From f3c7941ec8d3743718ae4513bd98612f90e5d840 Mon Sep 17 00:00:00 2001 From: Kai Song <50285351+USTCKAY@users.noreply.github.com> Date: Thu, 9 Apr 2026 05:47:48 +0800 Subject: [PATCH] [Bugfix]Fix EP precision for Qwen3.5, Qwen3-Next (#39181) Signed-off-by: Song Kai --- vllm/model_executor/models/qwen2_moe.py | 3 +++ vllm/model_executor/models/qwen3_next.py | 1 + 2 files changed, 4 insertions(+) diff --git a/vllm/model_executor/models/qwen2_moe.py b/vllm/model_executor/models/qwen2_moe.py index 4b0c756165..750835cccd 100644 --- a/vllm/model_executor/models/qwen2_moe.py +++ b/vllm/model_executor/models/qwen2_moe.py @@ -80,6 +80,7 @@ class Qwen2MoeMLP(nn.Module): quant_config: QuantizationConfig | None = None, reduce_results: bool = True, expert_gate: torch.nn.Linear | None = None, + is_sequence_parallel: bool = False, prefix: str = "", ) -> None: super().__init__() @@ -88,6 +89,7 @@ class Qwen2MoeMLP(nn.Module): [intermediate_size] * 2, bias=False, quant_config=quant_config, + disable_tp=is_sequence_parallel, prefix=f"{prefix}.gate_up_proj", ) self.down_proj = RowParallelLinear( @@ -96,6 +98,7 @@ class Qwen2MoeMLP(nn.Module): bias=False, quant_config=quant_config, reduce_results=reduce_results, + disable_tp=is_sequence_parallel, prefix=f"{prefix}.down_proj", ) if hidden_act != "silu": diff --git a/vllm/model_executor/models/qwen3_next.py b/vllm/model_executor/models/qwen3_next.py index 6cf386cc8b..3bd026d9dc 100644 --- a/vllm/model_executor/models/qwen3_next.py +++ b/vllm/model_executor/models/qwen3_next.py @@ -140,6 +140,7 @@ class Qwen3NextSparseMoeBlock(nn.Module): quant_config=quant_config, reduce_results=False, expert_gate=self.shared_expert_gate, + is_sequence_parallel=self.is_sequence_parallel, prefix=f"{prefix}.shared_expert", ) else: