Merge branch 'v0.9.2-dev_rms_rope' into 'v0.9.2-dev'

perf(qwen3): 融合 q/k RMSNorm + RoPE

See merge request dcutoolkit/deeplearing/vllm!381

vllm/model_executor/models/qwen3.py

@@ -52,6 +52,7 @@ from .qwen2 import Qwen2MLP as Qwen3MLP
 from .qwen2 import Qwen2Model
 from .utils import AutoWeightsLoader, PPMissingLayer, maybe_prefix
 import vllm.envs as envs
+from vllm.utils import direct_register_custom_op
 
 logger = init_logger(__name__)
 
@@ -129,6 +130,58 @@ class Qwen3Attention(nn.Module):
         self.q_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
         self.k_norm = RMSNorm(self.head_dim, eps=rms_norm_eps)
 
+    def rms_rotary_embedding_fuse(
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: Optional[torch.Tensor],
+        head_size: int,
+        cos_sin_cache: torch.Tensor,
+        is_neox_style: bool,
+        q_weight: torch.Tensor,
+        k_weight: torch.Tensor,
+        q_bias: Optional[torch.Tensor],
+        k_bias: Optional[torch.Tensor],
+        epsilon: float,
+    ) -> None:
+        from lightop import rms_rotary_embedding_fuse as fused_kernel
+        fused_kernel(
+            positions,
+            query,
+            key,
+            head_size,
+            cos_sin_cache,
+            is_neox_style,
+            q_weight,
+            k_weight,
+            q_bias,
+            k_bias,
+            epsilon,
+        )
+
+    def rms_rotary_embedding_fuse_fake(
+        positions: torch.Tensor,
+        query: torch.Tensor,
+        key: Optional[torch.Tensor],
+        head_size: int,
+        cos_sin_cache: torch.Tensor,
+        is_neox_style: bool,
+        q_weight: torch.Tensor,
+        k_weight: torch.Tensor,
+        q_bias: Optional[torch.Tensor],
+        k_bias: Optional[torch.Tensor],
+        epsilon: float,
+    ) -> None:
+        # Fake impl intentionally left as no-op for graph tracing modes.
+        pass
+
+    if not hasattr(torch.ops.vllm, "rms_rotary_embedding_fuse"):
+        direct_register_custom_op(
+            op_name="rms_rotary_embedding_fuse",
+            op_func=rms_rotary_embedding_fuse,
+            mutates_args=["query", "key"],
+            fake_impl=rms_rotary_embedding_fuse_fake,
+        )
+
     def forward(
         self,
         positions: torch.Tensor,
@@ -136,7 +189,34 @@ 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)
-        # Add qk-norm
+        if envs.VLLM_USE_FUSED_RMS_ROPE:
+            # Fused RMSNorm + RoPE path through custom op.
+            cos_sin_cache = self.rotary_emb.cos_sin_cache
+            if (cos_sin_cache.device != q.device
+                    or cos_sin_cache.dtype != q.dtype):
+                cos_sin_cache = cos_sin_cache.to(q.device,
+                                                 dtype=q.dtype,
+                                                 non_blocking=True)
+                # Persist the converted cache so we don't re-copy/re-allocate
+                # on every forward when the original buffer starts on CPU.
+                self.rotary_emb.cos_sin_cache = cos_sin_cache
+            q = q.contiguous()
+            k = k.contiguous()
+            torch.ops.vllm.rms_rotary_embedding_fuse(
+                positions,
+                q,
+                k,
+                self.head_dim,
+                cos_sin_cache,
+                self.rotary_emb.is_neox_style,
+                self.q_norm.weight,
+                self.k_norm.weight,
+                None,
+                None,
+                self.q_norm.variance_epsilon,
+            )
+        else:
+            # Add qk-norm then RoPE (original path).
             q_by_head = q.view(*q.shape[:-1], q.shape[-1] // self.head_dim,
                                self.head_dim)
             if envs.VLLM_USE_APEX_RN: