[EPLB] Optimize memory overhead in Nixl communicator (#40013)

Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: Markov Ilya <markovilya19@gmail.com>
Co-authored-by: Markov Ilya <markovilya19@gmail.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
This commit is contained in:
Ilya Markov
2026-04-30 07:46:49 -07:00
committed by GitHub
co-authored by Markov Ilya Nicolò Lucchesi
parent ff449b6426
commit a7fb008510
2 changed files with 212 additions and 204 deletions
+204 -199
View File
@@ -11,6 +11,7 @@ from abc import ABC, abstractmethod
from collections.abc import Sequence
from datetime import timedelta
import numpy as np
import torch
from torch.distributed import (
P2POp,
@@ -47,15 +48,25 @@ class EplbCommunicator(ABC):
"""Abstract EPLB communicator for expert weight transfers."""
@abstractmethod
def add_send(self, tensor: torch.Tensor, dst_rank: int) -> None:
def add_send(
self,
tensors: list[torch.Tensor],
dst_rank: int,
expert_id: int,
) -> None:
pass
@abstractmethod
def add_recv(self, tensor: torch.Tensor, src_rank: int) -> None:
def add_recv(
self,
tensors: list[torch.Tensor],
src_rank: int,
expert_id: int,
) -> None:
pass
@abstractmethod
def execute(self) -> None:
def execute(self, old_indices: np.ndarray | None = None) -> None:
pass
@property
@@ -85,27 +96,39 @@ class TorchDistNcclEplbCommunicator(EplbCommunicator):
self._p2p_ops: list[P2POp] = []
self._log_initialized()
def add_send(self, tensor: torch.Tensor, dst_rank: int) -> None:
self._p2p_ops.append(
P2POp(
torch.distributed.isend,
tensor,
dst_rank,
self._ep_group,
def add_send(
self,
tensors: list[torch.Tensor],
dst_rank: int,
expert_id: int, # unused by this backend
) -> None:
for tensor in tensors:
self._p2p_ops.append(
P2POp(
torch.distributed.isend,
tensor,
dst_rank,
self._ep_group,
)
)
)
def add_recv(self, tensor: torch.Tensor, src_rank: int) -> None:
self._p2p_ops.append(
P2POp(
torch.distributed.irecv,
tensor,
src_rank,
self._ep_group,
def add_recv(
self,
tensors: list[torch.Tensor],
src_rank: int,
expert_id: int, # unused by this backend
) -> None:
for tensor in tensors:
self._p2p_ops.append(
P2POp(
torch.distributed.irecv,
tensor,
src_rank,
self._ep_group,
)
)
)
def execute(self) -> None:
def execute(self, old_indices: np.ndarray | None = None) -> None:
if not self._p2p_ops:
return
try:
@@ -130,13 +153,25 @@ class TorchDistGlooStagedEplbCommunicator(EplbCommunicator):
self._ops: list[tuple[str, torch.Tensor, int]] = []
self._log_initialized()
def add_send(self, tensor: torch.Tensor, dst_rank: int) -> None:
self._ops.append(("send", tensor, dst_rank))
def add_send(
self,
tensors: list[torch.Tensor],
dst_rank: int,
expert_id: int, # unused by this backend
) -> None:
for tensor in tensors:
self._ops.append(("send", tensor, dst_rank))
def add_recv(self, tensor: torch.Tensor, src_rank: int) -> None:
self._ops.append(("recv", tensor, src_rank))
def add_recv(
self,
tensors: list[torch.Tensor],
src_rank: int,
expert_id: int, # unused by this backend
) -> None:
for tensor in tensors:
self._ops.append(("recv", tensor, src_rank))
def execute(self) -> None:
def execute(self, old_indices: np.ndarray | None = None) -> None:
if not self._ops:
return
@@ -207,17 +242,17 @@ class NixlEplbCommunicator(EplbCommunicator):
self._cuda_stream = cuda_stream
self._world_size = cpu_group.size()
self._rank = cpu_group.rank()
self._send_tensors: dict[torch.dtype, list[list[torch.Tensor]]] = {}
self._recv_tensors: dict[torch.dtype, list[list[torch.Tensor]]] = {}
self._dtypes: list[torch.dtype] = []
# expert_id -> weight tensors to pack into the send buffer.
self._expert_send_map: dict[int, list[torch.Tensor]] = {}
# src_rank -> expert_id -> weight tensors to unpack after transfer.
self._recv_map: dict[int, dict[int, list[torch.Tensor]]] = {}
self._num_local_experts: int = expert_weights[0].shape[0]
self._device = expert_weights[0].device
for tensor in expert_weights:
assert tensor.device == self._device, (
"All local EPLB tensors are expected to be on the same device: "
f"expected={self._device}, got={tensor.device}"
)
if tensor.dtype not in self._dtypes:
self._dtypes.append(tensor.dtype)
config = (
nixl_agent_config(capture_telemetry=False)
@@ -228,13 +263,12 @@ class NixlEplbCommunicator(EplbCommunicator):
self._nixl_memory_type = "VRAM"
self._registered_desc: object | None = None
self._remote_agents: dict[int, str] = {}
self._remote_send_meta: dict[int, tuple[int, int, int]] = {}
self._remote_send_meta: dict[int, tuple[int, int]] = {}
self._send_buffer: torch.Tensor = torch.empty(0)
self._recv_buffer: torch.Tensor = torch.empty(0)
self._peer_partition_bytes: int = 0
self._dtype_max_bytes: dict[torch.dtype, int] = {}
self._expert_bytes: int = 0
self._cuda_device_id = int(self._device.index or 0)
self._xfer_cache: dict[tuple[int, int, int], tuple[int, int, int]] = {}
self._init_step("buffers", self._init_registered_buffers, expert_weights)
self._init_step("agents", self._init_remote_agents)
self._init_step("send meta", self._exchange_remote_send_meta)
@@ -258,34 +292,33 @@ class NixlEplbCommunicator(EplbCommunicator):
uid = uuid.uuid4().hex[:8]
return f"eplb-{self._rank}{pp_suffix}-{uid}"
def _get_peer_buckets(
def add_send(
self,
bucket_map: dict[torch.dtype, list[list[torch.Tensor]]],
dtype: torch.dtype,
) -> list[list[torch.Tensor]]:
peer_buckets = bucket_map.get(dtype)
if peer_buckets is None:
peer_buckets = [[] for _ in range(self._world_size)]
bucket_map[dtype] = peer_buckets
return peer_buckets
def add_send(self, tensor: torch.Tensor, dst_rank: int) -> None:
tensors: list[torch.Tensor],
dst_rank: int,
expert_id: int,
) -> None:
assert dst_rank != self._rank, (
"EPLB communicator should not enqueue same-rank sends: "
f"rank={self._rank}, dst_rank={dst_rank}"
)
self._get_peer_buckets(self._send_tensors, tensor.dtype)[dst_rank].append(
tensor
)
# An expert sent to multiple peers is packed only once; skip duplicates.
if expert_id not in self._expert_send_map:
self._expert_send_map[expert_id] = tensors
def add_recv(self, tensor: torch.Tensor, src_rank: int) -> None:
def add_recv(
self,
tensors: list[torch.Tensor],
src_rank: int,
expert_id: int,
) -> None:
assert src_rank != self._rank, (
"EPLB communicator should not enqueue same-rank recvs: "
f"rank={self._rank}, src_rank={src_rank}"
)
self._get_peer_buckets(self._recv_tensors, tensor.dtype)[src_rank].append(
tensor
)
recv_experts = self._recv_map.setdefault(src_rank, {})
if expert_id not in recv_experts:
recv_experts[expert_id] = tensors
def _init_remote_agents(self) -> None:
local_metadata = self._nixl_wrapper.get_agent_metadata()
@@ -303,30 +336,18 @@ class NixlEplbCommunicator(EplbCommunicator):
)
def _init_registered_buffers(self, expert_weights: Sequence[torch.Tensor]) -> None:
total_max_bytes = 0
for dtype in self._dtypes:
max_numel = max(
sum(t.numel() for t in expert_weights if t.dtype == dtype), 1
)
max_bytes = max_numel * dtype.itemsize
self._dtype_max_bytes[dtype] = max_bytes
total_max_bytes += max_bytes
self._peer_partition_bytes = total_max_bytes
# The send buffer needs world_size partitions because remote peers
# READ from fixed offsets (rank * partition_bytes).
# This allocates world_size * partition_bytes
# which can cause OOM on large models.
# TODO(ilmarkov): shrink to const * partition_bytes and execute
# communication in multiple steps dealing with the worst case.
send_total_bytes = self._peer_partition_bytes * self._world_size
total_bytes = max(sum(t.nbytes for t in expert_weights), 1)
assert total_bytes % self._num_local_experts == 0, (
f"Number of bytes in moe layer {total_bytes} is not divisible "
f"by number of local experts {self._num_local_experts}"
)
self._expert_bytes = total_bytes // self._num_local_experts
self._send_buffer = torch.empty(
send_total_bytes, device=self._device, dtype=torch.uint8
total_bytes, device=self._device, dtype=torch.uint8
)
self._recv_buffer = torch.empty(
self._peer_partition_bytes, device=self._device, dtype=torch.uint8
total_bytes, device=self._device, dtype=torch.uint8
)
descs = self._nixl_wrapper.get_reg_descs([self._send_buffer, self._recv_buffer])
@@ -336,12 +357,11 @@ class NixlEplbCommunicator(EplbCommunicator):
def _exchange_remote_send_meta(self) -> None:
"""Exchange send-buffer metadata so each rank can build dynamic
descriptors at execute time."""
local_meta: tuple[int, int, int] = (
local_meta: tuple[int, int] = (
self._send_buffer.data_ptr(),
self._peer_partition_bytes,
self._cuda_device_id,
)
gathered_meta: list[tuple[int, int, int] | None] = [None] * self._world_size
gathered_meta: list[tuple[int, int] | None] = [None] * self._world_size
torch.distributed.all_gather_object(
gathered_meta, local_meta, group=self._cpu_group
)
@@ -353,14 +373,11 @@ class NixlEplbCommunicator(EplbCommunicator):
@staticmethod
def _pack_send_buffer(
peer_tensors: list[torch.Tensor],
in_tensors: list[torch.Tensor],
send_buffer: torch.Tensor,
byte_offset: int,
) -> int:
"""
Returns the byte offset after the last written byte.
"""
for tensor in peer_tensors:
) -> None:
for tensor in in_tensors:
raw = tensor.reshape(-1).view(torch.uint8)
if raw.numel() == 0:
continue
@@ -368,18 +385,14 @@ class NixlEplbCommunicator(EplbCommunicator):
raw, non_blocking=True
)
byte_offset += raw.numel()
return byte_offset
@staticmethod
def _unpack_recv_buffer(
recv_buffer: torch.Tensor,
peer_tensors: list[torch.Tensor],
out_tensors: list[torch.Tensor],
byte_offset: int,
) -> int:
"""
Returns the byte offset after the last read byte.
"""
for tensor in peer_tensors:
) -> None:
for tensor in out_tensors:
num_bytes = tensor.numel() * tensor.element_size()
if num_bytes == 0:
continue
@@ -388,19 +401,6 @@ class NixlEplbCommunicator(EplbCommunicator):
non_blocking=True,
)
byte_offset += num_bytes
return byte_offset
def _release_all_cached_handles(self) -> None:
"""Best-effort release of every cached dlist and xfer handle."""
for local_dlist, remote_dlist, xfer in self._xfer_cache.values():
for release_fn, handle in (
(self._nixl_wrapper.release_xfer_handle, xfer),
(self._nixl_wrapper.release_dlist_handle, local_dlist),
(self._nixl_wrapper.release_dlist_handle, remote_dlist),
):
with contextlib.suppress(Exception):
release_fn(handle)
self._xfer_cache.clear()
def _wait_for_all_transfers(self, handles: list[int]) -> None:
pending = set(handles)
@@ -418,82 +418,68 @@ class NixlEplbCommunicator(EplbCommunicator):
if pending:
time.sleep(0.0005)
def _get_or_create_xfer(self, src: int, total_bytes: int, recv_offset: int) -> int:
"""Return a cached xfer handle or create and cache a new one."""
key = (src, total_bytes, recv_offset)
cached = self._xfer_cache.get(key)
if cached is not None:
return cached[2]
def _create_peer_xfer(
self,
src: int,
local_descs: list[tuple[int, int, int]],
remote_descs: list[tuple[int, int, int]],
) -> tuple[int, int, int]:
"""Create a batched xfer for multiple descriptors from one peer.
recv_base = self._recv_buffer.data_ptr()
Each element in *local_descs* / *remote_descs* is an
``(address, size, device_id)`` tuple.
Returns ``(local_dlist, remote_dlist, xfer_handle)``.
"""
local_desc = self._nixl_wrapper.get_xfer_descs(
[
(
recv_base + recv_offset,
total_bytes,
self._cuda_device_id,
)
],
self._nixl_memory_type,
local_descs, self._nixl_memory_type
)
local_handle = self._nixl_wrapper.prep_xfer_dlist(
"NIXL_INIT_AGENT",
local_desc,
)
remote_base, remote_part_bytes, remote_dev = self._remote_send_meta[src]
agent_name = self._remote_agents[src]
remote_desc = self._nixl_wrapper.get_xfer_descs(
[
(
remote_base + self._rank * remote_part_bytes,
total_bytes,
remote_dev,
)
],
self._nixl_memory_type,
remote_descs, self._nixl_memory_type
)
remote_handle = self._nixl_wrapper.prep_xfer_dlist(
agent_name,
self._remote_agents[src],
remote_desc,
)
indices = list(range(len(local_descs)))
xfer_handle = self._nixl_wrapper.make_prepped_xfer(
"READ",
local_handle,
[0],
indices,
remote_handle,
[0],
indices,
)
self._xfer_cache[key] = (local_handle, remote_handle, xfer_handle)
return xfer_handle
return (local_handle, remote_handle, xfer_handle)
def execute(self) -> None:
xfer_handles: list[int] = []
def execute(self, old_indices: np.ndarray | None = None) -> None:
assert old_indices is not None, (
"NixlEplbCommunicator.execute requires old_indices"
)
xfer_entries: list[tuple[int, int, int]] = []
try:
# Phase 1: pack send buffers.
n = self._num_local_experts
rank_experts = old_indices[: self._world_size * n].reshape(
self._world_size, n
)
# Build expert_id -> send slot mapping per rank.
expert_to_send_slot: list[dict[int, int]] = [
{int(eid): i for i, eid in enumerate(row) if eid != -1}
for row in rank_experts
]
# Phase 1: pack each expert at its slot offset in the send buffer.
with torch.cuda.stream(self._cuda_stream):
for dst in range(self._world_size):
byte_offset = dst * self._peer_partition_bytes
for dtype in self._dtypes:
peer_tensors = self._send_tensors.get(
dtype, [[] for _ in range(self._world_size)]
)[dst]
actual_bytes = sum(
t.numel() * t.element_size() for t in peer_tensors
)
if actual_bytes > self._dtype_max_bytes[dtype]:
raise RuntimeError(
"NIXL EPLB send overflow for dtype "
f"{dtype}: peer={dst}, "
f"required={actual_bytes}, "
f"capacity={self._dtype_max_bytes[dtype]}"
)
byte_offset = self._pack_send_buffer(
peer_tensors,
self._send_buffer,
byte_offset,
)
for expert_id, tensors in self._expert_send_map.items():
slot = expert_to_send_slot[self._rank][expert_id]
byte_offset = slot * self._expert_bytes
self._pack_send_buffer(tensors, self._send_buffer, byte_offset)
# Ensure all packed data is visible in device memory before pulls.
if self._cuda_stream is not None:
@@ -508,58 +494,65 @@ class NixlEplbCommunicator(EplbCommunicator):
timeout=timedelta(minutes=5),
)
# Phase 2: look up or create descriptors and issue all READs.
# Data from all peers is packed sequentially into the single
# partition-sized recv buffer at running offsets.
recv_offsets: dict[int, int] = {}
# Phase 2: issue one batched READ per peer.
recv_offsets: dict[tuple[int, int], int] = {}
recv_offset = 0
recv_base = self._recv_buffer.data_ptr()
for src in range(self._world_size):
if src == self._rank:
continue
actual_total_bytes = 0
for dtype in self._dtypes:
peer_tensors = self._recv_tensors.get(
dtype, [[] for _ in range(self._world_size)]
)[src]
actual_total_bytes += sum(
t.numel() * t.element_size() for t in peer_tensors
)
if actual_total_bytes == 0:
recv_experts = self._recv_map.get(src)
if not recv_experts:
continue
recv_offsets[src] = recv_offset
xfer_handle = self._get_or_create_xfer(
src, actual_total_bytes, recv_offset
expert_ids = list(recv_experts.keys())
remote_base, remote_dev = self._remote_send_meta[src]
local_descs: list[tuple[int, int, int]] = []
remote_descs: list[tuple[int, int, int]] = []
for expert_id in expert_ids:
slot = expert_to_send_slot[src][expert_id]
remote_off = slot * self._expert_bytes
recv_offsets[(src, expert_id)] = recv_offset
local_descs.append(
(
recv_base + recv_offset,
self._expert_bytes,
self._cuda_device_id,
)
)
remote_descs.append(
(remote_base + remote_off, self._expert_bytes, remote_dev)
)
recv_offset += self._expert_bytes
assert recv_offset <= self._recv_buffer.nbytes
local_h, remote_h, xfer_h = self._create_peer_xfer(
src, local_descs, remote_descs
)
self._nixl_wrapper.transfer(xfer_handle)
xfer_handles.append(xfer_handle)
recv_offset += actual_total_bytes
self._nixl_wrapper.transfer(xfer_h)
xfer_entries.append((local_h, remote_h, xfer_h))
# Phase 3: single wait for all in-flight transfers, then unpack.
self._wait_for_all_transfers(xfer_handles)
# Phase 3: wait for all in-flight transfers, then unpack.
self._wait_for_all_transfers([x[2] for x in xfer_entries])
with torch.cuda.stream(self._cuda_stream):
for src, offset in recv_offsets.items():
byte_offset = offset
for dtype in self._dtypes:
peer_tensors = self._recv_tensors.get(
dtype, [[] for _ in range(self._world_size)]
)[src]
byte_offset = self._unpack_recv_buffer(
self._recv_buffer,
peer_tensors,
byte_offset,
)
except Exception:
self._release_all_cached_handles()
raise
for (src, expert_id), offset in recv_offsets.items():
self._unpack_recv_buffer(
self._recv_buffer,
self._recv_map[src][expert_id],
offset,
)
finally:
self._send_tensors.clear()
self._recv_tensors.clear()
for local_h, remote_h, xfer_h in xfer_entries:
with contextlib.suppress(Exception):
self._nixl_wrapper.release_xfer_handle(xfer_h)
with contextlib.suppress(Exception):
self._nixl_wrapper.release_dlist_handle(local_h)
with contextlib.suppress(Exception):
self._nixl_wrapper.release_dlist_handle(remote_h)
self._expert_send_map.clear()
self._recv_map.clear()
def __del__(self) -> None:
try:
self._release_all_cached_handles()
if self._registered_desc is not None:
self._nixl_wrapper.deregister_memory(self._registered_desc)
self._registered_desc = None
@@ -588,15 +581,27 @@ class PyNcclEplbCommunicator(EplbCommunicator):
self._pynccl_comm.group_start()
self._group_started = True
def add_send(self, tensor: torch.Tensor, dst_rank: int) -> None:
def add_send(
self,
tensors: list[torch.Tensor],
dst_rank: int,
expert_id: int, # unused by this backend
) -> None:
self._ensure_group_started()
self._pynccl_comm.send(tensor, dst_rank, stream=self._cuda_stream)
for tensor in tensors:
self._pynccl_comm.send(tensor, dst_rank, stream=self._cuda_stream)
def add_recv(self, tensor: torch.Tensor, src_rank: int) -> None:
def add_recv(
self,
tensors: list[torch.Tensor],
src_rank: int,
expert_id: int, # unused by this backend
) -> None:
self._ensure_group_started()
self._pynccl_comm.recv(tensor, src_rank, stream=self._cuda_stream)
for tensor in tensors:
self._pynccl_comm.recv(tensor, src_rank, stream=self._cuda_stream)
def execute(self) -> None:
def execute(self, old_indices: np.ndarray | None = None) -> None:
if self._group_started:
self._pynccl_comm.group_end()
self._group_started = False
+8 -5
View File
@@ -294,9 +294,9 @@ def move_to_buffer(
recver_pos = remainder_start + sender_pos
if recver_pos < len(ranks_to_recv):
recv_ranks.append(ranks_to_recv[recver_pos])
expert_tensors = [w[src] for w in expert_weights]
for dst in recv_ranks:
for w in expert_weights:
communicator.add_send(w[src], dst)
communicator.add_send(expert_tensors, dst, expert_id=int(expert))
# 3. Post recvs
if recv_count > 0:
@@ -325,11 +325,14 @@ def move_to_buffer(
src = ranks_to_send[recver_pos // num_dst_per_sender]
else:
src = ranks_to_send[recver_pos - remainder_start]
for b in expert_weights_buffers:
communicator.add_recv(b[dst], src)
communicator.add_recv(
[b[dst] for b in expert_weights_buffers],
src,
expert_id=int(expert),
)
# 4. Execute the P2P operations. The real communication happens here.
communicator.execute()
communicator.execute(old_indices=old_indices)
# wait for the communication to finish
return TransferMetadata(
is_unchanged=is_unchanged,