From 739096a028214991cff4af23a558795905ad35a1 Mon Sep 17 00:00:00 2001 From: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Date: Fri, 29 May 2026 14:55:00 -0400 Subject: [PATCH] [Bug] Fix torch device issue for MOE permute (#44005) Signed-off-by: yewentao256 --- vllm/model_executor/layers/fused_moe/moe_permute_unpermute.py | 3 +++ 1 file changed, 3 insertions(+) diff --git a/vllm/model_executor/layers/fused_moe/moe_permute_unpermute.py b/vllm/model_executor/layers/fused_moe/moe_permute_unpermute.py index df43068943..ad9fb509a1 100644 --- a/vllm/model_executor/layers/fused_moe/moe_permute_unpermute.py +++ b/vllm/model_executor/layers/fused_moe/moe_permute_unpermute.py @@ -74,6 +74,9 @@ class MoEPermuteScratch: self.sort_workspace = torch.empty( sorter_size, dtype=torch.int8, device=self.device ) + # torch.device("cuda") in config, after initialized, + # will be changed to cuda:{index}, so we need to refresh here. + self.device = self.token_expert_indices.device def validate(self, hidden_states: torch.Tensor, topk_ids: torch.Tensor) -> None: n_token, n_hidden = hidden_states.shape