support 1 shot allreduce in 1-node and 2-node using mscclpp (#6277)

benchmark/kernels/all_reduce/benchmark_mscclpp.py

+"""For Now, MSCCL is only supported on TP16 and TP8 case
+
+export WORLD_SIZE=1
+export RANK=0
+export MASTER_ADDR=127.0.0.1
+export MASTER_PORT=12345
+
+torchrun --nproc_per_node gpu \
+--nnodes $WORLD_SIZE \
+--node_rank $RANK \
+--master_addr $MASTER_ADDR \
+--master_port $MASTER_PORT benchmark/kernels/all_reduce/benchmark_mscclpp.py
+"""
+
+import os
+from contextlib import nullcontext
+from typing import List
+
+import torch
+import torch.distributed as dist
+from torch.distributed import ProcessGroup
+
+from sglang.srt.distributed import init_distributed_environment
+from sglang.srt.distributed.device_communicators.pymscclpp import PyMscclppCommunicator
+from sglang.srt.distributed.device_communicators.pynccl import PyNcclCommunicator
+from sglang.srt.distributed.parallel_state import (
+    get_tensor_model_parallel_group,
+    graph_capture,
+    initialize_model_parallel,
+    set_mscclpp_all_reduce,
+)
+
+
+def torch_allreduce(torch_input: torch.Tensor, group: ProcessGroup) -> torch.Tensor:
+    dist.all_reduce(torch_input, group=group)
+    return torch_input
+
+
+def msccl_allreduce(
+    msccl_input: torch.Tensor, msccl_comm: PyMscclppCommunicator
+) -> torch.Tensor:
+    return msccl_comm.all_reduce(msccl_input)
+
+
+def pynccl_allreduce(
+    msccl_input: torch.Tensor, pynccl_comm: PyNcclCommunicator
+) -> torch.Tensor:
+    pynccl_comm.all_reduce(msccl_input)
+    return msccl_input
+
+
+def _bench_graph_time(func, inp_randn, warmup_loop=2, graph_loop=10, test_loop=10):
+    graph_input = inp_randn.clone()
+    with graph_capture() as graph_capture_context:
+        graph = torch.cuda.CUDAGraph()
+        with torch.cuda.graph(graph, stream=graph_capture_context.stream):
+            for _ in range(graph_loop):
+                graph_out = func(graph_input)
+
+    graph.replay()
+    func_output = graph_out.clone()
+
+    for _ in range(warmup_loop):
+        graph.replay()
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    latencies: List[float] = []
+    for _ in range(test_loop):
+        torch.cuda.synchronize()
+        dist.barrier()
+        start_event.record()
+        graph.replay()
+        end_event.record()
+        end_event.synchronize()
+        latencies.append(start_event.elapsed_time(end_event))
+    func_cost_us = sum(latencies) / len(latencies) / graph_loop * 1000
+    graph.reset()
+    return func_output, func_cost_us
+
+
+def _bench_eager_time(func, inp_randn, warmup_loop=2, test_loop=10):
+    eager_input = inp_randn.clone()
+    eager_output = func(eager_input)
+    func_output = eager_output.clone()
+
+    for _ in range(warmup_loop):
+        func(eager_input)
+    torch.cuda.synchronize()
+
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+    torch.cuda.synchronize()
+    start_event.record()
+    for _ in range(test_loop):
+        func(eager_input)
+    end_event.record()
+    torch.cuda.synchronize()
+    func_cost_us = start_event.elapsed_time(end_event) / test_loop * 1000
+
+    return func_output, func_cost_us
+
+
+def get_torch_prof_ctx(do_prof: bool):
+    ctx = (
+        torch.profiler.profile(
+            activities=[
+                torch.profiler.ProfilerActivity.CPU,
+                torch.profiler.ProfilerActivity.CUDA,
+            ],
+            record_shapes=True,
+            with_stack=True,
+        )
+        if do_prof
+        else nullcontext()
+    )
+    return ctx
+
+
+def human_readable_size(size, decimal_places=1):
+    for unit in ["B", "KiB", "MiB", "GiB", "TiB", "PiB"]:
+        if size < 1024.0 or unit == "PiB":
+            break
+        size /= 1024.0
+    return f"{size:.{decimal_places}f} {unit}"
+
+
+try:
+    from tabulate import tabulate
+except ImportError:
+    print("tabulate not installed, skipping table printing")
+    tabulate = None
+
+
+def print_markdown_table(data):
+    if tabulate is not None:
+        print(tabulate(data, headers="keys", tablefmt="github"))
+        return
+    headers = data[0].keys()
+    header_row = "| " + " | ".join(headers) + " |"
+    separator = "| " + " | ".join(["---"] * len(headers)) + " |"
+    rows = []
+    for item in data:
+        row = "| " + " | ".join(str(item[key]) for key in headers) + " |"
+        rows.append(row)
+    markdown_table = "\n".join([header_row, separator] + rows)
+    print(markdown_table)
+
+
+if __name__ == "__main__":
+    import logging
+
+    logging.basicConfig(
+        level=logging.INFO,
+        format="%(asctime)s - %(levelname)s - %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+        force=True,
+    )
+    if not dist.is_initialized():
+        dist.init_process_group(backend="nccl")
+    world, world_size = dist.group.WORLD, dist.get_world_size()
+    rank = dist.get_rank()
+    torch.cuda.set_device(rank % 8)
+    device = torch.cuda.current_device()
+    set_mscclpp_all_reduce(True)
+    init_distributed_environment(
+        world_size=world_size,
+        rank=rank,
+        local_rank=rank % 8,
+    )
+    initialize_model_parallel(tensor_model_parallel_size=world_size)
+    group = get_tensor_model_parallel_group().device_group
+    cpu_group = get_tensor_model_parallel_group().cpu_group
+    pynccl_comm = get_tensor_model_parallel_group().pynccl_comm
+    pymscclpp_comm = get_tensor_model_parallel_group().pymscclpp_comm
+    dist.barrier()
+    profile = False
+    dtype = torch.bfloat16
+    ctx = get_torch_prof_ctx(profile)
+    result = []
+
+    with ctx:
+        for i in range(10, 20):
+            sz = 2**i
+            if sz * dtype.itemsize > 2**20:
+                break
+            inp_randn = torch.randint(1, 16, (sz,), dtype=dtype, device=device)
+
+            memory = torch.empty_like(inp_randn)
+            memory_out = torch.empty_like(memory)
+            torch_eager_output, torch_eager_time = _bench_eager_time(
+                lambda inp: torch_allreduce(inp, group), inp_randn
+            )
+            msccl_eager_output, msccl_eager_time = _bench_eager_time(
+                lambda inp: msccl_allreduce(inp, pymscclpp_comm), inp_randn
+            )
+            msccl_graph_output, msccl_graph_time = _bench_graph_time(
+                lambda inp: msccl_allreduce(inp, pymscclpp_comm), inp_randn
+            )
+            # since pynccl is inplace op, this return result is not correct if graph loop > 1
+            _, pynccl_graph_time = _bench_graph_time(
+                lambda inp: pynccl_allreduce(inp, pynccl_comm), inp_randn
+            )
+            torch.testing.assert_close(torch_eager_output, msccl_graph_output)
+            torch.testing.assert_close(torch_eager_output, msccl_eager_output)
+            result.append(
+                {
+                    "msg_size": human_readable_size(inp_randn.nbytes),
+                    "torch eager time": torch_eager_time,
+                    "msccl eager time": msccl_eager_time,
+                    "msccl graph time": msccl_graph_time,
+                    "pynccl graph time": pynccl_graph_time,
+                }
+            )
+            if rank == 0:
+                print(f"sz={sz}, dtype={dtype}: correctness check PASS!")
+    if rank == 0:
+        print_markdown_table(result)
+    if profile:
+        prof_dir = f"prof/msccl"
+        os.makedirs(prof_dir, exist_ok=True)
+        ctx.export_chrome_trace(f"{prof_dir}/trace_rank{dist.get_rank()}.json.gz")

benchmark/lora/launch_server.py

@@ -26,6 +26,8 @@ def launch_server(args):
     cmd += f"--tp-size {args.tp_size} "
     if args.disable_custom_all_reduce:
         cmd += "--disable-custom-all-reduce"
+    if args.enable_mscclpp:
+        cmd += "--enable-mscclpp"
     print(cmd)
     os.system(cmd)
 
@@ -63,6 +65,11 @@ if __name__ == "__main__":
         action="store_true",
         help="Disable custom all reduce when device does not support p2p communication",
     )
+    parser.add_argument(
+        "--enable-mscclpp",
+        action="store_true",
+        help="Enable using mscclpp for small messages for all-reduce kernel and fall back to NCCL.",
+    )
     args = parser.parse_args()
 
     launch_server(args)

docs/backend/server_arguments.md

@@ -201,6 +201,7 @@ Please consult the documentation below and [server_args.py](https://github.com/s
 | `disable_cuda_graph_padding` | Disable CUDA Graph when padding is needed; otherwise, still use CUDA Graph. | `False` |
 | `disable_outlines_disk_cache` | Disable disk cache for outlines grammar backend. | `False` |
 | `disable_custom_all_reduce` | Disable usage of custom all-reduce kernel. | `False` |
+| `enable_mscclpp` | Enable usage of mscclpp kernel for small message all-reduce. | `False` |
 | `disable_overlap_schedule` | Disable the [Overhead-Scheduler](https://lmsys.org/blog/2024-12-04-sglang-v0-4/#zero-overhead-batch-scheduler). | `False` |
 | `enable_nan_detection` | Enable warning if the logits contain `NaN`. | `False` |
 | `enable_p2p_check` | Turns off the default of always allowing P2P checks when accessing GPU. | `False` |

python/sglang/srt/_custom_ops.py

@@ -113,3 +113,37 @@ else:
 
     def get_meta_buffer_ipc_handle(inp: torch.Tensor) -> torch.Tensor:
         return sgl_kernel.allreduce.get_meta_buffer_ipc_handle(inp)
+
+
+def mscclpp_generate_unique_id() -> bytes:
+    return sgl_kernel.allreduce.mscclpp_generate_unique_id()
+
+
+def mscclpp_init_context(
+    unique_id: bytes,
+    rank: int,
+    world_size: int,
+    scratch: torch.Tensor,
+    put_buffer: torch.Tensor,
+    nranks_per_node: int,
+    rank_to_node: List[int],
+    rank_to_ib: List[int],
+    context_selection: int,
+) -> int:
+    return sgl_kernel.allreduce.mscclpp_init_context(
+        unique_id,
+        rank,
+        world_size,
+        scratch,
+        put_buffer,
+        nranks_per_node,
+        rank_to_node,
+        rank_to_ib,
+        context_selection,
+    )
+
+
+def mscclpp_allreduce(
+    context: int, inp: torch.Tensor, out: torch.Tensor, nthreads: int, nblocks: int
+) -> None:
+    return sgl_kernel.allreduce.mscclpp_allreduce(context, inp, out, nthreads, nblocks)

python/sglang/srt/distributed/device_communicators/pymscclpp.py

+import bisect
+import logging
+import math
+import os
+from contextlib import contextmanager
+from enum import IntEnum
+from typing import Any, Callable, List, Optional, TypeVar, Union
+
+import torch
+import torch.distributed as dist
+from torch.distributed import ProcessGroup, ReduceOp
+
+from sglang.srt import _custom_ops as ops
+from sglang.srt.utils import is_cuda, is_hip
+
+logger = logging.getLogger(__name__)
+
+_is_cuda = is_cuda()
+_is_hip = is_hip()
+
+mscclpp_is_available = False
+if _is_hip:
+    # TODO(zyksir): mscclpp is untested on AMD and therefore disabled.
+    mscclpp_is_available = False
+if _is_cuda:
+    try:
+        import sgl_kernel
+
+        mscclpp_is_available = True
+    except:
+        mscclpp_is_available = False
+
+
+class MscclContextSelection(IntEnum):
+    MSCCL1SHOT1NODELL = 1
+    MSCCL1SHOT2NODELL = 2
+
+
+def mscclpp_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()
+    )
+
+
+def mscclpp_convert_to_bytes(size_str):
+    """
+    Converts a human-readable size string (e.g., "1MB", "2.5kb", "3 GB")
+    into the equivalent number of bytes using binary units.
+
+    Args:
+        size_str (str): A string representing size with unit (KB, MB, GB).
+
+    Returns:
+        int: Number of bytes.
+    """
+    size_str = size_str.strip().lower()
+
+    if not size_str:
+        raise ValueError("Empty input string")
+
+    # Extract numeric part and unit
+    for i in range(len(size_str)):
+        if not size_str[i].isdigit() and size_str[i] != ".":
+            break
+    num_str = size_str[:i]
+    unit = size_str[i:].strip()
+
+    try:
+        num = float(num_str)
+    except ValueError:
+        raise ValueError(f"Invalid numeric value in '{size_str}'")
+
+    # Conversion factors
+    if unit == "b":
+        return int(num)
+    elif unit == "kb":
+        return int(num * 1024)
+    elif unit == "mb":
+        return int(num * 1024 * 1024)
+    elif unit == "gb":
+        return int(num * 1024 * 1024 * 1024)
+    else:
+        raise ValueError(f"Unsupported unit: {unit}, support B, KB, MB, GB only")
+
+
+def mscclpp_bench_time(func, test_niter: int = 10, warmup_niter: int = 2):
+    # warmup
+    for _ in range(warmup_niter):
+        func()
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+    torch.cuda.synchronize()
+    dist.barrier()
+    start_event.record()
+    for _ in range(test_niter):
+        func()
+    end_event.record()
+    end_event.synchronize()
+    func_cost_us = start_event.elapsed_time(end_event) / test_niter * 1000
+    return func_cost_us
+
+
+class PyMscclppCommunicator:
+    _SUPPORTED_WORLD_SIZES = [8, 16]
+    _MAX_BYTES = mscclpp_convert_to_bytes(os.getenv("SGLANG_MSCCLPP_MAX_BYTES", "1MB"))
+    _SUPPORTED_DTYPE = [torch.float, torch.float16, torch.bfloat16]
+
+    # max_bytes: max supported mscclpp allreduce size
+    # in A100 mscclpp is faster than nccl only under condition of msg size smaller than1MB
+    def __init__(
+        self,
+        group: ProcessGroup,
+        device: Union[int, str, torch.device],
+        max_bytes=_MAX_BYTES,
+    ) -> None:
+        """
+        Args:
+            group: the process group to work on. If None, it will use the
+                default process group.
+            device: the device to bind the CustomAllreduce to. If None,
+                it will be bind to f"cuda:{local_rank}".
+        It is the caller's responsibility to make sure each communicator
+        is bind to a unique device, and all communicators in this group
+        are in the same node.
+        """
+        self._IS_CAPTURING = False
+        self.disabled = True
+
+        if not mscclpp_is_available:
+            # disable because of missing mscclpp library
+            # e.g. in a non-cuda environment
+            return
+
+        self.group = group
+
+        assert (
+            dist.get_backend(group) != dist.Backend.NCCL
+        ), "CustomAllreduce should be attached to a non-NCCL group."
+
+        rank = dist.get_rank(group=self.group)
+        world_size = dist.get_world_size(group=self.group)
+        if world_size == 1:
+            # No need to initialize mscclpp for single GPU case.
+            return
+
+        if world_size not in PyMscclppCommunicator._SUPPORTED_WORLD_SIZES:
+            logger.warning(
+                "PyMscclpp is disabled due to an unsupported world"
+                " size: %d. Supported world sizes: %s. To silence this "
+                "warning, specify disable_mscclpp=True explicitly.",
+                world_size,
+                str(PyMscclppCommunicator._SUPPORTED_WORLD_SIZES),
+            )
+            return
+
+        self.ranks = torch.distributed.get_process_group_ranks(group)
+        self.nranks_per_node = torch.cuda.device_count()
+        # for now mscclpp with stride in the communicator is not tested
+        if not (abs(self.ranks[-1] - self.ranks[0]) == world_size - 1):
+            logger.warning(
+                "PyMscclpp is disabled due to an unsupported group %s."
+                "Please ensure all ranks in the group are consecutive."
+                "To silence this warning, specify disable_mscclpp=True explicitly.",
+                str(self.ranks),
+            )
+            return
+
+        if isinstance(device, int):
+            device = torch.device(f"cuda:{device}")
+        elif isinstance(device, str):
+            device = torch.device(device)
+        # now `device` is a `torch.device` object
+        assert isinstance(device, torch.device)
+        self.device = device
+
+        self.max_bytes = max_bytes
+        self.rank = rank
+        self.world_size = world_size
+
+        if dist.get_rank(group) == 0:
+            unique_id = [ops.mscclpp_generate_unique_id()]
+        else:
+            unique_id = [None]
+        dist.broadcast_object_list(unique_id, src=self.ranks[0], group=self.group)
+        self.unique_id = unique_id[0]
+        self.rank_to_node, self.rank_to_ib = list(range(world_size)), list(
+            range(world_size)
+        )
+        for r in range(world_size):
+            self.rank_to_node[r] = r // 8
+            self.rank_to_ib[r] = self.rank % 8
+
+        self._context = None
+        self.context_selection = None
+        self.msg_size_for_finetune = [
+            2**i for i in range(10, math.floor(math.log2(self.max_bytes)) + 1)
+        ]
+        self.msg_size2best_config = {}
+        if world_size == 8:
+            self.context_selection = MscclContextSelection.MSCCL1SHOT1NODELL
+        elif world_size == 16:
+            self.context_selection = MscclContextSelection.MSCCL1SHOT2NODELL
+        if not _is_hip:
+            self.scratch = torch.empty(
+                self.max_bytes * 8,
+                dtype=torch.uint8,
+                device=self.device,
+            )
+            self.put_buffer = torch.empty(
+                self.max_bytes * 8 // self.nranks_per_node,
+                dtype=torch.uint8,
+                device=self.device,
+            )
+            self._context = ops.mscclpp_init_context(
+                self.unique_id,
+                self.rank,
+                self.world_size,
+                self.scratch,
+                self.put_buffer,
+                self.nranks_per_node,
+                self.rank_to_node,
+                self.rank_to_ib,
+                int(self.context_selection),
+            )
+        else:
+            raise NotImplementedError("HIP Mscclpp is not supported yet.")
+
+        self.msg_size2best_config = {}
+        self.pre_tune_config()
+        if dist.get_rank(group) == 0:
+            msg_size2best_config = [self.msg_size2best_config]
+        else:
+            msg_size2best_config = [None]
+        dist.broadcast_object_list(
+            msg_size2best_config, src=self.ranks[0], group=self.group
+        )
+        self.msg_size2best_config = msg_size2best_config[0]
+
+        # PyMscclpp is enabled only in cuda graph
+        self.disabled = True
+
+    def pre_tune_config(self, dtype=torch.bfloat16) -> bool:
+        logger.debug(f"start to pre-tune configs for rank {self.rank}")
+        nthreads_to_try = [256, 512, 1024]
+        nblocks_to_try = [21, 42, 84]
+        inp_randn = torch.ones(
+            self.msg_size_for_finetune[-1] // dtype.itemsize, dtype=dtype, device="cuda"
+        )
+        oup_randn = torch.empty_like(inp_randn)
+        for msg_size in self.msg_size_for_finetune:
+            mock_inp, mock_outp = (
+                inp_randn[: msg_size // dtype.itemsize],
+                oup_randn[: msg_size // dtype.itemsize],
+            )
+            best_config, best_time = None, None
+            for nthreads in nthreads_to_try:
+                for nblocks in nblocks_to_try:
+                    cur_cost = mscclpp_bench_time(
+                        lambda: ops.mscclpp_allreduce(
+                            self._context, mock_inp, mock_outp, nthreads, nblocks
+                        )
+                    )
+                    if best_time is None or cur_cost < best_time:
+                        best_config = (nthreads, nblocks)
+                        best_time = cur_cost
+            self.msg_size2best_config[msg_size] = best_config
+            if self.rank == 0:
+                logger.debug(
+                    f"for msg_size {msg_size}, best_config: {best_config}, best_time: {best_time}us"
+                )
+
+    def should_mscclpp_allreduce(
+        self, inp: torch.Tensor, op: ReduceOp = ReduceOp.SUM
+    ) -> bool:
+        if self.disabled or self._context is None:
+            return False
+        if inp.dtype not in PyMscclppCommunicator._SUPPORTED_DTYPE:
+            return False
+        if not mscclpp_is_weak_contiguous(inp):
+            return False
+        # only support sum op
+        if op != ReduceOp.SUM:
+            return False
+        if inp.numel() * inp.element_size() > self.max_bytes:
+            return False
+        return True
+
+    def all_reduce(self, tensor: torch.Tensor, op: ReduceOp = ReduceOp.SUM):
+        if self._IS_CAPTURING:
+            if torch.cuda.is_current_stream_capturing():
+                self.graph_input_set.add((tensor.dtype, tensor.numel()))
+        msg_size = tensor.numel() * tensor.itemsize
+        index = bisect.bisect_left(self.msg_size_for_finetune, msg_size)
+        msg_size_finetune = self.msg_size_for_finetune[index]
+        nthreads, nblocks = self.msg_size2best_config[msg_size_finetune]
+        result = torch.empty_like(tensor)
+        ops.mscclpp_allreduce(self._context, tensor, result, nthreads, nblocks)
+        return result
+
+    @contextmanager
+    def change_state(
+        self,
+        enable: Optional[bool] = None,
+    ):
+        if enable is None:
+            # guess a default value when not specified
+            enable = self.available
+
+        old_disable = self.disabled
+        self.disabled = not enable
+
+        yield
+
+        self.disabled = old_disable

python/sglang/srt/distributed/parallel_state.py

@@ -190,6 +190,7 @@ class GroupCoordinator:
     cpu_group: ProcessGroup  # group for CPU communication
     device_group: ProcessGroup  # group for device communication
     use_pynccl: bool  # a hint of whether to use PyNccl
+    use_pymscclpp: bool  # a hint of whether to use PyMsccl
     use_custom_allreduce: bool  # a hint of whether to use CustomAllreduce
     use_message_queue_broadcaster: (
         bool  # a hint of whether to use message queue broadcaster
@@ -205,6 +206,7 @@ class GroupCoordinator:
         local_rank: int,
         torch_distributed_backend: Union[str, Backend],
         use_pynccl: bool,
+        use_pymscclpp: bool,
         use_custom_allreduce: bool,
         use_hpu_communicator: bool,
         use_xpu_communicator: bool,
@@ -244,6 +246,7 @@ class GroupCoordinator:
             self.device = torch.device("cpu")
 
         self.use_pynccl = use_pynccl
+        self.use_pymscclpp = use_pymscclpp
         self.use_custom_allreduce = use_custom_allreduce
         self.use_hpu_communicator = use_hpu_communicator
         self.use_xpu_communicator = use_xpu_communicator
@@ -265,6 +268,17 @@ class GroupCoordinator:
                 device=self.device,
             )
 
+        from sglang.srt.distributed.device_communicators.pymscclpp import (
+            PyMscclppCommunicator,
+        )
+
+        self.pymscclpp_comm: Optional[PyMscclppCommunicator] = None
+        if use_pymscclpp and self.world_size > 1:
+            self.pymscclpp_comm = PyMscclppCommunicator(
+                group=self.cpu_group,
+                device=self.device,
+            )
+
         self.ca_comm: Optional[CustomAllreduce] = None
         if use_custom_allreduce and self.world_size > 1:
             # Initialize a custom fast all-reduce implementation.
@@ -373,11 +387,15 @@ class GroupCoordinator:
             # --------------------------------------------
             # custom allreduce       | enabled | enabled |
             # PyNccl                 | disabled| enabled |
+            # PyMscclpp              | disabled| enabled |
             # torch.distributed      | enabled | disabled|
             #
             # Note that custom allreduce will have a runtime check, if the
             #  tensor size is too large, it will fallback to the next
             #  available option.
+            # Note that the PyMsccl needs to register the tensor in ahead,
+            #  which will introduce large overhead in the eager case,
+            #  therefore it is only supported in the graph case.
             # In summary: When using CUDA graph, we use
             #  either custom all-reduce kernel or pynccl. When not using
             #  CUDA graph, we use either custom all-reduce kernel or
@@ -392,7 +410,14 @@ class GroupCoordinator:
                 maybe_pynccl_context = pynccl_comm.change_state(
                     enable=True, stream=torch.cuda.current_stream()
                 )
-            with maybe_pynccl_context:
+
+            pymscclpp_comm = self.pymscclpp_comm
+            maybe_pymscclpp_context: Any
+            if not pymscclpp_comm:
+                maybe_pymscclpp_context = nullcontext()
+            else:
+                maybe_pymscclpp_context = pymscclpp_comm.change_state(enable=True)
+            with maybe_pynccl_context, maybe_pymscclpp_context:
                 yield graph_capture_context
 
     def all_reduce(self, input_: torch.Tensor) -> torch.Tensor:
@@ -437,6 +462,10 @@ class GroupCoordinator:
             self.ca_comm is not None
             and not self.ca_comm.disabled
             and self.ca_comm.should_custom_ar(input_)
+        ) or (
+            self.pymscclpp_comm is not None
+            and not self.pymscclpp_comm.disabled
+            and self.pymscclpp_comm.should_mscclpp_allreduce(input_)
         ):
             return torch.ops.sglang.outplace_all_reduce(
                 input_, group_name=self.unique_name
@@ -447,9 +476,13 @@ class GroupCoordinator:
 
     def _all_reduce_out_place(self, input_: torch.Tensor) -> torch.Tensor:
         ca_comm = self.ca_comm
-        assert ca_comm is not None
-        assert not ca_comm.disabled
-        out = ca_comm.custom_all_reduce(input_)
+        pymscclpp_comm = self.pymscclpp_comm
+        assert ca_comm is not None or pymscclpp_comm is not None
+        if ca_comm is not None and not ca_comm.disabled:
+            out = ca_comm.custom_all_reduce(input_)
+        else:
+            assert not pymscclpp_comm.disabled
+            out = pymscclpp_comm.all_reduce(input_)
         assert out is not None
         return out
 
@@ -958,6 +991,7 @@ def init_world_group(
         local_rank=local_rank,
         torch_distributed_backend=backend,
         use_pynccl=False,
+        use_pymscclpp=False,
         use_custom_allreduce=False,
         use_hpu_communicator=False,
         use_xpu_communicator=False,
@@ -973,14 +1007,18 @@ def init_model_parallel_group(
     use_custom_allreduce: Optional[bool] = None,
     use_message_queue_broadcaster: bool = False,
     group_name: Optional[str] = None,
+    use_mscclpp_allreduce: Optional[bool] = None,
 ) -> GroupCoordinator:
     if use_custom_allreduce is None:
         use_custom_allreduce = _ENABLE_CUSTOM_ALL_REDUCE
+    if use_mscclpp_allreduce is None:
+        use_mscclpp_allreduce = _ENABLE_MSCCLPP_ALL_REDUCE
     return GroupCoordinator(
         group_ranks=group_ranks,
         local_rank=local_rank,
         torch_distributed_backend=backend,
         use_pynccl=not is_npu(),
+        use_pymscclpp=use_mscclpp_allreduce,
         use_custom_allreduce=use_custom_allreduce,
         use_hpu_communicator=True,
         use_xpu_communicator=True,
@@ -1037,6 +1075,7 @@ def graph_capture():
 logger = logging.getLogger(__name__)
 
 _ENABLE_CUSTOM_ALL_REDUCE = True
+_ENABLE_MSCCLPP_ALL_REDUCE = False
 
 
 def set_custom_all_reduce(enable: bool):
@@ -1044,6 +1083,11 @@ def set_custom_all_reduce(enable: bool):
     _ENABLE_CUSTOM_ALL_REDUCE = enable
 
 
+def set_mscclpp_all_reduce(enable: bool):
+    global _ENABLE_MSCCLPP_ALL_REDUCE
+    _ENABLE_MSCCLPP_ALL_REDUCE = enable
+
+
 def init_distributed_environment(
     world_size: int = -1,
     rank: int = -1,

python/sglang/srt/layers/dp_attention.py

@@ -98,11 +98,12 @@ def initialize_dp_attention(
         ],
         local_rank,
         torch.distributed.get_backend(tp_group.device_group),
-        SYNC_TOKEN_IDS_ACROSS_TP,
-        False,
-        False,
-        False,
-        False,
+        use_pynccl=SYNC_TOKEN_IDS_ACROSS_TP,
+        use_pymscclpp=False,
+        use_custom_allreduce=False,
+        use_hpu_communicator=False,
+        use_xpu_communicator=False,
+        use_npu_communicator=False,
         group_name="attention_tp",
     )
 

python/sglang/srt/model_executor/model_runner.py

@@ -35,6 +35,7 @@ from sglang.srt.distributed import (
     init_distributed_environment,
     initialize_model_parallel,
     set_custom_all_reduce,
+    set_mscclpp_all_reduce,
 )
 from sglang.srt.distributed.parallel_state import monkey_patch_vllm_parallel_state
 from sglang.srt.layers.attention.tbo_backend import TboAttnBackend
@@ -460,6 +461,7 @@ class ModelRunner:
         else:
             dist_init_method = f"tcp://127.0.0.1:{self.dist_port}"
         set_custom_all_reduce(not self.server_args.disable_custom_all_reduce)
+        set_mscclpp_all_reduce(self.server_args.enable_mscclpp)
 
         if not self.is_draft_worker:
             # Only initialize the distributed environment on the target model worker.

python/sglang/srt/server_args.py

@@ -165,6 +165,7 @@ class ServerArgs:
     enable_tokenizer_batch_encode: bool = False
     disable_outlines_disk_cache: bool = False
     disable_custom_all_reduce: bool = False
+    enable_mscclpp: bool = False
     disable_overlap_schedule: bool = False
     enable_mixed_chunk: bool = False
     enable_dp_attention: bool = False
@@ -1168,6 +1169,11 @@ class ServerArgs:
             action="store_true",
             help="Disable the custom all-reduce kernel and fall back to NCCL.",
         )
+        parser.add_argument(
+            "--enable-mscclpp",
+            action="store_true",
+            help="Enable using mscclpp for small messages for all-reduce kernel and fall back to NCCL.",
+        )
         parser.add_argument(
             "--disable-overlap-schedule",
             action="store_true",

sgl-kernel/CMakeLists.txt

@@ -73,6 +73,14 @@ FetchContent_Declare(
     GIT_SHALLOW    OFF
 )
 FetchContent_Populate(repo-flash-attention)
+# mscclpp
+FetchContent_Declare(
+    repo-mscclpp
+    GIT_REPOSITORY https://github.com/microsoft/mscclpp.git
+    GIT_TAG        51eca89d20f0cfb3764ccd764338d7b22cd486a6
+    GIT_SHALLOW    OFF
+)
+FetchContent_Populate(repo-mscclpp)
 
 # ccache option
 option(ENABLE_CCACHE "Whether to use ccache" ON)
@@ -99,6 +107,7 @@ include_directories(
     ${repo-cutlass_SOURCE_DIR}/tools/util/include
     ${repo-flashinfer_SOURCE_DIR}/include
     ${repo-flashinfer_SOURCE_DIR}/csrc
+    ${repo-mscclpp_SOURCE_DIR}/include
 )
 
 set(SGL_KERNEL_CUDA_FLAGS
@@ -196,6 +205,7 @@ string(REPLACE "-D__CUDA_NO_BFLOAT16_CONVERSIONS__" "" CMAKE_CUDA_FLAGS "${CMAKE
 string(REPLACE "-D__CUDA_NO_HALF2_OPERATORS__"      "" CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS}")
 
 set(SOURCES
+    "csrc/allreduce/mscclpp_allreduce.cu"
     "csrc/allreduce/custom_all_reduce.cu"
     "csrc/attention/cascade.cu"
     "csrc/attention/merge_attn_states.cu"
@@ -250,7 +260,27 @@ target_include_directories(common_ops PRIVATE
     ${repo-cutlass_SOURCE_DIR}/examples/common
     ${repo-flash-attention_SOURCE_DIR}/csrc/flash_attn/src
 )
-target_link_libraries(common_ops PRIVATE ${TORCH_LIBRARIES} c10 cuda cublas cublasLt)
+
+find_package(Python3 COMPONENTS Interpreter REQUIRED)
+execute_process(
+    COMMAND ${Python3_EXECUTABLE} -c "import torch; print(int(torch._C._GLIBCXX_USE_CXX11_ABI))"
+    OUTPUT_VARIABLE TORCH_CXX11_ABI
+    OUTPUT_STRIP_TRAILING_WHITESPACE
+)
+if(TORCH_CXX11_ABI STREQUAL "0")
+    message(STATUS "Using old C++ ABI (-D_GLIBCXX_USE_CXX11_ABI=0)")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D_GLIBCXX_USE_CXX11_ABI=0")
+    set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D_GLIBCXX_USE_CXX11_ABI=0")
+else()
+    message(STATUS "Using new C++11 ABI (-D_GLIBCXX_USE_CXX11_ABI=1)")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D_GLIBCXX_USE_CXX11_ABI=1")
+    set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} -D_GLIBCXX_USE_CXX11_ABI=1")
+endif()
+set(MSCCLPP_USE_CUDA ON)
+set(MSCCLPP_BYPASS_GPU_CHECK ON)
+set(MSCCLPP_BUILD_TESTS OFF)
+add_subdirectory(${repo-mscclpp_SOURCE_DIR})
+target_link_libraries(common_ops PRIVATE ${TORCH_LIBRARIES} c10 cuda cublas cublasLt mscclpp_static)
 
 target_compile_definitions(common_ops PRIVATE
     FLASHATTENTION_DISABLE_BACKWARD

sgl-kernel/Makefile

@@ -19,14 +19,14 @@ submodule: ## Initialize and update git submodules
 	@git submodule update --init --recursive
 
 ln: submodule ## Create compilation database
-	@rm -rf build && mkdir build && cd build && cmake .. -DCMAKE_EXPORT_COMPILE_COMMANDS=YES
+	@rm -rf build && mkdir build && cd build && cmake .. -DCMAKE_EXPORT_COMPILE_COMMANDS=YES -DCMAKE_POLICY_VERSION_MINIMUM=3.5
 
 
 install: submodule ## Install package in development mode
 	@pip install -e . --no-build-isolation
 
 build: install-deps submodule ## Build and install wheel package
-	@rm -rf dist/* || true && export MAX_JOBS=$(nproc) && CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) uv build --wheel -Cbuild-dir=build . --verbose --color=always --no-build-isolation && pip3 install dist/*whl --force-reinstall --no-deps
+	@rm -rf dist/* || true && export MAX_JOBS=$(nproc) && CMAKE_POLICY_VERSION_MINIMUM=3.5 CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) uv build --wheel -Cbuild-dir=build . --verbose --color=always --no-build-isolation && pip3 install dist/*whl --force-reinstall --no-deps
 
 clean: ## Remove build artifacts
 	@rm -rf build dist *.egg-info

sgl-kernel/build.sh

@@ -50,6 +50,9 @@ docker run --rm \
    which cmake
    cmake --version
 
+   yum install numactl-devel -y && \
+   yum install libibverbs -y && \
+   ln -sv /usr/lib64/libibverbs.so.1 /usr/lib64/libibverbs.so && \
    ${PYTHON_ROOT_PATH}/bin/${TORCH_INSTALL} && \
    ${PYTHON_ROOT_PATH}/bin/pip install --no-cache-dir ninja setuptools==75.0.0 wheel==0.41.0 numpy uv scikit-build-core && \
    export TORCH_CUDA_ARCH_LIST='7.5 8.0 8.9 9.0+PTX' && \

sgl-kernel/csrc/allreduce/mscclpp_allreduce.cu

+#include <c10/cuda/CUDAGuard.h>
+#include <c10/cuda/CUDAStream.h>
+#include <torch/all.h>
+#include <torch/library.h>
+
+#include "mscclpp_allreduce.cuh"
+
+enum MscclContextSelection {
+  MSCCL1NODELL = 1,
+  MSCCL2NODELL = 2,
+};
+
+class MscclContext {
+ public:
+  MscclContextSelection selection_;
+  std::shared_ptr<sglang::Msccl1NodeLLcontext> msccl_1nodeLL_context;
+  std::shared_ptr<sglang::Msccl2NodeLLcontext> msccl_2nodeLL_context;
+  MscclContext(MscclContextSelection selection) : selection_(selection) {}
+  template <typename T>
+  void allreduce(
+      cudaStream_t stream, T* input, T* output, const size_t input_numel, int threads = 512, int block_limit = 21) {
+    if (selection_ == MSCCL1NODELL) {
+      msccl_1nodeLL_context->allreduce<T>(stream, input, output, input_numel, threads, block_limit);
+    } else if (selection_ == MSCCL2NODELL) {
+      msccl_2nodeLL_context->allreduce<T>(stream, input, output, input_numel, threads, block_limit);
+    }
+  }
+};
+
+using fptr_t = int64_t;
+static_assert(sizeof(void*) == sizeof(fptr_t));
+
+torch::Tensor _unique_id2tensor(const mscclpp::UniqueId& unique_id) {
+  auto options = torch::TensorOptions().dtype(torch::kByte).device(torch::kCPU);
+  auto tensor = torch::empty({static_cast<int64_t>(unique_id.size())}, options);
+  std::memcpy(tensor.data_ptr<uint8_t>(), unique_id.data(), unique_id.size());
+  return tensor;
+}
+
+// Function to convert vector of int32_t back to array of uint8_t
+mscclpp::UniqueId _tensor2unique_id(const torch::Tensor& tensor) {
+  mscclpp::UniqueId unique_id;
+  std::memcpy(unique_id.data(), tensor.data_ptr<uint8_t>(), unique_id.size());
+  return unique_id;
+}
+
+torch::Tensor mscclpp_generate_unique_id() {
+  mscclpp::UniqueId unique_id = mscclpp::TcpBootstrap::createUniqueId();
+  return _unique_id2tensor(unique_id);
+}
+
+fptr_t mscclpp_init_context(
+    const torch::Tensor& unique_id,
+    const int64_t rank,
+    const int64_t world_size,
+    torch::Tensor& scratch,
+    torch::Tensor& put_buffer,
+    const int64_t nranks_per_node,
+    const std::vector<int64_t>& rank_to_node,
+    const std::vector<int64_t>& rank_to_ib,
+    const int64_t context_selection) {
+  MscclContext* context_ptr = new MscclContext(static_cast<MscclContextSelection>(context_selection));
+  mscclpp::UniqueId uid = _tensor2unique_id(unique_id);
+  if (context_selection == MSCCL1NODELL) {
+    void* scratch_ptr = reinterpret_cast<void*>(scratch.data_ptr());
+    const size_t scratch_bytes = scratch.numel() * scratch.element_size();
+    context_ptr->msccl_1nodeLL_context = std::make_shared<sglang::Msccl1NodeLLcontext>(
+        uid, rank, world_size, scratch_ptr, scratch_bytes, nranks_per_node, rank_to_node, rank_to_ib);
+  } else if (context_selection == MSCCL2NODELL) {
+    void* scratch_ptr = reinterpret_cast<void*>(scratch.data_ptr());
+    const size_t scratch_bytes = scratch.numel() * scratch.element_size();
+    void* put_buffer_ptr = reinterpret_cast<void*>(put_buffer.data_ptr());
+    const size_t put_buffer_bytes = put_buffer.numel() * put_buffer.element_size();
+    context_ptr->msccl_2nodeLL_context = std::make_shared<sglang::Msccl2NodeLLcontext>(
+        uid,
+        rank,
+        world_size,
+        scratch_ptr,
+        scratch_bytes,
+        put_buffer_ptr,
+        put_buffer_bytes,
+        nranks_per_node,
+        rank_to_node,
+        rank_to_ib);
+  } else {
+    throw std::runtime_error("invalid context selection");
+  }
+  return (fptr_t)context_ptr;
+}
+
+bool _mscclpp_is_weak_contiguous(torch::Tensor& t) {
+  return t.is_contiguous() ||
+         (t.storage().nbytes() - t.storage_offset() * t.element_size() == t.numel() * t.element_size());
+}
+void mscclpp_allreduce(fptr_t _context, torch::Tensor& inp, torch::Tensor& out, int64_t nthreads, int64_t nblocks) {
+  MscclContext* context = reinterpret_cast<MscclContext*>(_context);
+  const at::cuda::OptionalCUDAGuard device_guard(device_of(inp));
+  auto stream = c10::cuda::getCurrentCUDAStream().stream();
+
+  TORCH_CHECK_EQ(inp.scalar_type(), out.scalar_type());
+  TORCH_CHECK_EQ(inp.numel(), out.numel());
+  TORCH_CHECK(_mscclpp_is_weak_contiguous(out));
+  TORCH_CHECK(_mscclpp_is_weak_contiguous(inp));
+  switch (out.scalar_type()) {
+    case at::ScalarType::Float: {
+      context->allreduce<float>(
+          stream,
+          reinterpret_cast<float*>(inp.data_ptr()),
+          reinterpret_cast<float*>(out.data_ptr()),
+          inp.numel(),
+          nthreads,
+          nblocks);
+      break;
+    }
+    case at::ScalarType::Half: {
+      context->allreduce<half>(
+          stream,
+          reinterpret_cast<half*>(inp.data_ptr()),
+          reinterpret_cast<half*>(out.data_ptr()),
+          inp.numel(),
+          nthreads,
+          nblocks);
+      break;
+    }
+#if (__CUDA_ARCH__ >= 800 || !defined(__CUDA_ARCH__))
+    case at::ScalarType::BFloat16: {
+      context->allreduce<__nv_bfloat16>(
+          stream,
+          reinterpret_cast<__nv_bfloat16*>(inp.data_ptr()),
+          reinterpret_cast<__nv_bfloat16*>(out.data_ptr()),
+          inp.numel(),
+          nthreads,
+          nblocks);
+      break;
+    }
+#endif
+    default:
+      throw std::runtime_error("custom allreduce only supports float32, float16 and bfloat16");
+  }
+}

sgl-kernel/csrc/allreduce/test_mscclpp_allreduce.cu

+/*
+ * this file is used to test mscclpp_allreduce.cu using mpirun
+ * this file is adapted from https://github.com/flashinfer-ai/flashinfer/blob/v0.2.5/src/test_sum_all_reduce.cu
+usage:
+cd PATH-TO-THIS-FILE
+export MPI_HOME=/usr/local/mpi
+# export MPI_HOME=/opt/hpcx/ompi/
+export MSCCLPP_HOME=/workspace/test/mscclpp
+nvcc -O2 -arch=native -std=c++17 test_mscclpp_allreduce.cu \
+  -o test_mscclpp_allreduce -D_GLIBCXX_USE_CXX11_ABI=0 \
+  -I${MSCCLPP_HOME}/include -L${MSCCLPP_HOME}/build -lmscclpp \
+  -lnccl -I${MPI_HOME}/include -L${MPI_HOME}/lib -lmpi
+
+/opt/hpcx/ompi/bin/
+mpirun --allow-run-as-root -H 127.0.0.1:8 -np 8 \
+  --map-by ppr:8:node \
+  --mca btl_openib_warn_no_device_params_found 0 \
+  --mca btl_tcp_if_include bond0 \
+  --allow-run-as-root -np 8 \
+  -x NCCL_RUNTIME_CONNECT=0 -x NCCL_IB_GID_INDEX=3 -x NCCL_DEBUG=WARN \
+  -x LD_PRELOAD=${MSCCLPP_HOME}/build/libmscclpp.so ./test_mscclpp_allreduce
+ */
+#include <mpi.h>
+#include <thrust/detail/raw_pointer_cast.h>
+#include <thrust/device_vector.h>
+#include <thrust/host_vector.h>
+
+#ifndef CHECK_CUDA_SUCCESS
+#define CHECK_CUDA_SUCCESS(cmd)                                                             \
+  do {                                                                                      \
+    cudaError_t e = cmd;                                                                    \
+    if (e != cudaSuccess) {                                                                 \
+      printf("Failed: Cuda error %s:%d '%s'\n", __FILE__, __LINE__, cudaGetErrorString(e)); \
+      exit(EXIT_FAILURE);                                                                   \
+    }                                                                                       \
+  } while (0)
+#endif
+
+#include <cstdint>
+
+#include "mscclpp_allreduce.cuh"
+
+template <typename T>
+bool isclose(T a, T b, float rtol = 1e-5, float atol = 1e-8) {
+  return fabs(a - b) <= (atol + rtol * fabs(b));
+}
+
+int main(int argc, char* argv[]) {
+  // init mpi
+  MPI_Init(&argc, &argv);
+  printf("MPI Initialized.\n");
+  int nranks, rank;
+
+  // get work size and rank id
+  MPI_Comm_size(MPI_COMM_WORLD, &nranks);
+  MPI_Comm_rank(MPI_COMM_WORLD, &rank);
+  cudaSetDevice(rank);
+  printf("nranks: %d, rank: %d\n", nranks, rank);
+
+  // init host and device buffers
+  using T = float;
+  using ReduceT = float;
+  const size_t num_elems = 2 * 1024 * 1024;
+  std::vector<T> host_buf(num_elems);
+  for (uint32_t i = 0; i < num_elems; ++i) {
+    host_buf[i] = T(i + rank);
+  }
+  thrust::device_vector<T> device_buf(host_buf);
+  const size_t buf_size_in_bytes = num_elems * sizeof(T);
+  std::vector<T> host_result_buf(num_elems);
+  thrust::device_vector<T> device_result_buf(host_result_buf);
+
+  std::vector<T> host_scratch_buf(num_elems * 8);
+  for (uint32_t i = 0; i < num_elems; ++i) {
+    host_scratch_buf[i] = 1;
+  }
+  thrust::device_vector<T> device_scratch_buf(host_scratch_buf);
+  std::vector<T> host_put_buf(num_elems);
+  thrust::device_vector<T> device_put_buf(host_put_buf);
+
+  mscclpp::UniqueId unique_id;
+  if (rank == 0) unique_id = mscclpp::TcpBootstrap::createUniqueId();
+  MPI_Bcast(&unique_id, sizeof(unique_id), MPI_BYTE, 0, MPI_COMM_WORLD);
+
+  std::vector<int64_t> rank_to_node(nranks);
+  std::vector<int64_t> rank_to_ib(nranks);
+  for (int i = 0; i < nranks; i++) {
+    rank_to_node[i] = i / 8;
+    rank_to_ib[i] = i % 8;
+  }
+
+  cudaStream_t s;
+  CHECK_CUDA_SUCCESS(cudaStreamCreate(&s));
+  CHECK_CUDA_SUCCESS(cudaStreamSynchronize(s));
+  if (nranks == 8) {
+    auto context = std::make_shared<sglang::Msccl1NodeLLcontext>(
+        unique_id,
+        rank,
+        nranks,
+        thrust::raw_pointer_cast(device_scratch_buf.data()),
+        buf_size_in_bytes * 8,
+        rank_to_node,
+        rank_to_ib);
+    printf("rank: %d, Msccl1NodeLLcontext setup.\n", rank);
+    MPI_Barrier(MPI_COMM_WORLD);
+    context->allreduce<T>(
+        s,
+        thrust::raw_pointer_cast(device_buf.data()),
+        thrust::raw_pointer_cast(device_result_buf.data()),
+        device_buf.size());
+  } else if (nranks == 16) {
+    // TODO: this branch is untested since there is something wrong with mpirun in my test machince
+    auto context = std::make_shared<sglang::Msccl2NodeLLcontext>(
+        unique_id,
+        rank,
+        nranks,
+        thrust::raw_pointer_cast(device_scratch_buf.data()),
+        buf_size_in_bytes * 8,
+        thrust::raw_pointer_cast(device_put_buf.data()),
+        buf_size_in_bytes,
+        rank_to_node,
+        rank_to_ib);
+    printf("rank: %d, Msccl2NodeLLcontext setup.\n", rank);
+    MPI_Barrier(MPI_COMM_WORLD);
+    context->allreduce<T>(
+        s,
+        thrust::raw_pointer_cast(device_buf.data()),
+        thrust::raw_pointer_cast(device_result_buf.data()),
+        device_buf.size());
+  }
+
+  // check result correctness
+  thrust::host_vector<T> host_buf_result = device_result_buf;
+  size_t num_results_error_atol_1e_3_rtol_1e_3 = 0;
+  bool nan_detected = false;
+
+  for (uint32_t i = 0; i < num_elems; ++i) {
+    T expected = T(i * nranks + (nranks - 1) * nranks / 2);
+    if (std::isnan(float(host_buf_result[i]))) {
+      nan_detected = true;
+    }
+    if (!isclose(float(host_buf_result[i]), float(expected), 1e-3, 1e-3)) {
+      num_results_error_atol_1e_3_rtol_1e_3++;
+    }
+  }
+  float result_accuracy = 1. - float(num_results_error_atol_1e_3_rtol_1e_3) / float(num_elems);
+
+  printf("rank: %d, nan_detected: %d accuracy: %f\n", rank, nan_detected, result_accuracy);
+
+  CHECK_CUDA_SUCCESS(cudaStreamDestroy(s));
+  MPI_Finalize();
+  return 0;
+}

sgl-kernel/csrc/common_extension.cc

@@ -38,6 +38,15 @@ TORCH_LIBRARY_FRAGMENT(sgl_kernel, m) {
       "all_reduce(int fa, Tensor inp, Tensor! out, int reg_buffer, "
       "int reg_buffer_sz_bytes) -> ()");
   m.impl("all_reduce", torch::kCUDA, &all_reduce);
+
+  m.def("mscclpp_generate_unique_id", &mscclpp_generate_unique_id);
+  m.def(
+      "mscclpp_init_context(Tensor unique_id, int rank, int world_size, Tensor scratch, Tensor put_buffer, "
+      "int nranks_per_node, int[] rank_to_node, int[] rank_to_ib, int context_selection) -> int");
+  m.impl("mscclpp_init_context", torch::kCUDA, &mscclpp_init_context);
+
+  m.def("mscclpp_allreduce(int context, Tensor inp, Tensor! out, int nthreads, int nblocks) -> ()");
+  m.impl("mscclpp_allreduce", torch::kCUDA, &mscclpp_allreduce);
   /*
    * From csrc/attention
    */

sgl-kernel/include/sgl_kernel_ops.h

@@ -74,6 +74,18 @@ std::tuple<std::vector<int64_t>, std::vector<int64_t>> get_graph_buffer_ipc_meta
 void register_buffer(fptr_t _fa, const std::vector<fptr_t>& fake_ipc_ptrs);
 void register_graph_buffers(
     fptr_t _fa, const std::vector<std::vector<int64_t>>& handles, const std::vector<std::vector<int64_t>>& offsets);
+torch::Tensor mscclpp_generate_unique_id();
+fptr_t mscclpp_init_context(
+    const torch::Tensor& unique_id,
+    const int64_t rank,
+    const int64_t world_size,
+    torch::Tensor& scratch,
+    torch::Tensor& put_buffer,
+    const int64_t nranks_per_node,
+    const std::vector<int64_t>& rank_to_node,
+    const std::vector<int64_t>& rank_to_ib,
+    const int64_t context_selection);
+void mscclpp_allreduce(fptr_t _context, torch::Tensor& inp, torch::Tensor& out, int64_t nthreads, int64_t nblocks);
 #endif
 
 /*

sgl-kernel/python/sgl_kernel/allreduce.py

@@ -49,6 +49,27 @@ if torch.version.hip is not None:
     def get_meta_buffer_ipc_handle(inp: torch.Tensor) -> torch.Tensor:
         return torch.ops.sgl_kernel.get_meta_buffer_ipc_handle.default(inp)
 
+    def mscclpp_generate_unique_id() -> bytes:
+        raise NotImplementedError()
+
+    def mscclpp_init_context(
+        unique_id: bytes,
+        rank: int,
+        world_size: int,
+        scratch: torch.Tensor,
+        put_buffer: torch.Tensor,
+        nranks_per_node: int,
+        rank_to_node: List[int],
+        rank_to_ib: List[int],
+        context_selection: int,
+    ) -> int:
+        raise NotImplementedError()
+
+    def mscclpp_allreduce(
+        context: int, inp: torch.Tensor, out: torch.Tensor, nthreads: int, nblocks: int
+    ) -> None:
+        raise NotImplementedError()
+
 else:
 
     def init_custom_ar(
@@ -85,3 +106,36 @@ else:
 
     def meta_size() -> int:
         return torch.ops.sgl_kernel.meta_size.default()
+
+    def mscclpp_generate_unique_id() -> torch.Tensor:
+        return torch.ops.sgl_kernel.mscclpp_generate_unique_id.default()
+
+    def mscclpp_init_context(
+        unique_id: torch.Tensor,
+        rank: int,
+        world_size: int,
+        scratch: torch.Tensor,
+        put_buffer: torch.Tensor,
+        nranks_per_node: int,
+        rank_to_node: List[int],
+        rank_to_ib: List[int],
+        context_selection: int,
+    ) -> int:
+        return torch.ops.sgl_kernel.mscclpp_init_context.default(
+            unique_id,
+            rank,
+            world_size,
+            scratch,
+            put_buffer,
+            nranks_per_node,
+            rank_to_node,
+            rank_to_ib,
+            context_selection,
+        )
+
+    def mscclpp_allreduce(
+        context: int, inp: torch.Tensor, out: torch.Tensor, nthreads: int, nblocks: int
+    ) -> None:
+        torch.ops.sgl_kernel.mscclpp_allreduce.default(
+            context, inp, out, nthreads, nblocks
+        )

sgl-kernel/tests/test_mscclpp.py

+import multiprocessing as mp
+import os
+import socket
+import unittest
+from enum import IntEnum
+from typing import Any
+
+import sgl_kernel.allreduce as custom_ops
+import torch
+import torch.distributed as dist
+
+
+class MscclContextSelection(IntEnum):
+    MSCCL1SHOT1NODELL = 1
+    MSCCL1SHOT2NODELL = 2
+
+
+def _run_correctness_worker(world_size, rank, distributed_init_port, test_sizes):
+    device = torch.device(f"cuda:{rank % torch.cuda.device_count()}")
+    torch.cuda.set_device(device)
+    distributed_init_method = f"tcp://localhost:{distributed_init_port}"
+    dist.init_process_group(
+        backend="nccl",
+        init_method=distributed_init_method,
+        rank=rank,
+        world_size=world_size,
+    )
+    group = dist.group.WORLD
+    cpu_group = torch.distributed.new_group(list(range(world_size)), backend="gloo")
+    if rank == 0:
+        unique_id = [custom_ops.mscclpp_generate_unique_id()]
+    else:
+        unique_id = [None]
+    dist.broadcast_object_list(
+        unique_id, src=0, device=torch.device("cpu"), group=cpu_group
+    )
+    unique_id = unique_id[0]
+    rank_to_node, rank_to_ib = list(range(world_size)), list(range(world_size))
+    for r in range(world_size):
+        rank_to_node[r] = r // 8
+        rank_to_ib[r] = rank % 8
+    MAX_BYTES = 2**20
+    scratch = torch.empty(
+        MAX_BYTES * 8, dtype=torch.bfloat16, device=torch.cuda.current_device()
+    )
+    put_buffer = torch.empty(
+        MAX_BYTES, dtype=torch.bfloat16, device=torch.cuda.current_device()
+    )
+    print(f"[{rank}] start mscclpp_context init")
+    nranks_per_node = torch.cuda.device_count()
+    selection = int(MscclContextSelection.MSCCL1SHOT1NODELL)
+    mscclpp_context = custom_ops.mscclpp_init_context(
+        unique_id,
+        rank,
+        world_size,
+        scratch,
+        put_buffer,
+        nranks_per_node,
+        rank_to_node,
+        rank_to_ib,
+        selection,
+    )
+    try:
+        test_loop = 10
+        for sz in test_sizes:
+            for dtype in [torch.float32, torch.float16, torch.bfloat16]:
+                if sz * dtype.itemsize > MAX_BYTES:
+                    continue
+                if rank == 0:
+                    print(f"mscclpp allreduce test sz {sz}, dtype {dtype}")
+                for _ in range(test_loop):
+                    inp1 = torch.randint(1, 16, (sz,), dtype=dtype, device=device)
+                    inp1_ref = inp1.clone()
+                    out1 = torch.empty_like(inp1)
+                    custom_ops.mscclpp_allreduce(
+                        mscclpp_context, inp1, out1, nthreads=512, nblocks=21
+                    )
+                    dist.all_reduce(inp1_ref, group=group)
+                    torch.testing.assert_close(out1, inp1_ref)
+    finally:
+        dist.barrier(group=group)
+        dist.destroy_process_group(group=group)
+
+
+def get_open_port() -> int:
+    try:
+        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+            s.bind(("127.0.0.1", 0))
+            return s.getsockname()[1]
+    except OSError:
+        with socket.socket(socket.AF_INET6, socket.SOCK_STREAM) as s:
+            s.bind(("::1", 0))
+            return s.getsockname()[1]
+
+
+def multi_process_parallel(
+    world_size: int, test_target: Any, target_args: tuple = ()
+) -> None:
+    mp.set_start_method("spawn", force=True)
+
+    procs = []
+    distributed_init_port = get_open_port()
+    for i in range(world_size):
+        proc_args = (world_size, i, distributed_init_port) + target_args
+        proc = mp.Process(target=test_target, args=proc_args, name=f"Worker-{i}")
+        proc.start()
+        procs.append(proc)
+
+    for i in range(world_size):
+        procs[i].join()
+        assert (
+            procs[i].exitcode == 0
+        ), f"Process {i} failed with exit code {procs[i].exitcode}"
+
+
+class TestMSCCLAllReduce(unittest.TestCase):
+    test_sizes = [
+        512,
+        2560,
+        4096,
+        5120,
+        7680,
+        32768,
+        262144,
+        524288,
+    ]
+    world_sizes = [8]
+
+    def test_correctness(self):
+        for world_size in self.world_sizes:
+            available_gpus = torch.cuda.device_count()
+            if world_size > available_gpus:
+                print(
+                    f"Skipping world_size={world_size}, found {available_gpus} and now ray is not supported here"
+                )
+                continue
+
+            print(f"Running test for world_size={world_size}")
+            multi_process_parallel(
+                world_size, _run_correctness_worker, target_args=(self.test_sizes,)
+            )
+            print(f"custom allreduce tp = {world_size}: OK")
+
+
+if __name__ == "__main__":
+    unittest.main()

test/srt/test_mscclpp.py

+"""For Now, MSCCL is only supported on TP16 and TP8 case
+
+if [[ $RANK -eq 0  ]]; then
+    ray start --block --head --port=6379 &
+    python3 test_mscclpp.py;
+else
+    ray start --block --address=${MASTER_ADDR}:6379;
+fi
+"""
+
+import itertools
+import os
+import random
+import socket
+import unittest
+from contextlib import contextmanager, nullcontext
+from typing import Any, List, Optional, Union
+
+import ray
+import torch
+import torch.distributed as dist
+from torch.distributed import ProcessGroup, ReduceOp
+
+from sglang.srt.distributed import init_distributed_environment
+from sglang.srt.distributed.communication_op import (  # noqa
+    tensor_model_parallel_all_reduce,
+)
+from sglang.srt.distributed.device_communicators.custom_all_reduce import (
+    CustomAllreduce,
+)
+from sglang.srt.distributed.device_communicators.pymscclpp import PyMscclppCommunicator
+from sglang.srt.distributed.device_communicators.pynccl import PyNcclCommunicator
+from sglang.srt.distributed.parallel_state import (
+    get_tensor_model_parallel_group,
+    graph_capture,
+    initialize_model_parallel,
+    set_custom_all_reduce,
+    set_mscclpp_all_reduce,
+)
+from sglang.srt.distributed.utils import StatelessProcessGroup
+from sglang.test.test_utils import CustomTestCase
+
+
+def get_open_port() -> int:
+    # try ipv4
+    try:
+        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
+            s.bind(("", 0))
+            return s.getsockname()[1]
+    except OSError:
+        # try ipv6
+        with socket.socket(socket.AF_INET6, socket.SOCK_STREAM) as s:
+            s.bind(("", 0))
+            return s.getsockname()[1]
+
+
+def multi_process_parallel(
+    world_size: int,
+    master_addr: str,
+    cls: Any,
+    test_target: Any,
+) -> None:
+
+    # Using ray helps debugging the error when it failed
+    # as compared to multiprocessing.
+    # NOTE: We need to set working_dir for distributed tests,
+    # otherwise we may get import errors on ray workers
+
+    ray.init(log_to_driver=True)
+
+    distributed_init_port = get_open_port()
+    refs = []
+    for rank in range(world_size):
+        refs.append(
+            test_target.remote(
+                cls, world_size, master_addr, rank, distributed_init_port
+            )
+        )
+    ray.get(refs)
+
+    ray.shutdown()
+
+
+class TestMSCCLAllReduce(CustomTestCase):
+    @classmethod
+    def setUpClass(cls):
+        random.seed(42)
+        # 1KB to 1MB
+        cls.test_sizes = [512, 4096, 32768, 262144, 524288]
+        cls.world_sizes = [8]
+        TEST_TP16 = int(os.getenv("SGL_MSCCLPP_TEST_TP16", "0"))
+        if TEST_TP16:
+            cls.world_sizes = [16]
+        cls.test_loop = 10
+
+    def test_graph_allreduce(self):
+        TEST_MASTER_ADDR = os.getenv("SGL_MSCCLPP_TEST_MASTER_ADDR", "localhost")
+        for world_size in self.world_sizes:
+            if world_size not in [8, 16]:
+                continue
+            multi_process_parallel(
+                world_size, TEST_MASTER_ADDR, self, self.graph_allreduce
+            )
+
+    def test_eager_allreduce(self):
+        TEST_MASTER_ADDR = os.getenv("SGL_MSCCLPP_TEST_MASTER_ADDR", "localhost")
+        for world_size in self.world_sizes:
+            if world_size not in [8, 16]:
+                continue
+            multi_process_parallel(
+                world_size, TEST_MASTER_ADDR, self, self.eager_allreduce
+            )
+
+    @ray.remote(num_gpus=1, max_calls=1)
+    def graph_allreduce(self, world_size, master_addr, rank, distributed_init_port):
+        del os.environ["CUDA_VISIBLE_DEVICES"]
+        device = torch.device(f"cuda:{rank % torch.cuda.device_count()}")
+        torch.cuda.set_device(device)
+        distributed_init_method = f"tcp://{master_addr}:{distributed_init_port}"
+        set_mscclpp_all_reduce(True)
+        set_custom_all_reduce(False)
+        init_distributed_environment(
+            world_size=world_size,
+            rank=rank,
+            distributed_init_method=distributed_init_method,
+            local_rank=rank % torch.cuda.device_count(),
+        )
+        initialize_model_parallel(tensor_model_parallel_size=world_size)
+        group = get_tensor_model_parallel_group().device_group
+
+        # A small all_reduce for warmup.
+        # this is needed because device communicators might be created lazily
+        # (e.g. NCCL). This will ensure that the communicator is initialized
+        # before any communication happens, so that this group can be used for
+        # graph capture immediately.
+        data = torch.zeros(1)
+        data = data.to(device=device)
+        torch.distributed.all_reduce(data, group=group)
+        torch.cuda.synchronize()
+        del data
+
+        for sz in self.test_sizes:
+            for dtype in [torch.float32, torch.float16, torch.bfloat16]:
+                for _ in range(self.test_loop):
+                    with graph_capture() as graph_capture_context:
+                        # use integers so result matches NCCL exactly
+                        inp1 = torch.randint(
+                            1,
+                            16,
+                            (sz,),
+                            dtype=dtype,
+                            device=torch.cuda.current_device(),
+                        )
+                        inp2 = torch.randint(
+                            1,
+                            16,
+                            (sz,),
+                            dtype=dtype,
+                            device=torch.cuda.current_device(),
+                        )
+                        torch.cuda.synchronize()
+                        graph = torch.cuda.CUDAGraph()
+                        with torch.cuda.graph(
+                            graph, stream=graph_capture_context.stream
+                        ):
+                            out1 = tensor_model_parallel_all_reduce(inp1)
+                            # the input buffer is immediately modified to test
+                            # synchronization
+                            dist.all_reduce(inp1, group=group)
+                            out2 = tensor_model_parallel_all_reduce(inp2)
+                            dist.all_reduce(inp2, group=group)
+                    graph.replay()
+                    torch.testing.assert_close(out1, inp1)
+                    torch.testing.assert_close(out2, inp2)
+
+    @ray.remote(num_gpus=1, max_calls=1)
+    def eager_allreduce(self, world_size, master_addr, rank, distributed_init_port):
+        del os.environ["CUDA_VISIBLE_DEVICES"]
+        device = torch.device(f"cuda:{rank % torch.cuda.device_count()}")
+        torch.cuda.set_device(device)
+        distributed_init_method = f"tcp://{master_addr}:{distributed_init_port}"
+        set_mscclpp_all_reduce(True)
+        set_custom_all_reduce(False)
+        init_distributed_environment(
+            world_size=world_size,
+            rank=rank,
+            distributed_init_method=distributed_init_method,
+            local_rank=rank,
+        )
+        initialize_model_parallel(tensor_model_parallel_size=world_size)
+        group = get_tensor_model_parallel_group().device_group
+
+        for sz in self.test_sizes:
+            for dtype in [torch.float32, torch.float16, torch.bfloat16]:
+                for _ in range(self.test_loop):
+                    inp1 = torch.randint(
+                        1, 16, (sz,), dtype=dtype, device=torch.cuda.current_device()
+                    )
+                    out1 = tensor_model_parallel_all_reduce(inp1)
+                    dist.all_reduce(inp1, group=group)
+                    torch.testing.assert_close(out1, inp1)
+
+
+if __name__ == "__main__":
+    unittest.main()