xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-07-30 05:27:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

[fix]: disable cutlass block scaled group gemm for EP (#20781)

Browse Source 
Signed-off-by: Duncan Moss <djm.moss@gmail.com>
This commit is contained in:
Duncan Moss 2025-07-10 19:39:18 -07:00 committed by GitHub
parent 0cf893cae1
commit 5923ab9524
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 34 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu
View File

@ -201,11 +201,10 @@ void run_blockwise_scaled_group_mm(
reinterpret_cast<typename ScheduleConfig::LayoutSFB*>( reinterpret_cast<typename ScheduleConfig::LayoutSFB*>(
layout_sfb.data_ptr())}; layout_sfb.data_ptr())};
cutlass::KernelHardwareInfo hw_info; int device_id = a_ptrs.device().index();
hw_info.device_id = a_ptrs.get_device(); static const cutlass::KernelHardwareInfo hw_info{
hw_info.sm_count = device_id, cutlass::KernelHardwareInfo::query_device_multiprocessor_count(
cutlass::KernelHardwareInfo::query_device_multiprocessor_count( device_id)};
hw_info.device_id);
// Epilogue Arguments // Epilogue Arguments
typename GemmKernel::EpilogueArguments epilogue_args{ typename GemmKernel::EpilogueArguments epilogue_args{

29
vllm/model_executor/layers/fused_moe/cutlass_moe.py
View File

@ -553,8 +553,10 @@ def cutlass_moe_fp4(a: torch.Tensor,
return out.to(dtype=out_dtype) return out.to(dtype=out_dtype)
def _valid_cutlass_block_scaled_grouped_gemm(w1: torch.Tensor, def _valid_cutlass_block_scaled_grouped_gemm(
w2: torch.Tensor) -> bool: w1: torch.Tensor, w2: torch.Tensor, inplace: bool, activation: str,
apply_router_weight_on_input: bool,
expert_map: Optional[torch.Tensor]) -> bool:
def _valid_cutlass_block_scaled_grouped_gemm_shape(N: int, K: int): def _valid_cutlass_block_scaled_grouped_gemm_shape(N: int, K: int):
return N % 128 == 0 and K % 128 == 0 return N % 128 == 0 and K % 128 == 0
@ -570,6 +572,29 @@ def _valid_cutlass_block_scaled_grouped_gemm(w1: torch.Tensor,
"CutlassBlockScaledGroupedGemm disabled: invalid weight dtype(s).") "CutlassBlockScaledGroupedGemm disabled: invalid weight dtype(s).")
return False return False
if expert_map is not None:
logger.debug(
"CutlassBlockScaledGroupedGemm disabled: expert_parallel is"
" not supported.")
return False
if activation != "silu":
logger.debug(
"CutlassBlockScaledGroupedGemm disabled: only activation silu is"
" supported.")
return False
if apply_router_weight_on_input:
logger.debug("CutlassBlockScaledGroupedGemm disabled:"
" apply_router_weight_on_input is not supported.")
return False
if inplace:
logger.debug(
"CutlassBlockScaledGroupedGemm disabled: inplace is not supported."
)
return False
return True return True

5
vllm/model_executor/layers/fused_moe/fused_moe.py
View File

@ -1192,8 +1192,9 @@ def fused_experts(
apply_router_weight_on_input=apply_router_weight_on_input, apply_router_weight_on_input=apply_router_weight_on_input,
) )
elif (allow_cutlass_block_scaled_grouped_gemm and use_fp8_w8a8 elif (allow_cutlass_block_scaled_grouped_gemm and use_fp8_w8a8
and _valid_cutlass_block_scaled_grouped_gemm(w1, w2)): and _valid_cutlass_block_scaled_grouped_gemm(
assert apply_router_weight_on_input is False w1, w2, inplace, activation, apply_router_weight_on_input,
expert_map)):
return run_cutlass_block_scaled_fused_experts( return run_cutlass_block_scaled_fused_experts(
a=hidden_states, a=hidden_states,
w1=w1, w1=w1,