Strata

arc

[SYCL] prompt attention on the int8 DPAS (ESIMD, Battlemage): hi/lo int8 Q and P, 2D block loads for the four-cell selection blocks; default on for Battlemage, STRATA_PROMPT_ATTN_ESIMD=0 off

2026-10-30 13:22:33 +01:00

dflash2

metal: DFlash2 Q8_0 GEMM kernel over a repacked pack (df2_mm: 12.6 ms a round against 36), Q8_0 draft head, STRATA_DF2_DUMP and tools/dflash2/check_metal_dump.py (final rows cos 0.9999 against the torch port)

2026-11-01 23:12:16 +01:00

dsv41

pin_coder.py reads one folder level of a repository (the official Coder repository keeps its model shards in IQ2_XXS/ and the Engram and picture file at the root): subdir, engram_path, mmproj_path

2026-11-02 15:19:14 +01:00

dsv41_lab

dsv41 cudalab: gfx12 WMMA port recipe of the trellis GEMV: host emulation of the data flow (matches the reference), HIP kernel source, layout probe

2026-10-20 18:45:22 +02:00

hip

AMD community tuning: gfx1010 SDWA dot4 and two gfx1030 rocBLAS tables under one switch

2026-10-31 23:30:43 +01:00

image

Images (I2): NVIDIA mma.sync flash attention for the native runner, fused q/k/v GEMM

2026-10-13 05:53:21 +02:00

par

tools: parallelism load harness (par_load, par_analyze, par_run, sweep, recon), E00 capability probe, batch_test --co/--seq/--pad and the phase-1/2 variant lists

2026-10-30 11:08:53 +01:00

qwen35

Merge f-145-q36speed (Qwen3.6 speed: decode/prompt kernels, bigger prompt chunk by default)

2026-10-14 22:23:29 +02:00

trellis

trellis: computed codebook on the GPU encoder, README for the K5/K4 versions

2026-10-20 19:59:48 +02:00

vision

vision: the image-token limit has a default and is enforced (I7, #3690 )

2026-10-22 13:10:12 +02:00

webtest

Web page: German and dark screenshots in tools/webtest/shots.mjs

2026-10-26 20:23:06 +01:00

windows

tools/windows/enable_large_pages.ps1: grant "Lock pages in memory" to the current account (#2185 )

2026-10-13 04:36:04 +02:00

_paths.py

tools: gguf-py is also looked up in the install folder (STRATA_INSTALL_DIR, the app's folder), so tools run by hand from the Control package find what setup.py downloaded; the error names STRATA_GGUF_PY (#3269 )

2026-10-19 23:26:32 +02:00

ab_engine.py

bench/results: V100 decode kernels (GPU work per verify window 22.8 -> 21.2 ms); docs/NVIDIA_V100.md: the decode side; tools/ab_engine.py

2026-10-07 20:42:39 +02:00

agent_turns_bench.py

tools/agent_turns_bench.py: decode A/B on coding-agent turns over this repository, per-turn ratios

2026-10-09 15:23:32 +02:00

batch_groups_cache_test.py

tools: batch_groups_cache_test.py reads the engine log only

2026-10-09 11:51:43 +02:00

batch_interleave_test.py

tools: batch_interleave_test runs on a pipelined batch (no give-way required there) and checks a turn checkpoint taken from a slot

2026-10-14 07:00:09 +02:00

batch_test.py

tools: parallelism load harness (par_load, par_analyze, par_run, sweep, recon), E00 capability probe, batch_test --co/--seq/--pad and the phase-1/2 variant lists

2026-10-30 11:08:53 +01:00

bench_eviction.py

tools: bench_eviction replays --dump-routing traces against candidate eviction policies

2026-10-07 04:07:42 +02:00

build_standard_corpus.py

Standard benchmark: frozen fixtures (bench/prompts/standard-v1) and the tool that builds them

2026-10-19 21:23:19 +02:00

calibrate.py

Merge f-027-int2 (2a2264c21) into f-027-gh-eng

2026-10-27 01:11:52 +01:00

cancel_park_test.py

tools: cancel_park_test.py, a cancelled request must not lose the conversation it restored (#1620 )

2026-10-09 10:01:59 +02:00

canonical_xcheck.py

Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

2026-09-24 18:42:20 +02:00

check_cuda_spills.py

cuda_spill_guard table: gr_down_v3_kernel on sm_120 (8 B, df2-more's register-blocked instance) and DFlash2's rope_table_inv_kernel (32 B) measured on l620

2026-11-02 13:28:35 +01:00

check_isa_guard.py

check_isa_guard: iq512_reference_gu_rows and iq512x_rows are listed (main already reports them as unlisted; both are reached only behind cpu_avx512_ok / cpu_vnni_ok)

2026-11-06 00:51:10 +01:00

check_lookup_window_serve.py

Gate check: a Qwen serve engine must answer requests that make prompt lookup propose long windows (the Metal 12-row refusal went unseen because no gate request repeated text)

2026-11-02 01:12:02 +01:00

collect_profile_traces.py

tools: make_profile_mix.py (an expert profile from several routing traces, each trace counting the same, with weights), collect_profile_traces.py (the traces of bench/profile_corpus.json's 45 varied prompts through the engine)

2026-10-11 07:58:34 +02:00

conversation_cache_disabled.py

tools: the cache-disabled gate compares every resolved setting except the engine version (release vs previous release)

2026-09-30 17:52:12 +02:00

conversation_cache_growth.py

STATE_HASH: pooled= keeps its 0.1.29 meaning (the completed indexer rows); the extent with the spare row is pooled_full=

2026-09-30 16:43:28 +02:00

conversation_cache_http_smoke.py

Integrate validated conversation snapshot core with upstream 0.1.25

2026-09-30 16:31:44 +02:00

conversation_cache_isolation.py

STATE_HASH: pooled= keeps its 0.1.29 meaning (the completed indexer rows); the extent with the spare row is pooled_full=

2026-09-30 16:43:28 +02:00

conversation_cache_parity.py

STATE_HASH: pooled= keeps its 0.1.29 meaning (the completed indexer rows); the extent with the spare row is pooled_full=

2026-09-30 16:43:28 +02:00

conversation_cache_soak.py

STATE_HASH: pooled= keeps its 0.1.29 meaning (the completed indexer rows); the extent with the spare row is pooled_full=

2026-09-30 16:43:28 +02:00

conversation_restore_speed.py

tools: conversation_restore_speed.py, decode after a conversation-cache restore vs live (#528 )

2026-10-03 19:29:26 +02:00

dflash2_pack.py

DFlash2 pack with the Q4_0 selector, made on the person's computer at setup / Load from the checked Q4_0 pack and used when it verifies

2026-11-02 11:46:16 +01:00

dgx_128k.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

dgx_bench.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

dgx_compare.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

dgx_concurrent.py

dgx_concurrent: --parallel N and --distinct (a prompt and a hidden code for every client, the answers checked: no error, readable, its own code and nobody else's)

2026-10-27 18:30:16 +01:00

dgx_gate_ref.py

dgx_gate_ref.py: the 24-prompt identity gate against a recorded answer file (the handover's quality24 JSON)

2026-10-27 17:22:14 +01:00

dgx_kvresident2.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

dgx_kvresident.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

dgx_longctx.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

dgx_needle.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

dgx_quality24.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

dgx_quality.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

dgx_smoke.py

dgx_smoke: the setup check is read case-insensitively (setup prints ARM64) and reports the GPU line

2026-10-27 17:15:51 +01:00

dgx_turns.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

dgx_warm.py

The DGX Spark benchmark family and guide (from the owner's handover, 2026-10-26)

2026-10-27 17:12:36 +01:00

draft_vocab.py

setup: --draft-vocab fr - the English/code subset plus the tokens of French text (#597 )

2026-10-03 18:19:49 +02:00

early_close_test.py

server: a client that stops reading early keeps the engine in step

2026-10-03 08:04:53 +02:00

embd_bf16_pack.py

APR port, stages 1-2: map, trellis type tables, prism.hadamard packing

2026-10-15 21:28:47 +02:00

force_check.py

force_check.py: absolute path for the profile copy

2026-11-06 15:47:43 +01:00

gguf_reader.py

Naming: the DeepSeek sets are ISTA-DAS on screen and in the docs; the ids (--model IST-DAS-2.25bpw) are unchanged and the new spelling is accepted too

2026-10-31 09:51:11 +01:00

gguf_writer.py

Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

2026-09-24 18:42:20 +02:00

h2d_bench.cu

STRATA_ADAPT_OVERLAP=1 (opt-in): the adaptive tier's copies land while the next decode window runs, a layer waits for its own copies

2026-10-10 18:49:32 +02:00

harness_reuse_probe.py

serve: the conversation cache is sized from the free RAM when --conversation-cache-mib is not given (#3493 ): 40% of what is left above the parking floor, at most 8192, off under 1024, one log line; an explicit value (0 too) is kept; harness_reuse_probe.py; docs

2026-10-22 16:06:17 +02:00

hf_pin_fetch.py

tls: downloads check the server against the operating system's certificates; a failed check gives a plain message (#3722 )

2026-10-22 15:43:12 +02:00

idle_gpu_ab.py

2507-0: measure the idle GPUs with the existing switches (z440 2x A4000, P2003 RTX 3060)

2026-10-20 12:35:59 +02:00

iq_fixture.py

tests (#264 follow-up): the fixture step skips (exit 3) instead of failing when the vendored gguf-py is absent; no unused SCALE_F16

2026-09-30 20:36:29 +02:00

iq_pack.py

iq_pack: an interrupted experts.bin write continues at the next layer (finished layers are flushed and recorded, the last kept one is checked against the GGUF) instead of starting from zero; says so

2026-10-20 16:03:23 +02:00

logit_bias_english_example.py

docs: demonstrate Hello token bias with measured before and after requests

2026-10-09 15:21:30 +02:00

logit_bias_http_probe.py

fix(api): apply request logit biases in CUDA and HIP sampling

2026-10-09 15:21:29 +02:00

make_profile_mix.py

tools: make_profile_mix.py (an expert profile from several routing traces, each trace counting the same, with weights), collect_profile_traces.py (the traces of bench/profile_corpus.json's 45 varied prompts through the engine)

2026-10-11 07:58:34 +02:00

make_profile.py

layer split: score every four-way placement instead of guessing it

2026-10-05 19:57:36 +02:00

minimal_cli_smoke.py

CI smokes: serve smoke expects the server-package preflight, server smoke accepts OK (skipped=N), nested-telemetry fixture deep enough for Python 3.12+

2026-10-19 16:23:26 +02:00

minimal_server_smoke.py

CI smokes: serve smoke expects the server-package preflight, server smoke accepts OK (skipped=N), nested-telemetry fixture deep enough for Python 3.12+

2026-10-19 16:23:26 +02:00

monitor_soak.py

terminal: strata_cli, strata_control, bin/strata and the terminal tests

2026-10-15 10:23:23 +02:00

mtp_fetch.py

tls: downloads check the server against the operating system's certificates; a failed check gives a plain message (#3722 )

2026-10-22 15:43:12 +02:00

mtp_pack.py

Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

2026-09-24 18:42:20 +02:00

mtp_rt.py

Prompt lookup that pays, sampling fixes, shallow-path setup fix (engine 0.1.7)

2026-09-27 10:18:14 +02:00

needle_bench.py

tools/needle_bench.py: needle-in-a-haystack recall through a running server (#33 )

2026-09-28 14:53:30 +02:00

nvfp4_convert.py

NVFP4 routed experts: converter, pack, decode, prompt path, AVX-512 CPU rows

2026-10-12 20:29:45 +02:00

nvfp4_local_pack.py

nvfp4 local pack: read the checkpoint with cached sequential reads, not a process-lifetime np.memmap (#3377 , #3378 )

2026-10-26 18:56:54 +01:00

nvfp4_verify_gguf.py

NVFP4 routed experts: converter, pack, decode, prompt path, AVX-512 CPU rows

2026-10-12 20:29:45 +02:00

nvfp4_verify.py

NVFP4 routed experts: converter, pack, decode, prompt path, AVX-512 CPU rows

2026-10-12 20:29:45 +02:00

pack_index.py

Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

2026-09-24 18:42:20 +02:00

pack_layer.py

Prompt lookup that pays, sampling fixes, shallow-path setup fix (engine 0.1.7)

2026-09-27 10:18:14 +02:00

par_analyze.py

tools: parallelism load harness (par_load, par_analyze, par_run, sweep, recon), E00 capability probe, batch_test --co/--seq/--pad and the phase-1/2 variant lists

2026-10-30 11:08:53 +01:00

par_load.py

tools: parallelism load harness (par_load, par_analyze, par_run, sweep, recon), E00 capability probe, batch_test --co/--seq/--pad and the phase-1/2 variant lists

2026-10-30 11:08:53 +01:00

par_run.py

tools: parallelism load harness (par_load, par_analyze, par_run, sweep, recon), E00 capability probe, batch_test --co/--seq/--pad and the phase-1/2 variant lists

2026-10-30 11:08:53 +01:00

parking_test.py

Parking with a layer split: the draft ring with the last stage, on its GPU; guards; trim log

2026-10-03 18:14:59 +02:00

piped_read_join_bench.py

docs/BATCHING.md: reading beside the decoding slots, and what it measured

2026-10-14 11:58:16 +02:00

ple_fp8_pack.py

PLE: the n-gram table in FP8 E4M3, as Qwen ships it

2026-10-01 10:31:23 +02:00

portable_monitor_pty_smoke.py

Portable monitor bundle: carries consoleinfo.py and overview.py (manifest schema 6), the plain Monitor runs when launch is not bundled, smokes accept the Overview's /metrics read, test derives the bundle's import closure

2026-10-19 16:35:51 +02:00

portable_monitor_smoke.py

Portable monitor bundle: carries consoleinfo.py and overview.py (manifest schema 6), the plain Monitor runs when launch is not bundled, smokes accept the Overview's /metrics read, test derives the bundle's import closure

2026-10-19 16:35:51 +02:00

q8_dense_gguf.py

tools/q8_dense_gguf.py: convert every shard that holds dense matrices

2026-10-14 07:08:59 +02:00

q8_sm60_check.cu

sm_60: a Q8_0 decode GEMV for Pascal GP100 (STRATA_Q8_SM60=1, opt-in)

2026-10-14 07:08:58 +02:00

quant_bound.py

kq_cpu_bench: the CPU pool's expert rows per format on T threads in GB/s of weights; tools/quant_bound.py: decode bounds from the per-layer formats; setup copies libstrata_kq next to a compiled engine

2026-10-12 21:16:33 +02:00

quant_formats.py

quant_formats --hf: the header takes the file's size from the tree listing, so gguf_reader's bounds checks apply (#5075 )

2026-11-04 10:38:09 +01:00

quant_quality.py

Quality ids files: a token id outside the model's vocabulary fails loudly (the engine's --score / --profile-run / --spec-eval, and quant_quality.py run)

2026-11-01 01:54:24 +01:00

quantscope.py

Terminal: one discovery of existing installs and model files for the screens, models list, own-files import and launch; Gyro-S/M GGUFs are known; installed models show what is really missing (T03, T04, T05)

2026-10-18 22:22:31 +02:00

replay_cache_policy.py

replay_cache_policy: policy options on the command line

2026-10-09 21:09:03 +02:00

research_run.py

R1/R3: tools/research_run.py, docs/RESEARCH_RUNS.md (measured numbers), DETAILS.md strata_prefix section

2026-10-07 00:01:12 +02:00

serve_load.py

Concurrency stage 1: --batch-groups auto (one pipeline group per GPU stage), INFO batch_groups; tools/serve_load.py load generator

2026-10-07 21:07:59 +02:00

session_pressure_test.py

tools: session_pressure_test.py and batch_groups_cache_test.py

2026-10-09 10:30:16 +02:00

sm60_dp4a_check.cu

tools: sm60_dp4a_check (STRATA_DP4A vs byte reference, bit-exact + timing)

2026-10-06 14:58:47 +02:00

split_mtp_batch_parity.py

split: batched draft-layer prompt K/V on the last stage (STRATA_SPLIT_MTP_BATCH=1, opt-in)

2026-10-09 03:06:46 +02:00

split_serve_check.py

Split serving: docs with the measured results, check tool no longer hashes the state in speed runs

2026-10-27 01:38:32 +01:00

stop_resend_test.py

tools: stop_resend_test.py, the stop / edit / resend loop of #879 and #606 as a repeatable check

2026-10-06 16:35:09 +02:00

strata_bench.py

bench: results.json gets the same redaction as the other report files; a credential of "none" is no key; the key is replaced as a whole word in its own case (#4484 , #4483 )

2026-11-04 10:34:11 +01:00

strata_inspect.py

Own model files: the checker in front of setup accepts what setup --gguf-dir accepted (Unsloth split sets with a metadata-only first shard, links, several sizes in a folder, a folder above the model folder, more than 128 files); Unsloth Q5/Q6/Q8 and Swift IQ3_S are in the table

2026-10-20 14:07:31 +02:00

strata_mcp.py

Mac app: Unsloth's sets are offered and installed as the Terminal's setup_mac does (metal_repack to one GGUF, --ple-gguf the table's shard, disk counts the repacked file); the repack tools ship in the installer bundle; tests

2026-11-05 09:05:01 +01:00

strata_pack.py

strata_pack.py build --force: unmake the old pack before writing, so a build that stops part-way is an unfinished build, not a pack (#1176 )

2026-10-07 03:02:55 +02:00

strata_tokenizer.py

serve: DeepSeek-V4.1 through the server (deepseek-v3 pre-tokenizer, DSML tool calls, effort words, end token, card sampling)

2026-10-20 18:45:21 +02:00

stream_ownership_probe.py

tools: stream_ownership_probe.py - interleaved streaming clients with canaries and random disconnects against a running server (#4451 , #2892 )

2026-11-04 11:35:16 +01:00

test_appimage_elf.py

appimage_elf: a shared library is not a static PIE; the check tells them apart without name rules (#4337 checker)

2026-11-05 05:24:48 +01:00

test_bench_eviction.py

tools: bench_eviction replays --dump-routing traces against candidate eviction policies

2026-10-07 04:07:42 +02:00

test_bench_host.py

tools: poll the server address the config binds, not loopback (#2330 )

2026-10-14 06:52:37 +02:00

test_calibrate.py

calibrate: an engine that dies mid-measurement drops only its candidate (#3317 )

2026-10-26 18:43:46 +01:00

test_check_cuda_spills.py

tests: cuda_spill_guard fails the CUDA test build when a Strata kernel spills or keeps a per-thread array in local memory (#2683 , #2682 )

2026-10-26 18:47:35 +01:00

test_check_isa_guard.py

Fix #3358 : the ISA guard knows the trellis AVX-512 rows (tq_avx512.cpp), so isa_guard_check passes on a normal build

2026-10-20 13:59:25 +02:00

test_check_lookup_window_serve.py

Gate check: a Qwen serve engine must answer requests that make prompt lookup propose long windows (the Metal 12-row refusal went unseen because no gate request repeated text)

2026-11-02 01:12:02 +01:00

test_conversation_cache_disabled.py

tests: three suite gaps a live PC showed (#3236 )

2026-10-26 18:38:42 +01:00

test_conversation_cache_growth.py

tests: three suite gaps a live PC showed (#3236 )

2026-10-26 18:38:42 +01:00

test_conversation_cache_http_smoke.py

tests: three suite gaps a live PC showed (#3236 )

2026-10-26 18:38:42 +01:00

test_conversation_cache_isolation.py

tests: three suite gaps a live PC showed (#3236 )

2026-10-26 18:38:42 +01:00

test_conversation_cache_parity.py

tests: three suite gaps a live PC showed (#3236 )

2026-10-26 18:38:42 +01:00

test_conversation_cache_soak.py

tests: three suite gaps a live PC showed (#3236 )

2026-10-26 18:38:42 +01:00

test_conversation_restore_speed.py

tools: conversation_restore_speed.py, decode after a conversation-cache restore vs live (#528 )

2026-10-03 19:29:26 +02:00

test_cyrillic_paths.py

Windows with a non-UTF-8 code page (Russian): PowerShell output is read as UTF-8 (Documents folder, save dialog, fonts, GPU names), Python runs in UTF-8 mode from the launchers and the host's children, and the engine and picture reader get a UTF-8 manifest so a Cyrillic model folder arrives intact (#3202 )

2026-10-19 23:38:49 +02:00

test_dflash2_pack.py

DFlash2: opt-in Q4_0 drafter pack (dtype q4_0, tools/dflash2_pack.py --q4_0): nibble-packed tiled weights, gemm4_kernel on the existing dp4a activations path, df2_q4_tile at load, CPU parity test dflash2_q4_parity (40 of 40 cases within 2.4e-6 of 19)

2026-10-31 23:33:22 +01:00

test_dflash2_vocab.py

DFlash2-only installs use the draft-vocabulary subset the person chose in setup: the choice (en / fr / cyrillic) is copied beside the DFlash2 pack, refreshed when it changes and on update and start, kept by the drafter's fetch; tests for DFlash2-only and MTP-folder installs

2026-11-02 13:37:26 +01:00

test_dll_closure.py

Windows HIP: setup and the package put the whole dependency closure of amdhip64_7.dll and amd_comgr.dll beside strata.exe, and the engine says why the bundled runtime did not load (#461 )

2026-10-07 17:34:36 +02:00

test_docker_entrypoint.py

Docker: RETUNE=1 passes --retune to the launcher, so a tuned preset is measured again (#4779 )

2026-10-31 12:58:46 +01:00

test_draft_vocab.py

setup: --draft-vocab fr - the English/code subset plus the tokens of French text (#597 )

2026-10-03 18:19:49 +02:00

test_effort_end_args.py

tools tests: follow #1792 (no version range without CUDA, STRATA_VCVARS), and the effort-end test no longer collides on a case-insensitive disk

2026-10-10 11:11:20 +02:00

test_func_attr_scope.py

fused_gr: the shared-memory opt-in of the v3 down kernels is set once per DEVICE, not once per process (a layer split on two sm_120 cards answered 0 tokens)

2026-11-02 06:47:11 +01:00

test_gfx906_build_files.py

gfx906 (HIP wave64) builds again: no Qwen3.6 engine (cuBLAS calls the compat layer does not map), GGML_USE_HIP on the W4A4 target, the trellis overlay sources and the moe_f16_kq stub in the HIP source list; all targets compile for gfx906 (#3172 #3271 )

2026-10-20 00:41:38 +02:00

test_gguf_paths.py

tools: gguf-py is also looked up in the install folder (STRATA_INSTALL_DIR, the app's folder), so tools run by hand from the Control package find what setup.py downloaded; the error names STRATA_GGUF_PY (#3269 )

2026-10-19 23:26:32 +02:00

test_gguf_reader_malformed.py

tools/gguf_reader.py: bound the header reads and report malformed metadata as ValueError (#2288 , #1919 )

2026-10-26 18:55:57 +01:00

test_gguf_tqp.py

Gyro-L: read the pair-trellis types TQP5/TQP6/TQP7 (ggml 147-149) on the CPU, CUDA and the MMQ prompt path (#3806 )

2026-10-27 01:57:57 +01:00

test_iq_pack.py

catalog: Unsloth v3.1 Qwen3.8-Flash-Next sets (UD-IQ1_M, UD-Q2_0, UD-IQ2_XS, UD-Q2_K_XL, UD-IQ3_XXS, UD-Q3_K_XL, UD-IQ4_XS-v31) at d9620119 with a per-model revision; UD-Q4_K_XL's four-shard layout accepted from local files; manual-load refusal lifted (UD-IQ1_S stays refused); mac_fit switch for the Q5_0 down sets; header-only fixtures and tests (about lines pending wording)

2026-11-05 07:23:41 +01:00

test_kfd_gpus.py

Hardware list: AMD cards are counted from the kernel's KFD files when there is no setup module and no rocm-smi (ROCm from a Python venv), instead of not at all

2026-10-30 19:45:23 +01:00

test_make_profile_mix.py

SYCL: data/expert-profile-sycl.bin, the experts' ranking for a card on xe, built with make_profile_mix.py from the routing of 45 varied prompts and eight document summaries; setup_intel writes it instead of the shipped profile there (i915 keeps the shipped one). B70: +5..64% decode (mean +21%) on 12 unseen prompts, the same tokens. The copy-engine queue is Battlemage-only by default (it hung the A750's prompt)

2026-10-11 08:25:30 +02:00

test_make_profile.py

make_profile: --reorder so a trace can rank ahead of a complete base

2026-10-03 18:14:52 +02:00

test_mtp_fetch.py

Downloads: the MTP fetch shows progress in the app and survives a flaky link (runs again, fewer connections, one writer per folder); a cancel and a killed setup end the whole tree on Windows (snapshot + the fetch ends with its parent); Gyro's low-RAM pack gets --compat-bf16 (#3257 #3214 #3193 )

2026-10-19 23:38:38 +02:00

test_nvfp4_local_pack.py

nvfp4 local pack: read the checkpoint with cached sequential reads, not a process-lifetime np.memmap (#3377 , #3378 )

2026-10-26 18:56:54 +01:00

test_par_load.py

tools: parallelism load harness (par_load, par_analyze, par_run, sweep, recon), E00 capability probe, batch_test --co/--seq/--pad and the phase-1/2 variant lists

2026-10-30 11:08:53 +01:00

test_picture_defaults.py

Mac pictures: strata-vision with Metal built in, Mac setup and launch turn pictures on

2026-10-19 20:59:29 +02:00

test_pool_plan.py

CPU pool plan: hybrid CPUs with more E-cores than P-cores default to the P-cores but the host's plus half of the E-cores (Core Ultra 7 265: 12-16 workers +12.6% decode over 19; 24 workers -44%)

2026-10-18 13:18:38 +02:00

test_qimage.py

Images (I2): setup builds the native DiT runner into sd-cli on CUDA / HIP; --image-native on|off (off by default) and the config's "native" add --model-args qwen_image_2_1_native=1

2026-10-13 01:07:28 +02:00

test_quant_formats_hf.py

quant_formats --hf: the header takes the file's size from the tree listing, so gguf_reader's bounds checks apply (#5075 )

2026-11-04 10:38:09 +01:00

test_research_run.py

R1/R3: tools/research_run.py, docs/RESEARCH_RUNS.md (measured numbers), DETAILS.md strata_prefix section

2026-10-07 00:01:12 +02:00

test_setup_amd_cards.py

AMD cards fix, review round: setup's pinned one-card pick is still written (with gpus_by), only DeepSeek on AMD is judged when a copy adopts an earlier card, the launcher uses the plan's card rule and remembers a named card, Windows uses setup's amd_rank and no Linux arch list, message names strata launch --gpu

2026-10-31 03:05:41 +01:00

test_setup_amd.py

setup (AMD, Linux): the engine probes get the engine's lib_dirs on LD_LIBRARY_PATH (#3305 )

2026-10-26 18:23:22 +01:00

test_setup_bench_tips.py

setup: no "--prefill auto:32768" tip where a bare auto already takes 32768 (one NVIDIA card of compute capability 8.0+, no STRATA_NV_CHUNK=0 / STRATA_DEFAULTS=026, no layer split); docs: DETAILS.md describes the NVIDIA ceiling

2026-11-03 20:37:45 +01:00

test_setup_build_archs.py

Setup: a compiled engine gets code for every usable NVIDIA card in the PC, not only the card setup picked (a TITAN RTX + RTX 3080 build for sm_75 only failed every long prompt with unspecified launch failure) (#4662 )

2026-10-30 19:18:01 +01:00

test_setup_build_dir.py

setup: a build folder configured with another generator is configured again instead of stopping cmake (#4038 )

2026-10-26 18:28:23 +01:00

test_setup_cards_029.py

Setup: say which card is not used when a card under 8 GB sits beside a bigger NVIDIA card (a line only, no change to the cards used)

2026-10-31 03:10:24 +01:00

test_setup_choices.py

catalog: Unsloth v3.1 Qwen3.8-Flash-Next sets (UD-IQ1_M, UD-Q2_0, UD-IQ2_XS, UD-Q2_K_XL, UD-IQ3_XXS, UD-Q3_K_XL, UD-IQ4_XS-v31) at d9620119 with a per-model revision; UD-Q4_K_XL's four-shard layout accepted from local files; manual-load refusal lifted (UD-IQ1_S stays refused); mac_fit switch for the Q5_0 down sets; header-only fixtures and tests (about lines pending wording)

2026-11-05 07:23:41 +01:00

test_setup_coder.py

Setup: a picture file the person already has is used, not fetched again; the Mac's download total drops the picture file once the engine says no

2026-11-02 20:03:42 +01:00

test_setup_component_start.py

setup: hand component engines to the server for validation (#4027 , #4026 )

2026-10-26 18:20:42 +01:00

test_setup_compute_mode.py

setup: warn on a non-Default GPU compute mode; vision encoder start error names the busy/unavailable cause (#1445 )

2026-10-08 02:31:56 +02:00

test_setup_config.py

Your own engine settings are kept: user_args, a reserve of 0 on a compute-only card, control vectors survive a load (#3215 #3270 #3216 )

2026-10-20 00:11:11 +02:00

test_setup_configs.py

setup: the shared Chat settings file is no model config (#346 )

2026-10-05 15:58:35 +02:00

test_setup_cuda_repo.py

setup: Ubuntu keeps its checked cuda-toolkit-13-0 install; Debian stops with a message when no 13.x installs (#2172 follow-up)

2026-10-13 00:49:11 +02:00

test_setup_disk_room.py

setup: the low-RAM mode's copy of the experts is checked for room where it is written (the data folder's packs folder), and the refusal names that folder in the approved words instead of the model folder; test

2026-11-04 17:35:55 +01:00

test_setup_dl_fix.py

Test: torch has one pin per Python range since #3474 ; compare the package set

2026-10-21 00:23:52 +02:00

test_setup_download_length.py

setup: use the published size when the download length probes fail (#2287 )

2026-10-26 18:48:49 +01:00

test_setup_download.py

setup: wording for the kept .bad file (NEEDS OWNER APPROVAL): the mismatch line says the file was kept, and says when the server sends the very same bytes; one line when the kept file is removed for room

2026-11-06 15:47:44 +01:00

test_setup_draft_vocab.py

setup --update: an MTP folder shared by several configs keeps the draft subset of the most recently used one (the oldest config's default subset replaced cyrillic: Russian decode 68 -> 54 tok/s on an RTX 5060 Ti + RTX 3060, IQ3_XXS)

2026-10-13 04:36:28 +02:00

test_setup_dsv41.py

rc1 merge ds-coder (d6a5a7837)

2026-11-02 09:51:07 +01:00

test_setup_engine_hash.py

[Other] engine hash tests: mock the release probe, keep the engine folder in a temp dir

2026-10-19 20:38:04 +02:00

test_setup_engine_keep.py

setup: checksum verified line, .previous keeps its BUILD.json; update scripts drop the backup branch when the move fails

2026-10-07 20:46:11 +02:00

test_setup_feed_engine.py

Fix #3469 , #3437 : UPDATE.bat and setup fetch the engine from the signed update feed when the release has no flat zip, and the minimum engine is 0.1.46

2026-10-21 00:21:24 +02:00

test_setup_find_tool.py

setup: the build tools come from the virtualenv first and an older system cmake (below the 3.24 CMakeLists.txt needs) counts as missing, so pip installs the right one (#3822 ); tests

2026-10-24 13:28:02 +02:00

test_setup_gbk_adapters.py

Control: the tuning probe's child output is read with errors replaced too, the last place of #4675 's kind (#4629 : the same crash on a Russian-locale PC was fixed in 0.2.9 by 7146d6983)

2026-10-31 12:52:12 +01:00

test_setup_golden.json

Setup golden baseline re-recorded for the approved 128K default: only --max-context, --kv-resident and --resident-budget-gib moved (12 of 25 configs)

2026-11-02 10:59:20 +01:00

test_setup_golden.py

Golden setup test: the vision reader's thread count is mocked (6, the 12-thread baseline PC), so the baseline no longer depends on the runner's CPU count (it failed 10 cases x2 on any Linux runner with more than 12 threads, v0.2.9 included)

2026-11-02 11:07:43 +01:00

test_setup_gpu_fit.py

GPU roles: two identical cards recommend the highest number as Main, so the engine order stays [0, 1]; a model that keeps one card keeps setup's old pick; golden back to [0, 1] with roles only

2026-10-20 19:48:20 +02:00

test_setup_gyro.py

Fix #3425 : the memory tiers count RAM + VRAM less 5 GB that Windows and the system keep (setup.py and tools/strata_mcp.py, one constant each; the app and strata launch read the engine's pick; the Mac keeps its own macOS share)

2026-10-21 07:03:33 +02:00

test_setup_hotfix_tag.py

Fix #3469 , #3437 : UPDATE.bat and setup fetch the engine from the signed update feed when the release has no flat zip, and the minimum engine is 0.1.46

2026-10-21 00:21:24 +02:00

test_setup_hybrid.py

CPU pool: the LP-E cores (/sys/devices/cpu_lowpower/cpus) are efficiency cores, in the engine and in setup (#3262 )

2026-10-19 23:25:46 +02:00

test_setup_llama_archive.py

tools tests: the llama.cpp archive test passes a plain path (a file:// URI with a drive letter is not found on Windows)

2026-10-10 11:04:37 +02:00

test_setup_lowram.py

Fit checks and slot sizing count the draft layer's ~1 GB of graphics memory while MTP is on

2026-10-19 19:56:01 +02:00

test_setup_mac.py

tests: the Mac app's running-engine hint

2026-11-05 09:25:00 +01:00

test_setup_menu.py

catalog: Unsloth v3.1 Qwen3.8-Flash-Next sets (UD-IQ1_M, UD-Q2_0, UD-IQ2_XS, UD-Q2_K_XL, UD-IQ3_XXS, UD-Q3_K_XL, UD-IQ4_XS-v31) at d9620119 with a per-model revision; UD-Q4_K_XL's four-shard layout accepted from local files; manual-load refusal lifted (UD-IQ1_S stays refused); mac_fit switch for the Q5_0 down sets; header-only fixtures and tests (about lines pending wording)

2026-11-05 07:23:41 +01:00

test_setup_model_pins.py

model_pins guard: the credit link to the community DFlash2 GGUF in a comment and a docstring is not a model host the runtime reaches

2026-11-02 12:55:28 +01:00

test_setup_modelscope.py

Pin SHA-256 and size for every model file setup downloads; stop when the pinned version is gone

2026-10-19 08:41:24 +02:00

test_setup_nvfp4.py

Pictures on by default on every make of card: MCP install, Qwen3.6, Unsloth and the original Flash-Next families follow the one reader rule (CPU under 24 GB, GPU from 24 GB); the CPU reader keeps no VRAM reserve

2026-10-19 07:23:44 +02:00

test_setup_oldcpu.py

older CPUs (experimental): STRATA_ISA_FLOOR=avx|none, and setup compiles it instead of stopping (#394 #595 #623 )

2026-10-03 23:35:52 +02:00

test_setup_older_gpus.py

CUDA 12 driver floor: setup takes it from the engine package, not from a constant (#4247 , #4188 )

2026-10-28 23:53:00 +01:00

test_setup_ownfiles.py

test: the own-GGUF check loads its helper by path (another tests package can already be loaded in a full tools run)

2026-10-24 16:43:46 +02:00

test_setup_page_file.py

setup: the page-file warning advises a fixed size, and says why when Windows is set to grow the file

2026-10-09 15:16:03 +02:00

test_setup_parallel.py

#465 setup: recommend "parallel" only where the experts mostly fit in VRAM

2026-10-04 03:36:22 +02:00

test_setup_pcie.py

setup: step 1 shows the GPU's PCIe link and says when it reads fewer lanes than the card has (#912 )

2026-10-06 15:57:38 +02:00

test_setup_pins.py

setup.py on a DGX Spark: unified-memory GB10, its engine from the signed feed, the measured run-config settings

2026-10-27 17:12:36 +01:00

test_setup_prompts.py

setup: stop on EOF instead of accepting prompt defaults

2026-10-01 20:46:19 +02:00

test_setup_qwen36.py

Tests that pinned old values follow the merged decisions: native-fit (MTP layer 1.08 GB), tune-outcome (calm running step, GUI feedback 2 item 3), setup tests (128K default context on every card; the picture-file host in strata_control/pictures.py is behind acquisition.require)

2026-11-02 10:45:03 +01:00

test_setup_remote_opt.py

setup: the multi-GPU default follows the fit rule (largest card + RAM first, the split recommended only when the model does not fit one card), goldens re-recorded, setup tests

2026-10-17 20:16:02 +02:00

test_setup_rerun_size.py

setup: run again over an installed model keeps its size; the three env entries 0.1.44 made setup's own stay the user's in a GSQ-RCO config

2026-10-14 01:16:56 +02:00

test_setup_risk.py

Tests that pinned old values follow the merged decisions: native-fit (MTP layer 1.08 GB), tune-outcome (calm running step, GUI feedback 2 item 3), setup tests (128K default context on every card; the picture-file host in strata_control/pictures.py is behind acquisition.require)

2026-11-02 10:45:03 +01:00

test_setup_rocm_indexes.json

setup: the ROCm wheel is chosen per GPU family from what its TheRock index offers (a pinned version, else the newest of the same line or major, with a warning); gfx103X-all takes 7.13.0a20260515 and gfx1151 7.14.0a20260608 (#1103 , #1267 )

2026-10-07 03:57:38 +02:00

test_setup_rocm_pins.py

tests (#2147 ): test_setup_rocm_pins skips an index with no saved version list (the gfx1150 opt-in has none yet)

2026-10-14 10:48:36 +02:00

test_setup_rope.py

setup: an explicit --rope-scale must be finite and at least 1 (#3393 )

2026-10-26 18:47:46 +01:00

test_setup_shard_check.py

Naming: the DeepSeek sets are ISTA-DAS on screen and in the docs; the ids (--model IST-DAS-2.25bpw) are unchanged and the new spelling is accepted too

2026-10-31 09:51:11 +01:00

test_setup_small_card.py

setup: UD-* on a card of about 8 GB - estimate the VRAM the engine needs, write the least harmful config that starts (4-bit KV, smaller context, then no draft layer on the GPU), say so, warn that UD-Q4 is very slow there and suggest IQ3_XXS / IQ3_S; tests

2026-10-13 13:51:01 +02:00

test_setup_stale_stamp.py

Tests for the AMD card rules (4 cards, APU, mixed architectures, Windows, --gpu, --gpus, old configs, Intel, NVIDIA unchanged); stale-stamp test stubs setup.gpus

2026-10-31 02:56:19 +01:00

test_setup_stall_note.py

setup: a download that gets no data for two minutes says so with the percentage and the one next step, instead of a line that never moves (#3567 ); test

2026-10-24 13:35:43 +02:00

test_setup_start_gpus_amd.py

setup: --gpus when starting an installed model on a PC with AMD cards and no NVIDIA ones names the AMD cards (#1594 )

2026-10-09 03:15:12 +02:00

test_setup_strix_halo.py

Test fixes after the 0.2.1 merges: the gfx1151 table parse stops before the iGPU set, the benchmark's model-card link is not a model-host fetch, docker/build.sh copies bench/prompts/standard-v1 for the Terminal package, the Docker script tests use Git bash with unix paths on Windows

2026-10-20 00:05:48 +02:00

test_setup_sycl.py

Intel Arc B-series (Battlemage, e.g. Arc Pro B70) runs like any Arc: pictures on by default with the CPU reader, one plain line that the card is new and how to report a problem

2026-10-22 19:39:27 +02:00

test_setup_udlow.py

Accept Unsloth's DeepSeek UD-IQ1_S and UD-IQ1_M sets (pins, models, app file check, fit numbers; opt-in names)

2026-10-30 18:05:31 +01:00

test_setup_uma_pool.py

Launch: a remembered card choice is said when the PC has more cards (NVIDIA DeepSeek after --gpu 0 stayed on one card without a word); Strix Halo: a fixed small expert cache lowers setup's own RAM budget (the person's cache and budget are only warned about) and a silent exit code 1 gets a plain next step

2026-11-01 09:51:26 +01:00

test_setup_unsloth_v31.py

catalog: approved wording (about lines, SUPPORTED_GGUFS, UD-IQ1_S reason, Unsloth family line); METAL_Q5_0_DOWN on (f-0214-metal cbef1a9ff) so the Mac lists all seven sets

2026-11-05 08:04:15 +01:00

test_setup_unsloth.py

setup: wording for the kept .bad file (NEEDS OWNER APPROVAL): the mismatch line says the file was kept, and says when the server sends the very same bytes; one line when the kept file is removed for room

2026-11-06 15:47:44 +01:00

test_setup_update_gpu.py

setup: an engine update compiles for the cards the installed models use, not the biggest one (#1485 )

2026-10-09 03:15:12 +02:00

test_setup_update.py

UPDATE.bat / update.sh: an engine that stayed older than the minimum is said as 'the engine was NOT updated' with one next step, not '[ok] Strata is updated' (#3653 ); test

2026-10-24 13:08:46 +02:00

test_shards.py

strata_pack.py build --force: unmake the old pack before writing, so a build that stops part-way is an unfinished build, not a pack (#1176 )

2026-10-07 03:02:55 +02:00

test_source_nul_bytes.py

Sources: no raw NUL bytes (expert_source.cpp, pinned.cu) + a test that scans for them

2026-10-13 00:26:51 +02:00

test_strata_bench.py

bench: tools/strata_bench.py reads the CUDA UMD version, the storage, the engine's build toolkit and GPU power and PCIe detail (#3250 , #2111 )

2026-10-26 19:01:17 +01:00

test_strata_inspect.py

One strata binary for both families: the Qwen3.6 engine path as a library

2026-10-13 02:06:01 +02:00

test_strata_mcp.py

Mac app: Unsloth's sets are offered and installed as the Terminal's setup_mac does (metal_repack to one GGUF, --ple-gguf the table's shard, disk counts the repacked file); the repack tools ship in the installer bundle; tests

2026-11-05 09:05:01 +01:00

test_strata_tokenizer.py

serve: tokenise the prompt from the last shared prefix, not from the start (#1852 , #567 )

2026-10-15 05:41:24 +02:00

test_sycl_build_guard.py

SYCL: build.sh stops an AOT Battlemage build before the compile when ocloc is the one that crashes on the int8 kernels (Compute Runtime 26.22.38646.x, IGC 2.36.3); I8_ESIMD=OFF / STRATA_SYCL_I8_ESIMD builds without them; docs say docker build --pull (#3219 )

2026-10-20 00:47:08 +02:00

test_tokenizer_plain_spans.py

tokenizer: sweep protected spans in match order

2026-10-09 03:04:21 +02:00

test_user_args.py

Your own engine settings are kept: user_args, a reserve of 0 on a compute-only card, control vectors survive a load (#3215 #3270 #3216 )

2026-10-20 00:11:11 +02:00

test_vision_reader.py

Qwen picture reader on demand by default on an NVIDIA card with one GPU: the first picture starts it (the expert cache gives its room), 120 s without one stops it; it needs no config key (vision.elastic false or vram_elastic false keeps it on the card). The on-demand reader has the same program without --gpu as its CPU reader, so a picture is read whatever the card has left (found on boston: reserve 700 and no CPU reader gave a 503 on every picture)

2026-11-02 00:18:52 +01:00

test_windows_amd_gpu.py

Windows AMD: the card's name and VRAM from the registry's display class, in the server's hardware_static and in strata_bench (#1587 )

2026-10-13 04:12:24 +02:00

tqenc.py

trellis: computed codebook on the GPU encoder, README for the K5/K4 versions

2026-10-20 19:59:48 +02:00

unsloth_v31_fixture.py

catalog: Unsloth v3.1 Qwen3.8-Flash-Next sets (UD-IQ1_M, UD-Q2_0, UD-IQ2_XS, UD-Q2_K_XL, UD-IQ3_XXS, UD-Q3_K_XL, UD-IQ4_XS-v31) at d9620119 with a per-model revision; UD-Q4_K_XL's four-shard layout accepted from local files; manual-load refusal lifted (UD-IQ1_S stays refused); mac_fit switch for the Q5_0 down sets; header-only fixtures and tests (about lines pending wording)

2026-11-05 07:23:41 +01:00

web_i18n.py

Web page: the chat view in eight languages, and the page's docs

2026-10-26 20:22:05 +01:00