sglang 0.4.9.post3__tar.gz → 0.4.9.post4__tar.gz

{sglang-0.4.9.post3/sglang.egg-info → sglang-0.4.9.post4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: sglang
-Version: 0.4.9.post3
+Version: 0.4.9.post4
 Summary: SGLang is yet another fast serving framework for large language models and vision language models.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -253,13 +253,13 @@ Requires-Dist: uvloop; extra == "runtime-common"
 Requires-Dist: xgrammar==0.1.21; extra == "runtime-common"
 Provides-Extra: srt
 Requires-Dist: sglang[runtime_common]; extra == "srt"
-Requires-Dist: sgl-kernel==0.2.6.post1; extra == "srt"
+Requires-Dist: sgl-kernel==0.2.7; extra == "srt"
 Requires-Dist: torch==2.7.1; extra == "srt"
 Requires-Dist: torchaudio==2.7.1; extra == "srt"
 Requires-Dist: torchvision==0.22.1; extra == "srt"
 Requires-Dist: cuda-python; extra == "srt"
 Requires-Dist: einops; extra == "srt"
-Requires-Dist: flashinfer_python==0.2.7.post1; extra == "srt"
+Requires-Dist: flashinfer_python==0.2.9rc1; extra == "srt"
 Provides-Extra: blackwell
 Requires-Dist: sglang[runtime_common]; extra == "blackwell"
 Requires-Dist: sgl-kernel; extra == "blackwell"
@@ -268,11 +268,11 @@ Requires-Dist: torchaudio==2.7.1; extra == "blackwell"
 Requires-Dist: torchvision==0.22.1; extra == "blackwell"
 Requires-Dist: cuda-python; extra == "blackwell"
 Requires-Dist: einops; extra == "blackwell"
-Requires-Dist: flashinfer_python==0.2.7.post1; extra == "blackwell"
+Requires-Dist: flashinfer_python==0.2.9rc1; extra == "blackwell"
 Provides-Extra: srt-hip
 Requires-Dist: sglang[runtime_common]; extra == "srt-hip"
 Requires-Dist: torch; extra == "srt-hip"
-Requires-Dist: petit_kernel; extra == "srt-hip"
+Requires-Dist: petit_kernel==0.0.2; extra == "srt-hip"
 Provides-Extra: srt-xpu
 Requires-Dist: sglang[runtime_common]; extra == "srt-xpu"
 Provides-Extra: srt-hpu

{sglang-0.4.9.post3 → sglang-0.4.9.post4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "sglang"
-version = "0.4.9.post3"
+version = "0.4.9.post4"
 description = "SGLang is yet another fast serving framework for large language models and vision language models."
 readme = "README.md"
 requires-python = ">=3.8"
@@ -54,13 +54,13 @@ runtime_common = [
 
 srt = [
     "sglang[runtime_common]",
-    "sgl-kernel==0.2.6.post1",
+    "sgl-kernel==0.2.7",
     "torch==2.7.1",
     "torchaudio==2.7.1",
     "torchvision==0.22.1",
     "cuda-python",
     "einops",
-    "flashinfer_python==0.2.7.post1",
+    "flashinfer_python==0.2.9rc1",
 ]
 
 blackwell = [
@@ -71,7 +71,7 @@ blackwell = [
     "torchvision==0.22.1",
     "cuda-python",
     "einops",
-    "flashinfer_python==0.2.7.post1",
+    "flashinfer_python==0.2.9rc1",
 ]
 
 # HIP (Heterogeneous-computing Interface for Portability) for AMD
@@ -79,7 +79,7 @@ blackwell = [
 srt_hip = [
     "sglang[runtime_common]",
     "torch",
-    "petit_kernel",
+    "petit_kernel==0.0.2",
 ]
 
 # xpu is not enabled in public vllm and torch whl,

{sglang-0.4.9.post3 → sglang-0.4.9.post4}/sglang/srt/_custom_ops.py

@@ -1,6 +1,6 @@
 # Adapted from https://github.com/vllm-project/vllm/blob/v0.6.4.post1/vllm/_custom_ops.py
 import logging
-from typing import List, Tuple
+from typing import List, Optional, Tuple
 
 import torch
 
@@ -114,6 +114,34 @@ else:
     def get_meta_buffer_ipc_handle(inp: torch.Tensor) -> torch.Tensor:
         return sgl_kernel.allreduce.get_meta_buffer_ipc_handle(inp)
 
+    # ROCM custom quick allreduce
+
+    def init_custom_qr(
+        rank: int, world_size: int, qr_max_size: Optional[int] = None
+    ) -> int:
+        return sgl_kernel.allreduce.init_custom_qr(world_size, rank, qr_max_size)
+
+    def qr_get_handle(fa: int) -> torch.Tensor:
+        return sgl_kernel.allreduce.qr_get_handle(fa)
+
+    def qr_open_handles(fa: int, handles: list[torch.Tensor]) -> None:
+        sgl_kernel.allreduce.qr_open_handles(fa, handles)
+
+    def qr_all_reduce(
+        fa: int,
+        inp: torch.Tensor,
+        out: torch.Tensor,
+        quant_level: int,
+        cast_bf2half: bool,
+    ) -> None:
+        sgl_kernel.allreduce.qr_all_reduce(fa, inp, out, quant_level, cast_bf2half)
+
+    def qr_destroy(fa: int) -> None:
+        sgl_kernel.allreduce.qr_destroy(fa)
+
+    def qr_max_size() -> int:
+        return sgl_kernel.allreduce.qr_max_size()
+
 
 def mscclpp_generate_unique_id() -> bytes:
     return sgl_kernel.allreduce.mscclpp_generate_unique_id()

{sglang-0.4.9.post3 → sglang-0.4.9.post4}/sglang/srt/configs/model_config.py

@@ -475,7 +475,7 @@ class ModelConfig:
 
     def get_hf_eos_token_id(self) -> Optional[Set[int]]:
         eos_ids = getattr(self.hf_config, "eos_token_id", None)
-        if eos_ids:
+        if eos_ids is not None:
             # it can be either int or list of int
             eos_ids = {eos_ids} if isinstance(eos_ids, int) else set(eos_ids)
         if eos_ids is None:

{sglang-0.4.9.post3 → sglang-0.4.9.post4}/sglang/srt/conversation.py

@@ -984,7 +984,7 @@ register_conv_template(
 
 @register_conv_template_matching_function
 def match_internvl(model_path: str):
-    if re.search(r"internvl2_5", model_path, re.IGNORECASE):
+    if re.search(r"internvl", model_path, re.IGNORECASE):
         return "internvl-2-5"
 
 

{sglang-0.4.9.post3 → sglang-0.4.9.post4}/sglang/srt/disaggregation/common/conn.py

@@ -23,7 +23,14 @@ from sglang.srt.disaggregation.base.conn import (
 )
 from sglang.srt.disaggregation.utils import DisaggregationMode
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils import get_free_port, get_ip, get_local_ip_by_remote
+from sglang.srt.utils import (
+    format_tcp_address,
+    get_free_port,
+    get_ip,
+    get_local_ip_by_remote,
+    is_valid_ipv6_address,
+    maybe_wrap_ipv6_address,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -65,11 +72,18 @@ class CommonKVManager(BaseKVManager):
     def _register_to_bootstrap(self):
         """Register KVSender to bootstrap server via HTTP POST."""
         if self.dist_init_addr:
-            ip_address = socket.gethostbyname(self.dist_init_addr.split(":")[0])
+            if self.dist_init_addr.startswith("["):  # [ipv6]:port or [ipv6]
+                if self.dist_init_addr.endswith("]"):
+                    host = self.dist_init_addr
+                else:
+                    host, _ = self.dist_init_addr.rsplit(":", 1)
+            else:
+                host = socket.gethostbyname(self.dist_init_addr.rsplit(":", 1)[0])
         else:
-            ip_address = get_ip()
+            host = get_ip()
+            host = maybe_wrap_ipv6_address(host)
 
-        bootstrap_server_url = f"{ip_address}:{self.bootstrap_port}"
+        bootstrap_server_url = f"{host}:{self.bootstrap_port}"
         url = f"http://{bootstrap_server_url}/route"
         payload = {
             "role": "Prefill",
@@ -92,8 +106,10 @@ class CommonKVManager(BaseKVManager):
             logger.error(f"Prefill Failed to register to bootstrap server: {e}")
 
     @cache
-    def _connect(self, endpoint: str):
+    def _connect(self, endpoint: str, is_ipv6: bool = False):
         socket = zmq.Context().socket(zmq.PUSH)
+        if is_ipv6:
+            socket.setsockopt(zmq.IPV6, 1)
         socket.connect(endpoint)
         return socket
 
@@ -263,15 +279,27 @@ class CommonKVReceiver(BaseKVReceiver):
             return None
 
     @classmethod
-    def _connect(cls, endpoint: str):
+    def _connect(cls, endpoint: str, is_ipv6: bool = False):
         with cls._global_lock:
             if endpoint not in cls._socket_cache:
                 sock = cls._ctx.socket(zmq.PUSH)
+                if is_ipv6:
+                    sock.setsockopt(zmq.IPV6, 1)
                 sock.connect(endpoint)
                 cls._socket_cache[endpoint] = sock
                 cls._socket_locks[endpoint] = threading.Lock()
             return cls._socket_cache[endpoint], cls._socket_locks[endpoint]
 
+    @classmethod
+    def _connect_to_bootstrap_server(cls, bootstrap_info: dict):
+        ip_address = bootstrap_info["rank_ip"]
+        port = bootstrap_info["rank_port"]
+        is_ipv6_address = is_valid_ipv6_address(ip_address)
+        sock, lock = cls._connect(
+            format_tcp_address(ip_address, port), is_ipv6=is_ipv6_address
+        )
+        return sock, lock
+
     def _register_kv_args(self):
         pass
 

{sglang-0.4.9.post3 → sglang-0.4.9.post4}/sglang/srt/disaggregation/mini_lb.py

@@ -17,6 +17,7 @@ from fastapi import FastAPI, HTTPException
 from fastapi.responses import ORJSONResponse, Response, StreamingResponse
 
 from sglang.srt.disaggregation.utils import PDRegistryRequest
+from sglang.srt.utils import maybe_wrap_ipv6_address
 
 AIOHTTP_STREAM_READ_CHUNK_SIZE = (
     1024 * 64
@@ -271,7 +272,7 @@ async def handle_generate_request(request_data: dict):
 
     # Parse and transform prefill_server for bootstrap data
     parsed_url = urllib.parse.urlparse(prefill_server)
-    hostname = parsed_url.hostname
+    hostname = maybe_wrap_ipv6_address(parsed_url.hostname)
     modified_request = request_data.copy()
 
     batch_size = _get_request_batch_size(modified_request)
@@ -309,7 +310,7 @@ async def _forward_to_backend(request_data: dict, endpoint_name: str):
 
     # Parse and transform prefill_server for bootstrap data
     parsed_url = urllib.parse.urlparse(prefill_server)
-    hostname = parsed_url.hostname
+    hostname = maybe_wrap_ipv6_address(parsed_url.hostname)
     modified_request = request_data.copy()
     modified_request.update(
         {

{sglang-0.4.9.post3 → sglang-0.4.9.post4}/sglang/srt/disaggregation/mooncake/conn.py

@@ -35,7 +35,15 @@ from sglang.srt.disaggregation.common.utils import (
 from sglang.srt.disaggregation.mooncake.transfer_engine import MooncakeTransferEngine
 from sglang.srt.disaggregation.utils import DisaggregationMode
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils import get_free_port, get_int_env_var, get_ip, get_local_ip_auto
+from sglang.srt.utils import (
+    format_tcp_address,
+    get_free_port,
+    get_int_env_var,
+    get_ip,
+    get_local_ip_auto,
+    is_valid_ipv6_address,
+    maybe_wrap_ipv6_address,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -148,6 +156,9 @@ class MooncakeKVManager(BaseKVManager):
         self.request_status: Dict[int, KVPoll] = {}
         self.rank_port = None
         self.server_socket = zmq.Context().socket(zmq.PULL)
+        if is_valid_ipv6_address(self.local_ip):
+            self.server_socket.setsockopt(zmq.IPV6, 1)
+
         self.register_buffer_to_engine()
         if self.disaggregation_mode == DisaggregationMode.PREFILL:
             self.transfer_infos: Dict[int, Dict[str, TransferInfo]] = {}
@@ -240,8 +251,10 @@ class MooncakeKVManager(BaseKVManager):
             self.engine.register(aux_data_ptr, aux_data_len)
 
     @cache
-    def _connect(self, endpoint: str):
+    def _connect(self, endpoint: str, is_ipv6: bool = False):
         socket = zmq.Context().socket(zmq.PUSH)
+        if is_ipv6:
+            socket.setsockopt(zmq.IPV6, 1)
         socket.connect(endpoint)
         return socket
 
@@ -471,9 +484,9 @@ class MooncakeKVManager(BaseKVManager):
     def sync_status_to_decode_endpoint(
         self, remote: str, dst_port: int, room: int, status: int, prefill_rank: int
     ):
-        if ":" in remote:
-            remote = remote.split(":")[0]
-        self._connect("tcp://" + remote + ":" + str(dst_port)).send_multipart(
+        self._connect(
+            format_tcp_address(remote, dst_port), is_ipv6=is_valid_ipv6_address(remote)
+        ).send_multipart(
             [
                 str(room).encode("ascii"),
                 str(status).encode("ascii"),
@@ -616,9 +629,12 @@ class MooncakeKVManager(BaseKVManager):
                     f"Transfer thread failed because of {e}. Prefill instance with bootstrap_port={self.bootstrap_port} is dead."
                 )
 
+    def _bind_server_socket(self):
+        self.server_socket.bind(format_tcp_address(self.local_ip, self.rank_port))
+
     def start_prefill_thread(self):
         self.rank_port = get_free_port()
-        self.server_socket.bind(f"tcp://{self.local_ip}:{self.rank_port}")
+        self._bind_server_socket()
 
         def bootstrap_thread():
             """This thread recvs pre-alloc notification from the decode engine"""
@@ -657,7 +673,7 @@ class MooncakeKVManager(BaseKVManager):
 
     def start_decode_thread(self):
         self.rank_port = get_free_port()
-        self.server_socket.bind(f"tcp://{self.local_ip}:{self.rank_port}")
+        self._bind_server_socket()
 
         def decode_thread():
             while True:
@@ -776,7 +792,7 @@ class MooncakeKVManager(BaseKVManager):
         # requests with the same dst_sessions will be added into the same
         # queue, which enables early abort with failed sessions.
         dst_infos = self.transfer_infos[bootstrap_room].keys()
-        session_port_sum = sum(int(session.split(":")[1]) for session in dst_infos)
+        session_port_sum = sum(int(session.rsplit(":", 1)[1]) for session in dst_infos)
         shard_idx = session_port_sum % len(self.transfer_queues)
 
         self.transfer_queues[shard_idx].put(
@@ -814,11 +830,18 @@ class MooncakeKVManager(BaseKVManager):
     def _register_to_bootstrap(self):
         """Register KVSender to bootstrap server via HTTP POST."""
         if self.dist_init_addr:
-            ip_address = socket.gethostbyname(self.dist_init_addr.split(":")[0])
+            if self.dist_init_addr.startswith("["):  # [ipv6]:port or [ipv6]
+                if self.dist_init_addr.endswith("]"):
+                    host = self.dist_init_addr
+                else:
+                    host, _ = self.dist_init_addr.rsplit(":", 1)
+            else:
+                host = socket.gethostbyname(self.dist_init_addr.rsplit(":", 1)[0])
         else:
-            ip_address = get_ip()
+            host = get_ip()
+            host = maybe_wrap_ipv6_address(host)
 
-        bootstrap_server_url = f"{ip_address}:{self.bootstrap_port}"
+        bootstrap_server_url = f"{host}:{self.bootstrap_port}"
         url = f"http://{bootstrap_server_url}/route"
         payload = {
             "role": "Prefill",
@@ -1163,9 +1186,6 @@ class MooncakeKVReceiver(BaseKVReceiver):
 
     def _register_kv_args(self):
         for bootstrap_info in self.bootstrap_infos:
-            self.prefill_server_url = (
-                f"{bootstrap_info['rank_ip']}:{bootstrap_info['rank_port']}"
-            )
             packed_kv_data_ptrs = b"".join(
                 struct.pack("Q", ptr) for ptr in self.kv_mgr.kv_args.kv_data_ptrs
             )
@@ -1179,7 +1199,7 @@ class MooncakeKVReceiver(BaseKVReceiver):
             dst_tp_size = str(tp_size).encode("ascii")
             dst_kv_item_len = str(kv_item_len).encode("ascii")
 
-            sock, lock = self._connect("tcp://" + self.prefill_server_url)
+            sock, lock = self._connect_to_bootstrap_server(bootstrap_info)
             with lock:
                 sock.send_multipart(
                     [
@@ -1196,23 +1216,32 @@ class MooncakeKVReceiver(BaseKVReceiver):
                 )
 
     @classmethod
-    def _connect(cls, endpoint: str):
+    def _connect(cls, endpoint: str, is_ipv6: bool = False):
         with cls._global_lock:
             if endpoint not in cls._socket_cache:
                 sock = cls._ctx.socket(zmq.PUSH)
+                if is_ipv6:
+                    sock.setsockopt(zmq.IPV6, 1)
                 sock.connect(endpoint)
                 cls._socket_cache[endpoint] = sock
                 cls._socket_locks[endpoint] = threading.Lock()
             return cls._socket_cache[endpoint], cls._socket_locks[endpoint]
 
+    @classmethod
+    def _connect_to_bootstrap_server(cls, bootstrap_info: dict):
+        ip_address = bootstrap_info["rank_ip"]
+        port = bootstrap_info["rank_port"]
+        is_ipv6_address = is_valid_ipv6_address(ip_address)
+        sock, lock = cls._connect(
+            format_tcp_address(ip_address, port), is_ipv6=is_ipv6_address
+        )
+        return sock, lock
+
     def init(self, kv_indices: npt.NDArray[np.int32], aux_index: Optional[int] = None):
         for bootstrap_info in self.bootstrap_infos:
-            self.prefill_server_url = (
-                f"{bootstrap_info['rank_ip']}:{bootstrap_info['rank_port']}"
-            )
+            sock, lock = self._connect_to_bootstrap_server(bootstrap_info)
             is_dummy = bootstrap_info["is_dummy"]
 
-            sock, lock = self._connect("tcp://" + self.prefill_server_url)
             with lock:
                 sock.send_multipart(
                     [

{sglang-0.4.9.post3 → sglang-0.4.9.post4}/sglang/srt/disaggregation/mooncake/transfer_engine.py

@@ -1,7 +1,7 @@
 import logging
 from typing import List, Optional
 
-from sglang.srt.utils import get_bool_env_var, get_free_port
+from sglang.srt.utils import get_bool_env_var, get_free_port, maybe_wrap_ipv6_address
 
 logger = logging.getLogger(__name__)
 
@@ -27,7 +27,9 @@ class MooncakeTransferEngine:
             hostname=self.hostname,
             device_name=self.ib_device,
         )
-        self.session_id = f"{self.hostname}:{self.engine.get_rpc_port()}"
+        self.session_id = (
+            f"{maybe_wrap_ipv6_address(self.hostname)}:{self.engine.get_rpc_port()}"
+        )
 
     def register(self, ptr, length):
         try:

{sglang-0.4.9.post3 → sglang-0.4.9.post4}/sglang/srt/disaggregation/nixl/conn.py

@@ -27,7 +27,11 @@ from sglang.srt.disaggregation.common.conn import (
 from sglang.srt.disaggregation.common.utils import group_concurrent_contiguous
 from sglang.srt.disaggregation.utils import DisaggregationMode
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils import get_local_ip_by_remote
+from sglang.srt.utils import (
+    format_tcp_address,
+    get_local_ip_auto,
+    is_valid_ipv6_address,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -124,7 +128,10 @@ class NixlKVManager(CommonKVManager):
                 "to run SGLang with NixlTransferEngine."
             ) from e
         self.agent = nixl_agent(str(uuid.uuid4()))
+        self.local_ip = get_local_ip_auto()
         self.server_socket = zmq.Context().socket(zmq.PULL)
+        if is_valid_ipv6_address(self.local_ip):
+            self.server_socket.setsockopt(zmq.IPV6, 1)
         self.register_buffer_to_engine()
 
         if self.disaggregation_mode == DisaggregationMode.PREFILL:
@@ -337,8 +344,11 @@ class NixlKVManager(CommonKVManager):
             return False
         return self.transfer_statuses[room].is_done()
 
+    def _bind_server_socket(self):
+        self.server_socket.bind(format_tcp_address(self.local_ip, self.rank_port))
+
     def _start_bootstrap_thread(self):
-        self.server_socket.bind(f"tcp://{get_local_ip_by_remote()}:{self.rank_port}")
+        self._bind_server_socket()
 
         def bootstrap_thread():
             """This thread recvs transfer info from the decode engine"""
@@ -452,23 +462,20 @@ class NixlKVReceiver(CommonKVReceiver):
 
     def init(self, kv_indices: npt.NDArray[np.int32], aux_index: Optional[int] = None):
         for bootstrap_info in self.bootstrap_infos:
-            self.prefill_server_url = (
-                f"{bootstrap_info['rank_ip']}:{bootstrap_info['rank_port']}"
-            )
             logger.debug(
                 f"Fetched bootstrap info: {bootstrap_info} for engine rank: {self.kv_mgr.kv_args.engine_rank}"
             )
+            sock, lock = self._connect_to_bootstrap_server(bootstrap_info)
             is_dummy = bootstrap_info["is_dummy"]
             logger.debug(
-                f"Sending to {self.prefill_server_url} with bootstrap room {self.bootstrap_room} {is_dummy=}"
+                f"Sending to prefill server with bootstrap room {self.bootstrap_room} {is_dummy=}"
             )
-            sock, lock = self._connect("tcp://" + self.prefill_server_url)
             with lock:
                 sock.send_multipart(
                     [
                         GUARD,
                         str(self.bootstrap_room).encode("ascii"),
-                        get_local_ip_by_remote().encode("ascii"),
+                        self.kv_mgr.local_ip.encode("ascii"),
                         str(self.kv_mgr.rank_port).encode("ascii"),
                         self.kv_mgr.agent.name.encode("ascii"),
                         kv_indices.tobytes() if not is_dummy else b"",
@@ -494,9 +501,7 @@ class NixlKVReceiver(CommonKVReceiver):
 
     def _register_kv_args(self):
         for bootstrap_info in self.bootstrap_infos:
-            self.prefill_server_url = (
-                f"{bootstrap_info['rank_ip']}:{bootstrap_info['rank_port']}"
-            )
+            sock, lock = self._connect_to_bootstrap_server(bootstrap_info)
             packed_kv_data_ptrs = b"".join(
                 struct.pack("Q", ptr) for ptr in self.kv_mgr.kv_args.kv_data_ptrs
             )
@@ -504,13 +509,12 @@ class NixlKVReceiver(CommonKVReceiver):
                 struct.pack("Q", ptr) for ptr in self.kv_mgr.kv_args.aux_data_ptrs
             )
 
-            sock, lock = self._connect("tcp://" + self.prefill_server_url)
             with lock:
                 sock.send_multipart(
                     [
                         GUARD,
                         "None".encode("ascii"),
-                        get_local_ip_by_remote().encode("ascii"),
+                        self.kv_mgr.local_ip.encode("ascii"),
                         str(self.kv_mgr.rank_port).encode("ascii"),
                         self.kv_mgr.agent.name.encode("ascii"),
                         self.kv_mgr.agent.get_agent_metadata(),

{sglang-0.4.9.post3 → sglang-0.4.9.post4}/sglang/srt/distributed/device_communicators/custom_all_reduce.py

@@ -4,18 +4,18 @@ import ctypes
 import logging
 import os
 from contextlib import contextmanager
-from functools import wraps
-from typing import Any, Callable, List, Optional, TypeVar, Union
+from typing import Any, List, Optional, Union
 
 import torch
 import torch.distributed as dist
 from torch.distributed import ProcessGroup
-from typing_extensions import ParamSpec
 
 from sglang.srt import _custom_ops as ops
 from sglang.srt.distributed.device_communicators.cuda_wrapper import CudaRTLibrary
 from sglang.srt.distributed.device_communicators.custom_all_reduce_utils import (
     gpu_p2p_access_check,
+    is_full_nvlink,
+    is_weak_contiguous,
 )
 from sglang.srt.distributed.parallel_state import in_the_same_node_as
 from sglang.srt.utils import is_cuda, is_hip
@@ -25,23 +25,6 @@ logger = logging.getLogger(__name__)
 _is_cuda = is_cuda()
 _is_hip = is_hip()
 
-if _is_cuda:
-    try:
-        import pynvml
-    except ImportError as e:
-        logger.warning("Failed to import pynvml with %r", e)
-
-if _is_hip:
-    try:
-        from amdsmi import (
-            AmdSmiException,
-            amdsmi_get_processor_handles,
-            amdsmi_init,
-            amdsmi_shut_down,
-            amdsmi_topo_get_link_type,
-        )
-    except ImportError as e:
-        logger.warning("Failed to import amdsmi with %r", e)
 
 try:
     if ops.use_vllm_custom_allreduce and not _is_hip:
@@ -57,70 +40,6 @@ except Exception:
 
 logger = logging.getLogger(__name__)
 
-_P = ParamSpec("_P")
-_R = TypeVar("_R")
-
-
-def with_nvml_context(fn: Callable[_P, _R]) -> Callable[_P, _R]:
-    @wraps(fn)
-    def wrapper(*args: _P.args, **kwargs: _P.kwargs) -> _R:
-        if _is_hip:
-            try:
-                amdsmi_init()
-                return fn(*args, **kwargs)
-            finally:
-                amdsmi_shut_down()
-        else:
-            pynvml.nvmlInit()
-            try:
-                return fn(*args, **kwargs)
-            finally:
-                pynvml.nvmlShutdown()
-
-    return wrapper
-
-
-@with_nvml_context
-def is_full_nvlink(physical_device_ids: List[int], world_size: int) -> bool:
-    if _is_hip:
-        """
-        query if the set of gpus are fully connected by xgmi (1 hop)
-        """
-        handles = [amdsmi_get_processor_handles()[i] for i in physical_device_ids]
-        for i, handle in enumerate(handles):
-            for j, peer_handle in enumerate(handles):
-                if i < j:
-                    try:
-                        link_type = amdsmi_topo_get_link_type(handle, peer_handle)
-                        # type is 2 for XGMI
-                        if link_type["hops"] != 1 or link_type["type"] != 2:
-                            return False
-                    except AmdSmiException as error:
-                        logger.error("AMD 1 hop XGMI detection failed.", exc_info=error)
-                        return False
-        return True
-    else:
-        """
-        query if the set of gpus are fully connected by nvlink (1 hop)
-        """
-        handles = [pynvml.nvmlDeviceGetHandleByIndex(i) for i in physical_device_ids]
-        for i, handle in enumerate(handles):
-            for j, peer_handle in enumerate(handles):
-                if i < j:
-                    try:
-                        p2p_status = pynvml.nvmlDeviceGetP2PStatus(
-                            handle, peer_handle, pynvml.NVML_P2P_CAPS_INDEX_NVLINK
-                        )
-                        if p2p_status != pynvml.NVML_P2P_STATUS_OK:
-                            return False
-                    except pynvml.NVMLError:
-                        logger.exception(
-                            "NVLink detection failed. This is normal if your"
-                            " machine has no NVLink equipped."
-                        )
-                        return False
-        return True
-
 
 def _can_p2p(rank: int, world_size: int) -> bool:
     # SGLANG_SKIP_P2P_CHECK can be set to False in sglang
@@ -136,13 +55,6 @@ def _can_p2p(rank: int, world_size: int) -> bool:
     return True
 
 
-def is_weak_contiguous(inp: torch.Tensor):
-    return inp.is_contiguous() or (
-        inp.storage().nbytes() - inp.storage_offset() * inp.element_size()
-        == inp.numel() * inp.element_size()
-    )
-
-
 class CustomAllreduce:
     _SUPPORTED_WORLD_SIZES = [2, 4, 6, 8]
     _MAX_CAR_SIZE = 8192 * 1024