From 3c01b4470070135585a46a617b75c8139fb3cb08 Mon Sep 17 00:00:00 2001
From: 1StepForever <wangww1Step@foxmail.com>
Date: Tue, 3 Mar 2026 14:56:06 +0800
Subject: [PATCH] [Fix] NPU deepep hccl buffer and fix IPC safe check (#17804)

---
 .../sglang/srt/distributed/parallel_state.py  | 24 ++++++++++++++++++-
 python/sglang/srt/utils/common.py             |  1 +
 2 files changed, 24 insertions(+), 1 deletion(-)

diff --git a/python/sglang/srt/distributed/parallel_state.py b/python/sglang/srt/distributed/parallel_state.py
index 5b8a82232..97a199d7d 100644
--- a/python/sglang/srt/distributed/parallel_state.py
+++ b/python/sglang/srt/distributed/parallel_state.py
@@ -69,6 +69,24 @@ TensorMetadata = namedtuple("TensorMetadata", ["device", "dtype", "size"])
 REDUCE_OP_SUM = int(torch.distributed.ReduceOp.SUM)
 
 
+def get_torch_distributed_pg_options(group_name=None):
+    if not _is_npu:
+        return None
+
+    # Only create HCCL options for default group or MoE-related groups
+    if group_name is not None and "moe" not in group_name:
+        return None
+
+    import torch_npu
+
+    options = torch_npu._C._distributed_c10d.ProcessGroupHCCL.Options()
+    hccl_buffer_size = int(
+        os.environ.get("DEEPEP_HCCL_BUFFSIZE") or os.environ.get("HCCL_BUFFSIZE") or 200
+    )
+    options.hccl_config = {"hccl_buffer_size": hccl_buffer_size}
+    return options
+
+
 @dataclass
 class GraphCaptureContext:
     stream: torch.get_device_module().Stream
@@ -274,8 +292,9 @@ class GroupCoordinator:
                     pg_options=MooncakeBackendOptions(active_ranks_cpu),
                 )
             else:
+                pg_options = get_torch_distributed_pg_options(group_name)
                 device_group = torch.distributed.new_group(
-                    ranks, backend=torch_distributed_backend
+                    ranks, backend=torch_distributed_backend, pg_options=pg_options
                 )
                 # a group with `gloo` backend, to allow direct coordination
                 # between processes through the CPU.
@@ -1629,6 +1648,8 @@ def init_distributed_environment(
             assert timeout > 0, "timeout must be positive"
             timeout = timedelta(seconds=timeout)
 
+        pg_options = get_torch_distributed_pg_options()
+
         # this backend is used for WORLD
         torch.distributed.init_process_group(
             backend=backend,
@@ -1636,6 +1657,7 @@ def init_distributed_environment(
             world_size=world_size,
             rank=rank,
             timeout=timeout,
+            pg_options=pg_options,
         )
 
     # set the local rank
diff --git a/python/sglang/srt/utils/common.py b/python/sglang/srt/utils/common.py
index 2e76227c4..53a8c6762 100644
--- a/python/sglang/srt/utils/common.py
+++ b/python/sglang/srt/utils/common.py
@@ -2359,6 +2359,7 @@ class SafeUnpickler(pickle.Unpickler):
         "sglang.srt.model_executor.model_runner.",
         "sglang.srt.layers.",
         "sglang.srt.utils.",
+        "torch_npu.",
     }
 
     DENY_CLASSES = {