[AMD][diffusion] fix: disable layernorm torch.compile decorator in eager mode on ROCm to avoid memory-access fault (#29673)

python/sglang/multimodal_gen/runtime/layers/layernorm.py

@@ -366,7 +366,10 @@ class LayerNorm(CustomOp):
         x = x.view(-1, self.hidden_size)
         return self.forward_triton(x).view(shape)
 
-    @torch.compile(backend="inductor", disable=current_platform.is_npu())
+    @torch.compile(
+        backend="inductor",
+        disable=current_platform.is_npu() or current_platform.is_rocm(),
+    )
     def forward_native(
         self,
         x: torch.Tensor,
@@ -566,7 +569,7 @@ class _ScaleResidualNormScaleShift(CustomOp):
         # so we fall back to the native PyTorch implementation.
         return self.forward_native(*args, **kwargs)
 
-    @torch.compile(disable=current_platform.is_npu())
+    @torch.compile(disable=current_platform.is_npu() or current_platform.is_rocm())
     def forward_native(
         self,
         residual: torch.Tensor,
@@ -734,7 +737,7 @@ class _NormScaleShift(CustomOp):
         # so we fall back to the native PyTorch implementation.
         return self.forward_native(*args, **kwargs)
 
-    @torch.compile(disable=current_platform.is_npu())
+    @torch.compile(disable=current_platform.is_npu() or current_platform.is_rocm())
     def forward_native(
         self, x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor
     ) -> torch.Tensor:
@@ -831,7 +834,7 @@ class _NormTanhMulAdd(CustomOp):
         # Fallback to native because ROCm does not support CuTeDSL.
         return self.forward_native(*args, **kwargs)
 
-    @torch.compile(disable=current_platform.is_npu())
+    @torch.compile(disable=current_platform.is_npu() or current_platform.is_rocm())
     def forward_native(
         self, x: torch.Tensor, scale: torch.Tensor, shift: torch.Tensor
     ) -> torch.Tensor: