diff --git a/python/sglang/multimodal_gen/runtime/layers/elementwise.py b/python/sglang/multimodal_gen/runtime/layers/elementwise.py index a53be8c753..4e8f6c64c2 100644 --- a/python/sglang/multimodal_gen/runtime/layers/elementwise.py +++ b/python/sglang/multimodal_gen/runtime/layers/elementwise.py @@ -37,7 +37,7 @@ class MulAdd(CustomOp): def forward_xpu( self, a: torch.Tensor, b: torch.Tensor, c: torch.Tensor, k: int = 0 ): - return self.forward_cuda(a, b, c, k=k) + return self.forward_native(a, b, c, k=k).to(a.dtype) @torch.compile def forward_musa( diff --git a/python/sglang/multimodal_gen/runtime/layers/layernorm.py b/python/sglang/multimodal_gen/runtime/layers/layernorm.py index 7b2d54c876..1585d1c528 100755 --- a/python/sglang/multimodal_gen/runtime/layers/layernorm.py +++ b/python/sglang/multimodal_gen/runtime/layers/layernorm.py @@ -915,10 +915,21 @@ class _NormScaleShift(CustomOp): # so we fall back to the native PyTorch implementation. return self.forward_native(*args, **kwargs) - def forward_xpu(self, *args, **kwargs): - # XPU does not support CUDA/CUTLASS-based fused kernels yet, - # so we fall back to the native PyTorch implementation. - return self.forward_native(*args, **kwargs) + def forward_xpu( + self, x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor + ) -> torch.Tensor: + normalized = self.norm(x) + if scale.dim() == 4: + # scale/shift: [batch_size, num_frames, 1, inner_dim] + num_frames = scale.shape[1] + frame_seqlen = normalized.shape[1] // num_frames + modulated = ( + normalized.unflatten(1, (num_frames, frame_seqlen)) * (1 + scale) + + shift + ).flatten(1, 2) + else: + modulated = normalized * (1 + scale) + shift + return modulated.to(x.dtype) @torch.compile(disable=current_platform.is_npu() or current_platform.is_rocm()) def forward_native(