[Qwen3.8] Support Qwen3.8 on branch v0.5.15

python/sglang/srt/configs/__init__.py

@@ -30,7 +30,12 @@
 )
 from sglang.srt.configs.nemotron_h import NemotronHConfig, NemotronHPuzzleConfig
 from sglang.srt.configs.olmo3 import Olmo3Config
-from sglang.srt.configs.qwen3_5 import Qwen3_5Config, Qwen3_5MoeConfig
+from sglang.srt.configs.qwen3_5 import (
+    Qwen3_5Config,
+    Qwen3_5MoeConfig,
+    Qwen3_5MoeTextConfig,
+    Qwen3_5TextConfig,
+)
 from sglang.srt.configs.qwen3_asr import Qwen3ASRConfig
 from sglang.srt.configs.qwen3_next import Qwen3NextConfig
 from sglang.srt.configs.step3_vl import (
@@ -64,6 +69,8 @@
     "Qwen3NextConfig",
     "Qwen3_5Config",
     "Qwen3_5MoeConfig",
+    "Qwen3_5TextConfig",
+    "Qwen3_5MoeTextConfig",
     "InternS2PreviewConfig",
     "DotsVLMConfig",
     "DotsOCRConfig",

python/sglang/srt/layers/layernorm.py

@@ -459,6 +459,9 @@ def forward_hip(
         residual: Optional[torch.Tensor] = None,
         post_residual_addition: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        if _use_aiter:
+            return self.forward_aiter(x, residual, post_residual_addition)
+
         # Fallback to native implementation if vllm is not available
         if not _has_vllm_rms_norm:
             return self.forward_native(x, residual, post_residual_addition)
@@ -480,14 +483,10 @@ def forward_hip(
             # NOTE: Remove this if aiter kernel supports discontinuous input
             x = x.contiguous()
         if residual is not None:
-            out = torch.empty_like(x)
-            residual_out = torch.empty_like(x)
             if post_residual_addition is not None:
                 residual = residual + post_residual_addition
-            fused_add_rms_norm(
-                out, x, residual_out, residual, self.weight.data, self.variance_epsilon
-            )
-            return out, residual_out
+            fused_add_rms_norm(x, residual, self.weight.data, self.variance_epsilon)
+            return x, residual
         out = torch.empty_like(x)
         rms_norm(out, x, self.weight.data, self.variance_epsilon)
         return out
@@ -788,6 +787,16 @@ def forward_hip(
         residual: Optional[torch.Tensor] = None,
         post_residual_addition: Optional[torch.Tensor] = None,
     ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        if x.numel() == 0:
+            if residual is not None:
+                if post_residual_addition is not None:
+                    residual = residual + post_residual_addition
+                return x, residual
+            return x
+
+        if is_batch_invariant_mode_enabled():
+            return self.forward_native(x, residual, post_residual_addition)
+
         if not _has_vllm_rms_norm:
             return self.forward_native(x, residual, post_residual_addition)
 
@@ -807,18 +816,14 @@ def forward_hip(
             return rms_norm(x, w, self.variance_epsilon)
         else:
             # vllm API: rms_norm(out, input, weight, eps) -> None (in-place)
-            #           fused_add_rms_norm(out, input, residual_out, residual, weight, eps)
+            #           fused_add_rms_norm(input, residual, weight, eps) -> None
             if not x.is_contiguous():
                 x = x.contiguous()
             if residual is not None:
-                out = torch.empty_like(x)
-                residual_out = torch.empty_like(x)
                 if post_residual_addition is not None:
                     residual = residual + post_residual_addition
-                fused_add_rms_norm(
-                    out, x, residual_out, residual, w, self.variance_epsilon
-                )
-                return out, residual_out
+                fused_add_rms_norm(x, residual, w, self.variance_epsilon)
+                return x, residual
             out = torch.empty_like(x)
             rms_norm(out, x, w, self.variance_epsilon)
             return out