class CustomAllreduce:
_SUPPORTED_WORLD_SIZES = [2, 4, 6, 8, 16]
_DEFAULT_ALL_GATHER_MAX_SIZE = 2 * 1024 * 1024
_DEFAULT_MNNVL_ALL_GATHER_MAX_SIZES = {
2: 8 * 1024 * 1024,
4: 4 * 1024 * 1024,
6: 2 * 1024 * 1024,
8: 2 * 1024 * 1024,
16: 2 * 1024 * 1024,
}
_DEFAULT_REDUCE_SCATTER_MAX_SIZE = 16 * 1024 * 1024
_DEFAULT_MNNVL_REDUCE_SCATTER_MAX_SIZE = 16 * 1024 * 1024
# max_size: max supported allreduce size
def __init__(
self,
group: ProcessGroup,
device: int | str | torch.device,
max_size=8192 * 1024,
max_all_gather_size=_DEFAULT_ALL_GATHER_MAX_SIZE,
max_mnnvl_all_gather_size=None,
max_reduce_scatter_size=_DEFAULT_REDUCE_SCATTER_MAX_SIZE,
max_mnnvl_reduce_scatter_size=_DEFAULT_MNNVL_REDUCE_SCATTER_MAX_SIZE,
symm_mem_enabled=False,
) -> 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 bound 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._ptr = 0
self.disabled = True
self.mnnvl_buffer = None
self.mnnvl_handle = None
self.mnnvl_peer_buffers: list[torch.Tensor] | None = None
self.mnnvl_multicast_ptr = 0
self.mnnvl_buffer_size = 0
self.mnnvl_lamport_ag_local_ptr = 0
self.mnnvl_lamport_ag_multicast_ptr = 0
self.mnnvl_lamport_rs_local_ptr = 0
self.mnnvl_lamport_epochs = None
self.mnnvl_lamport_ag_epoch_ptr = 0
self.mnnvl_lamport_rs_epoch_ptr = 0
self.mnnvl_only = False
if not custom_ar:
# disable because of missing custom allreduce library
# e.g. in a non-GPU environment
logger.info_once(
"Custom allreduce is disabled because "
"of missing custom allreduce library"
)
return
self.group = group
assert dist.get_backend(group) != dist.Backend.NCCL, (
"CustomAllreduce should be attached to a non-NCCL group."
)
same_node = all(in_the_same_node_as(group, source_rank=0))
self.mnnvl_only = not same_node
rank = dist.get_rank(group=self.group)
self.rank = rank
world_size = dist.get_world_size(group=self.group)
if world_size == 1:
# No need to initialize custom allreduce for single GPU case.
return
if world_size not in CustomAllreduce._SUPPORTED_WORLD_SIZES:
logger.warning_once(
"Custom allreduce is disabled due to an unsupported world"
" size: %d. Supported world sizes: %s. To silence this "
"warning, specify disable_custom_all_reduce=True explicitly.",
world_size,
str(CustomAllreduce._SUPPORTED_WORLD_SIZES),
)
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
device_capability = current_platform.get_device_capability()
if (
current_platform.is_cuda()
and symm_mem_enabled
and device_capability is not None
):
device_capability_str = device_capability.as_version_str()
if (
device_capability_str in CUSTOM_ALL_REDUCE_MAX_SIZES
and world_size in CUSTOM_ALL_REDUCE_MAX_SIZES[device_capability_str]
):
max_size = min(
CUSTOM_ALL_REDUCE_MAX_SIZES[device_capability_str][world_size],
max_size,
)
# device.index is a visible ordinal, not a logical local ID.
fully_connected = False
if same_node:
physical_device_id = (
current_platform.visible_device_id_to_physical_device_id(device.index)
)
tensor = torch.tensor([physical_device_id], dtype=torch.int, device="cpu")
gather_list = [
torch.tensor([0], dtype=torch.int, device="cpu")
for _ in range(world_size)
]
dist.all_gather(gather_list, tensor, group=self.group)
physical_device_ids = [t.item() for t in gather_list]
assert current_platform.is_cuda_alike()
fully_connected = current_platform.is_fully_connected(physical_device_ids)
if same_node and world_size > 2 and not fully_connected:
logger.warning(
"Custom allreduce is disabled because it's not supported on"
" more than two PCIe-only GPUs. To silence this warning, "
"specify disable_custom_all_reduce=True explicitly."
)
return
# test P2P capability, this checks software/cudaruntime support
# this is expensive to compute at the first time
# then we cache the result
# On AMD GPU, p2p is always enabled between XGMI connected GPUs
if (
same_node
and not current_platform.is_rocm()
and not _can_p2p(rank, world_size)
):
logger.warning(
"Custom allreduce is disabled because your platform lacks "
"GPU P2P capability or P2P test failed. To silence this "
"warning, specify disable_custom_all_reduce=True explicitly."
)
return
self.disabled = False
# Buffers memory are owned by this Python class and passed to C++.
# Metadata composes of two parts: metadata for synchronization and a
# temporary buffer for storing intermediate allreduce results.
if same_node:
self.meta_ptrs = self.create_shared_buffer(
ops.meta_size() + max_size, group=group, uncached=True
)
else:
meta_ptr, _ = ops.allocate_shared_buffer_and_handle(ops.meta_size())
self.meta_ptrs = [meta_ptr] * world_size
# This is a pre-registered IPC buffer. In eager mode, input tensors
# are first copied into this buffer before the operation is performed
legacy_buffer_size = max(max_size, max_all_gather_size, max_reduce_scatter_size)
if same_node:
self.buffer_ptrs = self.create_shared_buffer(
legacy_buffer_size,
group=group,
)
else:
buffer_ptr, _ = ops.allocate_shared_buffer_and_handle(legacy_buffer_size)
self.buffer_ptrs = [buffer_ptr] * world_size
# This stores tuples of pointers to IPC buffers from all ranks.
# Each registered tuple contains at most 16 addresses.
# Allocating 8MB is enough for 65536 such tuples. The largest model uses
# fewer than 10000 registered tuples.
self.rank_data = torch.empty(
8 * 1024 * 1024, dtype=torch.uint8, device=self.device
)
self.max_size = max_size
self.max_all_gather_size = max_all_gather_size
if max_mnnvl_all_gather_size is None:
max_mnnvl_all_gather_size = self._DEFAULT_MNNVL_ALL_GATHER_MAX_SIZES[
world_size
]
self.max_mnnvl_all_gather_size = max_mnnvl_all_gather_size
self.max_reduce_scatter_size = max_reduce_scatter_size
self.max_mnnvl_reduce_scatter_size = max_mnnvl_reduce_scatter_size
self.rank = rank
self.world_size = world_size
self.fully_connected = fully_connected
self._ptr = ops.init_custom_ar(
self.meta_ptrs, self.rank_data, rank, self.fully_connected
)
ops.register_buffer(self._ptr, self.buffer_ptrs)
self._init_mnnvl_buffer(
max(
max_mnnvl_all_gather_size * world_size,
max_mnnvl_reduce_scatter_size,
)
)
if not same_node and not self.mnnvl_multicast_ptr:
logger.warning(
"Custom collectives are disabled because this multi-node "
"group does not support MNNVL multicast."
)
self.close()
self.disabled = True
def _init_mnnvl_buffer(self, stage_size: int) -> None:
if torch_symm_mem is None or not current_platform.is_cuda():
return
try:
buffer_size = stage_size * 6
buffer = torch_symm_mem.empty(
buffer_size, dtype=torch.uint8, device=self.device
)
handle = torch_symm_mem.rendezvous(buffer, self.group.group_name)
if handle.multicast_ptr == 0:
return
peer_buffers = [
handle.get_buffer(
peer,
(buffer_size,),
torch.uint8,
storage_offset=0,
)
for peer in range(self.world_size)
]
ptrs = [peer_buffer.data_ptr() for peer_buffer in peer_buffers]
lamport_ag_offset = 0
lamport_rs_offset = stage_size * 3
lamport_ag_ptrs = [ptr + lamport_ag_offset for ptr in ptrs]
lamport_rs_ptrs = [ptr + lamport_rs_offset for ptr in ptrs]
ops.register_buffer(self._ptr, lamport_ag_ptrs)
ops.register_buffer(self._ptr, lamport_rs_ptrs)
buffer.view(torch.int32).fill_(-2147483648)
epochs = torch.zeros(
(2, 32),
dtype=torch.int32,
device=self.device,
)
torch.accelerator.synchronize()
dist.barrier(group=self.group)
self.mnnvl_buffer = buffer
self.mnnvl_handle = handle
self.mnnvl_peer_buffers = peer_buffers
self.mnnvl_multicast_ptr = handle.multicast_ptr
self.mnnvl_buffer_size = stage_size
self.mnnvl_lamport_ag_local_ptr = lamport_ag_ptrs[self.rank]
self.mnnvl_lamport_ag_multicast_ptr = (
handle.multicast_ptr + lamport_ag_offset
)
self.mnnvl_lamport_rs_local_ptr = lamport_rs_ptrs[self.rank]
self.mnnvl_lamport_epochs = epochs
self.mnnvl_lamport_ag_epoch_ptr = epochs[0].data_ptr()
self.mnnvl_lamport_rs_epoch_ptr = epochs[1].data_ptr()
except RuntimeError as error:
logger.debug("MNNVL AG/RS initialization failed: %s", error)
@contextmanager
def capture(self):
"""
The main responsibility of this context manager is the
`register_graph_buffers` call at the end of the context.
It records all the buffer addresses used in the CUDA graph.
"""
try:
self._IS_CAPTURING = True
yield
finally:
self._IS_CAPTURING = False
if not self.disabled:
self.register_graph_buffers()
def register_graph_buffers(self):
handle, offset = ops.get_graph_buffer_ipc_meta(self._ptr)
logger.debug("Registering %d cuda graph addresses", len(offset))
# We cannot directly use `dist.all_gather_object` here
# because it is incompatible with `gloo` backend under inference mode.
# see https://github.com/pytorch/pytorch/issues/126032 for details.
all_data: list[list[list[int] | None]]
all_data = [[None, None] for _ in range(dist.get_world_size(group=self.group))]
all_data[self.rank] = [handle, offset]
ranks = sorted(dist.get_process_group_ranks(group=self.group))
for i, rank in enumerate(ranks):
dist.broadcast_object_list(
all_data[i], src=rank, group=self.group, device="cpu"
)
# Unpack list of tuples to tuple of lists.
handles = cast(list[list[int]], [d[0] for d in all_data])
offsets = cast(list[list[int]], [d[1] for d in all_data])
ops.register_graph_buffers(self._ptr, handles, offsets)
def should_custom_ar(self, inp: torch.Tensor):
if self.disabled or self.world_size > 8:
return False
inp_size = inp.numel() * inp.element_size()
# custom allreduce requires input byte size to be multiples of 16
if inp_size % 16 != 0:
return False
if not is_weak_contiguous(inp):
return False
# for 4 or more non NVLink-capable GPUs, custom allreduce provides
# little performance improvement over NCCL.
if self.world_size == 2 or self.fully_connected:
return inp_size < self.max_size
return False
def all_reduce(
self, inp: torch.Tensor, *, out: torch.Tensor = None, registered: bool = False
):
"""Performs an out-of-place all reduce.
If registered is True, this assumes inp's pointer is already
IPC-registered. Otherwise, inp is first copied into a pre-registered
buffer.
"""
if out is None:
out = torch.empty_like(inp)
if registered:
ops.all_reduce(self._ptr, inp, out, 0, 0)
else:
ops.all_reduce(
self._ptr, inp, out, self.buffer_ptrs[self.rank], self.max_size
)
return out
def custom_all_reduce(self, input: torch.Tensor) -> torch.Tensor | None:
"""The main allreduce API that provides support for cuda graph."""
# When custom allreduce is disabled, this will be None.
if self.disabled or not self.should_custom_ar(input):
return None
if self._IS_CAPTURING:
if torch.cuda.is_current_stream_capturing():
return self.all_reduce(input, registered=True)
else:
# If warm up, mimic the allocation pattern since custom
# allreduce is out-of-place.
return torch.empty_like(input)
else:
# Note: outside of cuda graph context, custom allreduce incurs a
# cost of cudaMemcpy, which should be small (<=1% of overall
# latency) compared to the performance gain of using custom kernels
return self.all_reduce(input, registered=False)
def should_custom_all_gather(self, inp: torch.Tensor) -> bool:
if self.disabled or not current_platform.is_cuda():
return False
if self.world_size == 16 and not self.mnnvl_only:
return False
inp_size = inp.nbytes
if inp.dtype not in (
torch.float32,
torch.float16,
torch.bfloat16,
):
return False
max_size = (
self.max_mnnvl_all_gather_size
if self.mnnvl_multicast_ptr
else self.max_all_gather_size
)
return (
0 < inp_size <= max_size
and inp_size % 16 == 0
and is_weak_contiguous(inp)
and (self.fully_connected or bool(self.mnnvl_multicast_ptr))
)
def custom_all_gather(self, inp: torch.Tensor) -> torch.Tensor | None:
if not self.should_custom_all_gather(inp):
return None
out_shape = (inp.shape[0] * self.world_size,) + inp.shape[1:]
if self.mnnvl_multicast_ptr:
logger.info_once(
"Using the MNNVL Lamport all-gather kernel.",
scope="global",
)
out = torch.empty(out_shape, dtype=inp.dtype, device=inp.device)
ops.mnnvl_lamport_all_gather(
self._ptr,
inp,
out,
self.mnnvl_lamport_ag_local_ptr,
self.mnnvl_lamport_ag_multicast_ptr,
self.mnnvl_lamport_ag_epoch_ptr,
self.mnnvl_buffer_size,
)
else:
out = torch.empty(out_shape, dtype=inp.dtype, device=inp.device)
ops.custom_all_gather(
self._ptr,
inp,
out,
self.buffer_ptrs[self.rank],
self.max_all_gather_size,
)
return out
def should_custom_reduce_scatter(self, inp: torch.Tensor) -> bool:
if self.disabled or not current_platform.is_cuda():
return False
if self.world_size == 16 and not self.mnnvl_only:
return False
inp_size = inp.nbytes
if inp.dtype not in (torch.float32, torch.float16, torch.bfloat16):
return False
if inp.shape[0] % self.world_size != 0:
return False
output_size = inp_size // self.world_size
max_size = (
self.max_mnnvl_reduce_scatter_size
if self.mnnvl_multicast_ptr
else self.max_reduce_scatter_size
)
return (
0 < inp_size <= max_size
and output_size % 16 == 0
and is_weak_contiguous(inp)
and (self.fully_connected or bool(self.mnnvl_multicast_ptr))
)
def custom_reduce_scatter(self, inp: torch.Tensor) -> torch.Tensor | None:
if not self.should_custom_reduce_scatter(inp):
return None
out_shape = (inp.shape[0] // self.world_size,) + inp.shape[1:]
out = torch.empty(out_shape, dtype=inp.dtype, device=inp.device)
if self.mnnvl_multicast_ptr:
logger.info_once(
"Using the MNNVL Lamport reduce-scatter kernel.",
scope="global",
)
ops.mnnvl_lamport_reduce_scatter(
self._ptr,
inp,
out,
self.mnnvl_lamport_rs_local_ptr,
self.mnnvl_lamport_rs_epoch_ptr,
self.mnnvl_buffer_size,
)
else:
ops.custom_reduce_scatter(
self._ptr,
inp,
out,
self.buffer_ptrs[self.rank],
self.max_reduce_scatter_size,
)
return out
def close(self):
if not self.disabled and self._ptr:
if ops is not None:
ops.dispose(self._ptr)
self._ptr = 0
self.free_shared_buffer(self.meta_ptrs, rank=self.rank)
self.free_shared_buffer(self.buffer_ptrs, rank=self.rank)
self.mnnvl_peer_buffers = None
self.mnnvl_handle = None
self.mnnvl_buffer = None
self.mnnvl_lamport_epochs = None
def __del__(self):
self.close()
@staticmethod
def create_shared_buffer(
size_in_bytes: int,
group: ProcessGroup | None = None,
uncached: bool | None = False,
) -> list[int]:
pointer, handle = ops.allocate_shared_buffer_and_handle(size_in_bytes)
world_size = dist.get_world_size(group=group)
rank = dist.get_rank(group=group)
handles = [None] * world_size
dist.all_gather_object(handles, handle, group=group)
pointers: list[int] = []
for i, h in enumerate(handles):
if i == rank:
pointers.append(pointer) # type: ignore
else:
pointers.append(ops.open_mem_handle(h))
return pointers
@staticmethod
def free_shared_buffer(
pointers: list[int],
group: ProcessGroup | None = None,
rank: int | None = None,
) -> None:
if rank is None:
rank = dist.get_rank(group=group)
if ops is not None:
ops.free_shared_buffer(pointers[rank])