biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Misc] DP : Add ExpertTokensMetadata (#20332)

Browse Source 
Signed-off-by: Varun <vsundarr@redhat.com>
Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Co-authored-by: Varun <vsundarr@redhat.com>
This commit is contained in:
Varun Sundar Rabindranath
2025-07-09 20:33:14 -04:00
committed by GitHub
parent b7d9e9416f
commit 805d62ca88
12 changed files with 117 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm/model_executor/layers/fused_moe/cutlass_moe.py
View File

@@ -303,11 +303,17 @@ class CutlassExpertsFp8(mk.FusedMoEPermuteExpertsUnpermute):
a2_scale: Optional[torch.Tensor],
workspace13: torch.Tensor,
workspace2: torch.Tensor,
expert_num_tokens: Optional[torch.Tensor],
expert_tokens_meta: Optional[mk.ExpertTokensMetadata],
):
assert w1_zp is None, "w1_zp is not supported in CUTLASS MoE"
assert w2_zp is None, "w2_zp is not supported in CUTLASS MoE"
expert_num_tokens = None
if expert_tokens_meta is not None:
expert_num_tokens = expert_tokens_meta.expert_num_tokens
activation_callable = lambda o, i: self.activation(activation, o, i)
in_dtype = hidden_states.dtype
run_cutlass_moe_fp8(
output, hidden_states, w1, w2, topk_ids, activation_callable,