Convert formatting to use ruff instead of yapf + isort (#26247)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -11,7 +11,9 @@ from torch.distributed import ProcessGroup
|
||||
import vllm.envs as envs
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.distributed.device_communicators.all_reduce_utils import (
|
||||
CUSTOM_ALL_REDUCE_MAX_SIZES, gpu_p2p_access_check)
|
||||
CUSTOM_ALL_REDUCE_MAX_SIZES,
|
||||
gpu_p2p_access_check,
|
||||
)
|
||||
from vllm.distributed.parallel_state import in_the_same_node_as
|
||||
from vllm.logger import init_logger
|
||||
from vllm.platforms import current_platform
|
||||
@@ -32,8 +34,7 @@ def _can_p2p(rank: int, world_size: int) -> bool:
|
||||
if i == rank:
|
||||
continue
|
||||
if envs.VLLM_SKIP_P2P_CHECK:
|
||||
logger.info(
|
||||
"Skipping P2P check and trusting the driver's P2P report.")
|
||||
logger.info("Skipping P2P check and trusting the driver's P2P report.")
|
||||
return torch.cuda.can_device_access_peer(rank, i)
|
||||
if not gpu_p2p_access_check(rank, i):
|
||||
return False
|
||||
@@ -41,21 +42,23 @@ def _can_p2p(rank: int, world_size: int) -> bool:
|
||||
|
||||
|
||||
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())
|
||||
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_size: max supported allreduce size
|
||||
def __init__(self,
|
||||
group: ProcessGroup,
|
||||
device: Union[int, str, torch.device],
|
||||
max_size=8192 * 1024,
|
||||
symm_mem_enabled=False) -> None:
|
||||
def __init__(
|
||||
self,
|
||||
group: ProcessGroup,
|
||||
device: Union[int, str, torch.device],
|
||||
max_size=8192 * 1024,
|
||||
symm_mem_enabled=False,
|
||||
) -> None:
|
||||
"""
|
||||
Args:
|
||||
group: the process group to work on. If None, it will use the
|
||||
@@ -72,20 +75,24 @@ class CustomAllreduce:
|
||||
if not custom_ar:
|
||||
# disable because of missing custom allreduce library
|
||||
# e.g. in a non-GPU environment
|
||||
logger.info("Custom allreduce is disabled because "
|
||||
"of missing custom allreduce library")
|
||||
logger.info(
|
||||
"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.")
|
||||
"CustomAllreduce should be attached to a non-NCCL group."
|
||||
)
|
||||
|
||||
if not all(in_the_same_node_as(group, source_rank=0)):
|
||||
# No need to initialize custom allreduce for multi-node case.
|
||||
logger.warning(
|
||||
"Custom allreduce is disabled because this process group"
|
||||
" spans across nodes.")
|
||||
" spans across nodes."
|
||||
)
|
||||
return
|
||||
|
||||
rank = dist.get_rank(group=self.group)
|
||||
@@ -100,7 +107,9 @@ class CustomAllreduce:
|
||||
"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))
|
||||
world_size,
|
||||
str(CustomAllreduce._SUPPORTED_WORLD_SIZES),
|
||||
)
|
||||
return
|
||||
|
||||
if isinstance(device, int):
|
||||
@@ -110,13 +119,15 @@ class CustomAllreduce:
|
||||
# now `device` is a `torch.device` object
|
||||
assert isinstance(device, torch.device)
|
||||
self.device = device
|
||||
device_capability = current_platform.get_device_capability(
|
||||
).as_version_str()
|
||||
if (current_platform.is_cuda() and symm_mem_enabled
|
||||
and device_capability in CUSTOM_ALL_REDUCE_MAX_SIZES):
|
||||
device_capability = current_platform.get_device_capability().as_version_str()
|
||||
if (
|
||||
current_platform.is_cuda()
|
||||
and symm_mem_enabled
|
||||
and device_capability in CUSTOM_ALL_REDUCE_MAX_SIZES
|
||||
):
|
||||
max_size = min(
|
||||
CUSTOM_ALL_REDUCE_MAX_SIZES[device_capability][world_size],
|
||||
max_size)
|
||||
CUSTOM_ALL_REDUCE_MAX_SIZES[device_capability][world_size], max_size
|
||||
)
|
||||
cuda_visible_devices = envs.CUDA_VISIBLE_DEVICES
|
||||
if cuda_visible_devices:
|
||||
device_ids = list(map(int, cuda_visible_devices.split(",")))
|
||||
@@ -124,12 +135,9 @@ class CustomAllreduce:
|
||||
device_ids = list(range(cuda_device_count_stateless()))
|
||||
|
||||
physical_device_id = device_ids[device.index]
|
||||
tensor = torch.tensor([physical_device_id],
|
||||
dtype=torch.int,
|
||||
device="cpu")
|
||||
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)
|
||||
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]
|
||||
@@ -138,13 +146,13 @@ class CustomAllreduce:
|
||||
# where custom allreduce is not supported
|
||||
# this checks hardware and driver support for NVLink
|
||||
assert current_platform.is_cuda_alike()
|
||||
fully_connected = current_platform.is_fully_connected(
|
||||
physical_device_ids)
|
||||
fully_connected = current_platform.is_fully_connected(physical_device_ids)
|
||||
if 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.")
|
||||
"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
|
||||
@@ -154,16 +162,17 @@ class CustomAllreduce:
|
||||
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.")
|
||||
"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.
|
||||
self.meta_ptrs = self.create_shared_buffer(ops.meta_size() + max_size,
|
||||
group=group,
|
||||
uncached=True)
|
||||
self.meta_ptrs = self.create_shared_buffer(
|
||||
ops.meta_size() + max_size, group=group, uncached=True
|
||||
)
|
||||
# This is a pre-registered IPC buffer. In eager mode, input tensors
|
||||
# are first copied into this buffer before allreduce is performed
|
||||
self.buffer_ptrs = self.create_shared_buffer(max_size, group=group)
|
||||
@@ -172,21 +181,22 @@ class CustomAllreduce:
|
||||
# 8*world_size bytes where world_size is at most 8. Allocating 8MB
|
||||
# is enough for 131072 such tuples. The largest model I've seen only
|
||||
# needs less than 10000 of registered tuples.
|
||||
self.rank_data = torch.empty(8 * 1024 * 1024,
|
||||
dtype=torch.uint8,
|
||||
device=self.device)
|
||||
self.rank_data = torch.empty(
|
||||
8 * 1024 * 1024, dtype=torch.uint8, device=self.device
|
||||
)
|
||||
self.max_size = max_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)
|
||||
self._ptr = ops.init_custom_ar(
|
||||
self.meta_ptrs, self.rank_data, rank, self.fully_connected
|
||||
)
|
||||
ops.register_buffer(self._ptr, self.buffer_ptrs)
|
||||
|
||||
@contextmanager
|
||||
def capture(self):
|
||||
"""
|
||||
The main responsibility of this context manager is the
|
||||
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.
|
||||
"""
|
||||
@@ -204,15 +214,13 @@ class CustomAllreduce:
|
||||
# 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 = [[None, None]
|
||||
for _ in range(dist.get_world_size(group=self.group))]
|
||||
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")
|
||||
dist.broadcast_object_list(
|
||||
all_data[i], src=rank, group=self.group, device="cpu"
|
||||
)
|
||||
# Unpack list of tuples to tuple of lists.
|
||||
handles = [d[0] for d in all_data] # type: ignore
|
||||
offsets = [d[1] for d in all_data] # type: ignore
|
||||
@@ -233,13 +241,11 @@ class CustomAllreduce:
|
||||
return inp_size < self.max_size
|
||||
return False
|
||||
|
||||
def all_reduce(self,
|
||||
inp: torch.Tensor,
|
||||
*,
|
||||
out: torch.Tensor = None,
|
||||
registered: bool = 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.
|
||||
@@ -249,8 +255,9 @@ class CustomAllreduce:
|
||||
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)
|
||||
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) -> Optional[torch.Tensor]:
|
||||
@@ -283,9 +290,11 @@ class CustomAllreduce:
|
||||
self.close()
|
||||
|
||||
@staticmethod
|
||||
def create_shared_buffer(size_in_bytes: int,
|
||||
group: Optional[ProcessGroup] = None,
|
||||
uncached: Optional[bool] = False) -> list[int]:
|
||||
def create_shared_buffer(
|
||||
size_in_bytes: int,
|
||||
group: Optional[ProcessGroup] = None,
|
||||
uncached: Optional[bool] = False,
|
||||
) -> list[int]:
|
||||
pointer, handle = ops.allocate_shared_buffer_and_handle(size_in_bytes)
|
||||
|
||||
world_size = dist.get_world_size(group=group)
|
||||
@@ -302,9 +311,11 @@ class CustomAllreduce:
|
||||
return pointers
|
||||
|
||||
@staticmethod
|
||||
def free_shared_buffer(pointers: list[int],
|
||||
group: Optional[ProcessGroup] = None,
|
||||
rank: Optional[int] = None) -> None:
|
||||
def free_shared_buffer(
|
||||
pointers: list[int],
|
||||
group: Optional[ProcessGroup] = None,
|
||||
rank: Optional[int] = None,
|
||||
) -> None:
|
||||
if rank is None:
|
||||
rank = dist.get_rank(group=group)
|
||||
if ops is not None:
|
||||
|
||||
Reference in New Issue
Block a user