[Issue]: how to make sure the gemm operation using hipblasLtMatmul()

[zeyang12-jpg]

Problem Description

machine:MI308 I set the flag HIPBLASLT_LOG_LEVEL=2&TORCH_BLAS_PREFER_HIPBLASLT=1when I run my test,the output, it can find rocblaslt_matmul from my log,does it mean that the hipblaslt_matmul is not used.

[2024-12-12 07:19:50][HIPBLASLT][1085][Trace][rocblaslt_matmul] A=0x7f11c2c00000 Adesc=[type=R_32F rows=4096 cols=14336 ld=4096] B=0x7f0f8c600000 Bdesc=[type=R_32F rows=4096 cols=267 ld=4096] C=0x7f0f8f200000 Cdesc=[type=R_32F rows=14336 cols=267 ld=14336] D=0x7f0f8f200000 Ddesc=[type=R_32F rows=14336 cols=267 ld=14336] computeDesc=[computeType=COMPUTE_32F scaleType=R_32F transA=OP_T transB=OP_N epilogue=EPILOGUE_DEFAULT biasPointer=0x0] workSpace=0x7f1033a7a200 workSpaceSizeInBytes=1048576 alpha=1 beta=0 stream=0 [2024-12-12 07:19:50][HIPBLASLT][1085][Trace][rocblaslt_matmul] A=0x7f11b4a00000 Adesc=[type=R_32F rows=14336 cols=4096 ld=14336] B=0x7f0f90800000 Bdesc=[type=R_32F rows=14336 cols=267 ld=14336] C=0x7f0f8ea9a000 Cdesc=[type=R_32F rows=4096 cols=267 ld=4096] D=0x7f0f8ea9a000 Ddesc=[type=R_32F rows=4096 cols=267 ld=4096] computeDesc=[computeType=COMPUTE_32F scaleType=R_32F transA=OP_T transB=OP_N epilogue=EPILOGUE_DEFAULT biasPointer=0x0] workSpace=0x7f1033a7a200 workSpaceSizeInBytes=1048576 alpha=1 beta=0 stream=0

Operating System

ubuntu22.04

CPU

no

GPU

AMD Instinct MI250X

Other

No response

ROCm Version

ROCm 6.2.2

ROCm Component

hipBLASLt

Steps to Reproduce

No response

(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

No response

[ppanchad-amd]

Hi @zeyang12-jpg. Internal ticket has been created to investigate your issue. Thanks!

[tcgu-amd]

Hi @zeyang12-jpg, thanks for reaching out. Just in case, would you be able to confirm the outputs of

echo $TORCH_BLAS_PREFER_HIPBLASLT  

please? What kind of work load are you trying to run? A simple reproducer would be helpful.

Thanks!!

[tcgu-amd]

Hi @zeyang12-jpg! Due to inactivity, this issue will be closed for now. Please feel free to re-open in the future. Thanks!