[Fix] Enable graph capture and MSCCL++ for attention TP groups (#31629)

Co-authored-by: Cheng Wan <54331508+ch-wan@users.noreply.github.com>

python/sglang/srt/distributed/parallel_state.py

@@ -1971,7 +1971,7 @@ def graph_capture(stream=None):
     ):
         with contextlib.ExitStack() as stack:
             seen = {id(_TP), id(_PP)}
-            for group in (_DCP, _MOE_EP, _MOE_TP):
+            for group in (_DCP, _ATTN_TP, _MOE_EP, _MOE_TP):
                 if group is not None and id(group) not in seen:
                     seen.add(id(group))
                     stack.enter_context(group.graph_capture(context))
@@ -2404,7 +2404,6 @@ def initialize_model_parallel(
             get_world_group().local_rank,
             backend,
             use_pynccl=SYNC_TOKEN_IDS_ACROSS_TP or enable_symm_mem,
-            use_mscclpp_allreduce=False,
             use_custom_allreduce=False,
             use_torch_symm_mem_allreduce=False,
             use_message_queue_broadcaster=envs.SGLANG_USE_MESSAGE_QUEUE_BROADCASTER.get(),