Update flashinfer CUTLASS MoE Kernel (#21408)

Signed-off-by: Shu Wang. <shuw@nvidia.com>
This commit is contained in:
Shu Wang
2025-07-24 08:13:31 -07:00
committed by GitHub
parent e8cb0d0495
commit 1b25f1fe75
3 changed files with 8 additions and 8 deletions
@@ -11,7 +11,7 @@ from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.fused_moe.config import FusedMoEQuantConfig
from vllm.model_executor.layers.fused_moe.utils import (
extract_required_args, moe_kernel_quantize_input)
from vllm.utils.flashinfer import block_scale_interleave
from vllm.utils.flashinfer import nvfp4_block_scale_interleave
def get_local_sizes(local_tokens):
@@ -92,7 +92,7 @@ class FlashInferCutlassMoEPrepareAndFinalize(mk.FusedMoEPrepareAndFinalize):
dim=0,
sizes=get_local_sizes(local_tokens))
a1_m, a1_n = a1q.shape
a1q_scale = block_scale_interleave(a1q_scale)
a1q_scale = nvfp4_block_scale_interleave(a1q_scale)
return a1q, a1q_scale, None, topk_ids, topk_weights
@@ -1254,8 +1254,8 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
x, layer.w13_weight, layer.w2_weight), (
"Flashinfer CUTLASS Fused MoE not applicable!")
a1_gscale = torch.min(layer.w13_input_scale_quant)
a2_gscale = torch.min(layer.w2_input_scale_quant)
a1_gscale = layer.w13_input_scale_quant
a2_gscale = layer.w2_input_scale_quant
extra_expert_args = {
'g1_alphas': layer.g1_alphas,
'g2_alphas': layer.g2_alphas,