[Bugfix] Restore xpu_communicator for CUDA graph capture
parallel_state.graph_capture() imports xpu_communicator on CUDA too. The SM86 cut removed the file and capture crashed at startup. Co-authored-by: Hinata <hinata@idihore.id> Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
@@ -0,0 +1,296 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
|
||||
import torch
|
||||
import torch.distributed as dist
|
||||
from torch.distributed import ProcessGroup
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.logger import init_logger
|
||||
|
||||
from .base_device_communicator import DeviceCommunicatorBase
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class XpuCommunicator(DeviceCommunicatorBase):
|
||||
def __init__(
|
||||
self,
|
||||
cpu_group: ProcessGroup,
|
||||
device: torch.device | None = None,
|
||||
device_group: ProcessGroup | None = None,
|
||||
unique_name: str = "",
|
||||
use_all2all: bool = False,
|
||||
):
|
||||
super().__init__(
|
||||
cpu_group, device, device_group, unique_name, use_all2all=use_all2all
|
||||
)
|
||||
self.ca_comm: None = None
|
||||
if self.use_all2all:
|
||||
if self.all2all_backend in ("naive", "allgather_reducescatter"):
|
||||
from .all2all import AgRsAll2AllManager
|
||||
|
||||
self.all2all_manager = AgRsAll2AllManager(self.cpu_group)
|
||||
logger.info("Using AgRs manager on XPU device.")
|
||||
|
||||
else: # type: ignore[has-type]
|
||||
logger.warning(
|
||||
"`%s` all2all manager is not supported on XPU. "
|
||||
"Falling back to AgRs manager for XPU, "
|
||||
"which is the Default backend",
|
||||
self.all2all_backend, # type: ignore[has-type]
|
||||
)
|
||||
from .all2all import AgRsAll2AllManager
|
||||
|
||||
self.all2all_manager = AgRsAll2AllManager(self.cpu_group)
|
||||
logger.info("Using AgRs manager on XPU device.")
|
||||
|
||||
def _fixed_rank_sum(self, input_: torch.Tensor) -> torch.Tensor:
|
||||
flat_input = input_.reshape(-1)
|
||||
gathered = torch.empty(
|
||||
(self.world_size, flat_input.numel()),
|
||||
dtype=input_.dtype,
|
||||
device=input_.device,
|
||||
)
|
||||
dist.all_gather_single(gathered.view(-1), flat_input, group=self.device_group)
|
||||
output = gathered[0].clone()
|
||||
for rank in range(1, self.world_size):
|
||||
output.add_(gathered[rank])
|
||||
return output.view(input_.shape)
|
||||
|
||||
def all_reduce(self, input_: torch.Tensor) -> torch.Tensor:
|
||||
if envs.VLLM_BATCH_INVARIANT:
|
||||
# XCCL changes its reduction order with message size. Gather first
|
||||
# and accumulate in rank order, independent of the token batch.
|
||||
return self._fixed_rank_sum(input_)
|
||||
|
||||
output = input_.clone()
|
||||
dist.all_reduce(output, group=self.device_group)
|
||||
return output
|
||||
|
||||
def reduce_scatter(self, input_: torch.Tensor, dim: int = -1):
|
||||
world_size = self.world_size
|
||||
|
||||
if dim < 0:
|
||||
# Convert negative dim to positive.
|
||||
dim += input_.dim()
|
||||
|
||||
# Note: This will produce an incorrect answer if we don't make
|
||||
# the input_tensor contiguous. Possible bug in reduce_scatter_tensor?
|
||||
input_tensor = input_.movedim(0, dim).contiguous()
|
||||
|
||||
assert input_tensor.shape[0] % world_size == 0
|
||||
chunk_size = input_tensor.shape[0] // world_size
|
||||
output_shape = (chunk_size,) + input_tensor.shape[1:]
|
||||
|
||||
output = torch.empty(
|
||||
output_shape, dtype=input_tensor.dtype, device=input_tensor.device
|
||||
)
|
||||
|
||||
if envs.VLLM_BATCH_INVARIANT:
|
||||
reduced = self._fixed_rank_sum(input_tensor)
|
||||
start = self.rank_in_group * chunk_size
|
||||
output.copy_(reduced.narrow(0, start, chunk_size))
|
||||
else:
|
||||
dist.reduce_scatter_tensor(output, input_tensor, group=self.device_group)
|
||||
|
||||
# Reshape before returning
|
||||
return output.movedim(0, dim).contiguous()
|
||||
|
||||
def reduce_scatterv(
|
||||
self, input_: torch.Tensor, dim: int = -1, sizes: list[int] | None = None
|
||||
):
|
||||
world_size = self.world_size
|
||||
|
||||
if dim < 0:
|
||||
# Convert negative dim to positive.
|
||||
dim += input_.dim()
|
||||
|
||||
# Note: This will produce an incorrect answer if we don't make
|
||||
# the input_tensor contiguous. Possible bug in reduce_scatter_tensor?
|
||||
input_tensor = input_.movedim(0, dim).contiguous()
|
||||
|
||||
if sizes is not None:
|
||||
assert len(sizes) == world_size
|
||||
assert input_tensor.shape[0] == sum(sizes)
|
||||
chunk_size = sizes[self.rank_in_group]
|
||||
else:
|
||||
assert input_tensor.shape[0] % world_size == 0
|
||||
chunk_size = input_tensor.shape[0] // world_size
|
||||
output_shape = (chunk_size,) + input_tensor.shape[1:]
|
||||
|
||||
output = torch.empty(
|
||||
output_shape, dtype=input_tensor.dtype, device=input_tensor.device
|
||||
)
|
||||
if envs.VLLM_BATCH_INVARIANT:
|
||||
reduced = self._fixed_rank_sum(input_tensor)
|
||||
start = (
|
||||
self.rank_in_group * chunk_size
|
||||
if sizes is None
|
||||
else sum(sizes[: self.rank_in_group])
|
||||
)
|
||||
output.copy_(reduced.narrow(0, start, chunk_size))
|
||||
elif sizes is None or sizes.count(sizes[0]) == len(sizes):
|
||||
dist.reduce_scatter_tensor(output, input_tensor, group=self.device_group)
|
||||
elif torch.xpu.device_count() < self.world_size:
|
||||
# uneven reduce_scatter is not supported when the
|
||||
# process cannot see every device in the group (
|
||||
# e.g. external-LBDP).
|
||||
max_size = max(sizes)
|
||||
padded = input_tensor.new_zeros(
|
||||
(world_size * max_size,) + input_tensor.shape[1:]
|
||||
)
|
||||
for rank, split in enumerate(input_tensor.split(sizes, dim=0)):
|
||||
padded[rank * max_size : rank * max_size + sizes[rank]] = split
|
||||
padded_output = input_tensor.new_empty((max_size,) + input_tensor.shape[1:])
|
||||
dist.reduce_scatter_tensor(padded_output, padded, group=self.device_group)
|
||||
output = padded_output[:chunk_size]
|
||||
else:
|
||||
input_splits = list(input_tensor.split(sizes, dim=0))
|
||||
dist.reduce_scatter(output, input_splits, group=self.device_group)
|
||||
# Reshape before returning
|
||||
return output.movedim(0, dim).contiguous()
|
||||
|
||||
def all_gatherv(
|
||||
self,
|
||||
input_: torch.Tensor | list[torch.Tensor],
|
||||
dim: int = 0,
|
||||
sizes: list[int] | None = None,
|
||||
):
|
||||
if dim != 0:
|
||||
raise NotImplementedError("only dim 0 all-gatherv is supported")
|
||||
world_size = self.world_size
|
||||
|
||||
# 'sizes' is not needed if all inputs in the same group have the same
|
||||
# shape
|
||||
if sizes is not None and all(s == sizes[0] for s in sizes):
|
||||
sizes = None
|
||||
|
||||
def _all_gather_single(input_: torch.Tensor, sizes: list[int] | None = None):
|
||||
input_size = input_.size()
|
||||
if sizes is not None:
|
||||
assert len(sizes) == world_size
|
||||
assert input_.shape[dim] == sizes[self.rank_in_group], (
|
||||
f"{input_.shape[dim]} != {sizes[self.rank_in_group]}"
|
||||
)
|
||||
output_size = (sum(sizes),) + input_size[1:]
|
||||
else:
|
||||
output_size = (input_size[0] * world_size,) + input_size[1:]
|
||||
# Allocate output tensor.
|
||||
output_tensor = torch.empty(
|
||||
output_size, dtype=input_.dtype, device=input_.device
|
||||
)
|
||||
|
||||
if sizes is not None:
|
||||
all_gather_list = []
|
||||
for size in sizes:
|
||||
all_gather_list.append(
|
||||
torch.empty(
|
||||
(size,) + input_.shape[1:],
|
||||
dtype=input_.dtype,
|
||||
device=input_.device,
|
||||
)
|
||||
)
|
||||
dist.all_gather(all_gather_list, input_, group=self.device_group)
|
||||
output_tensor = torch.cat(all_gather_list, dim=0)
|
||||
else:
|
||||
dist.all_gather([output_tensor], input_, group=self.device_group)
|
||||
return output_tensor
|
||||
|
||||
if isinstance(input_, torch.Tensor):
|
||||
return _all_gather_single(input_, sizes)
|
||||
|
||||
output_list = []
|
||||
for inp in input_:
|
||||
output_list.append(_all_gather_single(inp, sizes=sizes))
|
||||
return output_list
|
||||
|
||||
def gather(
|
||||
self, input_: torch.Tensor, dst: int = 0, dim: int = -1
|
||||
) -> torch.Tensor | None:
|
||||
assert -input_.dim() <= dim < input_.dim(), (
|
||||
f"Invalid dim ({dim}) for input tensor with shape {input_.size()}"
|
||||
)
|
||||
if dim < 0:
|
||||
# Convert negative dim to positive.
|
||||
dim += input_.dim()
|
||||
# For xpu path, gather doesn't work properly together with ray
|
||||
# cluster so we use all_gather instead for now.
|
||||
input_size = input_.size()
|
||||
# Allocate output tensor.
|
||||
output_tensor = torch.empty(
|
||||
(self.world_size,) + input_size, dtype=input_.dtype, device=input_.device
|
||||
)
|
||||
# All-gather.
|
||||
dist.all_gather_single(output_tensor, input_, group=self.device_group)
|
||||
if self.rank_in_group == dst:
|
||||
# Reshape
|
||||
output_tensor = output_tensor.movedim(0, dim)
|
||||
output_tensor = output_tensor.reshape(
|
||||
input_size[:dim]
|
||||
+ (self.world_size * input_size[dim],)
|
||||
+ input_size[dim + 1 :]
|
||||
)
|
||||
else:
|
||||
output_tensor = None
|
||||
return output_tensor
|
||||
|
||||
def broadcast(self, input_: torch.Tensor, src: int = 0) -> None:
|
||||
dist.broadcast(input_, src=src, group=self.device_group)
|
||||
|
||||
def dispatch_router_logits(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
router_logits: torch.Tensor,
|
||||
is_sequence_parallel: bool = False,
|
||||
extra_tensors: list[torch.Tensor] | None = None,
|
||||
) -> (
|
||||
tuple[torch.Tensor, torch.Tensor]
|
||||
| tuple[torch.Tensor, torch.Tensor, list[torch.Tensor]]
|
||||
):
|
||||
"""Dispatch the hidden states and router logits to the appropriate device.
|
||||
This is a no-op in the base class.
|
||||
"""
|
||||
assert self.all2all_manager is not None
|
||||
return self.all2all_manager.dispatch_router_logits(
|
||||
hidden_states,
|
||||
router_logits,
|
||||
is_sequence_parallel,
|
||||
extra_tensors,
|
||||
)
|
||||
|
||||
def dispatch(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
topk_weights: torch.Tensor,
|
||||
topk_ids: torch.Tensor,
|
||||
is_sequence_parallel: bool = False,
|
||||
extra_tensors: list[torch.Tensor] | None = None,
|
||||
) -> (
|
||||
tuple[torch.Tensor, torch.Tensor, torch.Tensor]
|
||||
| tuple[torch.Tensor, torch.Tensor, torch.Tensor, list[torch.Tensor]]
|
||||
):
|
||||
"""Dispatch the hidden states and topk weights/ids to the appropriate device.
|
||||
This is a no-op in the base class.
|
||||
"""
|
||||
assert self.all2all_manager is not None
|
||||
return self.all2all_manager.dispatch(
|
||||
hidden_states,
|
||||
topk_weights,
|
||||
topk_ids,
|
||||
is_sequence_parallel,
|
||||
extra_tensors=extra_tensors,
|
||||
)
|
||||
|
||||
def combine(
|
||||
self, hidden_states: torch.Tensor, is_sequence_parallel: bool = False
|
||||
) -> torch.Tensor:
|
||||
"""Combine the hidden states and router logits from the appropriate device.
|
||||
This is a no-op in the base class.
|
||||
"""
|
||||
assert self.all2all_manager is not None
|
||||
return self.all2all_manager.combine(
|
||||
hidden_states,
|
||||
is_sequence_parallel,
|
||||
)
|
||||
@@ -1,130 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Encoder-cache connector backed by Mooncake TransferEngine."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from typing import TYPE_CHECKING, Any
|
||||
|
||||
from vllm.distributed.ec_transfer.ec_connector.base import (
|
||||
ECConnectorBase,
|
||||
ECConnectorMetadata,
|
||||
ECConnectorRole,
|
||||
ECConnectorWorkerMetadata,
|
||||
)
|
||||
from vllm.distributed.ec_transfer.ec_connector.mooncake.metadata import (
|
||||
ECMooncakeConnectorMetadata,
|
||||
)
|
||||
from vllm.distributed.ec_transfer.ec_connector.mooncake.scheduler import (
|
||||
ECMooncakeScheduler,
|
||||
)
|
||||
from vllm.distributed.ec_transfer.ec_connector.mooncake.worker import ECMooncakeWorker
|
||||
|
||||
if TYPE_CHECKING:
|
||||
import torch
|
||||
|
||||
from vllm.config import VllmConfig
|
||||
from vllm.v1.core.sched.output import SchedulerOutput
|
||||
from vllm.v1.outputs import ECConnectorOutput
|
||||
from vllm.v1.request import Request
|
||||
|
||||
|
||||
class ECMooncakeConnector(ECConnectorBase):
|
||||
"""Preserve the public API while delegating to one process-role component."""
|
||||
|
||||
def __init__(self, vllm_config: VllmConfig, role: ECConnectorRole):
|
||||
super().__init__(vllm_config=vllm_config, role=role)
|
||||
|
||||
self._scheduler: ECMooncakeScheduler | None = None
|
||||
self._worker: ECMooncakeWorker | None = None
|
||||
self._closed = False
|
||||
|
||||
if role == ECConnectorRole.SCHEDULER:
|
||||
self._scheduler = ECMooncakeScheduler(vllm_config)
|
||||
elif role == ECConnectorRole.WORKER:
|
||||
self._worker = ECMooncakeWorker(vllm_config)
|
||||
else:
|
||||
raise ValueError(f"Unknown EC connector role: {role}")
|
||||
|
||||
def start_worker_services(self) -> None:
|
||||
assert self._worker is not None
|
||||
self._worker.start_services()
|
||||
|
||||
def start_save_caches(self, **kwargs: Any) -> None:
|
||||
assert self._worker is not None
|
||||
metadata = self._get_connector_metadata()
|
||||
assert isinstance(metadata, ECMooncakeConnectorMetadata)
|
||||
self._worker.start_save_caches(metadata, **kwargs)
|
||||
|
||||
def start_load_caches(
|
||||
self, encoder_cache: dict[str, torch.Tensor], **kwargs: Any
|
||||
) -> None:
|
||||
assert self._worker is not None
|
||||
metadata = self._get_connector_metadata()
|
||||
assert isinstance(metadata, ECMooncakeConnectorMetadata)
|
||||
self._worker.start_load_caches(metadata, encoder_cache, **kwargs)
|
||||
|
||||
def save_caches(
|
||||
self, encoder_cache: dict[str, torch.Tensor], mm_hash: str, **kwargs: Any
|
||||
) -> None:
|
||||
assert self._worker is not None
|
||||
self._worker.save_caches(encoder_cache, mm_hash, **kwargs)
|
||||
|
||||
def get_finished(
|
||||
self, finished_req_ids: set[str]
|
||||
) -> tuple[set[str] | None, set[str] | None]:
|
||||
assert self._worker is not None
|
||||
return self._worker.get_finished(finished_req_ids)
|
||||
|
||||
def build_connector_worker_meta(self) -> ECConnectorWorkerMetadata | None:
|
||||
assert self._worker is not None
|
||||
return self._worker.build_connector_worker_meta()
|
||||
|
||||
def take_unavailable_requests(self) -> set[str]:
|
||||
assert self._scheduler is not None
|
||||
return self._scheduler.take_unavailable_requests()
|
||||
|
||||
def has_cache_item(self, identifier: str) -> bool:
|
||||
assert self._scheduler is not None
|
||||
return self._scheduler.has_cache_item(identifier)
|
||||
|
||||
def ensure_cache_available(
|
||||
self, request: Request, num_computed_tokens: int
|
||||
) -> bool:
|
||||
assert self._scheduler is not None
|
||||
return self._scheduler.ensure_cache_available(request, num_computed_tokens)
|
||||
|
||||
def update_state_after_alloc(self, request: Request, index: int) -> None:
|
||||
assert self._scheduler is not None
|
||||
self._scheduler.update_state_after_alloc(request, index)
|
||||
|
||||
def update_state_after_free(self, request: Request, index: int) -> None:
|
||||
assert self._scheduler is not None
|
||||
self._scheduler.update_state_after_free(request, index)
|
||||
|
||||
def build_connector_meta(
|
||||
self, scheduler_output: SchedulerOutput
|
||||
) -> ECConnectorMetadata:
|
||||
assert self._scheduler is not None
|
||||
return self._scheduler.build_connector_meta(scheduler_output)
|
||||
|
||||
def update_connector_output(self, connector_output: ECConnectorOutput) -> None:
|
||||
assert self._scheduler is not None
|
||||
self._scheduler.update_connector_output(connector_output)
|
||||
|
||||
def has_pending_push_work(self) -> bool:
|
||||
assert self._scheduler is not None
|
||||
return self._scheduler.has_pending_push_work()
|
||||
|
||||
def request_finished(self, request: Request) -> tuple[bool, dict[str, Any] | None]:
|
||||
assert self._scheduler is not None
|
||||
return self._scheduler.request_finished(request)
|
||||
|
||||
def shutdown(self) -> None:
|
||||
if self._closed:
|
||||
return
|
||||
self._closed = True
|
||||
if self._scheduler is not None:
|
||||
self._scheduler.close()
|
||||
if self._worker is not None:
|
||||
self._worker.close()
|
||||
Reference in New Issue
Block a user