[Neuron] Support inference with transformers-neuronx (#2569)

This commit is contained in:
Liangfu Chen
2024-02-28 09:34:34 -08:00
committed by GitHub
parent e46fa5d52e
commit 3b7178cfa4
18 changed files with 516 additions and 42 deletions
+7 -9
View File
@@ -44,7 +44,7 @@ class EngineArgs:
lora_extra_vocab_size: int = 256
lora_dtype = 'auto'
max_cpu_loras: Optional[int] = None
device: str = 'cuda'
device: str = 'auto'
def __post_init__(self):
if self.tokenizer is None:
@@ -171,7 +171,7 @@ class EngineArgs:
parser.add_argument('--block-size',
type=int,
default=EngineArgs.block_size,
choices=[8, 16, 32],
choices=[8, 16, 32, 128],
help='token block size')
parser.add_argument('--seed',
type=int,
@@ -264,13 +264,11 @@ class EngineArgs:
help=('Maximum number of LoRAs to store in CPU memory. '
'Must be >= than max_num_seqs. '
'Defaults to max_num_seqs.'))
parser.add_argument(
"--device",
type=str,
default=EngineArgs.device,
choices=["cuda"],
help=('Device type for vLLM execution. '
'Currently, only CUDA-compatible devices are supported.'))
parser.add_argument("--device",
type=str,
default=EngineArgs.device,
choices=["auto", "cuda", "neuron"],
help='Device type for vLLM execution.')
return parser
@classmethod
+18 -3
View File
@@ -3,6 +3,7 @@ from collections import defaultdict
import os
import time
import pickle
import importlib
from typing import (TYPE_CHECKING, Any, Dict, Iterable, List, Optional, Tuple,
Union)
@@ -20,7 +21,8 @@ from vllm.sequence import (SamplerOutput, Sequence, SequenceGroup,
SequenceGroupOutput, SequenceOutput, SequenceStatus)
from vllm.transformers_utils.tokenizer import (detokenize_incrementally,
TokenizerGroup)
from vllm.utils import Counter, set_cuda_visible_devices, get_ip, get_open_port, get_distributed_init_method
from vllm.utils import (Counter, set_cuda_visible_devices, get_ip,
get_open_port, get_distributed_init_method)
if ray:
from ray.util.scheduling_strategies import PlacementGroupSchedulingStrategy
@@ -31,6 +33,12 @@ if TYPE_CHECKING:
logger = init_logger(__name__)
_LOCAL_LOGGING_INTERVAL_SEC = 5
# A map between the device type (in device config) to its worker module.
DEVICE_TO_WORKER_MODULE_MAP = {
"cuda": "vllm.worker.worker",
"neuron": "vllm.worker.neuron_worker",
}
# If the env var is set, it uses the Ray's compiled DAG API
# which optimizes the control plane overhead.
# Run VLLM with VLLM_USE_RAY_COMPILED_DAG=1 to enable it.
@@ -138,10 +146,17 @@ class LLMEngine:
def get_tokenizer_for_seq(self, sequence: Sequence):
return self.tokenizer.get_lora_tokenizer(sequence.lora_request)
def _dispatch_worker(self):
worker_module = DEVICE_TO_WORKER_MODULE_MAP[
self.device_config.device_type]
imported_worker = importlib.import_module(worker_module)
Worker = imported_worker.Worker
return Worker
def _init_workers(self):
# Lazy import the Worker to avoid importing torch.cuda/xformers
# before CUDA_VISIBLE_DEVICES is set in the Worker
from vllm.worker.worker import Worker
Worker = self._dispatch_worker()
assert self.parallel_config.world_size == 1, (
"Ray is required if parallel_config.world_size > 1.")
@@ -243,7 +258,7 @@ class LLMEngine:
# Lazy import the Worker to avoid importing torch.cuda/xformers
# before CUDA_VISIBLE_DEVICES is set in the Worker
from vllm.worker.worker import Worker
Worker = self._dispatch_worker()
# Initialize torch distributed process group for the workers.
model_config = copy.deepcopy(self.model_config)