biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[core][executor] simplify instance id (#10976)

Browse Source 
Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
youkaichao
2024-12-07 09:33:45 -08:00
committed by GitHub
parent 78029b34ed
commit 1b62745b1d
10 changed files with 22 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/executor/cpu_executor.py
View File

@@ -10,8 +10,7 @@ from vllm.lora.request import LoRARequest
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.prompt_adapter.request import PromptAdapterRequest
from vllm.sequence import ExecuteModelRequest
from vllm.utils import (get_distributed_init_method, get_open_port,
get_vllm_instance_id, make_async)
from vllm.utils import get_distributed_init_method, get_open_port, make_async
from vllm.worker.worker_base import WorkerWrapperBase
logger = init_logger(__name__)
@@ -31,9 +30,6 @@ class CPUExecutor(ExecutorBase):
# Environment variables for CPU executor
#
# Ensure that VLLM_INSTANCE_ID is set, to be inherited by workers
os.environ["VLLM_INSTANCE_ID"] = get_vllm_instance_id()
# Disable torch async compiling which won't work with daemonic processes
os.environ["TORCHINDUCTOR_COMPILE_THREADS"] = "1"

5
vllm/executor/multiproc_gpu_executor.py
View File

@@ -16,7 +16,7 @@ from vllm.sequence import ExecuteModelRequest
from vllm.triton_utils.importing import HAS_TRITON
from vllm.utils import (_run_task_with_lock, cuda_device_count_stateless,
cuda_is_initialized, get_distributed_init_method,
get_open_port, get_vllm_instance_id, make_async,
get_open_port, make_async,
update_environment_variables)
if HAS_TRITON:
@@ -37,9 +37,6 @@ class MultiprocessingGPUExecutor(DistributedGPUExecutor):
world_size = self.parallel_config.world_size
tensor_parallel_size = self.parallel_config.tensor_parallel_size
# Ensure that VLLM_INSTANCE_ID is set, to be inherited by workers
os.environ["VLLM_INSTANCE_ID"] = get_vllm_instance_id()
# Disable torch async compiling which won't work with daemonic processes
os.environ["TORCHINDUCTOR_COMPILE_THREADS"] = "1"

7
vllm/executor/ray_gpu_executor.py
View File

@@ -15,8 +15,7 @@ from vllm.logger import init_logger
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.sequence import ExecuteModelRequest
from vllm.utils import (_run_task_with_lock, get_distributed_init_method,
get_ip, get_open_port, get_vllm_instance_id,
make_async)
get_ip, get_open_port, make_async)
if ray is not None:
from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
@@ -220,14 +219,10 @@ class RayGPUExecutor(DistributedGPUExecutor):
" environment variable, make sure it is unique for"
" each node.")
VLLM_INSTANCE_ID = get_vllm_instance_id()
# Set environment variables for the driver and workers.
all_args_to_update_environment_variables = [({
"CUDA_VISIBLE_DEVICES":
",".join(map(str, node_gpus[node_id])),
"VLLM_INSTANCE_ID":
VLLM_INSTANCE_ID,
"VLLM_TRACE_FUNCTION":
str(envs.VLLM_TRACE_FUNCTION),
**({

7
vllm/executor/ray_hpu_executor.py
View File

@@ -15,8 +15,7 @@ from vllm.logger import init_logger
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.sequence import ExecuteModelRequest
from vllm.utils import (_run_task_with_lock, get_distributed_init_method,
get_ip, get_open_port, get_vllm_instance_id,
make_async)
get_ip, get_open_port, make_async)
if ray is not None:
from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
@@ -196,12 +195,8 @@ class RayHPUExecutor(DistributedGPUExecutor):
"environment variable, make sure it is unique for"
" each node.")
VLLM_INSTANCE_ID = get_vllm_instance_id()
# Set environment variables for the driver and workers.
all_args_to_update_environment_variables = [({
"VLLM_INSTANCE_ID":
VLLM_INSTANCE_ID,
"VLLM_TRACE_FUNCTION":
str(envs.VLLM_TRACE_FUNCTION),
}, ) for (node_id, _) in worker_node_and_gpu_ids]

6
vllm/executor/ray_tpu_executor.py
View File

@@ -13,7 +13,7 @@ from vllm.logger import init_logger
from vllm.model_executor.layers.sampler import SamplerOutput
from vllm.sequence import ExecuteModelRequest
from vllm.utils import (get_distributed_init_method, get_ip, get_open_port,
get_vllm_instance_id, make_async)
make_async)
if ray is not None:
from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
@@ -144,12 +144,8 @@ class RayTPUExecutor(TPUExecutor):
for i, (node_id, _) in enumerate(worker_node_and_gpu_ids):
node_workers[node_id].append(i)
VLLM_INSTANCE_ID = get_vllm_instance_id()
# Set environment variables for the driver and workers.
all_args_to_update_environment_variables = [({
"VLLM_INSTANCE_ID":
VLLM_INSTANCE_ID,
"VLLM_TRACE_FUNCTION":
str(envs.VLLM_TRACE_FUNCTION),
}, ) for _ in worker_node_and_gpu_ids]

6
vllm/executor/ray_xpu_executor.py
View File

@@ -5,7 +5,7 @@ import vllm.envs as envs
from vllm.executor.ray_gpu_executor import RayGPUExecutor, RayGPUExecutorAsync
from vllm.executor.xpu_executor import XPUExecutor
from vllm.logger import init_logger
from vllm.utils import get_vllm_instance_id, make_async
from vllm.utils import make_async
logger = init_logger(__name__)
@@ -17,12 +17,8 @@ class RayXPUExecutor(RayGPUExecutor, XPUExecutor):
worker_node_and_gpu_ids = self._run_workers("get_node_and_gpu_ids",
use_dummy_driver=True)
VLLM_INSTANCE_ID = get_vllm_instance_id()
# Set environment variables for the driver and workers.
all_args_to_update_environment_variables = [({
"VLLM_INSTANCE_ID":
VLLM_INSTANCE_ID,
"VLLM_TRACE_FUNCTION":
str(envs.VLLM_TRACE_FUNCTION),
}, ) for (_, _) in worker_node_and_gpu_ids]