152 of its 7680 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-17",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-17 (CUDA0), 18-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.7% (6159 hits / 6240 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (3796 hits / 3840 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (3811 hits / 3840 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (27723 hits / 27840 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (3807 hits / 3840 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (142002 hits / 143040 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (139882 hits / 140160 lookups)",
+  "strata serve: decode expert cache hit rate: 99.9% (131807 hits / 132000 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (119278 hits / 120480 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (117410 hits / 117600 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (117422 hits / 117600 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k19.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k19",
+ "extra": [
+  "--layer-split",
+  "19"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 9875.2,
+   "pp_tok_s": 1626.9,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 15498.0,
+   "pp_tok_s": 1790.1,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4085.5,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 238.7,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 79.3,
+   "reps": [
+    74.8,
+    85.6,
+    77.4
+   ],
+   "hit_rate": 0.9977757017660928,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, the glass was vibrating like a plucked cello string. The sea, usually a rhythmic pulse against the granite base, was now a chaotic, frothing beast, throwing itself against the rocks with the force of a battering ram.\n\nElias stood at the top of the spiral staircase, his hand resting on the cold iron railing. He was not afraid. Fear was for the tourists who came in summer, wide-eyed and clutching their cameras. Elias was a keeper. He was a guardian.\n\nHe checked the logbook. The entry for the night was simple: *Wind, 60 knots. Visibility, zero. Light, operational.*\n\nHe walked down the stairs, his boots heavy on the metal grating. The smell of salt"
+  },
+  {
+   "name": "code",
+   "tok_s": 107.4,
+   "reps": [
+    98.9,
+    111.3,
+    112.1
+   ],
+   "hit_rate": 0.9986286196529353,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header row\n    and values from the corresponding row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 533 windows; per window: wait for the GPU 6.226 ms, pool + plan 0.294 ms, host staging 0.561 ms, commit 0.225 ms",
+  "strata serve: stage 1: 533 windows; per window: wait for the GPU 14.156 ms, pool + plan 0.373 ms, host staging 0.000 ms, commit 0.371 ms",
+  "strata serve: stage 0: 599 windows; per window: wait for the GPU 6.314 ms, pool + plan 0.292 ms, host staging 0.501 ms, commit 0.225 ms",
+  "strata serve: stage 1: 599 windows; per window: wait for the GPU 14.366 ms, pool + plan 0.375 ms, host staging 0.000 ms, commit 0.374 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 19)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4606 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 19-47, expert cache 7160 slots (14.14 GiB), 7160 of its 7424 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-18",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-18 (CUDA0), 19-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.5% (6141 hits / 6234 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (3797 hits / 3836 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (3814 hits / 3837 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (28659 hits / 28795 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (3814 hits / 3839 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (141378 hits / 142508 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (124963 hits / 125273 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (134574 hits / 134874 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (121132 hits / 122322 lookups)",
+  "strata serve: decode expert cache hit rate: 99.9% (114073 hits / 114235 lookups)",
+  "strata serve: decode expert cache hit rate: 99.9% (113598 hits / 113754 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k20.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k20",
+ "extra": [
+  "--layer-split",
+  "20"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 9044.3,
+   "pp_tok_s": 1776.4,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13928.7,
+   "pp_tok_s": 1991.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3743.5,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 236.8,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 76.4,
+   "reps": [
+    74.2,
+    78.4,
+    76.7
+   ],
+   "hit_rate": 0.9979911611088791,
+   "text": "The wind didn\u2019t just blow; it screamed. It clawed at the granite base of the tower, trying to pry it from the bedrock, and it howled through the narrow slits of the lantern room glass.\n\nElias Thorne was seventy years old, but he moved with the grace of a man who had spent fifty years listening to the sea. He knew the rhythm of the waves against the rocks below\u2014the *thud, thud, crash*\u2014and he knew when that rhythm changed. When the sea grew too quiet, or too violent, he knew a storm was coming, or a ship was in trouble.\n\nTonight, the storm was a Category 4. The rain lashed against the windows like gravel thrown by a giant. Elias stood in the center of the lantern room, his hand resting on the cold brass railing. He wasn\u2019t afraid. He was a guardian.\n\nHe checked the clockwork mechanism. It was old, a relic from 1"
+  },
+  {
+   "name": "code",
+   "tok_s": 105.9,
+   "reps": [
+    103.2,
+    106.8,
+    107.7
+   ],
+   "hit_rate": 0.9981849807154201,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 537 windows; per window: wait for the GPU 6.685 ms, pool + plan 0.403 ms, host staging 0.593 ms, commit 0.215 ms",
+  "strata serve: stage 1: 537 windows; per window: wait for the GPU 13.928 ms, pool + plan 0.282 ms, host staging 0.000 ms, commit 0.359 ms",
+  "strata serve: stage 0: 606 windows; per window: wait for the GPU 6.764 ms, pool + plan 0.399 ms, host staging 0.528 ms, commit 0.216 ms",
+  "strata serve: stage 1: 606 windows; per window: wait for the GPU 14.108 ms, pool + plan 0.288 ms, host staging 0.000 ms, commit 0.362 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 20)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4603 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 20-47, expert cache 7142 slots (14.14 GiB), 7142 of its 7168 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-19",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-19 (CUDA0), 20-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.9% (6162 hits / 6230 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (3803 hits / 3835 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (3810 hits / 3836 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (27185 hits / 27330 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (3807 hits / 3835 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (139825 hits / 140938 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (142122 hits / 142524 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (146556 hits / 146851 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (118026 hits / 119233 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (118729 hits / 118997 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (118791 hits / 119007 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k21.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k21",
+ "extra": [
+  "--layer-split",
+  "21"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8880.5,
+   "pp_tok_s": 1809.1,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13654.6,
+   "pp_tok_s": 2031.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3680.6,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 235.4,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 77.4,
+   "reps": [
+    80.2,
+    80.8,
+    71.2
+   ],
+   "hit_rate": 0.9973622550922132,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and shadow.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined his life for thirty years.\n\nElias was not a man of many words. He was a man of many tasks. His hands, mapped with scars from rope burns and broken glass, moved with a rhythm that predated his memory. He checked the clockwork mechanism of the Fresnel lens, ensuring the rotation was smooth, silent, and steady. *Tick. Tick. Tick.* The sound was the heartbeat of the coast.\n\n"
+  },
+  {
+   "name": "code",
+   "tok_s": 106.7,
+   "reps": [
+    100.6,
+    110.5,
+    109.0
+   ],
+   "hit_rate": 0.998218711927068,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 527 windows; per window: wait for the GPU 6.934 ms, pool + plan 0.620 ms, host staging 0.530 ms, commit 0.245 ms",
+  "strata serve: stage 1: 527 windows; per window: wait for the GPU 13.912 ms, pool + plan 0.365 ms, host staging 0.000 ms, commit 0.359 ms",
+  "strata serve: stage 0: 595 windows; per window: wait for the GPU 7.031 ms, pool + plan 0.595 ms, host staging 0.472 ms, commit 0.245 ms",
+  "strata serve: stage 1: 595 windows; per window: wait for the GPU 14.051 ms, pool + plan 0.361 ms, host staging 0.000 ms, commit 0.361 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 21)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4627 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 21-47, expert cache 6912 slots (13.78 GiB), 6912 of its 6912 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-20",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-20 (CUDA0), 21-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.7% (5672 hits / 5748 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (4287 hits / 4314 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (4287 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (29024 hits / 29229 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4278 hits / 4316 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (131499 hits / 133085 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (130974 hits / 131451 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (138767 hits / 139134 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (121454 hits / 122932 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (116809 hits / 117079 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (118803 hits / 119015 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k22.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k22",
+ "extra": [
+  "--layer-split",
+  "22"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8730.9,
+   "pp_tok_s": 1840.1,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13375.4,
+   "pp_tok_s": 2074.2,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3611.4,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 234.8,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 80.2,
+   "reps": [
+    78.4,
+    80.9,
+    81.5
+   ],
+   "hit_rate": 0.9977764197412695,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy-three years old, and he had been the keeper of the Blackwood Point Light for forty of those years. The other thirty-three had been spent in a small, drafty cottage on the mainland, waiting for the day when the sea would finally claim him.\n\nTonight, the storm was a Category 4. The waves weren\u2019t just crashing; they were climbing. They reached the gallery railing, ten feet above the base of the tower, spraying a fine mist of brine that coated Elias\u2019s glasses.\n\nHe climbed the spiral stairs, his knees protesting with every step. The iron steps were cold, slick with condensation. At the top, in the lantern room, the air was thin and sharp. The great Fresnel lens, a cathedral of glass and brass, stood silent. It was time to wake it.\n\n"
+  },
+  {
+   "name": "code",
+   "tok_s": 106.4,
+   "reps": [
+    100.7,
+    110.3,
+    108.1
+   ],
+   "hit_rate": 0.9971798370756316,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 529 windows; per window: wait for the GPU 7.229 ms, pool + plan 0.561 ms, host staging 0.529 ms, commit 0.233 ms",
+  "strata serve: stage 1: 529 windows; per window: wait for the GPU 13.058 ms, pool + plan 0.253 ms, host staging 0.000 ms, commit 0.334 ms",
+  "strata serve: stage 0: 598 windows; per window: wait for the GPU 7.325 ms, pool + plan 0.552 ms, host staging 0.469 ms, commit 0.235 ms",
+  "strata serve: stage 1: 598 windows; per window: wait for the GPU 13.238 ms, pool + plan 0.260 ms, host staging 0.000 ms, commit 0.337 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 22)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4640 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 22-47, expert cache 6656 slots (13.34 GiB), 6656 of its 6656 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-21",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-21 (CUDA0), 22-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 97.7% (5609 hits / 5739 lookups)",
+  "strata serve: decode expert cache hit rate: 98.7% (4256 hits / 4310 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4278 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (27008 hits / 27289 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (4260 hits / 4313 lookups)",
+  "strata serve: decode expert cache hit rate: 98.5% (135183 hits / 137274 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (136132 hits / 136708 lookups)",
+  "strata serve: decode expert cache hit rate: 99.8% (135515 hits / 135817 lookups)",
+  "strata serve: decode expert cache hit rate: 98.4% (121287 hits / 123280 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (117658 hits / 118026 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (120574 hits / 120915 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k23.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k23",
+ "extra": [
+  "--layer-split",
+  "23"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8742.6,
+   "pp_tok_s": 1837.7,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13084.2,
+   "pp_tok_s": 2120.3,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3566.1,
+  "turn2_tokens": 6660,
+  "reused": 6631,
+  "turn2_ms": 238.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 81.0,
+   "reps": [
+    76.8,
+    85.2,
+    81.1
+   ],
+   "hit_rate": 0.9961302681992337,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning black void, swallowing the horizon, throwing white foam against the granite base like a tantrum.\n\nInside the lantern room, the air was thick with the smell of kerosene and old brass. Elias wiped his hands on his wool trousers, the fabric stiff with salt and grease. He checked the clockwork mechanism that turned the lens. It was a beast of a thing, a Fresnel lens of eight feet in diameter, a cathedral of glass and crystal that focused the beam of the lamp into a single, piercing needle of light.\n\nHe wound the weights. *Click, click, click.* The gears engaged. The massive lens began its slow, hypnotic rotation.\n\nFor a moment, the"
+  },
+  {
+   "name": "code",
+   "tok_s": 108.4,
+   "reps": [
+    105.1,
+    109.0,
+    111.1
+   ],
+   "hit_rate": 0.9961698810290479,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 537 windows; per window: wait for the GPU 7.423 ms, pool + plan 0.953 ms, host staging 0.542 ms, commit 0.237 ms",
+  "strata serve: stage 1: 537 windows; per window: wait for the GPU 12.223 ms, pool + plan 0.236 ms, host staging 0.000 ms, commit 0.324 ms",
+  "strata serve: stage 0: 606 windows; per window: wait for the GPU 7.519 ms, pool + plan 0.904 ms, host staging 0.483 ms, commit 0.239 ms",
+  "strata serve: stage 1: 606 windows; per window: wait for the GPU 12.395 ms, pool + plan 0.242 ms, host staging 0.000 ms, commit 0.326 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 23)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4609 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 23-47, expert cache 6400 slots (12.82 GiB), 6400 of its 6400 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-22",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-22 (CUDA0), 23-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 96.8% (5528 hits / 5709 lookups)",
+  "strata serve: decode expert cache hit rate: 98.4% (4235 hits / 4303 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (4260 hits / 4308 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (30643 hits / 31089 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (4248 hits / 4309 lookups)",
+  "strata serve: decode expert cache hit rate: 98.3% (134776 hits / 137143 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (133024 hits / 133766 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (129995 hits / 130500 lookups)",
+  "strata serve: decode expert cache hit rate: 97.9% (115782 hits / 118287 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (117953 hits / 118464 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (117560 hits / 118012 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k24.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k24",
+ "extra": [
+  "--layer-split",
+  "24"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8122.5,
+   "pp_tok_s": 1978.0,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 12285.8,
+   "pp_tok_s": 2258.1,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3407.0,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 234.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.5,
+   "reps": [
+    84.9,
+    83.3,
+    82.3
+   ],
+   "hit_rate": 0.9949381327334084,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and shadow.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined Elias\u2019s life for thirty years. He was not merely a keeper; he was a guardian of silence.\n\nHe checked the clock. 11:42 PM. The rotation of the great Fresnel lens was due. With a practiced hand, he engaged the brake on the clockwork mechanism. The heavy brass gears groaned, then began their slow, rhythmic turn. *Tick. Tick. Tick.* It was a heartbeat"
+  },
+  {
+   "name": "code",
+   "tok_s": 109.5,
+   "reps": [
+    104.9,
+    111.0,
+    112.6
+   ],
+   "hit_rate": 0.9953796671318199,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 519 windows; per window: wait for the GPU 7.835 ms, pool + plan 0.767 ms, host staging 0.557 ms, commit 0.244 ms",
+  "strata serve: stage 1: 519 windows; per window: wait for the GPU 11.922 ms, pool + plan 0.232 ms, host staging 0.000 ms, commit 0.325 ms",
+  "strata serve: stage 0: 587 windows; per window: wait for the GPU 7.943 ms, pool + plan 0.744 ms, host staging 0.497 ms, commit 0.246 ms",
+  "strata serve: stage 1: 587 windows; per window: wait for the GPU 12.073 ms, pool + plan 0.238 ms, host staging 0.000 ms, commit 0.326 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 24)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4603 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 24-47, expert cache 6144 slots (12.35 GiB), 6144 of its 6144 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-23",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-23 (CUDA0), 24-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 95.9% (5916 hits / 6171 lookups)",
+  "strata serve: decode expert cache hit rate: 97.4% (3717 hits / 3815 lookups)",
+  "strata serve: decode expert cache hit rate: 98.4% (3761 hits / 3823 lookups)",
+  "strata serve: decode expert cache hit rate: 98.1% (27138 hits / 27664 lookups)",
+  "strata serve: decode expert cache hit rate: 97.8% (3742 hits / 3826 lookups)",
+  "strata serve: decode expert cache hit rate: 97.9% (126114 hits / 128875 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (133626 hits / 134658 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (132675 hits / 133350 lookups)",
+  "strata serve: decode expert cache hit rate: 96.8% (113593 hits / 117399 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (118372 hits / 118949 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (116981 hits / 117524 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k26.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k26",
+ "extra": [
+  "--layer-split",
+  "26"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 7880.9,
+   "pp_tok_s": 2038.6,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 11771.9,
+   "pp_tok_s": 2356.7,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3323.9,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 231.5,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.8,
+   "reps": [
+    85.9,
+    83.1,
+    82.4
+   ],
+   "hit_rate": 0.9933966961830739,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and green.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined Elias\u2019s life for thirty years. He was a keeper of the light, but more importantly, he was a keeper of time.\n\nHe checked his pocket watch. 11:42 PM. The rotation of the great Fresnel lens was due. With a practiced, rhythmic *clunk-clunk-clunk*, he wound the clockwork mechanism. The massive lens began its slow, hypnotic turn, sweeping its beam across"
+  },
+  {
+   "name": "code",
+   "tok_s": 109.7,
+   "reps": [
+    104.4,
+    112.7,
+    111.9
+   ],
+   "hit_rate": 0.9938670271352759,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 519 windows; per window: wait for the GPU 8.376 ms, pool + plan 1.030 ms, host staging 0.566 ms, commit 0.273 ms",
+  "strata serve: stage 1: 519 windows; per window: wait for the GPU 10.886 ms, pool + plan 0.214 ms, host staging 0.000 ms, commit 0.297 ms",
+  "strata serve: stage 0: 588 windows; per window: wait for the GPU 8.506 ms, pool + plan 0.985 ms, host staging 0.505 ms, commit 0.274 ms",
+  "strata serve: stage 1: 588 windows; per window: wait for the GPU 11.036 ms, pool + plan 0.219 ms, host staging 0.000 ms, commit 0.298 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 26)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4568 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 26-47, expert cache 5632 slots (11.32 GiB), 5632 of its 5632 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-25",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-25 (CUDA0), 26-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 92.8% (5259 hits / 5665 lookups)",
+  "strata serve: decode expert cache hit rate: 95.2% (4056 hits / 4259 lookups)",
+  "strata serve: decode expert cache hit rate: 97.2% (4164 hits / 4285 lookups)",
+  "strata serve: decode expert cache hit rate: 97.1% (27690 hits / 28510 lookups)",
+  "strata serve: decode expert cache hit rate: 97.5% (4184 hits / 4293 lookups)",
+  "strata serve: decode expert cache hit rate: 96.9% (121839 hits / 125724 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (132165 hits / 133648 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (137652 hits / 138567 lookups)",
+  "strata serve: decode expert cache hit rate: 95.3% (111307 hits / 116822 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (118436 hits / 119329 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (119109 hits / 119844 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k28.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k28",
+ "extra": [
+  "--layer-split",
+  "28"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 7989.3,
+   "pp_tok_s": 2010.9,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 12132.9,
+   "pp_tok_s": 2286.6,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3433.5,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 228.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.5,
+   "reps": [
+    83.7,
+    82.4,
+    84.5
+   ],
+   "hit_rate": 0.9900441970361987,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, the glass was vibrating like a plucked cello string. The sea, usually a rhythmic pulse against the granite base, was now a chaotic, frothing beast, throwing itself against the rocks with the force of a battering ram.\n\nElias climbed the spiral stairs, his hand gripping the cold iron railing. His knees ached, not from age, but from the constant vibration of the floor beneath his feet. He reached the lantern room, the heart of the lighthouse, and checked the clockwork mechanism. The great lens, a prism of Fresnel glass, rotated with a steady, hypnotic *whirrr-click, whirrr-click*. It was the only sound that mattered. The light swept out into the black void, cutting through"
+  },
+  {
+   "name": "code",
+   "tok_s": 112.5,
+   "reps": [
+    101.4,
+    112.1,
+    124.1
+   ],
+   "hit_rate": 0.9896636587366694,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv(file_path: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys being the header values\n    and values being the corresponding cell values.\n    \n    Args:\n        file_path: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(file_path, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n```<|im_end|>"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 527 windows; per window: wait for the GPU 8.765 ms, pool + plan 1.401 ms, host staging 0.593 ms, commit 0.281 ms",
+  "strata serve: stage 1: 527 windows; per window: wait for the GPU 9.893 ms, pool + plan 0.195 ms, host staging 0.000 ms, commit 0.276 ms",
+  "strata serve: stage 0: 578 windows; per window: wait for the GPU 8.908 ms, pool + plan 1.359 ms, host staging 0.559 ms, commit 0.284 ms",
+  "strata serve: stage 1: 578 windows; per window: wait for the GPU 10.025 ms, pool + plan 0.200 ms, host staging 0.000 ms, commit 0.278 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 28)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4537 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 28-47, expert cache 5120 slots (10.30 GiB), 5120 of its 5120 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-27",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-27 (CUDA0), 28-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 88.8% (5405 hits / 6086 lookups)",
+  "strata serve: decode expert cache hit rate: 93.8% (3539 hits / 3773 lookups)",
+  "strata serve: decode expert cache hit rate: 94.2% (3560 hits / 3781 lookups)",
+  "strata serve: decode expert cache hit rate: 95.7% (26702 hits / 27899 lookups)",
+  "strata serve: decode expert cache hit rate: 95.9% (3647 hits / 3804 lookups)",
+  "strata serve: decode expert cache hit rate: 95.7% (122618 hits / 128160 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (135035 hits / 136884 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (129476 hits / 130778 lookups)",
+  "strata serve: decode expert cache hit rate: 93.9% (111037 hits / 118290 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (118829 hits / 120140 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (90480 hits / 91425 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k30.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k30",
+ "extra": [
+  "--layer-split",
+  "30"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8301.3,
+   "pp_tok_s": 1935.4,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 12507.6,
+   "pp_tok_s": 2218.1,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3448.7,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 229.4,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 81.1,
+   "reps": [
+    78.0,
+    84.4,
+    81.0
+   ],
+   "hit_rate": 0.9798126997277784,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, the glass was vibrating like a plucked cello string. The sea, usually a rhythmic pulse against the granite base, was now a chaotic, frothing beast, throwing itself against the rocks with the force of a battering ram.\n\nElias stood at the top of the spiral staircase, his hand resting on the cold iron railing. He was not afraid. Fear was for the tourists who came in summer, wide-eyed and clutching their cameras. Elias was a keeper. He was a guardian.\n\nHe checked the logbook. The entry for the night was simple: *Wind, 60 knots. Visibility, zero. Light, operational.*\n\nHe walked down the stairs, his boots heavy on the metal grating. The smell of salt"
+  },
+  {
+   "name": "code",
+   "tok_s": 109.6,
+   "reps": [
+    100.7,
+    114.1,
+    114.1
+   ],
+   "hit_rate": 0.9825370343977666,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 540 windows; per window: wait for the GPU 9.206 ms, pool + plan 1.794 ms, host staging 0.525 ms, commit 0.298 ms",
+  "strata serve: stage 1: 540 windows; per window: wait for the GPU 8.767 ms, pool + plan 0.173 ms, host staging 0.000 ms, commit 0.242 ms",
+  "strata serve: stage 0: 608 windows; per window: wait for the GPU 9.390 ms, pool + plan 1.733 ms, host staging 0.471 ms, commit 0.301 ms",
+  "strata serve: stage 1: 608 windows; per window: wait for the GPU 8.910 ms, pool + plan 0.178 ms, host staging 0.000 ms, commit 0.244 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 30)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4479 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 30-47, expert cache 4608 slots (9.19 GiB), 4608 of its 4608 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-29",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-29 (CUDA0), 30-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 84.6% (5099 hits / 6028 lookups)",
+  "strata serve: decode expert cache hit rate: 90.7% (3403 hits / 3751 lookups)",
+  "strata serve: decode expert cache hit rate: 91.1% (3415 hits / 3747 lookups)",
+  "strata serve: decode expert cache hit rate: 94.1% (26108 hits / 27737 lookups)",
+  "strata serve: decode expert cache hit rate: 93.4% (3534 hits / 3784 lookups)",
+  "strata serve: decode expert cache hit rate: 94.6% (127532 hits / 134866 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (135828 hits / 137786 lookups)",
+  "strata serve: decode expert cache hit rate: 98.0% (132455 hits / 135184 lookups)",
+  "strata serve: decode expert cache hit rate: 91.0% (106351 hits / 116917 lookups)",
+  "strata serve: decode expert cache hit rate: 97.9% (115791 hits / 118246 lookups)",
+  "strata serve: decode expert cache hit rate: 98.3% (117198 hits / 119281 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-k32.json

@@ -0,0 +1,85 @@
+{
+ "label": "ks-k32",
+ "extra": [
+  "--layer-split",
+  "32"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 8837.1,
+   "pp_tok_s": 1818.0,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13449.8,
+   "pp_tok_s": 2062.7,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3565.7,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 227.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.0,
+   "reps": [
+    75.0,
+    86.1,
+    87.8
+   ],
+   "hit_rate": 0.9796934433020751,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning cauldron of black ink and white foam, throwing itself against the granite base with the force of a battering ram.\n\nInside the lantern room, the air was thick with the smell of kerosene and old brass. Elias wiped his hands on his woolen trousers, leaving faint, oily smears on the fabric. He checked the clock. 11:42 PM. The rotation was due.\n\nHe reached for the crank handle, a heavy iron wheel worn smooth by decades of his own palms. With a grunt that matched the strain of his shoulders, he began to turn it. *Click. Click. Click.* The gears engaged, and the massive Fresnel lens began its slow"
+  },
+  {
+   "name": "code",
+   "tok_s": 108.7,
+   "reps": [
+    98.8,
+    112.2,
+    115.1
+   ],
+   "hit_rate": 0.9777186846317358,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 533 windows; per window: wait for the GPU 9.839 ms, pool + plan 2.192 ms, host staging 0.585 ms, commit 0.318 ms",
+  "strata serve: stage 1: 533 windows; per window: wait for the GPU 8.014 ms, pool + plan 0.156 ms, host staging 0.000 ms, commit 0.231 ms",
+  "strata serve: stage 0: 602 windows; per window: wait for the GPU 10.014 ms, pool + plan 2.111 ms, host staging 0.520 ms, commit 0.321 ms",
+  "strata serve: stage 1: 602 windows; per window: wait for the GPU 8.094 ms, pool + plan 0.160 ms, host staging 0.000 ms, commit 0.232 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 32)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4488 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 32-47, expert cache 4096 slots (8.26 GiB), 4096 of its 4096 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-31",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-31 (CUDA0), 32-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 81.3% (4889 hits / 6010 lookups)",
+  "strata serve: decode expert cache hit rate: 87.7% (3272 hits / 3733 lookups)",
+  "strata serve: decode expert cache hit rate: 87.9% (3280 hits / 3733 lookups)",
+  "strata serve: decode expert cache hit rate: 92.3% (24238 hits / 26254 lookups)",
+  "strata serve: decode expert cache hit rate: 91.5% (3457 hits / 3778 lookups)",
+  "strata serve: decode expert cache hit rate: 93.7% (133254 hits / 142240 lookups)",
+  "strata serve: decode expert cache hit rate: 98.0% (130112 hits / 132812 lookups)",
+  "strata serve: decode expert cache hit rate: 98.0% (130069 hits / 132765 lookups)",
+  "strata serve: decode expert cache hit rate: 89.1% (105263 hits / 118190 lookups)",
+  "strata serve: decode expert cache hit rate: 97.2% (116970 hits / 120309 lookups)",
+  "strata serve: decode expert cache hit rate: 97.8% (115450 hits / 118081 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-ks-single.json

@@ -0,0 +1,76 @@
+{
+ "label": "ks-single",
+ "extra": [],
+ "gpus": "0",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 9310.4,
+   "pp_tok_s": 1725.6,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 14085.4,
+   "pp_tok_s": 1969.6,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3388.7,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 200.7,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 86.7,
+   "reps": [
+    78.2,
+    91.4,
+    90.4
+   ],
+   "hit_rate": 0.921882148773369,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy-three years old, and his hands were a map of the coastline he had guarded for fifty years. The knuckles were swollen, the skin cracked like dry earth, but they moved with a precision that no machine could replicate.\n\nHe stood on the gallery of the *Harrow Point* lighthouse, the iron railing cold against his palms. Below, the Atlantic churned in a frenzy, whitecaps exploding against the black rocks like fireworks. Above, the sky was a bruised purple, heavy with rain that hadn\u2019t yet fallen.\n\nElias checked the clock. 11:42 PM. The light was due.\n\nHe didn\u2019t look at the gauge. He didn\u2019t check the fuel levels. He knew. He always knew. The rhythm of the ocean, the pressure of the air, the subtle vibration"
+  },
+  {
+   "name": "code",
+   "tok_s": 105.2,
+   "reps": [
+    87.8,
+    113.5,
+    114.2
+   ],
+   "hit_rate": 0.9109811738648947,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 382 windows; per window: wait for the GPU 15.155 ms, pool + plan 4.331 ms, host staging 0.656 ms, commit 0.458 ms",
+  "strata serve: stage 0: 455 windows; per window: wait for the GPU 15.790 ms, pool + plan 5.552 ms, host staging 0.688 ms, commit 0.467 ms",
+  "strata serve: stage 0: 524 windows; per window: wait for the GPU 16.064 ms, pool + plan 5.529 ms, host staging 0.606 ms, commit 0.472 ms",
+  "strata serve: stage 0: 593 windows; per window: wait for the GPU 16.305 ms, pool + plan 5.462 ms, host staging 0.540 ms, commit 0.475 ms"
+ ],
+ "engine_lines": [
+  "strata generate: expert cache auto: 9.15 GiB free, 700 MiB reserved -> 3415 slots",
+  "strata generate: expert cache 4293 slots, 8.17 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2343 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 59.6% (2749 hits / 4610 lookups)",
+  "strata serve: decode expert cache hit rate: 67.8% (2299 hits / 3391 lookups)",
+  "strata serve: decode expert cache hit rate: 68.5% (2345 hits / 3424 lookups)",
+  "strata serve: decode expert cache hit rate: 75.6% (16605 hits / 21959 lookups)",
+  "strata serve: decode expert cache hit rate: 73.3% (2608 hits / 3558 lookups)",
+  "strata serve: decode expert cache hit rate: 83.7% (104878 hits / 125346 lookups)",
+  "strata serve: decode expert cache hit rate: 92.7% (119739 hits / 129184 lookups)",
+  "strata serve: decode expert cache hit rate: 92.2% (116867 hits / 126770 lookups)",
+  "strata serve: decode expert cache hit rate: 68.1% (64111 hits / 94116 lookups)",
+  "strata serve: decode expert cache hit rate: 89.0% (96370 hits / 108259 lookups)",
+  "strata serve: decode expert cache hit rate: 91.1% (102827 hits / 112875 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-rel-5080.json

@@ -0,0 +1,72 @@
+{
+ "label": "rel-5080",
+ "extra": [],
+ "gpus": "0",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2132.8,
+   "pp_tok_s": 1278.1,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5008.9,
+   "pp_tok_s": 1672.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 7855.4,
+   "pp_tok_s": 2045.2,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13833.9,
+   "pp_tok_s": 2005.4,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3298.2,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 197.9,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.2,
+   "hit_rate": 0.8418683557684931,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning cauldron of black ink and white foam, throwing itself against the granite base with the force of a battering ram.\n\nInside the lantern room, the air was thick with the smell of kerosene and old brass. Elias wiped his hands on his woolen trousers, leaving faint, oily smears on the fabric. He checked the clock. 11:42 PM. The rotation was due.\n\nHe reached for the crank handle, a heavy iron wheel worn smooth by decades of his own palms. With a grunt that matched the strain of his shoulders, he began to turn it. *Click. Click. Click.* The gears engaged, and the massive Fresnel lens began its slow"
+  },
+  {
+   "name": "code",
+   "tok_s": 88.2,
+   "hit_rate": 0.7011385968111186,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: expert cache auto: 9.15 GiB free, 700 MiB reserved -> 3415 slots",
+  "strata generate: expert cache 4293 slots, 8.17 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2343 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 59.6% (2749 hits / 4610 lookups)",
+  "strata serve: decode expert cache hit rate: 64.2% (2136 hits / 3329 lookups)",
+  "strata serve: decode expert cache hit rate: 67.3% (2309 hits / 3432 lookups)",
+  "strata serve: decode expert cache hit rate: 72.5% (2539 hits / 3504 lookups)",
+  "strata serve: decode expert cache hit rate: 77.3% (2817 hits / 3642 lookups)",
+  "strata serve: decode expert cache hit rate: 75.3% (16500 hits / 21916 lookups)",
+  "strata serve: decode expert cache hit rate: 80.0% (2978 hits / 3723 lookups)",
+  "strata serve: decode expert cache hit rate: 84.2% (98752 hits / 117301 lookups)",
+  "strata serve: decode expert cache hit rate: 70.1% (69215 hits / 98718 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-t75-5080.json

@@ -0,0 +1,78 @@
+{
+ "label": "t75-5080",
+ "extra": [],
+ "gpus": "0",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2158.8,
+   "pp_tok_s": 1262.7,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5060.0,
+   "pp_tok_s": 1655.9,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3389.7,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 213.5,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 80.4,
+   "reps": [
+    80.4
+   ],
+   "hit_rate": 0.855538332528651,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning black void, punctuated by whitecaps that looked like teeth.\n\nHe climbed the spiral stairs, his boots heavy with salt and exhaustion. The iron steps groaned under his weight, a sound that had become so familiar to him that he could sleep to it. But tonight, he wasn\u2019t sleeping. He was working.\n\nThe lens was the heart of the lighthouse. A massive, rotating Fresnel lens, a cathedral of glass and brass that turned with a mechanical hum. Elias checked the clockwork mechanism. It was smooth, rhythmic, a heartbeat in the dark. But the light itself was failing.\n\nHe climbed higher, into the lantern room. The wind battered the glass panes,"
+  },
+  {
+   "name": "code",
+   "tok_s": 87.5,
+   "reps": [
+    87.5
+   ],
+   "hit_rate": 0.7075461363025227,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 39 windows; per window: wait for the GPU 17.252 ms, pool + plan 8.473 ms, host staging 0.611 ms, commit 0.525 ms",
+  "strata serve: stage 0: 47 windows; per window: wait for the GPU 18.018 ms, pool + plan 8.616 ms, host staging 0.607 ms, commit 0.534 ms",
+  "strata serve: stage 0: 169 windows; per window: wait for the GPU 15.223 ms, pool + plan 5.717 ms, host staging 0.688 ms, commit 0.467 ms",
+  "strata serve: stage 0: 243 windows; per window: wait for the GPU 16.500 ms, pool + plan 7.316 ms, host staging 0.742 ms, commit 0.481 ms"
+ ],
+ "engine_lines": [
+  "strata generate: expert cache auto: 9.15 GiB free, 700 MiB reserved -> 3415 slots",
+  "strata generate: expert cache 4294 slots, 8.18 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2343 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 59.6% (2749 hits / 4610 lookups)",
+  "strata serve: decode expert cache hit rate: 64.2% (2136 hits / 3329 lookups)",
+  "strata serve: decode expert cache hit rate: 67.3% (2309 hits / 3432 lookups)",
+  "strata serve: decode expert cache hit rate: 75.5% (16316 hits / 21602 lookups)",
+  "strata serve: decode expert cache hit rate: 74.4% (2660 hits / 3573 lookups)",
+  "strata serve: decode expert cache hit rate: 83.4% (92664 hits / 111086 lookups)",
+  "strata serve: decode expert cache hit rate: 72.0% (70946 hits / 98601 lookups)",
+  "strata generate: expert cache auto: 9.15 GiB free, 700 MiB reserved -> 3415 slots",
+  "strata generate: expert cache 4293 slots, 8.17 GiB of VRAM; policy is",
+  "strata serve: the prompt path borrows 2343 cache slots (4.42 GiB)",
+  "strata serve: decode expert cache hit rate: 59.6% (2749 hits / 4610 lookups)",
+  "strata serve: decode expert cache hit rate: 64.2% (2136 hits / 3329 lookups)",
+  "strata serve: decode expert cache hit rate: 67.3% (2309 hits / 3432 lookups)",
+  "strata serve: decode expert cache hit rate: 75.5% (16316 hits / 21602 lookups)",
+  "strata serve: decode expert cache hit rate: 74.4% (2659 hits / 3572 lookups)",
+  "strata serve: decode expert cache hit rate: 85.6% (101004 hits / 118059 lookups)",
+  "strata serve: decode expert cache hit rate: 70.8% (71044 hits / 100409 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-v-x2-auto.json

@@ -0,0 +1,103 @@
+{
+ "label": "v-x2-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2577.0,
+   "pp_tok_s": 1057.8,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 4433.5,
+   "pp_tok_s": 1889.9,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 7885.8,
+   "pp_tok_s": 2037.3,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13384.6,
+   "pp_tok_s": 2072.8,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3667.5,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 253.8,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 79.5,
+   "reps": [
+    78.9,
+    79.0,
+    80.7
+   ],
+   "hit_rate": 0.9969903443250137,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the rough fabric scratching against his stubble. Outside the thick, salt-crusted glass of the lantern room, the Atlantic was a churning cauldron of black and white foam. Rain lashed against the panes in horizontal sheets, blurring the world into a smear of gray and shadow.\n\nInside, however, it was warm. The smell of kerosene and old wood hung heavy in the air, a scent that had defined his life for thirty years.\n\nElias was not a man of many words. He was a man of many tasks. His hands, mapped with scars from rope burns and broken glass, moved with a rhythm that predated his memory. He checked the clockwork mechanism of the Fresnel lens, listening for the faint, rhythmic *tick-tick-tick* that signaled the rotation was true. If the rhythm faltered, ships"
+  },
+  {
+   "name": "code",
+   "tok_s": 108.9,
+   "reps": [
+    103.3,
+    113.7,
+    109.5
+   ],
+   "hit_rate": 0.9971795386830036,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 542 windows; per window: wait for the GPU 7.167 ms, pool + plan 0.572 ms, host staging 0.550 ms, commit 0.232 ms",
+  "strata serve: stage 1: 542 windows; per window: wait for the GPU 12.855 ms, pool + plan 0.250 ms, host staging 0.000 ms, commit 0.334 ms",
+  "strata serve: stage 0: 610 windows; per window: wait for the GPU 7.276 ms, pool + plan 0.559 ms, host staging 0.493 ms, commit 0.234 ms",
+  "strata serve: stage 1: 610 windows; per window: wait for the GPU 13.048 ms, pool + plan 0.258 ms, host staging 0.000 ms, commit 0.337 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: layer split auto: CUDA0 84 SMs at 2.81 GHz -> 0.31 ms per layer, 8.30 GiB for experts",
+  "strata generate: layer split auto: CUDA1 82 SMs at 1.86 GHz -> 0.48 ms per layer, 14.14 GiB for experts",
+  "strata generate: layer split auto: K=22 - predicted 19.7 ms per decode window; the caches hold 11298 of 12288 profiled pairs (~99.7% of the routed mass)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4640 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 22-47, expert cache 6656 slots (13.34 GiB), 6656 of its 6656 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-21",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-21 (CUDA0), 22-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 97.7% (5609 hits / 5739 lookups)",
+  "strata serve: decode expert cache hit rate: 98.5% (4241 hits / 4306 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (4262 hits / 4309 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (4290 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (4303 hits / 4319 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (28379 hits / 28721 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (4285 hits / 4314 lookups)",
+  "strata serve: decode expert cache hit rate: 98.5% (136232 hits / 138256 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (135540 hits / 136216 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (136812 hits / 137225 lookups)",
+  "strata serve: decode expert cache hit rate: 98.4% (118437 hits / 120406 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (112836 hits / 113230 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-v-x2r-auto.json

@@ -0,0 +1,88 @@
+{
+ "label": "v-x2r-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "2,0",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 9610.0,
+   "pp_tok_s": 1671.8,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 14786.1,
+   "pp_tok_s": 1876.3,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3850.8,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 245.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 90.3,
+   "reps": [
+    91.0,
+    97.0,
+    82.8
+   ],
+   "hit_rate": 0.993905472636816,
+   "text": "The wind didn\u2019t just blow; it screamed. It tore through the iron gallery of the Blackwood Point Lighthouse, rattling the glass panes like a prisoner shaking its bars.\n\nElias Thorne was seventy years old, and he had been the keeper for forty of them. He knew every creak of the spiral staircase, every draft that slipped under the door, and every shift in the barometric pressure that preceded a storm. He was not merely a worker; he was the heartbeat of the coast.\n\nIt was November, the month of grey seas and shorter days. The fog had rolled in thick, a wall of white so dense that even the beam of the Fresnel lens seemed to struggle against it. Elias stood in the lantern room, his hands roughened by salt and rope, adjusting the rotation mechanism. The great lens, a cathedral of glass and brass, began its slow, rhythmic sweep. *Flash. Dark. Flash. Dark.*\n\nHe checked the log"
+  },
+  {
+   "name": "code",
+   "tok_s": 105.1,
+   "reps": [
+    98.4,
+    109.9,
+    107.0
+   ],
+   "hit_rate": 0.9949900793650793,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 492 windows; per window: wait for the GPU 15.016 ms, pool + plan 0.297 ms, host staging 0.517 ms, commit 0.404 ms",
+  "strata serve: stage 1: 492 windows; per window: wait for the GPU 5.775 ms, pool + plan 0.854 ms, host staging 0.001 ms, commit 0.210 ms",
+  "strata serve: stage 0: 561 windows; per window: wait for the GPU 15.257 ms, pool + plan 0.306 ms, host staging 0.455 ms, commit 0.408 ms",
+  "strata serve: stage 1: 561 windows; per window: wait for the GPU 5.852 ms, pool + plan 0.820 ms, host staging 0.001 ms, commit 0.210 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 57.7 GB/s -> pcie_frac 0.55",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 9.98 GiB free",
+  "strata generate: layer split auto: CUDA0 82 SMs at 1.86 GHz -> 0.48 ms per layer, 16.45 GiB for experts",
+  "strata generate: layer split auto: CUDA1 84 SMs at 2.81 GHz -> 0.31 ms per layer, 5.99 GiB for experts",
+  "strata generate: layer split auto: K=30 - predicted 20.8 ms per decode window; the caches hold 10684 of 12288 profiled pairs (~99.5% of the routed mass)",
+  "strata generate: expert cache auto: 18.62 GiB free, 700 MiB reserved -> 6634 slots",
+  "strata generate: expert cache 7680 slots, 14.23 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 30-47, expert cache 3004 slots (5.98 GiB), 3004 of its 4608 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-29",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-29 (CUDA0), 30-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 93.2% (5367 hits / 5760 lookups)",
+  "strata serve: decode expert cache hit rate: 93.7% (4049 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 96.7% (4178 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 97.7% (26272 hits / 26880 lookups)",
+  "strata serve: decode expert cache hit rate: 97.1% (4194 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 97.8% (110748 hits / 113280 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (113787 hits / 114240 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (127856 hits / 128640 lookups)",
+  "strata serve: decode expert cache hit rate: 96.3% (116528 hits / 120960 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (116423 hits / 117120 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (120354 hits / 120960 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-v-x3-auto.json

@@ -0,0 +1,89 @@
+{
+ "label": "v-x3-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,1,2",
+ "needles": [
+  {
+   "tokens": 16066,
+   "prompt_ms": 9525.4,
+   "pp_tok_s": 1686.6,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 15017.1,
+   "pp_tok_s": 1847.4,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4039.6,
+  "turn2_tokens": 6660,
+  "reused": 6631,
+  "turn2_ms": 360.2,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 67.6,
+   "reps": [
+    59.8,
+    83.1,
+    59.9
+   ],
+   "hit_rate": 0.9967581085879247,
+   "text": "The wind didn\u2019t just blow; it screamed. It tore through the iron gallery of the Blackwood Point Lighthouse, rattling the glass panes like a prisoner shaking its bars.\n\nElias Thorne was seventy years old, and he had been the keeper for forty of them. He knew every creak in the spiral staircase, every draft that slipped under the door, and every shift in the weather that the barometer couldn\u2019t quite predict. He knew that when the fog rolled in thick enough to swallow the horizon, the world ceased to exist. There was only the light, the stone, and him.\n\nIt was November, the month of storms. The sea was a churning cauldron of grey and white, throwing itself against the granite base with the force of a battering ram. Elias stood in the lantern room, his hands rough as sandpaper, adjusting the massive Fresnel lens. The rotation mechanism groaned, a sound like a ship\u2019s anchor chain being h"
+  },
+  {
+   "name": "code",
+   "tok_s": 89.6,
+   "reps": [
+    81.0,
+    104.8,
+    83.0
+   ],
+   "hit_rate": 0.9975041387598007,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 2: 522 windows; per window: wait for the GPU 12.405 ms, pool + plan 0.242 ms, host staging 0.006 ms, commit 0.375 ms",
+  "strata serve: stage 0: 590 windows; per window: wait for the GPU 7.387 ms, pool + plan 0.549 ms, host staging 0.517 ms, commit 0.248 ms",
+  "strata serve: stage 1: 590 windows; per window: wait for the GPU 3.071 ms, pool + plan 0.011 ms, host staging 0.000 ms, commit 0.606 ms",
+  "strata serve: stage 2: 590 windows; per window: wait for the GPU 12.592 ms, pool + plan 0.250 ms, host staging 0.006 ms, commit 0.377 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 3 GPUs: CUDA0, then CUDA1,2 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 3.2 GB/s -> pcie_frac 0.00",
+  "strata generate: layer split: CUDA1 holds its weights, session; 16.67 GiB free",
+  "strata generate: layer split: CUDA2 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA2 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: layer split auto: CUDA0 84 SMs at 2.81 GHz -> 0.31 ms per layer, 8.30 GiB for experts",
+  "strata generate: layer split auto: CUDA1 68 SMs at 1.54 GHz -> 0.69 ms per layer, 13.50 GiB for experts",
+  "strata generate: layer split auto: CUDA2 82 SMs at 1.86 GHz -> 0.48 ms per layer, 14.14 GiB for experts",
+  "strata generate: layer split auto: K=22,23 - predicted 19.9 ms per decode window; the caches hold 11298 of 12288 profiled pairs (~99.7% of the routed mass)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4640 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 22-22, expert cache 256 slots (0.52 GiB), 256 of its 256 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA2 runs layers 23-47, expert cache 6400 slots (12.82 GiB), 6400 of its 6400 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-21",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-21 (CUDA0), 22-22 (CUDA1), 23-47 (CUDA2), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 97.7% (5609 hits / 5739 lookups)",
+  "strata serve: decode expert cache hit rate: 98.7% (4256 hits / 4310 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4278 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (29862 hits / 30170 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (4264 hits / 4313 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (134418 hits / 136339 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (116986 hits / 117525 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (129134 hits / 129554 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2-auto-c1024.json

@@ -0,0 +1,84 @@
+{
+ "label": "x2-auto-c1024",
+ "extra": [
+  "--layer-split",
+  "auto",
+  "--prefill",
+  "1024"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2648.7,
+   "pp_tok_s": 1029.2,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5547.0,
+   "pp_tok_s": 1510.5,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 10123.3,
+   "pp_tok_s": 1587.0,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 17652.0,
+   "pp_tok_s": 1571.7,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4379.9,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 235.5,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 82.7,
+   "hit_rate": 1.0,
+   "text": "The wind didn\u2019t just blow; it screamed. It tore through the iron gallery of the Blackwood Point Lighthouse, rattling the glass panes like a prisoner shaking its bars.\n\nElias Thorne was seventy years old, and he had been the keeper for forty of them. He knew every creak in the spiral staircase, every draft that slipped under the door, and every shift in the weather that the barometer couldn\u2019t quite predict. He knew that when the fog rolled in thick enough to swallow the horizon, the world ceased to exist. There was only the light, the stone, and him.\n\nIt was November, the month of storms. The sea was a churning cauldron of grey and white, throwing itself against the granite base with a violence that made the floorboards tremble. Elias stood by the lantern room, his hands wrapped around a mug of coffee that had gone cold hours ago. He wasn\u2019t drinking it. He was holding it for warmth."
+  },
+  {
+   "name": "code",
+   "tok_s": 111.1,
+   "hit_rate": 1.0,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv(file_path: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys being the header values\n    and values being the corresponding cell values.\n    \n    Args:\n        file_path: Path to the CSV file.\n        \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(file_path, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_basic():\n    \"\"\"Test basic CSV parsing with simple data.\"\"\"\n    import tempfile\n    import os\n    \n    # Create a temporary CSV"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: layer split auto: K=19 - the caches hold 12288 of 12288 profiled pairs (fullest device 99%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3613 slots",
+  "strata generate: expert cache 4864 slots, 8.76 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 19-47, expert cache 7424 slots (14.66 GiB), 7424 of its 7424 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-18",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-18 (CUDA0), 19-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 100.0% (5760 hits / 5760 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (4320 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (4320 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (4320 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (4320 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (26400 hits / 26400 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (4320 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (130080 hits / 130080 lookups)",
+  "strata serve: decode expert cache hit rate: 100.0% (110400 hits / 110400 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2-auto.json

@@ -0,0 +1,82 @@
+{
+ "label": "x2-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2711.7,
+   "pp_tok_s": 1005.3,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5010.3,
+   "pp_tok_s": 1672.4,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 8994.8,
+   "pp_tok_s": 1786.1,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 15497.2,
+   "pp_tok_s": 1790.2,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4050.3,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 249.7,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 83.0,
+   "hit_rate": 0.9920220447608565,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias had lived in the tower for thirty years, and he knew the difference between a gale and a hurricane by the way the glass rattled in its frame. Tonight, it was the latter. The sea was a churning cauldron of black ink and white foam, throwing itself against the granite base with the force of a battering ram.\n\nInside the lantern room, the air was thick with the smell of kerosene and old brass. Elias wiped his hands on his woolen trousers, leaving faint, oily smears on the fabric. He checked the clock. 11:42 PM. The rotation was due.\n\nHe reached for the crank handle, a heavy iron wheel worn smooth by decades of his own palms. With a grunt that matched the strain of his shoulders, he began to turn it. The massive Fresnel lens, a cathedral of glass and crystal, began its slow, hypnotic sweep"
+  },
+  {
+   "name": "code",
+   "tok_s": 101.8,
+   "hit_rate": 0.991767530342175,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: layer split auto: K=19 - the caches hold 11767 of 12288 profiled pairs (fullest device 100%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4606 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 19-47, expert cache 7160 slots (14.14 GiB), 7160 of its 7424 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-18",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-18 (CUDA0), 19-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.5% (6141 hits / 6234 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (3797 hits / 3838 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (3804 hits / 3838 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (3816 hits / 3839 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (3825 hits / 3839 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (27693 hits / 27832 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (3815 hits / 3839 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (127081 hits / 128103 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (119386 hits / 120377 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2-k16.json

@@ -0,0 +1,81 @@
+{
+ "label": "x2-k16",
+ "extra": [
+  "--layer-split",
+  "16"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2987.1,
+   "pp_tok_s": 912.6,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5908.6,
+   "pp_tok_s": 1418.1,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 10655.8,
+   "pp_tok_s": 1507.7,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 18484.6,
+   "pp_tok_s": 1500.9,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4790.8,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 243.3,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 71.4,
+   "hit_rate": 0.9814261642156863,
+   "text": "The wind didn\u2019t just blow; it screamed. It clawed at the granite base of the tower, trying to pry the lighthouse from its bedrock. Inside, however, the air was still, smelling faintly of kerosene and old paper.\n\nElias had been the keeper of the Point for thirty-two years. He knew every creak in the floorboards, every draft that slipped through the window seals, and exactly how much oil the great Fresnel lens needed to turn its slow, hypnotic rotation.\n\nIt was 2:00 AM. The storm was at its peak. The sea was a chaotic black mirror, whipped into whitecaps that looked like jagged teeth.\n\nElias climbed the spiral stairs, his hand gripping the cold iron railing. His boots, heavy with salt and wear, thumped a steady rhythm against the stone. *Thump. Thump. Thump.*\n\nAt the top, the lantern room hummed with the vibration of"
+  },
+  {
+   "name": "code",
+   "tok_s": 86.4,
+   "hit_rate": 0.9798276439089693,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 16)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4096 slots, 7.17 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 16-47, expert cache 7124 slots (14.14 GiB), 7124 of its 8192 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-15",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-15 (CUDA0), 16-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 97.7% (5626 hits / 5760 lookups)",
+  "strata serve: decode expert cache hit rate: 98.1% (4237 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (4269 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4279 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (4298 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (28471 hits / 28800 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4282 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 98.1% (128135 hits / 130560 lookups)",
+  "strata serve: decode expert cache hit rate: 98.0% (117109 hits / 119520 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2-k22.json

@@ -0,0 +1,81 @@
+{
+ "label": "x2-k22",
+ "extra": [
+  "--layer-split",
+  "22"
+ ],
+ "gpus": "0,2",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2537.7,
+   "pp_tok_s": 1074.2,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 4440.0,
+   "pp_tok_s": 1887.2,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 7872.5,
+   "pp_tok_s": 2040.8,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 13405.0,
+   "pp_tok_s": 2069.6,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 3610.4,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 236.0,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 78.1,
+   "hit_rate": 0.9851122208995093,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy-three years old, and he had been the keeper of the Blackwood Point Light for forty of those years. The other thirty-three had been spent in a hospital bed, waiting for the end.\n\nEvery night, Elias climbed the one hundred and twelve steps of the spiral staircase. His knees, worn smooth by the friction of his own boots over decades, protested with a sharp ache. He counted them. *One, two, three\u2026* The rhythm was his heartbeat. If he lost the count, he lost the light.\n\nAt the top, the lantern room hummed with the low, electric thrum of the Fresnel lens. It was a cathedral of glass and brass, a machine that had never failed him. Not once. Not in four decades.\n\nHe polished the brass railing with a rag that smelled of lemon and"
+  },
+  {
+   "name": "code",
+   "tok_s": 104.3,
+   "hit_rate": 0.9842543225870746,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split 22)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 18.13 GiB free",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4640 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 22-47, expert cache 6656 slots (13.34 GiB), 6656 of its 6656 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-21",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-21 (CUDA0), 22-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 97.7% (5609 hits / 5739 lookups)",
+  "strata serve: decode expert cache hit rate: 98.5% (4241 hits / 4306 lookups)",
+  "strata serve: decode expert cache hit rate: 98.9% (4262 hits / 4309 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (4290 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (4303 hits / 4319 lookups)",
+  "strata serve: decode expert cache hit rate: 98.8% (28379 hits / 28721 lookups)",
+  "strata serve: decode expert cache hit rate: 99.3% (4285 hits / 4314 lookups)",
+  "strata serve: decode expert cache hit rate: 98.5% (135713 hits / 137764 lookups)",
+  "strata serve: decode expert cache hit rate: 98.4% (118518 hits / 120414 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2r-auto.json

@@ -0,0 +1,82 @@
+{
+ "label": "x2r-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "2,0",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 2804.2,
+   "pp_tok_s": 972.1,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 5641.7,
+   "pp_tok_s": 1485.2,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 10643.4,
+   "pp_tok_s": 1509.5,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 18363.5,
+   "pp_tok_s": 1510.8,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 4517.2,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 245.6,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 78.9,
+   "hit_rate": 0.9913117283950618,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his woolen collar up to his chin, the fabric stiff with salt and ice. Outside the thick glass of the lantern room, the Atlantic was a churning cauldron of black and white, the waves crashing against the granite base with a violence that shook the very foundations of the tower. It was November, 1924, and the storm had been raging for three days.\n\nElias was not a man who feared the dark. He was a man who feared the silence.\n\nHe checked the clock on the wall. 11:42 PM. The rotation of the great Fresnel lens was steady, a rhythmic *whirrr-click, whirrr-click* that seemed to sync with his own heartbeat. He had been doing this for forty years. His hands were mapped with scars from rope burns and broken glass, his eyes permanently squinted against the glare of the bulb."
+  },
+  {
+   "name": "code",
+   "tok_s": 93.4,
+   "hit_rate": 0.990286044973545,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 57.7 GB/s -> pcie_frac 0.55",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 9.98 GiB free",
+  "strata generate: layer split auto: K=36 - the caches hold 11775 of 12288 profiled pairs (fullest device 100%)",
+  "strata generate: expert cache auto: 18.62 GiB free, 700 MiB reserved -> 6634 slots",
+  "strata generate: expert cache 8828 slots, 16.45 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 36-47, expert cache 2945 slots (5.99 GiB), 2945 of its 3072 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-35",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-35 (CUDA0), 36-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 98.6% (5678 hits / 5760 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (4283 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.2% (4285 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.6% (4303 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.7% (4306 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.5% (27700 hits / 27840 lookups)",
+  "strata serve: decode expert cache hit rate: 99.4% (4295 hits / 4320 lookups)",
+  "strata serve: decode expert cache hit rate: 99.1% (128474 hits / 129600 lookups)",
+  "strata serve: decode expert cache hit rate: 99.0% (119785 hits / 120960 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x2t-auto.json

@@ -0,0 +1,99 @@
+{
+ "label": "x2t-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,1",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 4201.1,
+   "pp_tok_s": 648.9,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 9655.6,
+   "pp_tok_s": 867.8,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 17573.8,
+   "pp_tok_s": 914.2,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 30615.6,
+   "pp_tok_s": 906.2,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 7864.7,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 366.9,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 45.5,
+   "reps": [
+    45.5
+   ],
+   "hit_rate": 0.9703345151405784,
+   "text": "The wind didn\u2019t just blow; it screamed.\n\nElias pulled his wool coat tighter around his shoulders, the fabric stiff with salt and age. He was seventy years old, but he moved with the grace of a man who had spent decades on a narrow spiral staircase. His hands, rough as sandpaper, knew every bolt, every lever, and every creak in the iron bones of the tower.\n\nFor forty years, Elias had been the guardian of the Point. Not just of the light, but of the silence between the waves.\n\nIt was November. The sky was a bruised purple, heavy with rain that hadn\u2019t yet fallen. The sea was a churning cauldron of grey and white, throwing itself against the granite base with a violence that shook the floorboards beneath Elias\u2019s boots.\n\nHe climbed the stairs. One hundred and twelve steps. He counted them by feel, his left hand trailing along the cold iron railing. At the top, the air was"
+  },
+  {
+   "name": "code",
+   "tok_s": 58.0,
+   "reps": [
+    58.0
+   ],
+   "hit_rate": 0.9700542735219787,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv_to_dicts(filepath: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys from the header and values from the row.\n    \n    Args:\n        filepath: Path to the CSV file.\n    \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result: List[Dict[str, Any]] = []\n    with open(filepath, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_to_dicts():\n    import tempfile\n    import os\n    \n    # Create a temporary CSV file\n    with tempfile.NamedTemporaryFile"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 0: 168 windows; per window: wait for the GPU 9.683 ms, pool + plan 0.509 ms, host staging 0.714 ms, commit 0.320 ms",
+  "strata serve: stage 1: 168 windows; per window: wait for the GPU 25.253 ms, pool + plan 0.881 ms, host staging 0.000 ms, commit 0.559 ms",
+  "strata serve: stage 0: 242 windows; per window: wait for the GPU 9.860 ms, pool + plan 0.550 ms, host staging 0.750 ms, commit 0.328 ms",
+  "strata serve: stage 1: 242 windows; per window: wait for the GPU 26.029 ms, pool + plan 0.980 ms, host staging 0.000 ms, commit 0.571 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 3.2 GB/s -> pcie_frac 0.00",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 16.18 GiB free",
+  "strata generate: layer split auto: K=20 - the caches hold 10761 of 12288 profiled pairs (fullest device 100%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4603 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 20-47, expert cache 6157 slots (12.19 GiB), 6157 of its 7168 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-19",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-19 (CUDA0), 20-47 (CUDA1), one hand-off per window",
+  "strata generate: layer split across 2 GPUs: CUDA0, then CUDA1 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 3.2 GB/s -> pcie_frac 0.00",
+  "strata generate: layer split: CUDA1 holds its weights, session and head; 16.18 GiB free",
+  "strata generate: layer split auto: K=20 - the caches hold 10761 of 12288 profiled pairs (fullest device 100%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3345 slots",
+  "strata generate: expert cache 4603 slots, 8.29 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 20-47, expert cache 6157 slots (12.19 GiB), 6157 of its 7168 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-19",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-19 (CUDA0), 20-47 (CUDA1), one hand-off per window",
+  "strata serve: decode expert cache hit rate: 96.6% (5559 hits / 5754 lookups)",
+  "strata serve: decode expert cache hit rate: 97.5% (4205 hits / 4315 lookups)",
+  "strata serve: decode expert cache hit rate: 97.8% (4217 hits / 4312 lookups)",
+  "strata serve: decode expert cache hit rate: 98.6% (4257 hits / 4318 lookups)"
+ ]
+}

bench/results/2026-09-29-layer-split/data/pp-x3-auto.json

@@ -0,0 +1,99 @@
+{
+ "label": "x3-auto",
+ "extra": [
+  "--layer-split",
+  "auto"
+ ],
+ "gpus": "0,2,1",
+ "needles": [
+  {
+   "tokens": 2726,
+   "prompt_ms": 3657.4,
+   "pp_tok_s": 745.3,
+   "found": true,
+   "answer": "meadow-254<|im_end|>"
+  },
+  {
+   "tokens": 8379,
+   "prompt_ms": 7594.1,
+   "pp_tok_s": 1103.4,
+   "found": true,
+   "answer": "juniper-766<|im_end|>"
+  },
+  {
+   "tokens": 16066,
+   "prompt_ms": 13996.4,
+   "pp_tok_s": 1147.9,
+   "found": true,
+   "answer": "falcon-174<|im_end|>"
+  },
+  {
+   "tokens": 27743,
+   "prompt_ms": 24255.2,
+   "pp_tok_s": 1143.8,
+   "found": true,
+   "answer": "willow-474<|im_end|>"
+  }
+ ],
+ "chat": {
+  "turn1_tokens": 6584,
+  "turn1_ms": 6104.9,
+  "turn2_tokens": 6660,
+  "reused": 6632,
+  "turn2_ms": 393.4,
+  "answer": "lantern-417<|im_end|>",
+  "ok": true
+ },
+ "decode": [
+  {
+   "name": "story",
+   "tok_s": 51.9,
+   "reps": [
+    51.9
+   ],
+   "hit_rate": 1.0,
+   "text": "The wind didn\u2019t just blow; it screamed. It tore through the iron gallery of the Blackwood Point Lighthouse, rattling the glass panes like a prisoner shaking its bars.\n\nElias Thorne was seventy years old, and he had been the keeper for forty of them. He knew every creak in the spiral staircase, every draft that slipped under the door, and every shift in the weather that the barometer couldn\u2019t quite predict. He knew that when the fog rolled in thick enough to swallow the horizon, the world ceased to exist. There was only the light, the stone, and him.\n\nIt was November, the month of storms. The sea was a churning cauldron of grey and white, throwing itself against the granite base with a violence that made the floorboards tremble. Elias stood by the lantern room, his hands wrapped around a mug of coffee that had gone cold hours ago. He wasn\u2019t drinking it. He was holding it for warmth."
+  },
+  {
+   "name": "code",
+   "tok_s": 69.7,
+   "reps": [
+    69.7
+   ],
+   "hit_rate": 1.0,
+   "text": "```python\nimport csv\nfrom typing import List, Dict, Any\n\n\ndef parse_csv(file_path: str) -> List[Dict[str, Any]]:\n    \"\"\"\n    Parse a CSV file into a list of dictionaries.\n    \n    Each dictionary represents a row, with keys being the header values\n    and values being the corresponding cell values.\n    \n    Args:\n        file_path: Path to the CSV file.\n        \n    Returns:\n        A list of dictionaries, one per row (excluding the header).\n    \"\"\"\n    result = []\n    with open(file_path, 'r', newline='', encoding='utf-8') as f:\n        reader = csv.DictReader(f)\n        for row in reader:\n            result.append(dict(row))\n    return result\n\n\n# Tests\ndef test_parse_csv_basic():\n    \"\"\"Test basic CSV parsing with simple data.\"\"\"\n    import tempfile\n    import os\n    \n    # Create a temporary CSV"
+  }
+ ],
+ "stage_timing": [
+  "strata serve: stage 2: 165 windows; per window: wait for the GPU 13.452 ms, pool + plan 0.162 ms, host staging 0.000 ms, commit 0.328 ms",
+  "strata serve: stage 0: 234 windows; per window: wait for the GPU 6.235 ms, pool + plan 0.145 ms, host staging 0.704 ms, commit 0.214 ms",
+  "strata serve: stage 1: 234 windows; per window: wait for the GPU 15.592 ms, pool + plan 0.246 ms, host staging 0.000 ms, commit 0.657 ms",
+  "strata serve: stage 2: 234 windows; per window: wait for the GPU 13.916 ms, pool + plan 0.177 ms, host staging 0.000 ms, commit 0.335 ms"
+ ],
+ "engine_lines": [
+  "strata generate: layer split across 3 GPUs: CUDA0, then CUDA1,2 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session; 18.62 GiB free",
+  "strata generate: layer split: CUDA2 PCIe probe 3.2 GB/s -> pcie_frac 0.00",
+  "strata generate: layer split: CUDA2 holds its weights, session and head; 16.18 GiB free",
+  "strata generate: layer split auto: K=12,33 - the caches hold 12288 of 12288 profiled pairs (fullest device 66%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3072 slots",
+  "strata generate: expert cache 3072 slots, 5.41 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 12-32, expert cache 5376 slots (10.22 GiB), 5376 of its 5376 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA2 runs layers 33-47, expert cache 3840 slots (7.79 GiB), 3840 of its 3840 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-11",
+  "strata serve: the prompt path allocates its own buffers (too few cache slots to borrow)",
+  "strata serve: layer split: layers 0-11 (CUDA0), 12-32 (CUDA1), 33-47 (CUDA2), one hand-off per window",
+  "strata generate: layer split across 3 GPUs: CUDA0, then CUDA1,2 (split auto)",
+  "strata generate: layer split: CUDA1 PCIe probe 6.4 GB/s -> pcie_frac 0.14",
+  "strata generate: layer split: CUDA1 holds its weights, session; 18.62 GiB free",
+  "strata generate: layer split: CUDA2 PCIe probe 3.2 GB/s -> pcie_frac 0.00",
+  "strata generate: layer split: CUDA2 holds its weights, session and head; 16.18 GiB free",
+  "strata generate: layer split auto: K=12,33 - the caches hold 12288 of 12288 profiled pairs (fullest device 66%)",
+  "strata generate: expert cache auto: 10.46 GiB free, 700 MiB reserved -> 3072 slots",
+  "strata generate: expert cache 3072 slots, 5.41 GiB of VRAM; policy is",
+  "strata generate: layer split: CUDA1 runs layers 12-32, expert cache 5376 slots (10.22 GiB), 5376 of its 5376 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA2 runs layers 33-47, expert cache 3840 slots (7.79 GiB), 3840 of its 3840 profiled pairs; slot 0 verified",
+  "strata generate: layer split: CUDA0 runs layers 0-11"
+ ]
+}

bench/results/2026-09-29-rtx3090-epyc-milan/README.md

@@ -0,0 +1,46 @@
+# Preliminary native RTX 3090 / EPYC 7453 results
+
+Issue: [#165](https://github.com/Niko1221/Strata/issues/165)
+Engine: Strata v0.1.26, source build at `ac8b251`
+Status: preliminary; prompt set and harness guidance requested upstream.
+
+## Hardware
+
+| Component | Guest-visible details |
+|---|---|
+| CPU | AMD EPYC 7453, 20 vCPU, AVX2; AVX-512 not exposed |
+| RAM | 165 GiB visible to Ubuntu 24.04.5 KVM guest; physical DIMM layout and bandwidth not verified |
+| GPUs | 2x RTX 3090 24 GB; PCIe Gen4 x16 maximum reported per card; driver 595.91.07 |
+| Power limit | 280 W configured on each GPU; defaults are 350 W / 420 W, max-adjustable limits 366 W / 450 W |
+| Storage | 541 GiB ext4 guest data disk; host storage layout not verified |
+
+The power values above are configured caps, not measured workload draw. Idle draw during the hardware capture was approximately 35 W / 29 W.
+
+## Method
+
+Native Linux setup, local build from v0.1.26, `--prefill auto`, shipped expert profile, automatic expert cache, int8 KV, MTP on, images off, experimental speed projection off. The GPU power limit was 280 W per card. Native runs used a systemd service with `LimitMEMLOCK=infinity`; other llama-swap model containers were stopped.
+
+The local API harness used the public Strata source tree as its prompt corpus, with generated tasks, `temperature: 0`, `reasoning_effort: none`, and `max_tokens: 256`. It is not the upstream code-agent prompt set or harness. Runs use unique request IDs and no conversation reuse in speed mode. Three runs per speed cell are summarized by medians; actual `usage.prompt_tokens` are recorded because nominal context labels can drift. The first decode run is also shown. Several requests ended early on EOS, so actual generated counts are in the raw JSON.
+
+No calibration comparison, upstream prompt comparison, `STRATA_SPLIT_TIMING` stage timing, or under-load power-draw capture has been done. Treat all prompt/decode comparisons with upstream as indicative only until the requested upstream method is available.
+
+## Results
+
+See [`matrix.md`](matrix.md) for summary tables and [`matrix.json`](matrix.json) for all successful run records. Per-run source JSON, logs, the local harness, and scrubbed hardware facts are in [`data/`](data/).
+
+Highlights:
+
+- IQ3_XXS on one 3090: 1K–262K cells, three runs each; 262K prompts were 232,183 / 259,589 / 261,068 tokens and fit within the 262,144 context.
+- Coder IQ1_M at 32K: single-GPU median decode 66.8 tok/s; dual-GPU auto (K=19) median 98.6 tok/s.
+- Fixed Coder K sweep: K=22 was the highest observed three-run median (103.3 tok/s); independently generated prompts mean this is not a paired ranking.
+- IQ3_XXS and IQ3_S dual-GPU auto both selected K=24 at 32K.
+- IQ3_XXS and IQ3_S needles passed 3/3 at 8K / 16K / 28K on the tested single/dual configurations; two-turn checkpoint reuse passed for both.
+
+## Limits and follow-up
+
+- Upstream prompts/harness are not yet available; all current speed data uses a local source-tree corpus.
+- The guest exposes one NUMA node and 165 GiB RAM, but physical DIMM topology and memory bandwidth are unknown.
+- PCIe x16 is the reported maximum link width; the engine probes were about 23–26 GB/s. Verify PCIe generation/width under sustained load before claiming a direct x16-vs-x4 comparison.
+- Coder dual-GPU needle/two-turn checks and dual-GPU timing breakdowns are not run yet.
+- K=18 has one successful run at 83.1 tok/s; a repeated startup failed while uploading a dense weight with an out-of-memory error, so it is not a three-run median.
+- Calibration/default comparisons and IQ3_S single-GPU results remain outstanding.

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-1x3090/speed-32k-1.json

@@ -0,0 +1,360 @@
+{
+ "label": "native-v0.1.26-Coder-1x3090",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 1,
+ "run_id": "6c1252339f83",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:11:07",
+ "prompt_tokens": 28317,
+ "completion_tokens": 48,
+ "ttft_s": 12.868,
+ "e2e_s": 13.739,
+ "api_prompt_tps": 2200.5,
+ "api_decode_tps": 54.0,
+ "engine_prompt_tps": 2229.2,
+ "engine_decode_tps": 52.7,
+ "log_lines": [
+  "strata serve: prompt 28317 tokens = 0 reused + 28317 read in 12703 ms (2229.2 tok/s), 48 generated in 911 ms (52.7 tok/s), drafts accepted 31 of 38, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 96.3% (25001 hits / 25956 lookups)"
+ ],
+ "answer_head": "I'll start by exploring the repository structure to understand what's there and identify the last file.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 8973,
+   "expert_cache_mib": 17503,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 369,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790716268.10927,
+    "duration_s": 13.6,
+    "finish": "stop",
+    "prompt_tokens": 28317,
+    "reused": 0,
+    "output_tokens": 48,
+    "engine_generated": 48,
+    "prompt_ms": 12702.6,
+    "decode_ms": 910.6,
+    "decode_tok_s": 52.7,
+    "hit_rate": 0.963
+   }
+  ],
+  "requests_kept": 1,
+  "totals": {
+   "since": 1790716256.7391467,
+   "requests": 1,
+   "prompt_tokens": 28317,
+   "reused": 0,
+   "output_tokens": 48,
+   "prompt_ms": 12702.6,
+   "decode_ms": 910.6
+  },
+  "hardware": {
+   "gpu_util": 88,
+   "gpu_mem_used": 25531711488,
+   "gpu_mem_total": 25769803776,
+   "gpu_temp": 46,
+   "gpu_power": 220.362,
+   "gpu_power_limit": 280.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 0.28515625,
+   "gpu_pcie_tx_mb": 0.4287109375,
+   "cpu": 81.3,
+   "ram_used": 43964985344,
+   "ram_total": 177175990272,
+   "disk_read_mb": 2.9148478336133063,
+   "disk_write_mb": 0.0,
+   "tok_s": 0.0,
+   "tok_s_mean": 0.0
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090",
+   "gpu_count": 1,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    1,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    99,
+    100,
+    100,
+    96,
+    100,
+    99,
+    100,
+    99,
+    99,
+    100,
+    100,
+    88
+   ],
+   "gpu_mem_used": [
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25407979520,
+    25418465280,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25531711488
+   ],
+   "gpu_temp": [
+    43,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    46,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    46
+   ],
+   "gpu_power": [
+    148.8,
+    140.48,
+    132.59,
+    119.03,
+    118.99,
+    119.03,
+    119.04,
+    119.03,
+    119.12,
+    119.17,
+    119.16,
+    119.15,
+    119.26,
+    191.57,
+    276.49,
+    276.88,
+    275.95,
+    277.04,
+    277.35,
+    267.91,
+    275.37,
+    276.16,
+    278.28,
+    279.07,
+    220.36
+   ],
+   "gpu_pcie_rx_mb": [
+    0.33,
+    0.33,
+    0.29,
+    0.29,
+    0.33,
+    0.33,
+    0.33,
+    0.29,
+    0.33,
+    0.29,
+    0.33,
+    0.29,
+    385.14,
+    19.88,
+    1892.23,
+    13280.39,
+    85.45,
+    16838.55,
+    680.06,
+    1060.72,
+    13.45,
+    15.74,
+    5013.75,
+    1194.52,
+    0.29
+   ],
+   "cpu": [
+    0.0,
+    0.7,
+    7.3,
+    15.7,
+    7.1,
+    2.0,
+    1.2,
+    2.0,
+    2.1,
+    1.6,
+    1.9,
+    7.4,
+    13.9,
+    18.3,
+    6.1,
+    6.0,
+    16.7,
+    24.9,
+    13.2,
+    12.0,
+    7.1,
+    6.2,
+    6.0,
+    8.8,
+    81.3
+   ],
+   "ram_used": [
+    43381886976,
+    43381694464,
+    43410796544,
+    43434872832,
+    43483377664,
+    43485900800,
+    43484897280,
+    43481210880,
+    43469135872,
+    43462295552,
+    43458363392,
+    43496140800,
+    43624353792,
+    43648593920,
+    43633405952,
+    43641139200,
+    43635249152,
+    43689267200,
+    43846070272,
+    43924541440,
+    43924467712,
+    43918565376,
+    43883311104,
+    44002877440,
+    43964985344
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.01,
+    0.0,
+    0.0,
+    0.0,
+    25.43,
+    204.93,
+    214.14,
+    7.41,
+    0.22,
+    184.19,
+    135.03,
+    0.0,
+    101.27,
+    0.0,
+    0.0,
+    0.0,
+    0.78,
+    2.91
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "time": 1790716282.234284
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-1x3090/speed-32k-2.json

@@ -0,0 +1,491 @@
+{
+ "label": "native-v0.1.26-Coder-1x3090",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 2,
+ "run_id": "894af0ad69b7",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:11:22",
+ "prompt_tokens": 31630,
+ "completion_tokens": 26,
+ "ttft_s": 12.792,
+ "e2e_s": 13.164,
+ "api_prompt_tps": 2472.7,
+ "api_decode_tps": 67.3,
+ "engine_prompt_tps": 2506.6,
+ "engine_decode_tps": 66.8,
+ "log_lines": [
+  "strata serve: prompt 31630 tokens = 0 reused + 31630 read in 12619 ms (2506.6 tok/s), 26 generated in 389 ms (66.8 tok/s), drafts accepted 12 of 21, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 99.1% (16646 hits / 16791 lookups)"
+ ],
+ "answer_head": "I don't see any file content in the conversation above. Could you please share the file you'd like me to analyze?",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 8973,
+   "expert_cache_mib": 17503,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 369,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790716282.3818104,
+    "duration_s": 13.0,
+    "finish": "stop",
+    "prompt_tokens": 31630,
+    "reused": 0,
+    "output_tokens": 26,
+    "engine_generated": 26,
+    "prompt_ms": 12618.8,
+    "decode_ms": 389.0,
+    "decode_tok_s": 66.8,
+    "hit_rate": 0.991
+   },
+   {
+    "projection": null,
+    "time": 1790716268.10927,
+    "duration_s": 13.6,
+    "finish": "stop",
+    "prompt_tokens": 28317,
+    "reused": 0,
+    "output_tokens": 48,
+    "engine_generated": 48,
+    "prompt_ms": 12702.6,
+    "decode_ms": 910.6,
+    "decode_tok_s": 52.7,
+    "hit_rate": 0.963
+   }
+  ],
+  "requests_kept": 2,
+  "totals": {
+   "since": 1790716256.7391467,
+   "requests": 2,
+   "prompt_tokens": 59947,
+   "reused": 0,
+   "output_tokens": 74,
+   "prompt_ms": 25321.4,
+   "decode_ms": 1299.6
+  },
+  "hardware": {
+   "gpu_util": 81,
+   "gpu_mem_used": 25531711488,
+   "gpu_mem_total": 25769803776,
+   "gpu_temp": 45,
+   "gpu_power": 256.109,
+   "gpu_power_limit": 280.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 9.7744140625,
+   "gpu_pcie_tx_mb": 25.1767578125,
+   "cpu": 46.1,
+   "ram_used": 43872071680,
+   "ram_total": 177175990272,
+   "disk_read_mb": 1.6222807345356267,
+   "disk_write_mb": 0.018689870213544086,
+   "tok_s": 0.0,
+   "tok_s_mean": 0.0
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090",
+   "gpu_count": 1,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    1,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    99,
+    100,
+    100,
+    96,
+    100,
+    99,
+    100,
+    99,
+    99,
+    100,
+    100,
+    88,
+    0,
+    100,
+    99,
+    100,
+    99,
+    100,
+    98,
+    99,
+    99,
+    100,
+    99,
+    100,
+    81
+   ],
+   "gpu_mem_used": [
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25407979520,
+    25418465280,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488
+   ],
+   "gpu_temp": [
+    43,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    46,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    46,
+    44,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    48,
+    47,
+    45
+   ],
+   "gpu_power": [
+    148.8,
+    140.48,
+    132.59,
+    119.03,
+    118.99,
+    119.03,
+    119.04,
+    119.03,
+    119.12,
+    119.17,
+    119.16,
+    119.15,
+    119.26,
+    191.57,
+    276.49,
+    276.88,
+    275.95,
+    277.04,
+    277.35,
+    267.91,
+    275.37,
+    276.16,
+    278.28,
+    279.07,
+    220.36,
+    189.89,
+    258.72,
+    276.75,
+    276.67,
+    277.52,
+    275.34,
+    268.93,
+    277.69,
+    277.77,
+    276.91,
+    278.53,
+    278.56,
+    256.11
+   ],
+   "gpu_pcie_rx_mb": [
+    0.33,
+    0.33,
+    0.29,
+    0.29,
+    0.33,
+    0.33,
+    0.33,
+    0.29,
+    0.33,
+    0.29,
+    0.33,
+    0.29,
+    385.14,
+    19.88,
+    1892.23,
+    13280.39,
+    85.45,
+    16838.55,
+    680.06,
+    1060.72,
+    13.45,
+    15.74,
+    5013.75,
+    1194.52,
+    0.29,
+    2531.19,
+    14.02,
+    3715.09,
+    10.68,
+    13440.37,
+    8568.48,
+    14.92,
+    15.31,
+    15.69,
+    6486.56,
+    17.31,
+    9674.45,
+    9.77
+   ],
+   "cpu": [
+    0.0,
+    0.7,
+    7.3,
+    15.7,
+    7.1,
+    2.0,
+    1.2,
+    2.0,
+    2.1,
+    1.6,
+    1.9,
+    7.4,
+    13.9,
+    18.3,
+    6.1,
+    6.0,
+    16.7,
+    24.9,
+    13.2,
+    12.0,
+    7.1,
+    6.2,
+    6.0,
+    8.8,
+    81.3,
+    4.2,
+    6.8,
+    6.4,
+    7.6,
+    5.5,
+    6.0,
+    20.6,
+    12.9,
+    6.2,
+    7.0,
+    5.7,
+    5.9,
+    46.1
+   ],
+   "ram_used": [
+    43381886976,
+    43381694464,
+    43410796544,
+    43434872832,
+    43483377664,
+    43485900800,
+    43484897280,
+    43481210880,
+    43469135872,
+    43462295552,
+    43458363392,
+    43496140800,
+    43624353792,
+    43648593920,
+    43633405952,
+    43641139200,
+    43635249152,
+    43689267200,
+    43846070272,
+    43924541440,
+    43924467712,
+    43918565376,
+    43883311104,
+    44002877440,
+    43964985344,
+    43959267328,
+    43943862272,
+    43937861632,
+    43930320896,
+    43922284544,
+    43915001856,
+    43817238528,
+    43819147264,
+    43818729472,
+    43828027392,
+    43824201728,
+    43825434624,
+    43872071680
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.01,
+    0.0,
+    0.0,
+    0.0,
+    25.43,
+    204.93,
+    214.14,
+    7.41,
+    0.22,
+    184.19,
+    135.03,
+    0.0,
+    101.27,
+    0.0,
+    0.0,
+    0.0,
+    0.78,
+    2.91,
+    0.61,
+    0.0,
+    0.0,
+    0.02,
+    0.0,
+    0.0,
+    79.64,
+    41.92,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    1.62
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "time": 1790716295.9009006
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-1x3090/speed-32k-3.json

@@ -0,0 +1,622 @@
+{
+ "label": "native-v0.1.26-Coder-1x3090",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 3,
+ "run_id": "a4c63bd599ab",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:11:35",
+ "prompt_tokens": 31601,
+ "completion_tokens": 45,
+ "ttft_s": 12.796,
+ "e2e_s": 13.243,
+ "api_prompt_tps": 2469.6,
+ "api_decode_tps": 98.5,
+ "engine_prompt_tps": 2502.4,
+ "engine_decode_tps": 97.4,
+ "log_lines": [
+  "strata serve: prompt 31601 tokens = 0 reused + 31601 read in 12628 ms (2502.4 tok/s), 45 generated in 462 ms (97.4 tok/s), drafts accepted 33 of 37, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 98.6% (23985 hits / 24333 lookups)"
+ ],
+ "answer_head": "I'll start by exploring the repository structure to understand what we're working with.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 8973,
+   "expert_cache_mib": 17503,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 369,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790716296.0458817,
+    "duration_s": 13.1,
+    "finish": "stop",
+    "prompt_tokens": 31601,
+    "reused": 0,
+    "output_tokens": 45,
+    "engine_generated": 45,
+    "prompt_ms": 12628.3,
+    "decode_ms": 461.9,
+    "decode_tok_s": 97.4,
+    "hit_rate": 0.986
+   },
+   {
+    "projection": null,
+    "time": 1790716282.3818104,
+    "duration_s": 13.0,
+    "finish": "stop",
+    "prompt_tokens": 31630,
+    "reused": 0,
+    "output_tokens": 26,
+    "engine_generated": 26,
+    "prompt_ms": 12618.8,
+    "decode_ms": 389.0,
+    "decode_tok_s": 66.8,
+    "hit_rate": 0.991
+   },
+   {
+    "projection": null,
+    "time": 1790716268.10927,
+    "duration_s": 13.6,
+    "finish": "stop",
+    "prompt_tokens": 28317,
+    "reused": 0,
+    "output_tokens": 48,
+    "engine_generated": 48,
+    "prompt_ms": 12702.6,
+    "decode_ms": 910.6,
+    "decode_tok_s": 52.7,
+    "hit_rate": 0.963
+   }
+  ],
+  "requests_kept": 3,
+  "totals": {
+   "since": 1790716256.7391467,
+   "requests": 3,
+   "prompt_tokens": 91548,
+   "reused": 0,
+   "output_tokens": 119,
+   "prompt_ms": 37949.7,
+   "decode_ms": 1761.5
+  },
+  "hardware": {
+   "gpu_util": 96,
+   "gpu_mem_used": 25531711488,
+   "gpu_mem_total": 25769803776,
+   "gpu_temp": 47,
+   "gpu_power": 245.399,
+   "gpu_power_limit": 280.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 414.037109375,
+   "gpu_pcie_tx_mb": 303.36328125,
+   "cpu": 35.1,
+   "ram_used": 43772702720,
+   "ram_total": 177175990272,
+   "disk_read_mb": 0.6087260569540994,
+   "disk_write_mb": 0.20539836277592313,
+   "tok_s": 97.29124144311484,
+   "tok_s_mean": 96.62332449971758
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090",
+   "gpu_count": 1,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    1,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    0,
+    99,
+    100,
+    100,
+    96,
+    100,
+    99,
+    100,
+    99,
+    99,
+    100,
+    100,
+    88,
+    0,
+    100,
+    99,
+    100,
+    99,
+    100,
+    98,
+    99,
+    99,
+    100,
+    99,
+    100,
+    81,
+    99,
+    100,
+    100,
+    99,
+    100,
+    99,
+    17,
+    99,
+    100,
+    99,
+    99,
+    99,
+    96
+   ],
+   "gpu_mem_used": [
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25382813696,
+    25407979520,
+    25418465280,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25424756736,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488,
+    25531711488
+   ],
+   "gpu_temp": [
+    43,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    42,
+    46,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    46,
+    44,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    47,
+    48,
+    47,
+    45,
+    47,
+    47,
+    47,
+    48,
+    48,
+    48,
+    46,
+    47,
+    48,
+    48,
+    48,
+    48,
+    47
+   ],
+   "gpu_power": [
+    148.8,
+    140.48,
+    132.59,
+    119.03,
+    118.99,
+    119.03,
+    119.04,
+    119.03,
+    119.12,
+    119.17,
+    119.16,
+    119.15,
+    119.26,
+    191.57,
+    276.49,
+    276.88,
+    275.95,
+    277.04,
+    277.35,
+    267.91,
+    275.37,
+    276.16,
+    278.28,
+    279.07,
+    220.36,
+    189.89,
+    258.72,
+    276.75,
+    276.67,
+    277.52,
+    275.34,
+    268.93,
+    277.69,
+    277.77,
+    276.91,
+    278.53,
+    278.56,
+    256.11,
+    199.6,
+    275.35,
+    275.49,
+    276.31,
+    275.46,
+    277.78,
+    266.35,
+    267.73,
+    276.71,
+    276.41,
+    275.34,
+    278.17,
+    245.4
+   ],
+   "gpu_pcie_rx_mb": [
+    0.33,
+    0.33,
+    0.29,
+    0.29,
+    0.33,
+    0.33,
+    0.33,
+    0.29,
+    0.33,
+    0.29,
+    0.33,
+    0.29,
+    385.14,
+    19.88,
+    1892.23,
+    13280.39,
+    85.45,
+    16838.55,
+    680.06,
+    1060.72,
+    13.45,
+    15.74,
+    5013.75,
+    1194.52,
+    0.29,
+    2531.19,
+    14.02,
+    3715.09,
+    10.68,
+    13440.37,
+    8568.48,
+    14.92,
+    15.31,
+    15.69,
+    6486.56,
+    17.31,
+    9674.45,
+    9.77,
+    2795.22,
+    4473.69,
+    14.5,
+    12.97,
+    14.92,
+    7716.66,
+    2819.78,
+    4278.33,
+    12009.48,
+    17.07,
+    9916.78,
+    8518.17,
+    414.04
+   ],
+   "cpu": [
+    0.0,
+    0.7,
+    7.3,
+    15.7,
+    7.1,
+    2.0,
+    1.2,
+    2.0,
+    2.1,
+    1.6,
+    1.9,
+    7.4,
+    13.9,
+    18.3,
+    6.1,
+    6.0,
+    16.7,
+    24.9,
+    13.2,
+    12.0,
+    7.1,
+    6.2,
+    6.0,
+    8.8,
+    81.3,
+    4.2,
+    6.8,
+    6.4,
+    7.6,
+    5.5,
+    6.0,
+    20.6,
+    12.9,
+    6.2,
+    7.0,
+    5.7,
+    5.9,
+    46.1,
+    4.0,
+    6.9,
+    5.7,
+    6.2,
+    7.1,
+    5.5,
+    5.6,
+    9.9,
+    13.7,
+    8.4,
+    7.7,
+    6.4,
+    35.1
+   ],
+   "ram_used": [
+    43381886976,
+    43381694464,
+    43410796544,
+    43434872832,
+    43483377664,
+    43485900800,
+    43484897280,
+    43481210880,
+    43469135872,
+    43462295552,
+    43458363392,
+    43496140800,
+    43624353792,
+    43648593920,
+    43633405952,
+    43641139200,
+    43635249152,
+    43689267200,
+    43846070272,
+    43924541440,
+    43924467712,
+    43918565376,
+    43883311104,
+    44002877440,
+    43964985344,
+    43959267328,
+    43943862272,
+    43937861632,
+    43930320896,
+    43922284544,
+    43915001856,
+    43817238528,
+    43819147264,
+    43818729472,
+    43828027392,
+    43824201728,
+    43825434624,
+    43872071680,
+    43734614016,
+    43728027648,
+    43719413760,
+    43703148544,
+    43695366144,
+    43687342080,
+    43744829440,
+    43791024128,
+    43789787136,
+    43766198272,
+    43761831936,
+    43750072320,
+    43772702720
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.01,
+    0.0,
+    0.0,
+    0.0,
+    25.43,
+    204.93,
+    214.14,
+    7.41,
+    0.22,
+    184.19,
+    135.03,
+    0.0,
+    101.27,
+    0.0,
+    0.0,
+    0.0,
+    0.78,
+    2.91,
+    0.61,
+    0.0,
+    0.0,
+    0.02,
+    0.0,
+    0.0,
+    79.64,
+    41.92,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    1.62,
+    0.72,
+    0.0,
+    0.0,
+    0.05,
+    0.0,
+    0.0,
+    0.18,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.61
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    97.29
+   ]
+  },
+  "time": 1790716309.6464093
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K18/speed-32k-1.json

@@ -0,0 +1,423 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K18",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 1,
+ "run_id": "052cbcbb7af7",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:21:48",
+ "prompt_tokens": 28313,
+ "completion_tokens": 57,
+ "ttft_s": 12.346,
+ "e2e_s": 12.991,
+ "api_prompt_tps": 2293.3,
+ "api_decode_tps": 86.9,
+ "engine_prompt_tps": 2324.6,
+ "engine_decode_tps": 83.1,
+ "log_lines": [
+  "strata serve: prompt 28313 tokens = 0 reused + 28313 read in 12180 ms (2324.6 tok/s), 57 generated in 686 ms (83.1 tok/s), drafts accepted 39 of 41, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 99.7% (28717 hits / 28800 lookups)"
+ ],
+ "answer_head": "I'll start by exploring the repository structure to understand what's there and identify the last file.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 4608,
+   "expert_cache_mib": 8437,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 9557,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790716909.0337873,
+    "duration_s": 12.9,
+    "finish": "stop",
+    "prompt_tokens": 28313,
+    "reused": 0,
+    "output_tokens": 57,
+    "engine_generated": 57,
+    "prompt_ms": 12179.9,
+    "decode_ms": 685.7,
+    "decode_tok_s": 83.1,
+    "hit_rate": 0.997
+   }
+  ],
+  "requests_kept": 1,
+  "totals": {
+   "since": 1790716889.8200681,
+   "requests": 1,
+   "prompt_tokens": 28313,
+   "reused": 0,
+   "output_tokens": 57,
+   "prompt_ms": 12179.9,
+   "decode_ms": 685.7
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 0,
+     "mem_used": 15823994880,
+     "mem_total": 25769803776,
+     "temp": 44,
+     "power": 161.267
+    },
+    {
+     "index": 1,
+     "util": 90,
+     "mem_used": 24208408576,
+     "mem_total": 25769803776,
+     "temp": 51,
+     "power": 278.413
+    }
+   ],
+   "gpu_util": 45.0,
+   "gpu_mem_used": 40032403456,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 51,
+   "gpu_power": 439.68,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 5.33984375,
+   "gpu_pcie_tx_mb": 18.5009765625,
+   "cpu": 16.1,
+   "ram_used": 44574941184,
+   "ram_total": 177175990272,
+   "disk_read_mb": 0.9737281407762519,
+   "disk_write_mb": 0.11497896865254635,
+   "tok_s": 36.0,
+   "tok_s_mean": 64.01736582189179
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    0.5,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    41.5,
+    68.0,
+    50.0,
+    73.0,
+    87.5,
+    41.5,
+    47.0,
+    50.0,
+    78.0,
+    50.5,
+    45.0
+   ],
+   "gpu_mem_used": [
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39866728448,
+    39965294592,
+    39965294592,
+    39965294592,
+    39965294592,
+    39965294592,
+    39965294592,
+    39965294592,
+    39965294592,
+    39965294592,
+    39965294592,
+    40032403456
+   ],
+   "gpu_temp": [
+    45,
+    44,
+    44,
+    44,
+    44,
+    44,
+    45,
+    44,
+    44,
+    45,
+    45,
+    45,
+    45,
+    45,
+    45,
+    45,
+    42,
+    41,
+    41,
+    47,
+    50,
+    51,
+    51,
+    51,
+    52,
+    51,
+    52,
+    51,
+    52,
+    51
+   ],
+   "gpu_power": [
+    255.0,
+    240.14,
+    239.41,
+    239.38,
+    239.55,
+    239.75,
+    239.74,
+    239.65,
+    239.56,
+    239.63,
+    239.71,
+    239.58,
+    239.49,
+    239.62,
+    239.83,
+    239.9,
+    200.58,
+    138.26,
+    69.98,
+    230.69,
+    516.11,
+    485.77,
+    500.33,
+    518.82,
+    491.98,
+    467.22,
+    490.75,
+    506.65,
+    510.68,
+    439.68
+   ],
+   "gpu_pcie_rx_mb": [
+    0.71,
+    0.62,
+    0.67,
+    0.67,
+    0.67,
+    0.57,
+    0.71,
+    0.62,
+    0.62,
+    0.57,
+    0.67,
+    0.62,
+    0.71,
+    0.67,
+    0.67,
+    0.62,
+    0.76,
+    0.86,
+    0.71,
+    1243.59,
+    45.3,
+    2454.38,
+    77.06,
+    3382.78,
+    1099.87,
+    47.06,
+    2677.87,
+    79.25,
+    1143.26,
+    5.34
+   ],
+   "cpu": [
+    0.0,
+    1.0,
+    2.8,
+    3.4,
+    1.0,
+    1.6,
+    1.6,
+    1.0,
+    1.4,
+    0.7,
+    3.6,
+    8.3,
+    1.0,
+    8.2,
+    5.7,
+    2.6,
+    3.4,
+    1.4,
+    10.1,
+    12.9,
+    8.2,
+    10.3,
+    10.7,
+    11.6,
+    27.0,
+    14.9,
+    11.2,
+    18.3,
+    15.0,
+    16.1
+   ],
+   "ram_used": [
+    43962929152,
+    43965493248,
+    43981484032,
+    43968327680,
+    43968602112,
+    43968954368,
+    43964538880,
+    43964022784,
+    43964801024,
+    43964657664,
+    43961843712,
+    43958075392,
+    43958878208,
+    43967004672,
+    43967807488,
+    43951247360,
+    43945017344,
+    43945320448,
+    43971616768,
+    44103409664,
+    44112924672,
+    44112818176,
+    43979321344,
+    43978510336,
+    44145692672,
+    44276580352,
+    44275351552,
+    44327723008,
+    44296556544,
+    44574941184
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    64.21,
+    105.78,
+    24.45,
+    89.11,
+    67.32,
+    92.34,
+    145.4,
+    79.43,
+    98.92,
+    75.23,
+    0.0,
+    0.97
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    36.0
+   ]
+  },
+  "time": 1790716922.4098496
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K22/speed-32k-1.json

@@ -0,0 +1,270 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K22",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 1,
+ "run_id": "69d919ce1297",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:32:53",
+ "prompt_tokens": 28316,
+ "completion_tokens": 57,
+ "ttft_s": 10.575,
+ "e2e_s": 11.256,
+ "api_prompt_tps": 2677.7,
+ "api_decode_tps": 82.2,
+ "engine_prompt_tps": 2718.9,
+ "engine_decode_tps": 79.8,
+ "log_lines": [
+  "strata serve: prompt 28316 tokens = 0 reused + 28316 read in 10414 ms (2718.9 tok/s), 57 generated in 715 ms (79.8 tok/s), drafts accepted 37 of 48, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 100.0% (34080 hits / 34080 lookups)"
+ ],
+ "answer_head": "I'll start by exploring the repository structure to understand what's there and identify the last file.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 5632,
+   "expert_cache_mib": 10325,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 7669,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790717573.1737459,
+    "duration_s": 11.1,
+    "finish": "stop",
+    "prompt_tokens": 28316,
+    "reused": 0,
+    "output_tokens": 57,
+    "engine_generated": 57,
+    "prompt_ms": 10414.4,
+    "decode_ms": 714.5,
+    "decode_tok_s": 79.8,
+    "hit_rate": 1.0
+   }
+  ],
+  "requests_kept": 1,
+  "totals": {
+   "since": 1790717571.262619,
+   "requests": 1,
+   "prompt_tokens": 28316,
+   "reused": 0,
+   "output_tokens": 57,
+   "prompt_ms": 10414.4,
+   "decode_ms": 714.5
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 34,
+     "mem_used": 17822580736,
+     "mem_total": 25769803776,
+     "temp": 49,
+     "power": 184.581
+    },
+    {
+     "index": 1,
+     "util": 54,
+     "mem_used": 22828482560,
+     "mem_total": 25769803776,
+     "temp": 52,
+     "power": 232.993
+    }
+   ],
+   "gpu_util": 44.0,
+   "gpu_mem_used": 40651063296,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 52,
+   "gpu_power": 417.57399999999996,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 0.666015625,
+   "gpu_pcie_tx_mb": 0.953125,
+   "cpu": 6.2,
+   "ram_used": 44281556992,
+   "ram_total": 177175990272,
+   "disk_read_mb": 3.421090719626586,
+   "disk_write_mb": 0.0,
+   "tok_s": 0.0,
+   "tok_s_mean": 0.0
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    1.0,
+    0.0,
+    0.0,
+    73.5,
+    58.5,
+    96.0,
+    57.5,
+    95.5,
+    47.5,
+    93.0,
+    58.5,
+    50.0,
+    44.0
+   ],
+   "gpu_mem_used": [
+    40449736704,
+    40449736704,
+    40449736704,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40651063296
+   ],
+   "gpu_temp": [
+    48,
+    47,
+    47,
+    50,
+    53,
+    53,
+    54,
+    53,
+    54,
+    54,
+    54,
+    54,
+    52
+   ],
+   "gpu_power": [
+    256.98,
+    244.54,
+    244.03,
+    340.49,
+    532.78,
+    528.1,
+    525.31,
+    536.01,
+    504.78,
+    526.85,
+    529.86,
+    484.8,
+    417.57
+   ],
+   "gpu_pcie_rx_mb": [
+    0.67,
+    0.62,
+    0.67,
+    78.72,
+    82.54,
+    76.1,
+    85.97,
+    55.17,
+    62.23,
+    87.12,
+    83.54,
+    85.83,
+    0.67
+   ],
+   "cpu": [
+    0.0,
+    1.2,
+    5.4,
+    10.7,
+    10.0,
+    9.8,
+    14.0,
+    18.3,
+    17.8,
+    18.6,
+    9.0,
+    5.8,
+    6.2
+   ],
+   "ram_used": [
+    43564908544,
+    43556737024,
+    43564126208,
+    43739512832,
+    43734302720,
+    43734560768,
+    43735560192,
+    43898122240,
+    44013469696,
+    43991674880,
+    43978436608,
+    43974189056,
+    44281556992
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    73.94,
+    96.25,
+    39.8,
+    73.77,
+    144.08,
+    110.85,
+    127.27,
+    131.68,
+    42.81,
+    0.0,
+    3.42
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "time": 1790717584.813652
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K22/speed-32k-2.json

@@ -0,0 +1,383 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K22",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 2,
+ "run_id": "8506420b91e1",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:33:04",
+ "prompt_tokens": 31633,
+ "completion_tokens": 45,
+ "ttft_s": 10.85,
+ "e2e_s": 11.271,
+ "api_prompt_tps": 2915.4,
+ "api_decode_tps": 104.6,
+ "engine_prompt_tps": 2962.7,
+ "engine_decode_tps": 103.3,
+ "log_lines": [
+  "strata serve: prompt 31633 tokens = 0 reused + 31633 read in 10677 ms (2962.7 tok/s), 45 generated in 436 ms (103.3 tok/s), drafts accepted 32 of 37, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 100.0% (24960 hits / 24960 lookups)"
+ ],
+ "answer_head": "I'll start by exploring the repository structure to understand what the last file does.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 5632,
+   "expert_cache_mib": 10325,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 7669,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790717584.9637764,
+    "duration_s": 11.1,
+    "finish": "stop",
+    "prompt_tokens": 31633,
+    "reused": 0,
+    "output_tokens": 45,
+    "engine_generated": 45,
+    "prompt_ms": 10677.2,
+    "decode_ms": 435.6,
+    "decode_tok_s": 103.3,
+    "hit_rate": 1.0
+   },
+   {
+    "projection": null,
+    "time": 1790717573.1737459,
+    "duration_s": 11.1,
+    "finish": "stop",
+    "prompt_tokens": 28316,
+    "reused": 0,
+    "output_tokens": 57,
+    "engine_generated": 57,
+    "prompt_ms": 10414.4,
+    "decode_ms": 714.5,
+    "decode_tok_s": 79.8,
+    "hit_rate": 1.0
+   }
+  ],
+  "requests_kept": 2,
+  "totals": {
+   "since": 1790717571.262619,
+   "requests": 2,
+   "prompt_tokens": 59949,
+   "reused": 0,
+   "output_tokens": 102,
+   "prompt_ms": 21091.6,
+   "decode_ms": 1150.1
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 37,
+     "mem_used": 17822580736,
+     "mem_total": 25769803776,
+     "temp": 49,
+     "power": 187.064
+    },
+    {
+     "index": 1,
+     "util": 60,
+     "mem_used": 22828482560,
+     "mem_total": 25769803776,
+     "temp": 53,
+     "power": 243.674
+    }
+   ],
+   "gpu_util": 48.5,
+   "gpu_mem_used": 40651063296,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 53,
+   "gpu_power": 430.738,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 0.6181640625,
+   "gpu_pcie_tx_mb": 1.0,
+   "cpu": 6.1,
+   "ram_used": 44308209664,
+   "ram_total": 177175990272,
+   "disk_read_mb": 0.7959400039611647,
+   "disk_write_mb": 0.24130307812397303,
+   "tok_s": 0.0,
+   "tok_s_mean": 0.0
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    1.0,
+    0.0,
+    0.0,
+    73.5,
+    58.5,
+    96.0,
+    57.5,
+    95.5,
+    47.5,
+    93.0,
+    58.5,
+    50.0,
+    44.0,
+    18.0,
+    52.0,
+    95.5,
+    50.0,
+    97.0,
+    24.5,
+    95.5,
+    82.0,
+    96.5,
+    50.0,
+    48.5
+   ],
+   "gpu_mem_used": [
+    40449736704,
+    40449736704,
+    40449736704,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296
+   ],
+   "gpu_temp": [
+    48,
+    47,
+    47,
+    50,
+    53,
+    53,
+    54,
+    53,
+    54,
+    54,
+    54,
+    54,
+    52,
+    49,
+    54,
+    54,
+    54,
+    54,
+    52,
+    54,
+    54,
+    54,
+    54,
+    53
+   ],
+   "gpu_power": [
+    256.98,
+    244.54,
+    244.03,
+    340.49,
+    532.78,
+    528.1,
+    525.31,
+    536.01,
+    504.78,
+    526.85,
+    529.86,
+    484.8,
+    417.57,
+    353.2,
+    539.11,
+    526.73,
+    525.83,
+    531.55,
+    507.76,
+    523.24,
+    527.24,
+    526.07,
+    501.32,
+    430.74
+   ],
+   "gpu_pcie_rx_mb": [
+    0.67,
+    0.62,
+    0.67,
+    78.72,
+    82.54,
+    76.1,
+    85.97,
+    55.17,
+    62.23,
+    87.12,
+    83.54,
+    85.83,
+    0.67,
+    1765.25,
+    83.88,
+    85.4,
+    85.07,
+    88.59,
+    82.35,
+    78.3,
+    85.16,
+    61.94,
+    46.3,
+    0.62
+   ],
+   "cpu": [
+    0.0,
+    1.2,
+    5.4,
+    10.7,
+    10.0,
+    9.8,
+    14.0,
+    18.3,
+    17.8,
+    18.6,
+    9.0,
+    5.8,
+    6.2,
+    4.4,
+    6.6,
+    6.4,
+    5.8,
+    6.1,
+    7.4,
+    5.9,
+    6.8,
+    19.8,
+    14.2,
+    6.1
+   ],
+   "ram_used": [
+    43564908544,
+    43556737024,
+    43564126208,
+    43739512832,
+    43734302720,
+    43734560768,
+    43735560192,
+    43898122240,
+    44013469696,
+    43991674880,
+    43978436608,
+    43974189056,
+    44281556992,
+    44278607872,
+    44263776256,
+    44251099136,
+    44242841600,
+    44235354112,
+    44232339456,
+    44218642432,
+    44210057216,
+    44216344576,
+    44288196608,
+    44308209664
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    73.94,
+    96.25,
+    39.8,
+    73.77,
+    144.08,
+    110.85,
+    127.27,
+    131.68,
+    42.81,
+    0.0,
+    3.42,
+    0.48,
+    0.01,
+    0.04,
+    0.06,
+    0.02,
+    0.02,
+    0.02,
+    12.53,
+    104.78,
+    0.0,
+    0.8
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "time": 1790717596.5877995
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K22/speed-32k-3.json

@@ -0,0 +1,496 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K22",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 3,
+ "run_id": "831d87a1dacc",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:33:16",
+ "prompt_tokens": 31601,
+ "completion_tokens": 60,
+ "ttft_s": 10.877,
+ "e2e_s": 11.43,
+ "api_prompt_tps": 2905.4,
+ "api_decode_tps": 106.7,
+ "engine_prompt_tps": 2951.6,
+ "engine_decode_tps": 105.7,
+ "log_lines": [
+  "strata serve: prompt 31601 tokens = 0 reused + 31601 read in 10706 ms (2951.6 tok/s), 60 generated in 568 ms (105.7 tok/s), drafts accepted 41 of 46, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 100.0% (31680 hits / 31680 lookups)"
+ ],
+ "answer_head": "I'll start by examining the repository structure to understand what the last file does.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 5632,
+   "expert_cache_mib": 10325,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 7669,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790717596.735188,
+    "duration_s": 11.3,
+    "finish": "stop",
+    "prompt_tokens": 31601,
+    "reused": 0,
+    "output_tokens": 60,
+    "engine_generated": 60,
+    "prompt_ms": 10706.3,
+    "decode_ms": 567.6,
+    "decode_tok_s": 105.7,
+    "hit_rate": 1.0
+   },
+   {
+    "projection": null,
+    "time": 1790717584.9637764,
+    "duration_s": 11.1,
+    "finish": "stop",
+    "prompt_tokens": 31633,
+    "reused": 0,
+    "output_tokens": 45,
+    "engine_generated": 45,
+    "prompt_ms": 10677.2,
+    "decode_ms": 435.6,
+    "decode_tok_s": 103.3,
+    "hit_rate": 1.0
+   },
+   {
+    "projection": null,
+    "time": 1790717573.1737459,
+    "duration_s": 11.1,
+    "finish": "stop",
+    "prompt_tokens": 28316,
+    "reused": 0,
+    "output_tokens": 57,
+    "engine_generated": 57,
+    "prompt_ms": 10414.4,
+    "decode_ms": 714.5,
+    "decode_tok_s": 79.8,
+    "hit_rate": 1.0
+   }
+  ],
+  "requests_kept": 3,
+  "totals": {
+   "since": 1790717571.262619,
+   "requests": 3,
+   "prompt_tokens": 91550,
+   "reused": 0,
+   "output_tokens": 162,
+   "prompt_ms": 31797.899999999998,
+   "decode_ms": 1717.6999999999998
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 25,
+     "mem_used": 17822580736,
+     "mem_total": 25769803776,
+     "temp": 50,
+     "power": 201.214
+    },
+    {
+     "index": 1,
+     "util": 61,
+     "mem_used": 22828482560,
+     "mem_total": 25769803776,
+     "temp": 53,
+     "power": 238.993
+    }
+   ],
+   "gpu_util": 43.0,
+   "gpu_mem_used": 40651063296,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 53,
+   "gpu_power": 440.207,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 0.666015625,
+   "gpu_pcie_tx_mb": 0.9052734375,
+   "cpu": 4.4,
+   "ram_used": 44291919872,
+   "ram_total": 177175990272,
+   "disk_read_mb": 0.7281688058408962,
+   "disk_write_mb": 0.01802398034259644,
+   "tok_s": 0.0,
+   "tok_s_mean": 0.0
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    1.0,
+    0.0,
+    0.0,
+    73.5,
+    58.5,
+    96.0,
+    57.5,
+    95.5,
+    47.5,
+    93.0,
+    58.5,
+    50.0,
+    44.0,
+    18.0,
+    52.0,
+    95.5,
+    50.0,
+    97.0,
+    24.5,
+    95.5,
+    82.0,
+    96.5,
+    50.0,
+    48.5,
+    49.5,
+    96.0,
+    48.0,
+    96.0,
+    47.5,
+    94.5,
+    61.0,
+    97.5,
+    47.5,
+    49.5,
+    43.0
+   ],
+   "gpu_mem_used": [
+    40449736704,
+    40449736704,
+    40449736704,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296,
+    40651063296
+   ],
+   "gpu_temp": [
+    48,
+    47,
+    47,
+    50,
+    53,
+    53,
+    54,
+    53,
+    54,
+    54,
+    54,
+    54,
+    52,
+    49,
+    54,
+    54,
+    54,
+    54,
+    52,
+    54,
+    54,
+    54,
+    54,
+    53,
+    50,
+    54,
+    54,
+    54,
+    55,
+    54,
+    54,
+    54,
+    55,
+    55,
+    53
+   ],
+   "gpu_power": [
+    256.98,
+    244.54,
+    244.03,
+    340.49,
+    532.78,
+    528.1,
+    525.31,
+    536.01,
+    504.78,
+    526.85,
+    529.86,
+    484.8,
+    417.57,
+    353.2,
+    539.11,
+    526.73,
+    525.83,
+    531.55,
+    507.76,
+    523.24,
+    527.24,
+    526.07,
+    501.32,
+    430.74,
+    359.38,
+    510.67,
+    527.45,
+    526.85,
+    532.56,
+    501.47,
+    534.58,
+    526.16,
+    522.92,
+    475.95,
+    440.21
+   ],
+   "gpu_pcie_rx_mb": [
+    0.67,
+    0.62,
+    0.67,
+    78.72,
+    82.54,
+    76.1,
+    85.97,
+    55.17,
+    62.23,
+    87.12,
+    83.54,
+    85.83,
+    0.67,
+    1765.25,
+    83.88,
+    85.4,
+    85.07,
+    88.59,
+    82.35,
+    78.3,
+    85.16,
+    61.94,
+    46.3,
+    0.62,
+    87.45,
+    41.53,
+    89.07,
+    84.21,
+    76.25,
+    52.17,
+    78.29,
+    83.59,
+    51.45,
+    47.92,
+    0.67
+   ],
+   "cpu": [
+    0.0,
+    1.2,
+    5.4,
+    10.7,
+    10.0,
+    9.8,
+    14.0,
+    18.3,
+    17.8,
+    18.6,
+    9.0,
+    5.8,
+    6.2,
+    4.4,
+    6.6,
+    6.4,
+    5.8,
+    6.1,
+    7.4,
+    5.9,
+    6.8,
+    19.8,
+    14.2,
+    6.1,
+    7.3,
+    6.6,
+    6.3,
+    5.7,
+    6.0,
+    6.2,
+    6.3,
+    6.1,
+    7.3,
+    6.4,
+    4.4
+   ],
+   "ram_used": [
+    43564908544,
+    43556737024,
+    43564126208,
+    43739512832,
+    43734302720,
+    43734560768,
+    43735560192,
+    43898122240,
+    44013469696,
+    43991674880,
+    43978436608,
+    43974189056,
+    44281556992,
+    44278607872,
+    44263776256,
+    44251099136,
+    44242841600,
+    44235354112,
+    44232339456,
+    44218642432,
+    44210057216,
+    44216344576,
+    44288196608,
+    44308209664,
+    44011671552,
+    43989819392,
+    43978940416,
+    43970682880,
+    43962933248,
+    44045533184,
+    44039798784,
+    44039798784,
+    44047568896,
+    44044705792,
+    44291919872
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    73.94,
+    96.25,
+    39.8,
+    73.77,
+    144.08,
+    110.85,
+    127.27,
+    131.68,
+    42.81,
+    0.0,
+    3.42,
+    0.48,
+    0.01,
+    0.04,
+    0.06,
+    0.02,
+    0.02,
+    0.02,
+    12.53,
+    104.78,
+    0.0,
+    0.8,
+    0.61,
+    0.05,
+    0.03,
+    0.09,
+    0.08,
+    0.04,
+    0.05,
+    0.04,
+    0.15,
+    0.0,
+    0.73
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "time": 1790717608.5200934
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K24/speed-32k-1.json

@@ -0,0 +1,279 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K24",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 1,
+ "run_id": "87af67edfd2d",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:34:02",
+ "prompt_tokens": 28314,
+ "completion_tokens": 256,
+ "ttft_s": 10.029,
+ "e2e_s": 12.977,
+ "api_prompt_tps": 2823.1,
+ "api_decode_tps": 86.5,
+ "engine_prompt_tps": 2871.1,
+ "engine_decode_tps": 85.9,
+ "log_lines": [
+  "strata serve: prompt 28314 tokens = 0 reused + 28314 read in 9862 ms (2871.1 tok/s), 256 generated in 2981 ms (85.9 tok/s), drafts accepted 143 of 193, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 100.0% (147360 hits / 147360 lookups)"
+ ],
+ "answer_head": "Looking at the last file above, `docs/AMD_HIP_PERFORMANCE.md`, it documents the performance evidence and correctness limits for the experimental AMD HIP backend targeting gfx1100 (RX 7900 XTX). It records:\n\n- The exact build/runtime configuration used for the benchmark (env vars, CLI flags, model pa",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 6144,
+   "expert_cache_mib": 11337,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 6657,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790717642.434618,
+    "duration_s": 12.9,
+    "finish": "length",
+    "prompt_tokens": 28314,
+    "reused": 0,
+    "output_tokens": 256,
+    "engine_generated": 256,
+    "prompt_ms": 9861.8,
+    "decode_ms": 2981.5,
+    "decode_tok_s": 85.9,
+    "hit_rate": 1.0
+   }
+  ],
+  "requests_kept": 1,
+  "totals": {
+   "since": 1790717640.7776525,
+   "requests": 1,
+   "prompt_tokens": 28314,
+   "reused": 0,
+   "output_tokens": 256,
+   "prompt_ms": 9861.8,
+   "decode_ms": 2981.5
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 43,
+     "mem_used": 18898419712,
+     "mem_total": 25769803776,
+     "temp": 50,
+     "power": 244.874
+    },
+    {
+     "index": 1,
+     "util": 57,
+     "mem_used": 21758935040,
+     "mem_total": 25769803776,
+     "temp": 54,
+     "power": 254.71
+    }
+   ],
+   "gpu_util": 50.0,
+   "gpu_mem_used": 40657354752,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 54,
+   "gpu_power": 499.584,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 5.38671875,
+   "gpu_pcie_tx_mb": 87.7373046875,
+   "cpu": 11.7,
+   "ram_used": 44619116544,
+   "ram_total": 177175990272,
+   "disk_read_mb": 4.515214018548113,
+   "disk_write_mb": 0.17657261525048487,
+   "tok_s": 107.09643367230112,
+   "tok_s_mean": 83.26979350665239
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    1.0,
+    0.0,
+    1.0,
+    93.0,
+    94.0,
+    64.0,
+    94.5,
+    88.5,
+    95.5,
+    92.5,
+    48.5,
+    32.0,
+    48.0,
+    50.0
+   ],
+   "gpu_mem_used": [
+    40447639552,
+    40447639552,
+    40476999680,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40657354752,
+    40657354752,
+    40657354752
+   ],
+   "gpu_temp": [
+    49,
+    49,
+    49,
+    53,
+    54,
+    54,
+    55,
+    54,
+    55,
+    55,
+    55,
+    52,
+    54,
+    54
+   ],
+   "gpu_power": [
+    259.59,
+    247.18,
+    245.88,
+    414.53,
+    546.71,
+    547.76,
+    546.95,
+    533.57,
+    531.41,
+    545.46,
+    539.14,
+    402.04,
+    486.67,
+    499.58
+   ],
+   "gpu_pcie_rx_mb": [
+    0.67,
+    0.62,
+    247.86,
+    81.11,
+    78.2,
+    80.92,
+    83.88,
+    57.6,
+    82.87,
+    79.92,
+    40.29,
+    5.39,
+    9.77,
+    5.39
+   ],
+   "cpu": [
+    0.0,
+    1.9,
+    9.1,
+    8.8,
+    9.5,
+    10.7,
+    13.9,
+    10.5,
+    13.9,
+    11.8,
+    6.0,
+    6.3,
+    12.4,
+    11.7
+   ],
+   "ram_used": [
+    43887411200,
+    43859918848,
+    43956232192,
+    44034678784,
+    44030210048,
+    44028076032,
+    44015280128,
+    44252610560,
+    44247789568,
+    44243210240,
+    44230340608,
+    44558966784,
+    44619108352,
+    44619116544
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    126.09,
+    44.05,
+    80.0,
+    93.94,
+    138.6,
+    104.01,
+    156.43,
+    97.38,
+    0.0,
+    1.76,
+    4.66,
+    4.52
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    78.74,
+    81.4,
+    107.1
+   ]
+  },
+  "time": 1790717655.7892108
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K24/speed-32k-2.json

@@ -0,0 +1,392 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K24",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 2,
+ "run_id": "b38394f620b3",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:34:15",
+ "prompt_tokens": 31635,
+ "completion_tokens": 45,
+ "ttft_s": 10.269,
+ "e2e_s": 10.686,
+ "api_prompt_tps": 3080.5,
+ "api_decode_tps": 105.5,
+ "engine_prompt_tps": 3130.4,
+ "engine_decode_tps": 104.2,
+ "log_lines": [
+  "strata serve: prompt 31635 tokens = 0 reused + 31635 read in 10106 ms (3130.4 tok/s), 45 generated in 432 ms (104.2 tok/s), drafts accepted 33 of 37, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 100.0% (24480 hits / 24480 lookups)"
+ ],
+ "answer_head": "I'll start by exploring the repository structure to understand what we're working with.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 6144,
+   "expert_cache_mib": 11337,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 6657,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790717655.9301767,
+    "duration_s": 10.5,
+    "finish": "stop",
+    "prompt_tokens": 31635,
+    "reused": 0,
+    "output_tokens": 45,
+    "engine_generated": 45,
+    "prompt_ms": 10105.9,
+    "decode_ms": 431.8,
+    "decode_tok_s": 104.2,
+    "hit_rate": 1.0
+   },
+   {
+    "projection": null,
+    "time": 1790717642.434618,
+    "duration_s": 12.9,
+    "finish": "length",
+    "prompt_tokens": 28314,
+    "reused": 0,
+    "output_tokens": 256,
+    "engine_generated": 256,
+    "prompt_ms": 9861.8,
+    "decode_ms": 2981.5,
+    "decode_tok_s": 85.9,
+    "hit_rate": 1.0
+   }
+  ],
+  "requests_kept": 2,
+  "totals": {
+   "since": 1790717640.7776525,
+   "requests": 2,
+   "prompt_tokens": 59949,
+   "reused": 0,
+   "output_tokens": 301,
+   "prompt_ms": 19967.699999999997,
+   "decode_ms": 3413.3
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 20,
+     "mem_used": 18898419712,
+     "mem_total": 25769803776,
+     "temp": 50,
+     "power": 191.174
+    },
+    {
+     "index": 1,
+     "util": 53,
+     "mem_used": 21758935040,
+     "mem_total": 25769803776,
+     "temp": 54,
+     "power": 238.032
+    }
+   ],
+   "gpu_util": 36.5,
+   "gpu_mem_used": 40657354752,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 54,
+   "gpu_power": 429.206,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 0.6181640625,
+   "gpu_pcie_tx_mb": 0.857421875,
+   "cpu": 5.3,
+   "ram_used": 44575621120,
+   "ram_total": 177175990272,
+   "disk_read_mb": 2.4688909410966033,
+   "disk_write_mb": 0.23427432287777988,
+   "tok_s": 0.0,
+   "tok_s_mean": 0.0
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    1.0,
+    0.0,
+    1.0,
+    93.0,
+    94.0,
+    64.0,
+    94.5,
+    88.5,
+    95.5,
+    92.5,
+    48.5,
+    32.0,
+    48.0,
+    50.0,
+    0.0,
+    84.5,
+    96.0,
+    95.5,
+    79.0,
+    97.5,
+    97.0,
+    95.5,
+    96.0,
+    50.0,
+    36.5
+   ],
+   "gpu_mem_used": [
+    40447639552,
+    40447639552,
+    40476999680,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752
+   ],
+   "gpu_temp": [
+    49,
+    49,
+    49,
+    53,
+    54,
+    54,
+    55,
+    54,
+    55,
+    55,
+    55,
+    52,
+    54,
+    54,
+    50,
+    54,
+    55,
+    55,
+    55,
+    55,
+    55,
+    55,
+    55,
+    55,
+    54
+   ],
+   "gpu_power": [
+    259.59,
+    247.18,
+    245.88,
+    414.53,
+    546.71,
+    547.76,
+    546.95,
+    533.57,
+    531.41,
+    545.46,
+    539.14,
+    402.04,
+    486.67,
+    499.58,
+    409.59,
+    480.21,
+    551.08,
+    548.0,
+    547.82,
+    539.97,
+    537.72,
+    548.44,
+    548.33,
+    513.55,
+    429.21
+   ],
+   "gpu_pcie_rx_mb": [
+    0.67,
+    0.62,
+    247.86,
+    81.11,
+    78.2,
+    80.92,
+    83.88,
+    57.6,
+    82.87,
+    79.92,
+    40.29,
+    5.39,
+    9.77,
+    5.39,
+    45.78,
+    64.8,
+    1284.55,
+    129.94,
+    83.06,
+    82.63,
+    84.78,
+    73.15,
+    82.44,
+    0.57,
+    0.62
+   ],
+   "cpu": [
+    0.0,
+    1.9,
+    9.1,
+    8.8,
+    9.5,
+    10.7,
+    13.9,
+    10.5,
+    13.9,
+    11.8,
+    6.0,
+    6.3,
+    12.4,
+    11.7,
+    6.7,
+    8.3,
+    6.5,
+    7.2,
+    6.1,
+    6.0,
+    6.1,
+    6.5,
+    11.2,
+    8.0,
+    5.3
+   ],
+   "ram_used": [
+    43887411200,
+    43859918848,
+    43956232192,
+    44034678784,
+    44030210048,
+    44028076032,
+    44015280128,
+    44252610560,
+    44247789568,
+    44243210240,
+    44230340608,
+    44558966784,
+    44619108352,
+    44619116544,
+    44154675200,
+    44152926208,
+    44151242752,
+    44149854208,
+    44145594368,
+    44375150592,
+    44366180352,
+    44365656064,
+    44362940416,
+    44400578560,
+    44575621120
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    126.09,
+    44.05,
+    80.0,
+    93.94,
+    138.6,
+    104.01,
+    156.43,
+    97.38,
+    0.0,
+    1.76,
+    4.66,
+    4.52,
+    0.63,
+    0.02,
+    0.04,
+    0.03,
+    0.05,
+    0.03,
+    0.02,
+    13.87,
+    103.6,
+    0.0,
+    2.47
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    78.74,
+    81.4,
+    107.1,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "time": 1790717666.9788535
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K24/speed-32k-3.json

@@ -0,0 +1,496 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K24",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 3,
+ "run_id": "3bee1c1e482a",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:34:26",
+ "prompt_tokens": 31602,
+ "completion_tokens": 60,
+ "ttft_s": 10.281,
+ "e2e_s": 10.865,
+ "api_prompt_tps": 3073.8,
+ "api_decode_tps": 101.0,
+ "engine_prompt_tps": 3123.2,
+ "engine_decode_tps": 100.2,
+ "log_lines": [
+  "strata serve: prompt 31602 tokens = 0 reused + 31602 read in 10118 ms (3123.2 tok/s), 60 generated in 599 ms (100.2 tok/s), drafts accepted 40 of 52, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 100.0% (34560 hits / 34560 lookups)"
+ ],
+ "answer_head": "I'll start by examining the repository structure to understand what the last file does.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 6144,
+   "expert_cache_mib": 11337,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 6657,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790717667.1193056,
+    "duration_s": 10.7,
+    "finish": "stop",
+    "prompt_tokens": 31602,
+    "reused": 0,
+    "output_tokens": 60,
+    "engine_generated": 60,
+    "prompt_ms": 10118.5,
+    "decode_ms": 599.0,
+    "decode_tok_s": 100.2,
+    "hit_rate": 1.0
+   },
+   {
+    "projection": null,
+    "time": 1790717655.9301767,
+    "duration_s": 10.5,
+    "finish": "stop",
+    "prompt_tokens": 31635,
+    "reused": 0,
+    "output_tokens": 45,
+    "engine_generated": 45,
+    "prompt_ms": 10105.9,
+    "decode_ms": 431.8,
+    "decode_tok_s": 104.2,
+    "hit_rate": 1.0
+   },
+   {
+    "projection": null,
+    "time": 1790717642.434618,
+    "duration_s": 12.9,
+    "finish": "length",
+    "prompt_tokens": 28314,
+    "reused": 0,
+    "output_tokens": 256,
+    "engine_generated": 256,
+    "prompt_ms": 9861.8,
+    "decode_ms": 2981.5,
+    "decode_tok_s": 85.9,
+    "hit_rate": 1.0
+   }
+  ],
+  "requests_kept": 3,
+  "totals": {
+   "since": 1790717640.7776525,
+   "requests": 3,
+   "prompt_tokens": 91551,
+   "reused": 0,
+   "output_tokens": 361,
+   "prompt_ms": 30086.199999999997,
+   "decode_ms": 4012.3
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 43,
+     "mem_used": 18898419712,
+     "mem_total": 25769803776,
+     "temp": 51,
+     "power": 178.691
+    },
+    {
+     "index": 1,
+     "util": 51,
+     "mem_used": 21758935040,
+     "mem_total": 25769803776,
+     "temp": 54,
+     "power": 244.501
+    }
+   ],
+   "gpu_util": 47.0,
+   "gpu_mem_used": 40657354752,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 54,
+   "gpu_power": 423.192,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 5.1962890625,
+   "gpu_pcie_tx_mb": 103.138671875,
+   "cpu": 6.6,
+   "ram_used": 44596383744,
+   "ram_total": 177175990272,
+   "disk_read_mb": 1.1207609014346964,
+   "disk_write_mb": 0.0288298624163266,
+   "tok_s": 110.47676922161895,
+   "tok_s_mean": 99.0096070780873
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    1.0,
+    0.0,
+    1.0,
+    93.0,
+    94.0,
+    64.0,
+    94.5,
+    88.5,
+    95.5,
+    92.5,
+    48.5,
+    32.0,
+    48.0,
+    50.0,
+    0.0,
+    84.5,
+    96.0,
+    95.5,
+    79.0,
+    97.5,
+    97.0,
+    95.5,
+    96.0,
+    50.0,
+    36.5,
+    49.5,
+    96.5,
+    61.0,
+    94.5,
+    91.0,
+    97.0,
+    95.5,
+    98.5,
+    47.5,
+    47.0
+   ],
+   "gpu_mem_used": [
+    40447639552,
+    40447639552,
+    40476999680,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40546205696,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752,
+    40657354752
+   ],
+   "gpu_temp": [
+    49,
+    49,
+    49,
+    53,
+    54,
+    54,
+    55,
+    54,
+    55,
+    55,
+    55,
+    52,
+    54,
+    54,
+    50,
+    54,
+    55,
+    55,
+    55,
+    55,
+    55,
+    55,
+    55,
+    55,
+    54,
+    52,
+    55,
+    55,
+    55,
+    55,
+    56,
+    56,
+    55,
+    56,
+    54
+   ],
+   "gpu_power": [
+    259.59,
+    247.18,
+    245.88,
+    414.53,
+    546.71,
+    547.76,
+    546.95,
+    533.57,
+    531.41,
+    545.46,
+    539.14,
+    402.04,
+    486.67,
+    499.58,
+    409.59,
+    480.21,
+    551.08,
+    548.0,
+    547.82,
+    539.97,
+    537.72,
+    548.44,
+    548.33,
+    513.55,
+    429.21,
+    354.61,
+    553.75,
+    546.57,
+    547.59,
+    532.21,
+    535.5,
+    547.28,
+    548.49,
+    534.02,
+    423.19
+   ],
+   "gpu_pcie_rx_mb": [
+    0.67,
+    0.62,
+    247.86,
+    81.11,
+    78.2,
+    80.92,
+    83.88,
+    57.6,
+    82.87,
+    79.92,
+    40.29,
+    5.39,
+    9.77,
+    5.39,
+    45.78,
+    64.8,
+    1284.55,
+    129.94,
+    83.06,
+    82.63,
+    84.78,
+    73.15,
+    82.44,
+    0.57,
+    0.62,
+    81.82,
+    83.45,
+    75.34,
+    84.16,
+    82.11,
+    81.49,
+    70.95,
+    73.91,
+    52.55,
+    5.2
+   ],
+   "cpu": [
+    0.0,
+    1.9,
+    9.1,
+    8.8,
+    9.5,
+    10.7,
+    13.9,
+    10.5,
+    13.9,
+    11.8,
+    6.0,
+    6.3,
+    12.4,
+    11.7,
+    6.7,
+    8.3,
+    6.5,
+    7.2,
+    6.1,
+    6.0,
+    6.1,
+    6.5,
+    11.2,
+    8.0,
+    5.3,
+    6.7,
+    12.7,
+    10.0,
+    5.9,
+    6.4,
+    6.0,
+    7.1,
+    5.7,
+    6.8,
+    6.6
+   ],
+   "ram_used": [
+    43887411200,
+    43859918848,
+    43956232192,
+    44034678784,
+    44030210048,
+    44028076032,
+    44015280128,
+    44252610560,
+    44247789568,
+    44243210240,
+    44230340608,
+    44558966784,
+    44619108352,
+    44619116544,
+    44154675200,
+    44152926208,
+    44151242752,
+    44149854208,
+    44145594368,
+    44375150592,
+    44366180352,
+    44365656064,
+    44362940416,
+    44400578560,
+    44575621120,
+    44112637952,
+    44134821888,
+    44114948096,
+    44112486400,
+    44240445440,
+    44343197696,
+    44341800960,
+    44344549376,
+    44344029184,
+    44596383744
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    126.09,
+    44.05,
+    80.0,
+    93.94,
+    138.6,
+    104.01,
+    156.43,
+    97.38,
+    0.0,
+    1.76,
+    4.66,
+    4.52,
+    0.63,
+    0.02,
+    0.04,
+    0.03,
+    0.05,
+    0.03,
+    0.02,
+    13.87,
+    103.6,
+    0.0,
+    2.47,
+    0.63,
+    0.05,
+    0.08,
+    0.08,
+    0.04,
+    0.07,
+    0.04,
+    0.17,
+    0.0,
+    1.12
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    78.74,
+    81.4,
+    107.1,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    110.48
+   ]
+  },
+  "time": 1790717678.3471994
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K26/speed-32k-1.json

@@ -0,0 +1,261 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K26",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 1,
+ "run_id": "5ca39dba7489",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:35:22",
+ "prompt_tokens": 28314,
+ "completion_tokens": 48,
+ "ttft_s": 9.79,
+ "e2e_s": 10.297,
+ "api_prompt_tps": 2892.2,
+ "api_decode_tps": 92.7,
+ "engine_prompt_tps": 2940.2,
+ "engine_decode_tps": 88.8,
+ "log_lines": [
+  "strata serve: prompt 28314 tokens = 0 reused + 28314 read in 9630 ms (2940.2 tok/s), 48 generated in 540 ms (88.8 tok/s), drafts accepted 33 of 37, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 100.0% (25920 hits / 25920 lookups)"
+ ],
+ "answer_head": "I'll first look at the repository structure to understand what files exist and identify the last file.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 6656,
+   "expert_cache_mib": 12387,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 5607,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790717722.129571,
+    "duration_s": 10.2,
+    "finish": "stop",
+    "prompt_tokens": 28314,
+    "reused": 0,
+    "output_tokens": 48,
+    "engine_generated": 48,
+    "prompt_ms": 9630.1,
+    "decode_ms": 540.5,
+    "decode_tok_s": 88.8,
+    "hit_rate": 1.0
+   }
+  ],
+  "requests_kept": 1,
+  "totals": {
+   "since": 1790717720.412441,
+   "requests": 1,
+   "prompt_tokens": 28314,
+   "reused": 0,
+   "output_tokens": 48,
+   "prompt_ms": 9630.1,
+   "decode_ms": 540.5
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 47,
+     "mem_used": 19980550144,
+     "mem_total": 25769803776,
+     "temp": 51,
+     "power": 179.665
+    },
+    {
+     "index": 1,
+     "util": 28,
+     "mem_used": 20641153024,
+     "mem_total": 25769803776,
+     "temp": 53,
+     "power": 236.004
+    }
+   ],
+   "gpu_util": 37.5,
+   "gpu_mem_used": 40621703168,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 53,
+   "gpu_power": 415.669,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 0.666015625,
+   "gpu_pcie_tx_mb": 0.9052734375,
+   "cpu": 6.0,
+   "ram_used": 44539072512,
+   "ram_total": 177175990272,
+   "disk_read_mb": 2.97776661958663,
+   "disk_write_mb": 0.0,
+   "tok_s": 0.0,
+   "tok_s_mean": 0.0
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    0.5,
+    0.0,
+    0.5,
+    71.0,
+    95.5,
+    96.0,
+    91.5,
+    66.5,
+    95.5,
+    90.5,
+    84.5,
+    37.5
+   ],
+   "gpu_mem_used": [
+    40447639552,
+    40447639552,
+    40472805376,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40621703168
+   ],
+   "gpu_temp": [
+    50,
+    50,
+    50,
+    52,
+    55,
+    55,
+    56,
+    56,
+    56,
+    55,
+    56,
+    53
+   ],
+   "gpu_power": [
+    261.03,
+    248.02,
+    247.5,
+    351.95,
+    548.48,
+    553.77,
+    555.05,
+    536.77,
+    531.38,
+    552.37,
+    551.57,
+    415.67
+   ],
+   "gpu_pcie_rx_mb": [
+    0.76,
+    0.62,
+    224.64,
+    86.02,
+    83.16,
+    75.62,
+    86.69,
+    102.66,
+    65.66,
+    83.97,
+    38.86,
+    0.67
+   ],
+   "cpu": [
+    0.0,
+    1.1,
+    8.1,
+    9.1,
+    9.3,
+    12.2,
+    12.6,
+    12.6,
+    23.5,
+    14.7,
+    5.9,
+    6.0
+   ],
+   "ram_used": [
+    43871281152,
+    43867545600,
+    43958726656,
+    44040531968,
+    44041007104,
+    44033200128,
+    44032602112,
+    44244361216,
+    44250234880,
+    44236812288,
+    44233678848,
+    44539072512
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    122.8,
+    47.19,
+    79.99,
+    100.77,
+    127.09,
+    108.72,
+    156.54,
+    97.45,
+    0.0,
+    2.98
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "time": 1790717732.8116624
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K26/speed-32k-2.json

@@ -0,0 +1,365 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K26",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 2,
+ "run_id": "14c7210ce0e9",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:35:32",
+ "prompt_tokens": 31635,
+ "completion_tokens": 45,
+ "ttft_s": 10.029,
+ "e2e_s": 10.496,
+ "api_prompt_tps": 3154.3,
+ "api_decode_tps": 94.3,
+ "engine_prompt_tps": 3205.2,
+ "engine_decode_tps": 93.5,
+ "log_lines": [
+  "strata serve: prompt 31635 tokens = 0 reused + 31635 read in 9870 ms (3205.2 tok/s), 45 generated in 481 ms (93.5 tok/s), drafts accepted 31 of 36, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 100.0% (25440 hits / 25440 lookups)"
+ ],
+ "answer_head": "I'll start by exploring the repository structure to understand what the last file does.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 6656,
+   "expert_cache_mib": 12387,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 5607,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790717732.947479,
+    "duration_s": 10.4,
+    "finish": "stop",
+    "prompt_tokens": 31635,
+    "reused": 0,
+    "output_tokens": 45,
+    "engine_generated": 45,
+    "prompt_ms": 9870.0,
+    "decode_ms": 481.3,
+    "decode_tok_s": 93.5,
+    "hit_rate": 1.0
+   },
+   {
+    "projection": null,
+    "time": 1790717722.129571,
+    "duration_s": 10.2,
+    "finish": "stop",
+    "prompt_tokens": 28314,
+    "reused": 0,
+    "output_tokens": 48,
+    "engine_generated": 48,
+    "prompt_ms": 9630.1,
+    "decode_ms": 540.5,
+    "decode_tok_s": 88.8,
+    "hit_rate": 1.0
+   }
+  ],
+  "requests_kept": 2,
+  "totals": {
+   "since": 1790717720.412441,
+   "requests": 2,
+   "prompt_tokens": 59949,
+   "reused": 0,
+   "output_tokens": 93,
+   "prompt_ms": 19500.1,
+   "decode_ms": 1021.8
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 51,
+     "mem_used": 19991035904,
+     "mem_total": 25769803776,
+     "temp": 51,
+     "power": 190.866
+    },
+    {
+     "index": 1,
+     "util": 56,
+     "mem_used": 20653735936,
+     "mem_total": 25769803776,
+     "temp": 55,
+     "power": 249.813
+    }
+   ],
+   "gpu_util": 53.5,
+   "gpu_mem_used": 40644771840,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 55,
+   "gpu_power": 440.679,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 16.44921875,
+   "gpu_pcie_tx_mb": 82.0625,
+   "cpu": 6.0,
+   "ram_used": 44553875456,
+   "ram_total": 177175990272,
+   "disk_read_mb": 0.6053434938228842,
+   "disk_write_mb": 0.0,
+   "tok_s": 0.0,
+   "tok_s_mean": 0.0
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    0.5,
+    0.0,
+    0.5,
+    71.0,
+    95.5,
+    96.0,
+    91.5,
+    66.5,
+    95.5,
+    90.5,
+    84.5,
+    37.5,
+    1.5,
+    97.0,
+    96.0,
+    81.0,
+    86.5,
+    95.5,
+    95.0,
+    96.5,
+    95.0,
+    53.5
+   ],
+   "gpu_mem_used": [
+    40447639552,
+    40447639552,
+    40472805376,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40644771840
+   ],
+   "gpu_temp": [
+    50,
+    50,
+    50,
+    52,
+    55,
+    55,
+    56,
+    56,
+    56,
+    55,
+    56,
+    53,
+    51,
+    56,
+    56,
+    56,
+    56,
+    56,
+    56,
+    56,
+    56,
+    55
+   ],
+   "gpu_power": [
+    261.03,
+    248.02,
+    247.5,
+    351.95,
+    548.48,
+    553.77,
+    555.05,
+    536.77,
+    531.38,
+    552.37,
+    551.57,
+    415.67,
+    326.37,
+    532.86,
+    555.24,
+    552.15,
+    555.12,
+    507.8,
+    554.85,
+    554.0,
+    554.13,
+    440.68
+   ],
+   "gpu_pcie_rx_mb": [
+    0.76,
+    0.62,
+    224.64,
+    86.02,
+    83.16,
+    75.62,
+    86.69,
+    102.66,
+    65.66,
+    83.97,
+    38.86,
+    0.67,
+    3698.35,
+    132.46,
+    131.89,
+    82.06,
+    85.92,
+    83.02,
+    1279.88,
+    50.4,
+    86.21,
+    16.45
+   ],
+   "cpu": [
+    0.0,
+    1.1,
+    8.1,
+    9.1,
+    9.3,
+    12.2,
+    12.6,
+    12.6,
+    23.5,
+    14.7,
+    5.9,
+    6.0,
+    4.5,
+    6.1,
+    6.0,
+    6.4,
+    6.7,
+    5.9,
+    6.7,
+    9.9,
+    7.4,
+    6.0
+   ],
+   "ram_used": [
+    43871281152,
+    43867545600,
+    43958726656,
+    44040531968,
+    44041007104,
+    44033200128,
+    44032602112,
+    44244361216,
+    44250234880,
+    44236812288,
+    44233678848,
+    44539072512,
+    44064968704,
+    44058497024,
+    44058497024,
+    44058501120,
+    44175929344,
+    44294356992,
+    44294352896,
+    44295569408,
+    44291391488,
+    44553875456
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    122.8,
+    47.19,
+    79.99,
+    100.77,
+    127.09,
+    108.72,
+    156.54,
+    97.45,
+    0.0,
+    2.98,
+    0.47,
+    0.01,
+    0.04,
+    0.06,
+    0.03,
+    0.03,
+    12.57,
+    71.85,
+    33.12,
+    0.61
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0
+   ]
+  },
+  "time": 1790717743.8106396
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K26/speed-32k-3.json

@@ -0,0 +1,469 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K26",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 3,
+ "run_id": "3f243fdb63ee",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:35:43",
+ "prompt_tokens": 31601,
+ "completion_tokens": 60,
+ "ttft_s": 10.026,
+ "e2e_s": 10.61,
+ "api_prompt_tps": 3151.9,
+ "api_decode_tps": 101.0,
+ "engine_prompt_tps": 3208.3,
+ "engine_decode_tps": 100.2,
+ "log_lines": [
+  "strata serve: prompt 31601 tokens = 0 reused + 31601 read in 9850 ms (3208.3 tok/s), 60 generated in 599 ms (100.2 tok/s), drafts accepted 40 of 50, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 100.0% (34080 hits / 34080 lookups)"
+ ],
+ "answer_head": "I'll start by examining the repository structure to understand what the last file is.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 6656,
+   "expert_cache_mib": 12387,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 5607,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790717743.964174,
+    "duration_s": 10.5,
+    "finish": "stop",
+    "prompt_tokens": 31601,
+    "reused": 0,
+    "output_tokens": 60,
+    "engine_generated": 60,
+    "prompt_ms": 9849.9,
+    "decode_ms": 598.9,
+    "decode_tok_s": 100.2,
+    "hit_rate": 1.0
+   },
+   {
+    "projection": null,
+    "time": 1790717732.947479,
+    "duration_s": 10.4,
+    "finish": "stop",
+    "prompt_tokens": 31635,
+    "reused": 0,
+    "output_tokens": 45,
+    "engine_generated": 45,
+    "prompt_ms": 9870.0,
+    "decode_ms": 481.3,
+    "decode_tok_s": 93.5,
+    "hit_rate": 1.0
+   },
+   {
+    "projection": null,
+    "time": 1790717722.129571,
+    "duration_s": 10.2,
+    "finish": "stop",
+    "prompt_tokens": 28314,
+    "reused": 0,
+    "output_tokens": 48,
+    "engine_generated": 48,
+    "prompt_ms": 9630.1,
+    "decode_ms": 540.5,
+    "decode_tok_s": 88.8,
+    "hit_rate": 1.0
+   }
+  ],
+  "requests_kept": 3,
+  "totals": {
+   "since": 1790717720.412441,
+   "requests": 3,
+   "prompt_tokens": 91550,
+   "reused": 0,
+   "output_tokens": 153,
+   "prompt_ms": 29350.0,
+   "decode_ms": 1620.6999999999998
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 46,
+     "mem_used": 19991035904,
+     "mem_total": 25769803776,
+     "temp": 52,
+     "power": 189.922
+    },
+    {
+     "index": 1,
+     "util": 52,
+     "mem_used": 20653735936,
+     "mem_total": 25769803776,
+     "temp": 55,
+     "power": 246.398
+    }
+   ],
+   "gpu_util": 49.0,
+   "gpu_mem_used": 40644771840,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 55,
+   "gpu_power": 436.32,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 17.8330078125,
+   "gpu_pcie_tx_mb": 43.1533203125,
+   "cpu": 6.0,
+   "ram_used": 44614168576,
+   "ram_total": 177175990272,
+   "disk_read_mb": 0.4973320821900814,
+   "disk_write_mb": 0.05405783502066102,
+   "tok_s": 99.26679787882165,
+   "tok_s_mean": 97.997900113756
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    0.5,
+    0.0,
+    0.5,
+    71.0,
+    95.5,
+    96.0,
+    91.5,
+    66.5,
+    95.5,
+    90.5,
+    84.5,
+    37.5,
+    1.5,
+    97.0,
+    96.0,
+    81.0,
+    86.5,
+    95.5,
+    95.0,
+    96.5,
+    95.0,
+    53.5,
+    0.0,
+    95.0,
+    95.5,
+    82.0,
+    94.5,
+    93.0,
+    88.5,
+    96.0,
+    97.0,
+    49.0
+   ],
+   "gpu_mem_used": [
+    40447639552,
+    40447639552,
+    40472805376,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40535719936,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40621703168,
+    40644771840,
+    40644771840,
+    40644771840,
+    40644771840,
+    40644771840,
+    40644771840,
+    40644771840,
+    40644771840,
+    40644771840,
+    40644771840,
+    40644771840
+   ],
+   "gpu_temp": [
+    50,
+    50,
+    50,
+    52,
+    55,
+    55,
+    56,
+    56,
+    56,
+    55,
+    56,
+    53,
+    51,
+    56,
+    56,
+    56,
+    56,
+    56,
+    56,
+    56,
+    56,
+    55,
+    52,
+    55,
+    56,
+    56,
+    56,
+    56,
+    56,
+    56,
+    56,
+    55
+   ],
+   "gpu_power": [
+    261.03,
+    248.02,
+    247.5,
+    351.95,
+    548.48,
+    553.77,
+    555.05,
+    536.77,
+    531.38,
+    552.37,
+    551.57,
+    415.67,
+    326.37,
+    532.86,
+    555.24,
+    552.15,
+    555.12,
+    507.8,
+    554.85,
+    554.0,
+    554.13,
+    440.68,
+    363.75,
+    474.54,
+    552.72,
+    554.16,
+    554.3,
+    509.24,
+    555.11,
+    555.88,
+    555.36,
+    436.32
+   ],
+   "gpu_pcie_rx_mb": [
+    0.76,
+    0.62,
+    224.64,
+    86.02,
+    83.16,
+    75.62,
+    86.69,
+    102.66,
+    65.66,
+    83.97,
+    38.86,
+    0.67,
+    3698.35,
+    132.46,
+    131.89,
+    82.06,
+    85.92,
+    83.02,
+    1279.88,
+    50.4,
+    86.21,
+    16.45,
+    38.96,
+    81.63,
+    83.25,
+    84.4,
+    72.53,
+    127.22,
+    48.4,
+    96.51,
+    100.71,
+    17.83
+   ],
+   "cpu": [
+    0.0,
+    1.1,
+    8.1,
+    9.1,
+    9.3,
+    12.2,
+    12.6,
+    12.6,
+    23.5,
+    14.7,
+    5.9,
+    6.0,
+    4.5,
+    6.1,
+    6.0,
+    6.4,
+    6.7,
+    5.9,
+    6.7,
+    9.9,
+    7.4,
+    6.0,
+    11.7,
+    10.8,
+    6.2,
+    11.4,
+    7.0,
+    6.7,
+    5.6,
+    5.4,
+    6.0,
+    6.0
+   ],
+   "ram_used": [
+    43871281152,
+    43867545600,
+    43958726656,
+    44040531968,
+    44041007104,
+    44033200128,
+    44032602112,
+    44244361216,
+    44250234880,
+    44236812288,
+    44233678848,
+    44539072512,
+    44064968704,
+    44058497024,
+    44058497024,
+    44058501120,
+    44175929344,
+    44294356992,
+    44294352896,
+    44295569408,
+    44291391488,
+    44553875456,
+    44129808384,
+    44180946944,
+    44184666112,
+    44182659072,
+    44296114176,
+    44404658176,
+    44394999808,
+    44386783232,
+    44374441984,
+    44614168576
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    122.8,
+    47.19,
+    79.99,
+    100.77,
+    127.09,
+    108.72,
+    156.54,
+    97.45,
+    0.0,
+    2.98,
+    0.47,
+    0.01,
+    0.04,
+    0.06,
+    0.03,
+    0.03,
+    12.57,
+    71.85,
+    33.12,
+    0.61,
+    0.53,
+    0.06,
+    0.06,
+    0.09,
+    0.08,
+    0.04,
+    0.04,
+    0.04,
+    0.15,
+    0.5
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    99.27
+   ]
+  },
+  "time": 1790717754.9239163
+ }
+}

bench/results/2026-09-29-rtx3090-epyc-milan/data/native-v0.1.26-Coder-2x3090-K28/speed-32k-1.json

@@ -0,0 +1,261 @@
+{
+ "label": "native-v0.1.26-Coder-2x3090-K28",
+ "mode": "speed",
+ "ctx": "32k",
+ "target_tokens": 31488,
+ "run": 1,
+ "run_id": "0796a1ce12e4",
+ "corpus_repeated": false,
+ "time": "2026-09-29T21:36:29",
+ "prompt_tokens": 28316,
+ "completion_tokens": 57,
+ "ttft_s": 10.285,
+ "e2e_s": 10.86,
+ "api_prompt_tps": 2753.1,
+ "api_decode_tps": 97.4,
+ "engine_prompt_tps": 2798.5,
+ "engine_decode_tps": 93.7,
+ "log_lines": [
+  "strata serve: prompt 28316 tokens = 0 reused + 28316 read in 10118 ms (2798.5 tok/s), 57 generated in 608 ms (93.7 tok/s), drafts accepted 40 of 42, 2 checkpoints",
+  "strata serve: decode expert cache hit rate: 100.0% (28800 hits / 28800 lookups)"
+ ],
+ "answer_head": "I'll start by exploring the repository structure to understand what's there and identify the last file.",
+ "metrics_after": {
+  "engine": {
+   "model": "qwen3.8-flash-next-coder-iq1_m",
+   "max_context": 32768,
+   "images": false,
+   "version": "0.1.26",
+   "context": 32768,
+   "kv": "int8",
+   "kv_resident": 0,
+   "expert_slots": 7168,
+   "expert_cache_mib": 13437,
+   "spec": 6,
+   "mtp_max": 4,
+   "lookup": 3,
+   "vram_free_mib": 4557,
+   "cvec": 0,
+   "arena_mib": 23981,
+   "pool_workers": 19,
+   "pcie_frac": "0.55",
+   "spec_min_p": "0.50",
+   "engine": "0.1.26"
+  },
+  "live": {
+   "state": "idle",
+   "queued": 0,
+   "phase": null,
+   "prompt_tokens": null,
+   "prompt_read": null,
+   "prompt_total": null,
+   "generated": null,
+   "max_tokens": null,
+   "elapsed_s": null,
+   "tok_s": null,
+   "tok_s_mean": null,
+   "tok_s_window_s": null
+  },
+  "requests": [
+   {
+    "projection": null,
+    "time": 1790717789.1846194,
+    "duration_s": 10.7,
+    "finish": "stop",
+    "prompt_tokens": 28316,
+    "reused": 0,
+    "output_tokens": 57,
+    "engine_generated": 57,
+    "prompt_ms": 10118.2,
+    "decode_ms": 608.0,
+    "decode_tok_s": 93.8,
+    "hit_rate": 1.0
+   }
+  ],
+  "requests_kept": 1,
+  "totals": {
+   "since": 1790717787.625935,
+   "requests": 1,
+   "prompt_tokens": 28316,
+   "reused": 0,
+   "output_tokens": 57,
+   "prompt_ms": 10118.2,
+   "decode_ms": 608.0
+  },
+  "hardware": {
+   "gpus": [
+    {
+     "index": 0,
+     "util": 41,
+     "mem_used": 21104623616,
+     "mem_total": 25769803776,
+     "temp": 52,
+     "power": 189.832
+    },
+    {
+     "index": 1,
+     "util": 42,
+     "mem_used": 19548536832,
+     "mem_total": 25769803776,
+     "temp": 54,
+     "power": 230.563
+    }
+   ],
+   "gpu_util": 41.5,
+   "gpu_mem_used": 40653160448,
+   "gpu_mem_total": 51539607552,
+   "gpu_temp": 54,
+   "gpu_power": 420.395,
+   "gpu_power_limit": 560.0,
+   "gpu_pcie_gen": 4,
+   "gpu_pcie_gen_max": 4,
+   "gpu_pcie_width": 16,
+   "gpu_pcie_rx_mb": 17.1650390625,
+   "gpu_pcie_tx_mb": 53.119140625,
+   "cpu": 6.1,
+   "ram_used": 44503613440,
+   "ram_total": 177175990272,
+   "disk_read_mb": 1.946188226381293,
+   "disk_write_mb": 0.0036040522710764685,
+   "tok_s": 81.42212134334822,
+   "tok_s_mean": 78.79383849818437
+  },
+  "hardware_static": {
+   "gpu_name": "NVIDIA GeForce RTX 3090 + NVIDIA GeForce RTX 3090",
+   "gpu_count": 2,
+   "cpu_name": "AMD EPYC 7453 28-Core Processor",
+   "cores": 20,
+   "threads": 20,
+   "psutil": true
+  },
+  "history": {
+   "gpu_util": [
+    0.5,
+    0.0,
+    0.0,
+    46.5,
+    96.0,
+    72.5,
+    93.0,
+    90.0,
+    95.5,
+    93.5,
+    97.5,
+    41.5
+   ],
+   "gpu_mem_used": [
+    40447639552,
+    40447639552,
+    40495874048,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40542011392,
+    40653160448
+   ],
+   "gpu_temp": [
+    51,
+    50,
+    50,
+    52,
+    55,
+    54,
+    56,
+    54,
+    56,
+    56,
+    57,
+    54
+   ],
+   "gpu_power": [
+    261.46,
+    249.04,
+    248.59,
+    342.07,
+    531.07,
+    533.47,
+    545.68,
+    511.29,
+    533.34,
+    534.58,
+    541.51,
+    420.39
+   ],
+   "gpu_pcie_rx_mb": [
+    0.67,
+    0.62,
+    182.48,
+    55.88,
+    87.02,
+    80.58,
+    57.5,
+    78.87,
+    83.88,
+    1294.42,
+    46.78,
+    17.17
+   ],
+   "cpu": [
+    0.0,
+    6.6,
+    18.0,
+    11.3,
+    10.5,
+    9.8,
+    11.5,
+    13.9,
+    11.4,
+    12.9,
+    7.4,
+    6.1
+   ],
+   "ram_used": [
+    43799613440,
+    43777974272,
+    43950305280,
+    43962880000,
+    43939622912,
+    43930345472,
+    43932086272,
+    44071780352,
+    44183257088,
+    44182880256,
+    44179062784,
+    44503613440
+   ],
+   "disk_read_mb": [
+    null,
+    0.0,
+    129.59,
+    40.83,
+    79.94,
+    84.92,
+    115.02,
+    136.42,
+    117.79,
+    106.04,
+    29.9,
+    1.95
+   ],
+   "tok_s": [
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    0.0,
+    81.42
+   ]
+  },
+  "time": 1790717800.421838
+ }
+}