diff --git a/.github/ISSUE_TEMPLATE/amd_split_report.yml b/.github/ISSUE_TEMPLATE/amd_split_report.yml
new file mode 100644
index 000000000..13d3e8d53
--- /dev/null
+++ b/.github/ISSUE_TEMPLATE/amd_split_report.yml
@@ -0,0 +1,68 @@
+name: AMD split engine report
+description: Tell us whether the all-cards split engine works on your AMD cards (it prints a line asking for this on cards Strata has not run it on)
+title: "[Engine] AMD split engine: <your card> x <count>"
+labels: ["engine", "amd-split-report"]
+body:
+  - type: markdown
+    value: |
+      Strata runs the all-cards split engine by itself on Radeon AI PRO R9700 / RX 9070 (gfx1201). On other AMD cards (RX 9060, RX 7000, the 780M and similar) it only starts when you ask for it with `strata launch --split-engine on`, and it prints one line asking you to report whether it works. We cannot test those cards, so your report is how we learn.
+
+      Write only what you saw. A report that says "it did not start" is as useful as one that says "it works".
+  - type: input
+    id: card
+    attributes:
+      label: Card model and count
+      description: For example "2 x RX 7900 XTX 24 GB" or "3 x RX 9060 XT 16 GB"
+    validations:
+      required: true
+  - type: input
+    id: arch
+    attributes:
+      label: Architecture (the gfx name)
+      description: From the engine's line, or `rocminfo | grep gfx`. For example gfx1100.
+    validations:
+      required: true
+  - type: input
+    id: rocm
+    attributes:
+      label: ROCm / driver version
+      description: Strata's engine package brings its own ROCm libraries; the kernel driver is yours (`uname -r`, `amd-smi version`).
+    validations:
+      required: true
+  - type: input
+    id: model
+    attributes:
+      label: Model and quantization
+      description: The model Strata loaded, for example Qwen3.8-Flash-Next UD-Q4_K_XL or the IQ3_S set.
+    validations:
+      required: true
+  - type: textarea
+    id: selftest
+    attributes:
+      label: The self-test line
+      description: Paste the line from the engine's log that starts "The split engine's self-test" (passed or failed). The log is in the run folder shown at launch; `strata launch` also prints it.
+      render: text
+    validations:
+      required: true
+  - type: dropdown
+    id: result
+    attributes:
+      label: What happened
+      options:
+        - It started and the answers look right
+        - It started and the answers look wrong
+        - The self-test failed, so the normal layer split started
+        - It failed to start (not the self-test)
+        - It started and crashed or hung later
+    validations:
+      required: true
+  - type: input
+    id: speed
+    attributes:
+      label: Speed, tok/s (the split engine, then the layer split with `--split-engine off` if you tried both)
+      description: For example "decode 61 tok/s, layer split 48 tok/s, one user".
+  - type: textarea
+    id: more
+    attributes:
+      label: Anything else
+      description: Wrong answers (the prompt and what you got), crashes (the last lines of the engine log), or what you changed.
diff --git a/packaging/build_split_hip.sh b/packaging/build_split_hip.sh
index 94babc6af..a19fa930b 100644
--- a/packaging/build_split_hip.sh
+++ b/packaging/build_split_hip.sh
@@ -28,7 +28,7 @@ SYSDEPS=""; [ -d "$HP/lib/rocm_sysdeps" ] && SYSDEPS=";$HP/lib/rocm_sysdeps"
 BC=$HP/lib/llvm/amdgcn/bitcode; [ -d "$BC" ] || BC=$HP/amdgcn/bitcode
 rm -f "$B/CMakeCache.txt"
 cmake -G Ninja -S "$SRC" -B "$B" -DCMAKE_BUILD_TYPE=Release -DSTRATA_ENABLE_HIP=ON -DSTRATA_ENABLE_CUDA=OFF -DSTRATA_BUILD_TESTS=OFF -DSTRATA_BUILD_SPLIT=ON \
-  "-DSTRATA_SPLIT_CARDS=${CARDS// /;}" "-DCMAKE_HIP_ARCHITECTURES=$ARCHS" -DSTRATA_PREFILL_MMQ=ON -DSTRATA_NATIVE_EXPERTS=ON -DSTRATA_GGML_DIR="$GG" \
+  "-DSTRATA_SPLIT_CARDS=${CARDS// /;}" "-DCMAKE_HIP_ARCHITECTURES=gfx1201" "-DSTRATA_SPLIT_HIP_ARCHS=$ARCHS" -DSTRATA_PREFILL_MMQ=ON -DSTRATA_NATIVE_EXPERTS=ON -DSTRATA_GGML_DIR="$GG" \
   -DCMAKE_HIP_COMPILER="$HP/llvm/bin/clang++" -DCMAKE_HIP_COMPILER_ROCM_ROOT="$HP" "-DCMAKE_PREFIX_PATH=$HP;$LIBW$SYSDEPS" \
   "-DCMAKE_HIP_FLAGS=--rocm-path=$HP --rocm-device-lib-path=$BC" -DSTRATA_PORTABLE=ON \
   "-DCMAKE_EXE_LINKER_FLAGS=-Wl,-rpath,'\$ORIGIN' -Wl,--enable-new-dtags"
diff --git a/src/split/CMakeLists.txt b/src/split/CMakeLists.txt
index f09e84f70..ca8a21650 100644
--- a/src/split/CMakeLists.txt
+++ b/src/split/CMakeLists.txt
@@ -34,15 +34,23 @@ if(_fnx_hip)
   # gfx11-generic = gfx1100 gfx1101 gfx1102 gfx1103 in one code object (RX 7000 and the 780M iGPU; the gfx11 WMMA path, hip/mma.h FNX_WMMA_GFX11; the self-test passes on a 780M,
   # no multi-card PC has run it). Only gfx1201 is automatic
   # (strata_control/split_engine.py AUTO_SPLIT_AMD_ARCHS); on another card the engine runs src/selftest.cu before it loads the model
+  # STRATA_SPLIT_HIP_ARCHS names the split engine's code objects when they are not the tree's CMAKE_HIP_ARCHITECTURES (whose list the HIP backend checks for the main
+  # engine and does not know gfx11-generic); empty: the tree's list
+  set(STRATA_SPLIT_HIP_ARCHS "" CACHE STRING "code objects of strata-split-N (gfx1201;gfx1200;gfx11-generic), empty: CMAKE_HIP_ARCHITECTURES")
+  if(STRATA_SPLIT_HIP_ARCHS)
+    set(_fnx_hip_archs ${STRATA_SPLIT_HIP_ARCHS})
+  else()
+    set(_fnx_hip_archs ${STRATA_HIP_ARCH_LIST})
+  endif()
   set(_fnx_hip_ok gfx1100 gfx1101 gfx1102 gfx1103 gfx11-generic gfx1200 gfx1201)
   set(_fnx_hip_bad "")
-  foreach(_a ${STRATA_HIP_ARCH_LIST})
+  foreach(_a ${_fnx_hip_archs})
     if(NOT _a IN_LIST _fnx_hip_ok)
       list(APPEND _fnx_hip_bad ${_a})
     endif()
   endforeach()
-  if(NOT STRATA_HIP_ARCH_LIST OR _fnx_hip_bad)
-    message(WARNING "STRATA_BUILD_SPLIT (HIP): the split engine builds for ${_fnx_hip_ok} only (CMAKE_HIP_ARCHITECTURES=${STRATA_HIP_ARCH_LIST}), skipped")
+  if(NOT _fnx_hip_archs OR _fnx_hip_bad)
+    message(WARNING "STRATA_BUILD_SPLIT (HIP): the split engine builds for ${_fnx_hip_ok} only (got ${_fnx_hip_archs}), skipped")
     return()
   endif()
   set(STRATA_SPLIT_ARCH 1201)   # (FNX_ARCH is only compared with 80 and 70: any HIP value takes the WMMA paths)
@@ -95,7 +103,7 @@ foreach(_n ${STRATA_SPLIT_CARDS})
                                FNX_DIR="${_sp}")
     target_compile_options(${_t} PRIVATE "$<$<COMPILE_LANGUAGE:HIP>:SHELL:-include ${_sp}/src/hip/fnx_hip.h>")
     target_link_libraries(${_t} PRIVATE strata_hip_runtime ${STRATA_HIP_BLAS_TARGETS})
-    set_target_properties(${_t} PROPERTIES CXX_STANDARD 17 HIP_STANDARD 17 HIP_STANDARD_REQUIRED ON)
+    set_target_properties(${_t} PROPERTIES CXX_STANDARD 17 HIP_STANDARD 17 HIP_STANDARD_REQUIRED ON HIP_ARCHITECTURES "${_fnx_hip_archs}")
     find_package(Threads REQUIRED)
     target_link_libraries(${_t} PRIVATE Threads::Threads)
     continue()