[CPU] Enable data parallel for CPU backend (#23903)

Signed-off-by: jiang1.li <jiang1.li@intel.com>
This commit is contained in:
Li, Jiang
2025-08-29 02:19:58 -07:00
committed by GitHub
parent 2554b27baa
commit ad39106b16
6 changed files with 48 additions and 9 deletions
+8
View File
@@ -69,6 +69,7 @@ class CpuPlatform(Platform):
device_type: str = "cpu"
dispatch_key: str = "CPU"
dist_backend: str = "gloo"
device_control_env_var = "CPU_VISIBLE_MEMORY_NODES"
@property
def supported_dtypes(self) -> list[torch.dtype]:
@@ -297,6 +298,13 @@ class CpuPlatform(Platform):
allowed_numa_nodes.add(x.numa_node) # type: ignore
allowed_numa_nodes_list = sorted(allowed_numa_nodes)
env_key = CpuPlatform.device_control_env_var
if (env_key in os.environ and os.environ[env_key] != ""):
visible_nodes = [int(s) for s in os.environ[env_key].split(',')]
allowed_numa_nodes_list = [
x for x in visible_nodes if x in allowed_cpu_id_list
]
return allowed_numa_nodes_list, logical_cpu_list
@classmethod
+6 -1
View File
@@ -1,7 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from contextlib import contextmanager
from typing import TYPE_CHECKING, Any
from typing import TYPE_CHECKING, Any, Optional
import torch
import torch.nn as nn
@@ -113,6 +113,11 @@ class CPUModelRunner(GPUModelRunner):
def _to_list(self, sampled_token_ids: torch.Tensor) -> list[list[int]]:
return sampled_token_ids.tolist()
def get_dp_padding(self,
num_tokens: int) -> tuple[int, Optional[torch.Tensor]]:
# Note: For CPU backend, dp padding is not required for now.
return 0, None
@contextmanager
def _torch_cuda_wrapper():
+11 -2
View File
@@ -55,7 +55,14 @@ class CPUWorker(Worker):
else:
self.local_omp_cpuid = "all"
else:
self.local_omp_cpuid = omp_cpuids.split("|")[self.rank]
local_dp_rank = self.parallel_config.data_parallel_rank_local
omp_cpuids = omp_cpuids.split("|")
if local_dp_rank is not None:
world_size = self.parallel_config.world_size
omp_cpuids = omp_cpuids[local_dp_rank *
world_size:(local_dp_rank + 1) *
world_size]
self.local_omp_cpuid = omp_cpuids[self.rank]
if self.local_omp_cpuid != "all":
ret = torch.ops._C_utils.init_cpu_threads_env(self.local_omp_cpuid)
@@ -162,7 +169,9 @@ class CPUWorker(Worker):
# Reserve CPUs for other processes
reserve_cpu_num = envs.VLLM_CPU_NUM_OF_RESERVED_CPU
if reserve_cpu_num is None:
reserve_cpu_num = 1 if self.parallel_config.world_size > 1 else 0
need_reserve = (self.parallel_config.world_size > 1 or
self.parallel_config.data_parallel_size_local > 1)
reserve_cpu_num = 1 if need_reserve else 0
assert len(logical_cpu_list) > reserve_cpu_num, (
f"VLLM_CPU_NUM_OF_RESERVED_CPU ({reserve_cpu_num}) "
f"should less than {len(logical_cpu_list)}.")