[Perf] Change Trtllm fp8 MoE to use Shuffled Weights and BlockMajorK Layout (#38993)

Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
Wei Zhao
2026-04-05 10:54:31 -04:00
committed by GitHub
co-authored by Robert Shaw
parent 228023b3a5
commit 1af6f78ae5
3 changed files with 72 additions and 13 deletions
@@ -13,7 +13,7 @@ import vllm.envs as envs
from vllm.distributed.parallel_state import get_dp_group, is_global_first_rank
from vllm.model_executor.layers.fused_moe.deep_gemm_moe import DeepGemmExperts
from vllm.model_executor.layers.fused_moe.deep_gemm_utils import compute_aligned_M
from vllm.model_executor.layers.fused_moe.layer import FusedMoE, FusedMoEModularMethod
from vllm.model_executor.layers.fused_moe.layer import FusedMoE
from vllm.model_executor.layers.fused_moe.triton_deep_gemm_moe import (
TritonOrDeepGemmExperts,
)
@@ -168,14 +168,12 @@ def _fused_moe_grouped_gemm_may_use_deep_gemm(module: torch.nn.Module) -> bool:
):
return False
if not isinstance(module.quant_method, FusedMoEModularMethod):
# modular kernels could invoke deep_gemm_moe_fp8
return True
moe_kernel = getattr(module.quant_method, "moe_kernel", None)
if moe_kernel is None:
return False
# Further check if the ModularKernel implementation uses the DeepGemmExperts
return isinstance(
module.quant_method.moe_kernel, (DeepGemmExperts, TritonOrDeepGemmExperts)
)
fused_experts = moe_kernel.impl.fused_experts
return isinstance(fused_experts, (DeepGemmExperts, TritonOrDeepGemmExperts))
FP8_GEMM_NT_WARMUP_CACHE: set[torch.Size] = set()