[Core][Refactor] move parallel_utils into vllm/distributed (#3950)

[WIP][Core][Refactor] move vllm/model_executor/parallel_utils into vllm/distributed and vllm/device_communicators (#3950)
This commit is contained in:
youkaichao
2024-04-10 15:33:30 -07:00
committed by GitHub
parent 934d3662f7
commit 63e7176f26
52 changed files with 111 additions and 141 deletions
+1 -2
View File
@@ -11,8 +11,7 @@ from transformers import (AutoModelForCausalLM, AutoProcessor,
from vllm import LLM, SamplingParams
from vllm.config import TokenizerPoolConfig, VisionLanguageConfig
from vllm.model_executor.parallel_utils.parallel_state import (
destroy_model_parallel)
from vllm.distributed import destroy_model_parallel
from vllm.sequence import MultiModalData
from vllm.transformers_utils.tokenizer import get_tokenizer
+3 -3
View File
@@ -8,9 +8,9 @@ import pytest
import ray
import torch
from vllm.model_executor.parallel_utils.communication_op import (
broadcast_tensor_dict, tensor_model_parallel_all_gather,
tensor_model_parallel_all_reduce)
from vllm.distributed import (broadcast_tensor_dict,
tensor_model_parallel_all_gather,
tensor_model_parallel_all_reduce)
from vllm.test_utils import (init_test_distributed_environment,
multi_process_tensor_parallel)
+6 -7
View File
@@ -6,9 +6,8 @@ import ray
import torch
import torch.distributed as dist
from vllm.model_executor.parallel_utils import custom_all_reduce as custom_ar
from vllm.model_executor.parallel_utils.communication_op import (
tensor_model_parallel_all_reduce)
from vllm.distributed import tensor_model_parallel_all_reduce
from vllm.distributed.device_communicators import custom_all_reduce
from vllm.test_utils import (init_test_distributed_environment,
multi_process_tensor_parallel)
@@ -26,10 +25,10 @@ def graph_allreduce(world_size, rank, distributed_init_port):
init_test_distributed_environment(1, world_size, rank,
distributed_init_port)
custom_ar.init_custom_ar()
custom_all_reduce.init_custom_all_reduce()
for sz in test_sizes:
for dtype in [torch.float32, torch.float16, torch.bfloat16]:
with custom_ar.capture():
with custom_all_reduce.capture():
# use integers so result matches NCCL exactly
inp1 = torch.randint(1,
16, (sz, ),
@@ -62,8 +61,8 @@ def eager_allreduce(world_size, rank, distributed_init_port):
distributed_init_port)
sz = 1024
custom_ar.init_custom_ar()
fa = custom_ar.get_handle()
custom_all_reduce.init_custom_all_reduce()
fa = custom_all_reduce.get_handle()
inp = torch.ones(sz, dtype=torch.float32, device=device)
out = fa.all_reduce_unreg(inp)
assert torch.allclose(out, inp * world_size)
+2 -2
View File
@@ -4,8 +4,8 @@ import os
import pytest
import torch
from vllm.model_executor.parallel_utils.pynccl import (NCCLCommunicator,
ncclGetUniqueId)
from vllm.distributed.device_communicators.pynccl import (NCCLCommunicator,
ncclGetUniqueId)
def distributed_run(fn, world_size):
+1 -2
View File
@@ -12,6 +12,7 @@ from huggingface_hub import snapshot_download
import vllm
from vllm.config import LoRAConfig
from vllm.distributed import destroy_model_parallel, initialize_model_parallel
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
MergedColumnParallelLinear,
RowParallelLinear)
@@ -19,8 +20,6 @@ from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.sampler import Sampler
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.model_executor.model_loader import get_model
from vllm.model_executor.parallel_utils.parallel_state import (
destroy_model_parallel, initialize_model_parallel)
def cleanup():