[rl] Stateful Trainer Send: IPC [2/N] (#48981)

Signed-off-by: haoaaron <ahao@anyscale.com>
This commit is contained in:
Aaron Hao
2026-07-30 21:59:08 +08:00
committed by GitHub
parent 59e831c09a
commit 30b4e7f479
8 changed files with 391 additions and 421 deletions
+37 -12
View File
@@ -5,7 +5,15 @@
from abc import ABC, abstractmethod
from collections.abc import Iterator
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, Any, Generic, Protocol, TypeVar, runtime_checkable
from typing import (
TYPE_CHECKING,
Any,
ClassVar,
Generic,
Protocol,
TypeVar,
runtime_checkable,
)
import torch
from typing_extensions import Self
@@ -18,7 +26,7 @@ from vllm.config.weight_transfer import WeightTransferConfig
TInitInfo = TypeVar("TInitInfo", bound="WeightTransferInitInfo")
TUpdateInfo = TypeVar("TUpdateInfo", bound="WeightTransferUpdateInfo")
TConfig = TypeVar("TConfig", bound="WeightTransferConfig")
TTrainerInitInfo = TypeVar("TTrainerInitInfo", bound="TrainerInitInfo")
# A trainer supplies its parameters as a `WeightSource` (defined below): a
# re-iterable stream of materialized `(name, tensor)` pairs plus a `metadata()`
@@ -108,7 +116,7 @@ class WeightTransferInitInfo(ABC): # noqa: B024
@dataclass
class TrainerInitInfo(WeightTransferInitInfo):
class TrainerInitInfo:
"""Base trainer-side init info: which trainer rank drives the transfer.
`rank` is this trainer process's rank, provided **explicitly** by the
@@ -118,10 +126,25 @@ class TrainerInitInfo(WeightTransferInitInfo):
while every rank still runs the trainer-side collectives. Backend subclasses
add their own (positional) fields; `rank` is keyword-only so that ordering
never conflicts.
Every concrete subclass sets a class-level `backend` string (the same key it
registers under in `WeightTransferTrainerFactory`). The factory reads it to
dispatch, so callers pass only the init info/ It is a `ClassVar`
(a fixed per-backend constant), so it is not an ``__init__`` field.
"""
backend: ClassVar[str]
rank: int = field(kw_only=True)
def __init_subclass__(cls, **kwargs: Any) -> None:
super().__init_subclass__(**kwargs)
if not getattr(cls, "backend", None):
raise TypeError(
f"{cls.__name__} must set a class-level `backend` string "
"(the WeightTransferTrainerFactory registry key)."
)
@property
def is_sender(self) -> bool:
return self.rank == 0
@@ -373,7 +396,7 @@ class VLLMWeightSyncClient(Protocol):
def finish_weight_update(self, weight_version: str | None = None) -> None: ...
class TrainerWeightTransferEngine(ABC, Generic[TConfig, TInitInfo]):
class TrainerWeightTransferEngine(ABC, Generic[TTrainerInitInfo]):
"""Trainer-side weight transfer engine.
Symmetric to `WeightTransferEngine` but lives in the training process.
@@ -382,6 +405,13 @@ class TrainerWeightTransferEngine(ABC, Generic[TConfig, TInitInfo]):
plans) on `self`. The `WeightSource` is required at `trainer_init`,
then replayed each round by the no-argument `send_weights()`.
Unlike the worker engine, the trainer side does not take a
`WeightTransferConfig`: the backend is selected from the init info's
`backend` `ClassVar` (so callers pass only the init info), and the static
wire params (packed, buffer sizes) ride the backend-specific
`TrainerInitInfo`, which the sender also propagates to the worker at the init
handshake.
Multi-rank trainers: `trainer_init` and `send_weights` are
called on *every* trainer rank. Rank 0 is the sender, resolved once at
`trainer_init` into `is_sender`. Non-sender ranks still run every
@@ -392,22 +422,18 @@ class TrainerWeightTransferEngine(ABC, Generic[TConfig, TInitInfo]):
Subclasses should define:
init_info_cls: Type of backend-specific trainer init info
config_cls: Type of backend-specific config
"""
# Subclasses should override these class attributes
init_info_cls: type[TInitInfo]
config_cls: type[TConfig]
# Subclasses should override this class attribute
init_info_cls: type[TTrainerInitInfo]
def __init__(
self,
config: TConfig,
*,
client: "VLLMWeightSyncClient",
source: "WeightSource",
is_sender: bool = True,
) -> None:
self.config = config
self.is_sender = is_sender
# The real client is held on every rank; each engine only *calls* it when
# `is_sender`, so non-sender ranks never touch the wire.
@@ -418,8 +444,7 @@ class TrainerWeightTransferEngine(ABC, Generic[TConfig, TInitInfo]):
@abstractmethod
def trainer_init(
cls,
config: TConfig,
init_info: TInitInfo,
init_info: TTrainerInitInfo,
*,
client: "VLLMWeightSyncClient",
source: "WeightSource",
+21 -9
View File
@@ -18,9 +18,9 @@ if TYPE_CHECKING:
from vllm.config import VllmConfig
from vllm.config.weight_transfer import WeightTransferConfig
from vllm.distributed.weight_transfer.base import (
TrainerInitInfo,
VLLMWeightSyncClient,
WeightSource,
WeightTransferInitInfo,
)
logger = init_logger(__name__)
@@ -164,9 +164,7 @@ class WeightTransferTrainerFactory:
@classmethod
def trainer_init(
cls,
backend: str,
config: "WeightTransferConfig",
init_info: "WeightTransferInitInfo",
init_info: "TrainerInitInfo",
*,
client: "VLLMWeightSyncClient",
source: "WeightSource",
@@ -176,16 +174,22 @@ class WeightTransferTrainerFactory:
Called on every trainer rank (multi-rank trainers construct on all
ranks; the sender is resolved inside the engine's ``trainer_init``).
The trainer side takes no `WeightTransferConfig` and no separate
`backend` argument: the backend is read from ``init_info.backend`` (a
`ClassVar` on each `TrainerInitInfo` subclass), and the static wire
params ride `init_info`.
Args:
backend: Backend name (must be registered).
config: Backend-specific weight transfer config.
init_info: Backend-specific trainer init info.
init_info: Backend-specific trainer init info. Its `backend`
selects the engine; it also carries the wire params (e.g.
`packed`).
client: Inference-side control-plane client.
source: `WeightSource` of `(name, tensor)` pairs to send each round.
Raises:
ValueError: If the backend is not registered.
ValueError: If `init_info.backend` is not registered.
"""
backend = init_info.backend
if backend not in cls._registry:
available = list(cls._registry.keys())
raise ValueError(
@@ -200,7 +204,6 @@ class WeightTransferTrainerFactory:
)
return engine_cls.trainer_init(
config=config,
init_info=init_info,
client=client,
source=source,
@@ -228,3 +231,12 @@ WeightTransferEngineFactory.register_engine(
"vllm.distributed.weight_transfer.sparse_nccl_engine",
"SparseNCCLWeightTransferEngine",
)
# Trainer-side engines. Backends register here as they migrate to the stateful
# trainer engine; NCCL / sparse NCCL keep their static trainer path until then.
WeightTransferTrainerFactory.register_engine(
"ipc",
"vllm.distributed.weight_transfer.ipc_engine",
"IPCTrainerWeightTransferEngine",
)
+183 -222
View File
@@ -3,19 +3,21 @@
"""IPC-based weight transfer engine using CUDA IPC for communication."""
import pickle
from collections.abc import Callable, Iterator
from dataclasses import asdict, dataclass
from typing import TYPE_CHECKING, Any
from typing import TYPE_CHECKING, Any, ClassVar
import pybase64 as base64
import ray
import requests
import torch
from torch.multiprocessing.reductions import rebuild_cuda_tensor, reduce_tensor
from typing_extensions import Self
from vllm import envs
from vllm.config.weight_transfer import WeightTransferConfig
from vllm.distributed.weight_transfer.base import (
TrainerInitInfo,
TrainerWeightTransferEngine,
VLLMWeightSyncClient,
WeightSource,
WeightTransferEngine,
WeightTransferInitInfo,
WeightTransferUpdateInfo,
@@ -31,47 +33,35 @@ from vllm.distributed.weight_transfer.packed_tensor import (
@dataclass
class IPCTrainerSendWeightsArgs:
"""Arguments for IPC trainer_send_weights method."""
class IPCWeightTransferInitInfo(WeightTransferInitInfo):
"""Worker-side init info for IPC weight transfer. No rendezvous needed.
`packed` is a must-agree wire param: the trainer ships it here at the init
handshake so the worker decodes with the same setting the trainer encoded
with. The consumer rebuilds from the IPC handle + `tensor_sizes`, so it does
not need the buffer size (producer-only)."""
send_mode: str | Callable[["IPCWeightTransferUpdateInfo"], None]
"""How to send updates to vLLM. Either a string ('ray' or 'http') for
built-in transports, or a callable that receives an
IPCWeightTransferUpdateInfo and performs the send."""
llm_handle: Any = None
"""Ray actor handle or list of handles (required for 'ray' send_mode)."""
url: str | None = None
"""Base URL for HTTP endpoint (required for 'http' send_mode)."""
packed: bool = False
"""Whether to use packed tensor transfer for bounded-memory chunking."""
packed_buffer_size_bytes: int = DEFAULT_PACKED_BUFFER_SIZE_BYTES
"""Size in bytes for each packed tensor buffer when packed=True."""
def __post_init__(self):
"""Validate that required arguments are provided for the selected mode."""
if callable(self.send_mode):
return
if self.send_mode == "ray" and self.llm_handle is None:
raise ValueError("llm_handle is required for 'ray' send_mode")
if self.send_mode == "http" and self.url is None:
raise ValueError("url is required for 'http' send_mode")
if self.send_mode not in ("ray", "http"):
raise ValueError(
f"send_mode must be 'ray', 'http', or a callable, "
f"got {self.send_mode!r}"
)
@dataclass
class IPCWeightTransferInitInfo(WeightTransferInitInfo):
"""Initialization info for IPC weight transfer backend. No init needed for IPC."""
class IPCTrainerInitInfo(TrainerInitInfo):
"""Trainer-side init info for IPC weight transfer. No rendezvous needed;
`rank` (from `TrainerInitInfo`) identifies this trainer process — rank 0
ships the merged IPC handles. All ranks still join the handle all-gather.
pass
`packed` / `packed_buffer_size_bytes` are the transfer's wire params. The
trainer propagates `packed` to the worker at `trainer_init` so the two sides
cannot disagree. `backend` is the factory dispatch key."""
backend: ClassVar[str] = "ipc"
packed: bool = False
packed_buffer_size_bytes: int = DEFAULT_PACKED_BUFFER_SIZE_BYTES
@dataclass
class IPCWeightTransferUpdateInfo(WeightTransferUpdateInfo):
"""Update info for IPC weight transfer backend."""
"""Per-round update info for the IPC weight transfer backend."""
names: list[str]
dtype_names: list[str]
@@ -85,8 +75,6 @@ class IPCWeightTransferUpdateInfo(WeightTransferUpdateInfo):
tensor_sizes: list[int] | None = None
"""Per-parameter sizes in bytes within the packed buffer.
Required when packed=True, unused otherwise."""
packed: bool = False
"""Whether this update uses packed tensor format."""
def __post_init__(self):
if self.ipc_handles_pickled is not None:
@@ -119,17 +107,14 @@ class IPCWeightTransferUpdateInfo(WeightTransferUpdateInfo):
f"`shapes` should be of the same size as `names`: "
f"got {len(self.shapes)} and {len(self.names)}"
)
if (
not self.packed
and isinstance(self.ipc_handles, list)
and len(self.ipc_handles) != num_params
):
# Unpacked transfers carry a per-parameter handle dict (a list); packed
# transfers carry a single dict for the whole buffer, so the list check
# only applies to the list case.
if isinstance(self.ipc_handles, list) and len(self.ipc_handles) != num_params:
raise ValueError(
f"`ipc_handles` should be of the same size as `names`: "
f"got {len(self.ipc_handles)} and {len(self.names)}"
)
if self.packed and self.tensor_sizes is None:
raise ValueError("`tensor_sizes` is required when packed=True")
class IPCWeightTransferEngine(
@@ -154,57 +139,21 @@ class IPCWeightTransferEngine(
device: torch.device,
model: torch.nn.Module,
) -> None:
"""
Initialize the IPC weight transfer engine.
Args:
config: The configuration for the weight transfer engine
vllm_config: The full vLLM config
device: The device this worker's model lives on
model: The local model instance which will receive the weights
"""
super().__init__(config, vllm_config, device, model)
def parse_update_info(
self, update_dict: dict[str, Any]
) -> IPCWeightTransferUpdateInfo:
"""Parse update dict, deserializing pickled IPC handles if present.
HTTP transport sends IPC handles as a base64-encoded pickle under the
key ``ipc_handles_pickled``. This method deserializes them back into
``ipc_handles`` before constructing the typed dataclass, keeping
serialization concerns out of the dataclass itself.
Requires ``VLLM_ALLOW_INSECURE_SERIALIZATION=1`` because the
payload is deserialized via ``pickle.loads``.
"""
pickled = update_dict.pop("ipc_handles_pickled", None)
if pickled is not None:
if update_dict.get("ipc_handles") is not None:
raise ValueError(
"Cannot specify both `ipc_handles` and `ipc_handles_pickled`"
)
if not envs.VLLM_ALLOW_INSECURE_SERIALIZATION:
raise ValueError(
"Refusing to deserialize `ipc_handles_pickled` without "
"VLLM_ALLOW_INSECURE_SERIALIZATION=1"
)
update_dict["ipc_handles"] = pickle.loads(base64.b64decode(pickled))
return super().parse_update_info(update_dict)
# Set from the trainer-supplied init info at the handshake; defaults are
# only for the (unreachable) receive-before-init case.
self.packed = False
def init_transfer_engine(self, init_info: IPCWeightTransferInitInfo) -> None:
"""
Initialize the weight transfer mechanism.
This is called once at the beginning of training.
No initialization needed for IPC backend.
Initialize the weight transfer mechanism. No data-plane rendezvous is
needed for IPC; this just records the trainer-supplied wire params so the
worker decodes exactly as the trainer encoded.
Args:
init_info: IPC initialization info (empty)
init_info: IPC initialization info (carries `packed`).
"""
pass
self.packed = init_info.packed
def start_weight_update(self) -> None:
"""Initialize layerwise reloading for the incoming checkpoint weights."""
@@ -226,6 +175,10 @@ class IPCWeightTransferEngine(
"""
Receive weights from the trainer via CUDA IPC handles and load them.
Whether the transfer is packed is read from `self.packed`, set at the
init handshake from the trainer's init info, so it is guaranteed to
match how the trainer encoded.
Args:
update_info: IPC update info containing parameter names, dtypes, shapes,
and IPC handles. Each IPC handle is a mapping between physical
@@ -238,8 +191,9 @@ class IPCWeightTransferEngine(
# rebuilt on the device the model lives on.
device_index = self.device.index
if update_info.packed:
assert update_info.tensor_sizes is not None
if self.packed:
if update_info.tensor_sizes is None:
raise ValueError("`tensor_sizes` is required when packed=True")
assert isinstance(update_info.ipc_handles, dict)
weights = packed_ipc_consumer(
ipc_handle=update_info.ipc_handles,
@@ -285,65 +239,108 @@ class IPCWeightTransferEngine(
pass
@staticmethod
def trainer_send_weights(
iterator: Iterator[tuple[str, torch.Tensor]],
trainer_args: dict[str, Any] | IPCTrainerSendWeightsArgs,
) -> None:
"""Send weights from trainer to inference workers via CUDA IPC.
def trainer_send_weights(*args: Any, **kwargs: Any) -> None:
"""Removed. Use the stateful `IPCTrainerWeightTransferEngine` instead.
Supports two transport modes ('ray' and 'http') and two transfer
strategies:
- Non-packed (default): all weights in a single API call.
- Packed (packed=True): chunked transfer with bounded GPU memory.
For multi-GPU training, all ranks must call this method in
parallel. IPC handles are all-gathered across ranks and merged
so that each vLLM worker can find its own GPU UUID. Only rank 0
sends the payload to vLLM.
.. note::
This method calls ``update_weights`` internally. The caller must
call ``start_weight_update`` before and ``finish_weight_update``
after this method.
Args:
iterator: Iterator of (name, tensor) pairs. For multi-GPU,
each rank should yield the full tensor on its own GPU
(e.g. via FSDP full_tensor()).
trainer_args: IPCTrainerSendWeightsArgs or equivalent dict.
Transitional stub kept only to satisfy the (still abstract)
`WeightTransferEngine.trainer_send_weights`; that member is dropped from
the worker ABC once every backend has migrated to the trainer engine.
"""
args = (
IPCTrainerSendWeightsArgs(**trainer_args)
if isinstance(trainer_args, dict)
else trainer_args
raise NotImplementedError(
"The static IPC trainer path has been replaced by "
"IPCTrainerWeightTransferEngine. Build it via "
"WeightTransferTrainerFactory.trainer_init(IPCTrainerInitInfo(...), "
"client=..., source=...) and drive it with send_weights()."
)
device_index = torch.accelerator.current_device_index()
gpu_uuid = str(torch.cuda.get_device_properties(device_index).uuid)
if args.packed:
IPCWeightTransferEngine._send_packed(iterator, args, gpu_uuid)
else:
IPCWeightTransferEngine._send_unpacked(iterator, args, gpu_uuid)
@staticmethod
def _is_rank_zero() -> bool:
"""Return True if this is rank 0 or no distributed group exists."""
if not torch.distributed.is_initialized():
return True
return torch.distributed.get_rank() == 0
@staticmethod
class IPCTrainerWeightTransferEngine(TrainerWeightTransferEngine[IPCTrainerInitInfo]):
"""Trainer-side CUDA IPC weight transfer engine.
Called on every trainer rank. For multi-rank (e.g. FSDP) trainers all ranks
iterate the source (materializing each tensor) and contribute to the
IPC-handle all-gather; only the sender (rank 0) ships the merged handles to
the inference side. IPC transfer is straight-line (no concurrent broadcast
like NCCL): `update_weights` *is* the transfer, and it rides the client, so
it no-ops on non-senders.
`packed` / `packed_buffer_size_bytes` come from `IPCTrainerInitInfo`; the
sender propagates `packed` to the worker at `trainer_init`.
"""
init_info_cls = IPCTrainerInitInfo
def __init__(
self,
*,
client: VLLMWeightSyncClient,
source: WeightSource,
is_sender: bool = True,
packed: bool = False,
packed_buffer_size_bytes: int = DEFAULT_PACKED_BUFFER_SIZE_BYTES,
) -> None:
super().__init__(client=client, source=source, is_sender=is_sender)
self.packed = packed
self.packed_buffer_size_bytes = packed_buffer_size_bytes
self.device_index = torch.accelerator.current_device_index()
self.gpu_uuid = str(torch.cuda.get_device_properties(self.device_index).uuid)
@classmethod
def trainer_init(
cls,
init_info: IPCTrainerInitInfo,
*,
client: VLLMWeightSyncClient,
source: WeightSource,
) -> Self:
engine = cls(
client=client,
source=source,
is_sender=init_info.is_sender,
packed=init_info.packed,
packed_buffer_size_bytes=init_info.packed_buffer_size_bytes,
)
# IPC needs no data-plane rendezvous. The sender ships the must-agree
# `packed` flag so the worker decodes exactly as this trainer encodes.
if engine.is_sender:
engine.client.init_weight_transfer_engine({"packed": init_info.packed})
return engine
def send_weights(self) -> None:
source = self.source
if self.is_sender:
self.client.start_weight_update()
# Unpacked returns strong refs to its IPC-shared copies; they must stay
# alive across `finish_weight_update` too.
weight_refs = self._send(source)
if self.is_sender:
self.client.finish_weight_update()
self._post_send_sync()
del weight_refs
# ---- data plane (runs on all ranks; only the sender ships) ----
def _send(self, source: WeightSource) -> list[torch.Tensor] | None:
if self.packed:
self._send_packed(source)
return None
return self._send_unpacked(source)
def _all_gather_and_merge_handles(
self,
handles: list[dict[str, tuple]],
) -> list[dict[str, tuple]]:
"""All-gather and merge IPC handle dicts across ranks in one call.
Each rank contributes a list of {gpu_uuid: ipc_args} dicts (one
per parameter or one per chunk). A single all_gather_object
collects every rank's full list, then rank 0 merges per-index so
collects every rank's full list, then the sender merges per-index so
each dict maps every GPU UUID to its args.
Non-rank-0 returns a list of empty dicts.
No-op (returns handles unchanged) when no distributed group exists.
The all-gather runs over the *default* process group; this assumes the
default group is exactly the set of colocated trainer ranks and that the
sender is a member. No-op (returns handles unchanged) when no distributed
group exists.
"""
if (
not torch.distributed.is_initialized()
@@ -357,7 +354,7 @@ class IPCWeightTransferEngine(
torch.distributed.barrier()
torch.cuda.synchronize()
if torch.distributed.get_rank() == 0:
if self.is_sender:
merged: list[dict[str, tuple]] = []
for param_idx in range(len(handles)):
m: dict[str, tuple] = {}
@@ -378,25 +375,23 @@ class IPCWeightTransferEngine(
torch.distributed.barrier()
torch.cuda.ipc_collect()
@staticmethod
def _send_unpacked(
iterator: Iterator[tuple[str, torch.Tensor]],
args: IPCTrainerSendWeightsArgs,
gpu_uuid: str,
) -> None:
"""Send all weights in a single API call (non-packed mode)."""
def _send_unpacked(self, source: WeightSource) -> list[torch.Tensor]:
"""Iterate the source, build one IPC handle per param, all-gather the
handles across ranks, and (sender) ship them in one update call.
Returns the strong refs to every contiguous copy. reduce_tensor's args
do NOT keep storage alive, and non-contiguous inputs allocate fresh
storage in .contiguous(); the caller must keep these alive until the
post-send barrier (past `finish`) so the consumer's IPC views stay valid.
"""
names: list[str] = []
dtype_names: list[str] = []
shapes: list[list[int]] = []
ipc_handles: list[dict[str, tuple]] = []
# Hold strong refs to every contiguous copy until the send + post-send
# sync completes. reduce_tensor's returned args do NOT keep storage
# alive, and non-contiguous inputs allocate fresh storage in
# .contiguous() that would otherwise be GC'd before the consumer opens
# the IPC handle.
weight_refs: list[torch.Tensor] = []
for name, tensor in iterator:
for name, tensor in source:
# FSDP shards were gathered by the source; ensure contiguity here.
names.append(name)
dtype_names.append(str(tensor.dtype).split(".")[-1])
shapes.append(list(tensor.shape))
@@ -404,98 +399,64 @@ class IPCWeightTransferEngine(
weight = tensor.detach().contiguous()
weight_refs.append(weight)
_, ipc_args = reduce_tensor(weight)
ipc_handles.append({gpu_uuid: ipc_args})
ipc_handles.append({self.gpu_uuid: ipc_args})
ipc_handles = IPCWeightTransferEngine._all_gather_and_merge_handles(ipc_handles)
ipc_handles = self._all_gather_and_merge_handles(ipc_handles)
self._do_send(
names=names,
dtype_names=dtype_names,
shapes=shapes,
ipc_handles=ipc_handles,
)
return weight_refs
if IPCWeightTransferEngine._is_rank_zero():
IPCWeightTransferEngine._do_send(
args=args,
names=names,
dtype_names=dtype_names,
shapes=shapes,
ipc_handles=ipc_handles,
)
IPCWeightTransferEngine._post_send_sync()
@staticmethod
def _send_packed(
iterator: Iterator[tuple[str, torch.Tensor]],
args: IPCTrainerSendWeightsArgs,
gpu_uuid: str,
) -> None:
def _send_packed(self, source: WeightSource) -> None:
"""Send weights in bounded-memory chunks (packed mode)."""
post_iter_func: Callable = lambda item: item[1]
for chunk in packed_ipc_producer(
iterator=iterator,
gpu_uuid=gpu_uuid,
post_iter_func=post_iter_func,
buffer_size_bytes=args.packed_buffer_size_bytes,
iterator=iter(source),
gpu_uuid=self.gpu_uuid,
post_iter_func=lambda item: item[1],
buffer_size_bytes=self.packed_buffer_size_bytes,
):
ipc_handle = IPCWeightTransferEngine._all_gather_and_merge_handles(
[chunk.ipc_handle]
)[0]
ipc_handle = self._all_gather_and_merge_handles([chunk.ipc_handle])[0]
self._do_send(
names=chunk.names,
dtype_names=chunk.dtype_names,
shapes=chunk.shapes,
ipc_handles=ipc_handle,
tensor_sizes=chunk.tensor_sizes,
)
# Per-chunk barrier: the producer reuses a single IPC buffer across
# chunks, but only the sender waits for the consumers (via _do_send).
# Without syncing every rank here, non-sender ranks race ahead and
# overwrite their buffer while their colocated worker is still
# reading the current chunk, silently corrupting the transfer.
self._post_send_sync()
if IPCWeightTransferEngine._is_rank_zero():
IPCWeightTransferEngine._do_send(
args=args,
names=chunk.names,
dtype_names=chunk.dtype_names,
shapes=chunk.shapes,
ipc_handles=ipc_handle,
tensor_sizes=chunk.tensor_sizes,
packed=True,
)
IPCWeightTransferEngine._post_send_sync()
@staticmethod
def _do_send(
args: IPCTrainerSendWeightsArgs,
self,
names: list[str],
dtype_names: list[str],
shapes: list[list[int]],
ipc_handles: list[dict[str, tuple]] | dict[str, tuple],
tensor_sizes: list[int] | None = None,
packed: bool = False,
) -> None:
"""Send a single update payload via the configured transport."""
"""Build one update payload and ship it via the client. Only the sender
ships (non-sender ranks already contributed to the handle all-gather).
Emits raw `ipc_handles`; transports that cannot carry them natively
(HTTP/JSON) pickle them in their client (see `HTTPVLLMWeightSyncClient`).
"""
if not self.is_sender:
return
update_fields: dict[str, Any] = {
"names": names,
"dtype_names": dtype_names,
"shapes": shapes,
"packed": packed,
"ipc_handles": ipc_handles,
}
if tensor_sizes is not None:
update_fields["tensor_sizes"] = tensor_sizes
update_fields["ipc_handles"] = ipc_handles
update_info = IPCWeightTransferUpdateInfo(**update_fields)
if callable(args.send_mode):
args.send_mode(update_info)
elif args.send_mode == "ray":
handles = (
args.llm_handle
if isinstance(args.llm_handle, list)
else [args.llm_handle]
)
ray.get(
[
h.update_weights.remote(dict(update_info=asdict(update_info)))
for h in handles
]
)
elif args.send_mode == "http":
pickled_handles = base64.b64encode(pickle.dumps(ipc_handles)).decode(
"utf-8"
)
http_fields = {k: v for k, v in update_fields.items() if k != "ipc_handles"}
http_fields["ipc_handles_pickled"] = pickled_handles
url = f"{args.url}/update_weights"
payload = {"update_info": http_fields}
response = requests.post(url, json=payload, timeout=300)
response.raise_for_status()
self.client.update_weights(asdict(update_info))
@@ -10,8 +10,7 @@ from typing import Any
import torch
from torch.multiprocessing.reductions import reduce_tensor
# Default values for packed tensor configuration.
# These are imported by NCCLWeightTransferUpdateInfo and trainer_send_weights.
# Default values for packed tensor transfer.
DEFAULT_PACKED_BUFFER_SIZE_BYTES = 1024 * 1024 * 1024 # 1GB
DEFAULT_PACKED_NUM_BUFFERS = 2