use split_group for pytorch process group creation (#41980)
Signed-off-by: Tushar Jain <tushar00jain@users.noreply.github.com> Co-authored-by: Tushar Jain <tushar00jain@users.noreply.github.com>
This commit is contained in:
co-authored by
Tushar Jain
parent
a947f7a420
commit
38fd2405f3
@@ -10,6 +10,7 @@ import torch
|
||||
from torch.multiprocessing import spawn # pyright: ignore[reportPrivateImportUsage]
|
||||
from typing_extensions import ParamSpec
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.config import VllmConfig, set_current_vllm_config
|
||||
from vllm.distributed import (
|
||||
cleanup_dist_env_and_memory,
|
||||
@@ -60,7 +61,15 @@ def _set_vllm_config(
|
||||
tensor_model_parallel_size=vllm_config.parallel_config.tensor_parallel_size,
|
||||
pipeline_model_parallel_size=vllm_config.parallel_config.pipeline_parallel_size,
|
||||
)
|
||||
cpu_group = torch.distributed.new_group(list(range(world_size)), backend="gloo")
|
||||
if envs.VLLM_DISTRIBUTED_USE_SPLIT_GROUP:
|
||||
cpu_group = torch.distributed.split_group(
|
||||
split_ranks=[list(range(world_size))],
|
||||
group_desc="moe_test_cpu",
|
||||
)
|
||||
else:
|
||||
cpu_group = torch.distributed.new_group(
|
||||
list(range(world_size)), backend="gloo"
|
||||
)
|
||||
return cpu_group
|
||||
|
||||
|
||||
|
||||
@@ -14,6 +14,7 @@ import torch.distributed
|
||||
from torch.distributed import ProcessGroup
|
||||
from typing_extensions import ParamSpec
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.config import VllmConfig, set_current_vllm_config
|
||||
from vllm.forward_context import set_forward_context
|
||||
from vllm.model_executor.layers.fused_moe.activation import MoEActivation
|
||||
@@ -375,7 +376,13 @@ def _test_deepep_deepgemm_moe(
|
||||
w1_scale = w1_scale.to(device=device)
|
||||
w2_scale = w2_scale.to(device=device)
|
||||
|
||||
pg = torch.distributed.new_group(list(range(pgi.world_size)))
|
||||
if envs.VLLM_DISTRIBUTED_USE_SPLIT_GROUP:
|
||||
pg = torch.distributed.split_group(
|
||||
split_ranks=[list(range(pgi.world_size))],
|
||||
group_desc="deepep_deepgemm_test",
|
||||
)
|
||||
else:
|
||||
pg = torch.distributed.new_group(list(range(pgi.world_size)))
|
||||
test_tensors = TestTensors.make(config, pgi.rank)
|
||||
block_shape = [w1.size(1) // w1_scale.size(1), w1.size(2) // w1_scale.size(2)]
|
||||
|
||||
|
||||
@@ -10,6 +10,7 @@ import pytest
|
||||
import torch.distributed
|
||||
from torch.distributed import ProcessGroup
|
||||
|
||||
import vllm.envs as envs
|
||||
from tests.kernels.moe.utils import make_dummy_moe_config
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.config import VllmConfig, set_current_vllm_config
|
||||
@@ -375,7 +376,13 @@ def _deep_ep_moe(
|
||||
w1_scale = w1_scale.to(device=device_idx)
|
||||
w2_scale = w2_scale.to(device=device_idx)
|
||||
|
||||
pg = torch.distributed.new_group(list(range(pgi.world_size)))
|
||||
if envs.VLLM_DISTRIBUTED_USE_SPLIT_GROUP:
|
||||
pg = torch.distributed.split_group(
|
||||
split_ranks=[list(range(pgi.world_size))],
|
||||
group_desc="deepep_test",
|
||||
)
|
||||
else:
|
||||
pg = torch.distributed.new_group(list(range(pgi.world_size)))
|
||||
test_tensors = TestTensors.make(config, low_latency_mode)
|
||||
|
||||
with set_current_vllm_config(VllmConfig()):
|
||||
|
||||
Reference in New Issue
Block a user