diff --git a/python/sglang/multimodal_gen/runtime/distributed/group_coordinator.py b/python/sglang/multimodal_gen/runtime/distributed/group_coordinator.py
index d4d502ac4..85290850d 100644
--- a/python/sglang/multimodal_gen/runtime/distributed/group_coordinator.py
+++ b/python/sglang/multimodal_gen/runtime/distributed/group_coordinator.py
@@ -441,7 +441,10 @@ class GroupCoordinator:
         if current_platform.is_cpu() and is_shm_available(
             input_.dtype, self.world_size, len(self.ranks)
         ):
-            return torch.ops.sgl_kernel.shm_allgather(input_, dim)
+            output_tensor = torch.ops.sgl_kernel.shm_allgather(input_, dim)
+            if separate_tensors:
+                return list(output_tensor.chunk(world_size, dim=dim))
+            return output_tensor
         else:
             all_gather_single(output_tensor, input_, group=self.device_group)
 
diff --git a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/denoising.py b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/denoising.py
index a0211526f..93ec9c2f2 100644
--- a/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/denoising.py
+++ b/python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/minimax_h3/stages/denoising.py
@@ -676,11 +676,12 @@ class MiniMaxH3DenoisingStage(DenoisingStage):
 
         if not (
             current_platform.is_cuda()
+            or current_platform.is_cpu()
             or current_platform.is_mps()
             or current_platform.is_npu()
         ):
             raise RuntimeError(
-                "MiniMax H3 full-loop denoise requires CUDA, MPS, or Ascend NPU"
+                "MiniMax H3 full-loop denoise requires CPU, CUDA, MPS, or Ascend NPU"
             )
 
         device = current_platform.get_local_torch_device()