From b254e0456c9851414cf9aada98a9129235614429 Mon Sep 17 00:00:00 2001 From: Woosuk Kwon Date: Tue, 2 Jun 2026 17:54:27 -0700 Subject: [PATCH] [DSV4] Minor cleanup for DeepseekV4MegaMoEExperts (#44367) Signed-off-by: Woosuk Kwon --- vllm/models/deepseek_v4/nvidia/model.py | 19 +------------------ 1 file changed, 1 insertion(+), 18 deletions(-) diff --git a/vllm/models/deepseek_v4/nvidia/model.py b/vllm/models/deepseek_v4/nvidia/model.py index 33baed7279..bd1ccc4cfb 100644 --- a/vllm/models/deepseek_v4/nvidia/model.py +++ b/vllm/models/deepseek_v4/nvidia/model.py @@ -420,25 +420,7 @@ class DeepseekV4MegaMoEExperts(nn.Module): f"but the symmetric buffer was sized for {self.max_num_tokens}." ) y = torch.empty_like(hidden_states, dtype=torch.bfloat16) - self._run_mega_moe( - hidden_states, - topk_weights, - topk_ids, - y, - activation_clamp, - fast_math, - ) - return y - def _run_mega_moe( - self, - hidden_states: torch.Tensor, - topk_weights: torch.Tensor, - topk_ids: torch.Tensor, - y: torch.Tensor, - activation_clamp: float | None, - fast_math: bool, - ) -> None: from vllm.utils.deep_gemm import _import_deep_gemm deep_gemm = _import_deep_gemm() @@ -484,6 +466,7 @@ class DeepseekV4MegaMoEExperts(nn.Module): activation_clamp=activation_clamp, fast_math=fast_math, ) + return y DeepseekV4MegaMoEExperts.weight_loader.supports_moe_loading = True # type: ignore[attr-defined]