[ROCm][AMD] unify CUDA_VISIBLE_DEVICES usage in cuda/rocm (#6352)

This commit is contained in:
Hongxia Yang
2024-07-11 21:30:46 -07:00
committed by GitHub
parent d26a8b3f1f
commit b6c16cf8ff
5 changed files with 10 additions and 34 deletions
+1 -8
View File
@@ -6,7 +6,6 @@ from typing import TYPE_CHECKING, ClassVar, List, Optional, Tuple, Union
import torch
from transformers import PretrainedConfig
import vllm.envs as envs
from vllm.logger import init_logger
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
from vllm.model_executor.models import ModelRegistry
@@ -14,7 +13,7 @@ from vllm.tracing import is_otel_installed
from vllm.transformers_utils.config import get_config, get_hf_text_config
from vllm.utils import (cuda_device_count_stateless, get_cpu_memory, is_cpu,
is_hip, is_neuron, is_openvino, is_tpu, is_xpu,
print_warning_once, update_environment_variables)
print_warning_once)
if TYPE_CHECKING:
from ray.util.placement_group import PlacementGroup
@@ -695,12 +694,6 @@ class ParallelConfig:
self.distributed_executor_backend = backend
logger.info("Defaulting to use %s for distributed inference",
backend)
# If CUDA_VISIBLE_DEVICES is set on ROCm prior to vLLM init,
# propagate changes to HIP_VISIBLE_DEVICES (conversion handled by
# the update_environment_variables function)
if is_hip() and envs.CUDA_VISIBLE_DEVICES:
update_environment_variables(
{"CUDA_VISIBLE_DEVICES": envs.CUDA_VISIBLE_DEVICES})
self._verify_args()
self.rank = 0
-4
View File
@@ -386,10 +386,6 @@ def get_open_port() -> int:
def update_environment_variables(envs: Dict[str, str]):
if is_hip() and "CUDA_VISIBLE_DEVICES" in envs:
# Propagate changes to CUDA_VISIBLE_DEVICES to
# ROCm's HIP_VISIBLE_DEVICES as well
envs["HIP_VISIBLE_DEVICES"] = envs["CUDA_VISIBLE_DEVICES"]
for k, v in envs.items():
if k in os.environ and os.environ[k] != v:
logger.warning(
+1 -9
View File
@@ -11,7 +11,7 @@ from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
from vllm.sequence import (ExecuteModelRequest, IntermediateTensors,
SamplerOutput)
from vllm.utils import (enable_trace_function_call_for_thread, is_hip,
from vllm.utils import (enable_trace_function_call_for_thread,
update_environment_variables)
from vllm.worker.model_runner_base import ModelRunnerBase, ModelRunnerInputBase
@@ -309,14 +309,6 @@ class WorkerWrapperBase:
# overwriting CUDA_VISIBLE_DEVICES is desired behavior
# suppress the warning in `update_environment_variables`
del os.environ[key]
if is_hip():
hip_env_var = "HIP_VISIBLE_DEVICES"
if hip_env_var in os.environ:
logger.warning(
"Ignoring pre-set environment variable `%s=%s` as "
"%s has also been set, which takes precedence.",
hip_env_var, os.environ[hip_env_var], key)
os.environ.pop(hip_env_var, None)
update_environment_variables(envs)
def init_worker(self, *args, **kwargs):