Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
Samples/2_Performance/matmul_story 下部分 MXFP4 样例存在以下问题:
Samples/2_Performance/matmul_story
matmul_recipes/examples/quant_matmul_mxfp4
aclrtEvent
quant_matmul_mxfp4_a_full_load.cpp
quant_matmul_mxfp4_swat.cpp
matmul_tutorials
.cpp
verify_result.py
quant_matmul_mxfp4/README.md
[M, K]
[N, K]
requirements.txt
受影响文件包括但不限于:
Samples/2_Performance/matmul_story/matmul_recipes/examples/quant_matmul_mxfp4/quant_matmul_mxfp4_algorithm_recommend.py
Samples/2_Performance/matmul_story/matmul_recipes/examples/quant_matmul_mxfp4/quant_matmul_mxfp4_a_full_load.cpp
Samples/2_Performance/matmul_story/matmul_recipes/examples/quant_matmul_mxfp4/quant_matmul_mxfp4_swat.cpp
Samples/2_Performance/matmul_story/matmul_recipes/examples/quant_matmul_mxfp4/README.md
Samples/2_Performance/matmul_story/matmul_tutorials/*/*.cpp
可以按以下方式观察问题:
cmake -S . -B build cmake --build build --parallel cmake --install build --prefix ./build_out cd build_out/2_Performance/matmul_story/matmul_recipes/examples/quant_matmul_mxfp4 python3 quant_matmul_mxfp4_algorithm_recommend.py 256 256 256
/assign
Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
一、问题描述 (必填)
Samples/2_Performance/matmul_story下部分 MXFP4 样例存在以下问题:matmul_recipes/examples/quant_matmul_mxfp4中的算法推荐脚本原先依赖样例程序输出的aclrtEvent计时作为排序依据,但 event 计时容易受到共享 NPU 场景影响,不适合作为推荐依据。quant_matmul_mxfp4_a_full_load.cpp和quant_matmul_mxfp4_swat.cpp中保留了基于 event 的计时代码。matmul_tutorials目录下所有 tutorial.cpp都包含基于 event 的计时代码。matmul_tutorials下部分 tutorial 样例执行结束后没有自动调用verify_result.py做结果校验。quant_matmul_mxfp4/README.md中缺少当前样例的使用约束说明,例如当前 case 仅支持 A 不转置、B 转置,即 A 为[M, K]、B 为[N, K],两侧 K 轴位于内轴。requirements.txt中缺少样例脚本实际使用到的 Python 依赖声明及版本信息。受影响文件包括但不限于:
Samples/2_Performance/matmul_story/matmul_recipes/examples/quant_matmul_mxfp4/quant_matmul_mxfp4_algorithm_recommend.pySamples/2_Performance/matmul_story/matmul_recipes/examples/quant_matmul_mxfp4/quant_matmul_mxfp4_a_full_load.cppSamples/2_Performance/matmul_story/matmul_recipes/examples/quant_matmul_mxfp4/quant_matmul_mxfp4_swat.cppSamples/2_Performance/matmul_story/matmul_recipes/examples/quant_matmul_mxfp4/README.mdSamples/2_Performance/matmul_story/matmul_tutorials/*/*.cpprequirements.txt二、环境信息 (可选)
三、重现步骤 (可选)
可以按以下方式观察问题:
cmake -S . -B build cmake --build build --parallel cmake --install build --prefix ./build_out cd build_out/2_Performance/matmul_story/matmul_recipes/examples/quant_matmul_mxfp4 python3 quant_matmul_mxfp4_algorithm_recommend.py 256 256 256