wouter/d4r
1
0 Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Correct M attention precision and validate every temporal transformer block

Browse Source
This commit is contained in:
Codex committed 2026-09-30 21:30:05 +03:00
1 parent e15f6a73a6
commit b8d801ae4c
14 files changed
+623 -18
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

CMakeLists.txt
+8 -2
View File
@@ -50,7 +50,7 @@ foreach(family k m)
list(APPEND native_flags -DD4R_K_FP16_BASELINE)
endif()
if(family STREQUAL "m" AND D4R_M_FP16_BASELINE)
list(APPEND native_flags -DSWIN_EXACT)
list(APPEND native_flags -DSWIN_EXACT -DSWIN_EXACT_PV)
endif()
if(flags_line)
separate_arguments(source_flags NATIVE_COMMAND "${CMAKE_MATCH_1}")
@@ -68,7 +68,7 @@ foreach(family k m)
endforeach()
endforeach()
add_custom_target(d4r_native_gfx1201_compile ALL DEPENDS ${native_objects})
foreach(probe hip_gfx1201_probe gfx12_wmma_probe k_module_probe native_replay_probe cuda_driver_probe cuda_image_probe d3d12_hip_interop_probe hip_stream_lifecycle_probe ngx_cuda_init_probe d3d12_dlss_probe)
foreach(probe hip_gfx1201_probe gfx12_wmma_probe k_module_probe native_replay_probe cuda_driver_probe cuda_mma_probe cuda_image_probe d3d12_hip_interop_probe hip_stream_lifecycle_probe ngx_cuda_init_probe d3d12_dlss_probe)
add_executable(d4r_${probe} tools/windows/${probe}.cpp)
target_compile_definitions(d4r_${probe} PRIVATE WIN32_LEAN_AND_MEAN NOMINMAX __HIP_PLATFORM_AMD__)
target_include_directories(d4r_${probe} PRIVATE "${D4R_HIP_ROOT}/include")
@@ -147,6 +147,7 @@ install(FILES tools/windows/k_ptx_validate.py DESTINATION .)
install(FILES tools/windows/frame_compare.py DESTINATION .)
install(FILES tools/windows/m_replay_validate.py DESTINATION .)
install(FILES tools/windows/m_ptx_validate.py DESTINATION .)
install(FILES tools/windows/m_temporal_validate.py tools/windows/m_capture_compare.py DESTINATION .)
install(FILES scripts/windows/test-m-replay.ps1 DESTINATION .)
install(FILES scripts/windows/test-m-captures.ps1 DESTINATION .)
install(FILES kernels/tools/pwin_model.py kernels/tools/pwin_check.py kernels/tools/ptxsim.py
@@ -182,6 +183,11 @@ if(EXISTS "${D4R_ZLUDA_ROOT}/nvcuda.dll")
set_tests_properties(cuda_${context} PROPERTIES TIMEOUT 180 LABELS "hardware;zluda")
endforeach()
if(EXISTS "${D4R_ZLUDA_ROOT}/build-info.json")
add_test(NAME cuda_f16_mma_reference COMMAND d4r_cuda_mma_probe
--hip-root "${D4R_HIP_ROOT}" --cuda-dll "${D4R_ZLUDA_ROOT}/nvcuda.dll")
set_tests_properties(cuda_f16_mma_reference PROPERTIES TIMEOUT 120
LABELS "hardware;zluda;mma;reference"
ENVIRONMENT "D4R_ZLUDA_WMMA=1;D4R_ZLUDA_WMMA_F16_REFERENCE=1")
add_test(NAME cuda_images COMMAND d4r_cuda_image_probe
--hip-root "${D4R_HIP_ROOT}" --cuda-dll "${D4R_ZLUDA_ROOT}/nvcuda.dll" --iterations 32)
set_tests_properties(cuda_images PROPERTIES TIMEOUT 90 LABELS "hardware;zluda;descriptors")