Optimize attention in llama4 (#5127)

python/sglang/srt/models/llama4.py

@@ -240,9 +240,13 @@ class Llama4Attention(nn.Module):
     def _get_attn_scale(self, positions: torch.Tensor) -> torch.Tensor:
         floor = torch.floor((positions + 1.0) / self.floor_scale)
         attn_scale = torch.log(floor + 1.0) * self.attn_scale + 1.0
-
         return attn_scale.unsqueeze(-1)
 
+    @torch.compile(dynamic=True, backend=get_compiler_backend())
+    def _mul_attn_scale(self, positions, q):
+        attn_scale = self._get_attn_scale(positions)
+        return (q * attn_scale).to(q.dtype)
+
     def forward(
         self,
         positions: torch.Tensor,
@@ -250,27 +254,29 @@ class Llama4Attention(nn.Module):
         forward_batch: ForwardBatch,
     ) -> torch.Tensor:
         qkv, _ = self.qkv_proj(hidden_states)
-        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+
+        qk, v = qkv.split([self.q_size + self.kv_size, self.kv_size], dim=-1)
 
         if self.rotary_emb is not None:
-            q, k = self.rotary_emb(positions, q, k)
+            q_view, k_view = qk.split([self.q_size, self.kv_size], dim=-1)
+            q_out_unused, k_out_unused = self.rotary_emb(positions, q_view, k_view)
+            assert (q_out_unused is q_view) and (k_out_unused is k_view)
+            del q_view, k_view, q_out_unused, k_out_unused
 
         if self.qk_norm is not None:
-            # TODO: support float
-            q = q.reshape(-1, self.head_dim).contiguous().bfloat16()
-            k = k.reshape(-1, self.head_dim).contiguous().bfloat16()
-            q = self.qk_norm(q).to(q.dtype)
-            k = self.qk_norm(k).to(k.dtype)
-            q = q.reshape(-1, self.q_size)
-            k = k.reshape(-1, self.kv_size)
+            # TODO there are still 2 redundant direct_copy_kernel_cuda for this `reshape` and (in attn backend) q.contiguous(), maybe we can fuse them later
+            qk = qk.reshape(-1, self.head_dim).contiguous().bfloat16()
+            qk = self.qk_norm(qk).to(torch.bfloat16)
+            qk = qk.reshape(-1, self.q_size + self.kv_size)
+
+        q, k = qk.split([self.q_size, self.kv_size], dim=-1)
 
         # We are applying temperature tuning (https://arxiv.org/abs/2501.19399) to NoPE layers, where
         # the inference-time temperature tuning function is customized to not affect short context
         # while working at very long context
         # https://arxiv.org/abs/2501.19399
         if self.attn_temperature_tuning and not self.use_rope:
-            attn_scale = self._get_attn_scale(positions)
-            q = (q * attn_scale).to(q.dtype)
+            q = self._mul_attn_scale(positions=positions, q=q)
 
         attn_output = self.attn(q, k, v, forward_batch)
         output, _ = self.o_proj(attn_output)