kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-13 22:18:36 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

[TRTLLM-9752][fix] WAR: Disable PDL for quant kernels to fix accuracy issues (#10285)

Browse Source 
Signed-off-by: Bo Deng <deemod@nvidia.com>
This commit is contained in:
Bo Deng 2026-01-03 14:34:55 +08:00 committed by GitHub
parent 45ffbf1f21
commit 9e7b50aefb
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
cpp/tensorrt_llm/kernels/quantization.cu
View File

@ -178,7 +178,7 @@ void invokeFP4Quantization(int b, int m, int n, T const* input, float const* SFS
config.stream = stream;
cudaLaunchAttribute attrs[1];
attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
attrs[0].val.programmaticStreamSerializationAllowed = tensorrt_llm::common::getEnvEnablePDL();
attrs[0].val.programmaticStreamSerializationAllowed = false;
config.numAttrs = 1;
config.attrs = attrs;
cudaLaunchKernelEx(&config, kernel_instance, b, m, n, n, input, SFScale, reinterpret_cast<uint32_t*>(output),
@ -213,7 +213,7 @@ void invokeMxFP8Quantization(int b, int m, int n, int padded_n, T const* input,
config.stream = stream;
cudaLaunchAttribute attrs[1];
attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
attrs[0].val.programmaticStreamSerializationAllowed = tensorrt_llm::common::getEnvEnablePDL();
attrs[0].val.programmaticStreamSerializationAllowed = false;
config.numAttrs = 1;
config.attrs = attrs;
cudaLaunchKernelEx(&config,
@ -388,7 +388,7 @@ void computePerTokenGlobalScaleForFP4Quantization(int b, int m, int n, T const*
config.stream = stream;
cudaLaunchAttribute attrs[1];
attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
attrs[0].val.programmaticStreamSerializationAllowed = tensorrt_llm::common::getEnvEnablePDL();
attrs[0].val.programmaticStreamSerializationAllowed = false;
config.numAttrs = 1;
config.attrs = attrs;
TLLM_CUDA_CHECK(cudaLaunchKernelEx(