kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-14 06:27:45 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

[None][fix] fix trtllm moe backend error when running gptoss on b300

Browse Source 
Signed-off-by: Jiagan Cheng <jiaganc@nvidia.com>
This commit is contained in:
Jiagan Cheng 2025-08-27 20:25:11 -07:00 committed by Xiwen Yu
parent c2e1ad9d45
commit 6fd765f89e
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
cpp/tensorrt_llm/thop/mxFp4BlockScaleMoe.cpp
View File

@ -46,8 +46,8 @@ torch::Tensor dtype_mxe2m1_block_scale_moe_runner(torch::Tensor const& routing_l
std::optional<double> const routed_scaling_factor, int64_t const tile_tokens_dim, int64_t const routing_method_type,
btg::Dtype const dtype, MoeRunnerType& moe_runner, int64_t moeConfigIndex)
{
auto const sm = tensorrt_llm::common::getSMVersion();
TORCH_CHECK(sm == 100, "Only SM100 is supported by FP4 block scale MOE");
auto const sm = tensorrt_llm::common::getSMFamily();
TORCH_CHECK(sm == 100, "Only SM100f is supported by MXFP4 block scale MOE");
TORCH_CHECK(tile_tokens_dim == 8 || tile_tokens_dim == 16 || tile_tokens_dim == 32 || tile_tokens_dim == 64,
"tile_tokens_dim must be 8, 16, 32, 64");
TORCH_CHECK(routing_logits.scalar_type() == at::ScalarType::Float