Merge branch 'v0.15.1-dev' of http://10.16.6.30/dcutoolkit/deeplearing/vllm into v0.15.1-dev

vllm/model_executor/models/qwen3.py

@@ -196,7 +196,7 @@ class Qwen3Attention(nn.Module):
     ) -> torch.Tensor:
         qkv, _ = self.qkv_proj(hidden_states)
         q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
-        if envs.VLLM_USE_FUSED_RMS_ROPE:
+        if envs.VLLM_USE_FUSED_RMS_ROPE and positions.ndim == 1:
             # Fused RMSNorm + RoPE path through custom op.
             cos_sin_cache = self.rotary_emb.cos_sin_cache
             if (cos_sin_cache.device != q.device

vllm/model_executor/models/qwen3_moe.py

@@ -444,9 +444,9 @@ class Qwen3MoeAttention(nn.Module):
             mrope_interleaved,
             q_weight,
             k_weight,
-            epsilon,
             q_residual,
             k_residual,
+            epsilon,
         )
 
     def rms_mrope_fuse_fake(