[Core] remove cupy dependency (#3625)

This commit is contained in:
youkaichao
2024-03-27 00:33:26 -07:00
committed by GitHub
parent e66b629c04
commit 8f44facddd
17 changed files with 506 additions and 223 deletions
+14 -16
View File
@@ -10,7 +10,7 @@ from vllm.config import (CacheConfig, DeviceConfig, LoRAConfig, ModelConfig,
ParallelConfig, SchedulerConfig, VisionLanguageConfig)
from vllm.lora.request import LoRARequest
from vllm.model_executor import set_random_seed
from vllm.model_executor.parallel_utils import cupy_utils
from vllm.model_executor.parallel_utils import pynccl_utils
from vllm.model_executor.parallel_utils.communication_op import (
broadcast_tensor_dict)
from vllm.model_executor.parallel_utils.custom_all_reduce import init_custom_ar
@@ -75,7 +75,7 @@ class Worker:
self.cache_engine = None
self.gpu_cache = None
def init_device(self, cupy_port: Optional[int] = None) -> None:
def init_device(self) -> None:
if self.device_config.device.type == "cuda":
# torch.distributed.all_reduce does not free the input tensor until
# the synchronization point. This causes the memory usage to grow
@@ -98,7 +98,7 @@ class Worker:
f"Not support device type: {self.device_config.device}")
# Initialize the distributed environment.
init_distributed_environment(self.parallel_config, self.rank,
cupy_port, self.distributed_init_method)
self.distributed_init_method)
# Set random seed.
set_random_seed(self.model_config.seed)
@@ -250,7 +250,6 @@ class Worker:
def init_distributed_environment(
parallel_config: ParallelConfig,
rank: int,
cupy_port: Optional[int],
distributed_init_method: Optional[str] = None,
) -> None:
"""Initialize the distributed environment."""
@@ -273,28 +272,27 @@ def init_distributed_environment(
init_method=distributed_init_method,
)
if cupy_utils.is_initialized():
cupy_world_size = cupy_utils.get_world_size()
if cupy_world_size != parallel_config.world_size:
if pynccl_utils.is_initialized():
pynccl_world_size = pynccl_utils.get_world_size()
if pynccl_world_size != parallel_config.world_size:
raise RuntimeError(
"cupy.distributed is already initialized but the cupy world "
"pynccl is already initialized but the pynccl world "
"size does not match parallel_config.world_size "
f"({cupy_world_size} vs. {parallel_config.world_size}).")
elif (parallel_config.world_size > 1 and cupy_port is not None):
# NOTE(woosuk): We don't initialize CuPy process group when world size
f"({pynccl_world_size} vs. {parallel_config.world_size}).")
elif parallel_config.world_size > 1:
# NOTE(woosuk): We don't initialize pynccl process group when world size
# is 1.
# TODO(woosuk): Support multi-node connection.
cupy_utils.init_process_group(
pynccl_utils.init_process_group(
world_size=parallel_config.world_size,
rank=rank,
host="localhost",
port=cupy_port,
init_method=distributed_init_method,
)
# A small all_reduce for warmup.
torch.distributed.all_reduce(torch.zeros(1).cuda())
if cupy_utils.is_initialized():
cupy_utils.all_reduce(torch.zeros(1).cuda())
if pynccl_utils.is_initialized():
pynccl_utils.all_reduce(torch.zeros(1).cuda())
ensure_model_parallel_initialized(parallel_config.tensor_parallel_size,
parallel_config.pipeline_parallel_size)