[Codex] Diffusion handle non-contiguous CFG communication (#24332)

Co-authored-by: BBuf Codex <bbuf-codex@users.noreply.github.com>

python/sglang/multimodal_gen/runtime/distributed/communication_op.py

@@ -61,4 +61,6 @@ def cfg_model_parallel_all_reduce(
     op: torch._C._distributed_c10d.ReduceOp = torch._C._distributed_c10d.ReduceOp.SUM,
 ) -> torch.Tensor:
     """All-reduce the input tensor across CFG parallel group."""
+    if not input_.is_contiguous():
+        input_ = input_.contiguous()
     return get_cfg_group().all_reduce(input_, op=op)

python/sglang/multimodal_gen/runtime/distributed/group_coordinator.py

@@ -435,6 +435,8 @@ class GroupCoordinator:
         if self.world_size == 1:
             return input_
         # Broadcast.
+        if not input_.is_contiguous():
+            input_ = input_.contiguous()
         torch.distributed.broadcast(
             input_,
             src=self.ranks[src],