This commit is contained in:
2026-01-09 13:34:11 +08:00
parent dfa6476b58
commit b2ef04d792
538 changed files with 105693 additions and 2 deletions

0
vllm/engine/__init__.py Normal file
View File

649
vllm/engine/arg_utils.py Normal file
View File

@@ -0,0 +1,649 @@
import argparse
import dataclasses
from dataclasses import dataclass
from typing import List, Optional, Union
from vllm.config import (CacheConfig, DecodingConfig, DeviceConfig,
EngineConfig, LoadConfig, LoRAConfig, ModelConfig,
ParallelConfig, SchedulerConfig, SpeculativeConfig,
TokenizerPoolConfig, VisionLanguageConfig)
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
from vllm.utils import str_to_int_tuple
def nullable_str(val: str):
if not val or val == "None":
return None
return val
@dataclass
class EngineArgs:
"""Arguments for vLLM engine."""
model: str
served_model_name: Optional[Union[List[str]]] = None
tokenizer: Optional[str] = None
skip_tokenizer_init: bool = False
tokenizer_mode: str = 'auto'
trust_remote_code: bool = False
download_dir: Optional[str] = None
load_format: str = 'auto'
dtype: str = 'auto'
kv_cache_dtype: str = 'auto'
quantization_param_path: Optional[str] = None
seed: int = 0
max_model_len: Optional[int] = None
worker_use_ray: bool = False
pipeline_parallel_size: int = 1
tensor_parallel_size: int = 1
max_parallel_loading_workers: Optional[int] = None
block_size: int = 16
enable_prefix_caching: bool = False
use_v2_block_manager: bool = False
swap_space: int = 4 # GiB
gpu_memory_utilization: float = 0.90
max_num_batched_tokens: Optional[int] = None
max_num_seqs: int = 256
max_logprobs: int = 5 # OpenAI default value
disable_log_stats: bool = False
revision: Optional[str] = None
code_revision: Optional[str] = None
tokenizer_revision: Optional[str] = None
quantization: Optional[str] = None
enforce_eager: bool = False
max_context_len_to_capture: Optional[int] = None
max_seq_len_to_capture: int = 8192
disable_custom_all_reduce: bool = False
tokenizer_pool_size: int = 0
tokenizer_pool_type: str = "ray"
tokenizer_pool_extra_config: Optional[dict] = None
enable_lora: bool = False
max_loras: int = 1
max_lora_rank: int = 16
fully_sharded_loras: bool = False
lora_extra_vocab_size: int = 256
lora_dtype = 'auto'
max_cpu_loras: Optional[int] = None
device: str = 'auto'
ray_workers_use_nsight: bool = False
num_gpu_blocks_override: Optional[int] = None
num_lookahead_slots: int = 0
model_loader_extra_config: Optional[dict] = None
# Related to Vision-language models such as llava
image_input_type: Optional[str] = None
image_token_id: Optional[int] = None
image_input_shape: Optional[str] = None
image_feature_size: Optional[int] = None
scheduler_delay_factor: float = 0.0
enable_chunked_prefill: bool = False
guided_decoding_backend: str = 'outlines'
# Speculative decoding configuration.
speculative_model: Optional[str] = None
num_speculative_tokens: Optional[int] = None
speculative_max_model_len: Optional[int] = None
ngram_prompt_lookup_max: Optional[int] = None
ngram_prompt_lookup_min: Optional[int] = None
def __post_init__(self):
if self.tokenizer is None:
self.tokenizer = self.model
@staticmethod
def add_cli_args(
parser: argparse.ArgumentParser) -> argparse.ArgumentParser:
"""Shared CLI arguments for vLLM engine."""
# Model arguments
parser.add_argument(
'--model',
type=str,
default='facebook/opt-125m',
help='Name or path of the huggingface model to use.')
parser.add_argument(
'--tokenizer',
type=nullable_str,
default=EngineArgs.tokenizer,
help='Name or path of the huggingface tokenizer to use.')
parser.add_argument(
'--skip-tokenizer-init',
action='store_true',
help='Skip initialization of tokenizer and detokenizer')
parser.add_argument(
'--revision',
type=nullable_str,
default=None,
help='The specific model version to use. It can be a branch '
'name, a tag name, or a commit id. If unspecified, will use '
'the default version.')
parser.add_argument(
'--code-revision',
type=nullable_str,
default=None,
help='The specific revision to use for the model code on '
'Hugging Face Hub. It can be a branch name, a tag name, or a '
'commit id. If unspecified, will use the default version.')
parser.add_argument(
'--tokenizer-revision',
type=nullable_str,
default=None,
help='The specific tokenizer version to use. It can be a branch '
'name, a tag name, or a commit id. If unspecified, will use '
'the default version.')
parser.add_argument(
'--tokenizer-mode',
type=str,
default=EngineArgs.tokenizer_mode,
choices=['auto', 'slow'],
help='The tokenizer mode.\n\n* "auto" will use the '
'fast tokenizer if available.\n* "slow" will '
'always use the slow tokenizer.')
parser.add_argument('--trust-remote-code',
action='store_true',
help='Trust remote code from huggingface.')
parser.add_argument('--download-dir',
type=nullable_str,
default=EngineArgs.download_dir,
help='Directory to download and load the weights, '
'default to the default cache dir of '
'huggingface.')
parser.add_argument(
'--load-format',
type=str,
default=EngineArgs.load_format,
choices=[
'auto', 'pt', 'safetensors', 'npcache', 'dummy', 'tensorizer'
],
help='The format of the model weights to load.\n\n'
'* "auto" will try to load the weights in the safetensors format '
'and fall back to the pytorch bin format if safetensors format '
'is not available.\n'
'* "pt" will load the weights in the pytorch bin format.\n'
'* "safetensors" will load the weights in the safetensors format.\n'
'* "npcache" will load the weights in pytorch format and store '
'a numpy cache to speed up the loading.\n'
'* "dummy" will initialize the weights with random values, '
'which is mainly for profiling.\n'
'* "tensorizer" will load the weights using tensorizer from '
'CoreWeave which assumes tensorizer_uri is set to the location of '
'the serialized weights.')
parser.add_argument(
'--dtype',
type=str,
default=EngineArgs.dtype,
choices=[
'auto', 'half', 'float16', 'bfloat16', 'float', 'float32'
],
help='Data type for model weights and activations.\n\n'
'* "auto" will use FP16 precision for FP32 and FP16 models, and '
'BF16 precision for BF16 models.\n'
'* "half" for FP16. Recommended for AWQ quantization.\n'
'* "float16" is the same as "half".\n'
'* "bfloat16" for a balance between precision and range.\n'
'* "float" is shorthand for FP32 precision.\n'
'* "float32" for FP32 precision.')
parser.add_argument(
'--kv-cache-dtype',
type=str,
choices=['auto', 'fp8'],
default=EngineArgs.kv_cache_dtype,
help='Data type for kv cache storage. If "auto", will use model '
'data type. FP8_E5M2 (without scaling) is only supported on cuda '
'version greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is instead '
'supported for common inference criteria.')
parser.add_argument(
'--quantization-param-path',
type=nullable_str,
default=None,
help='Path to the JSON file containing the KV cache '
'scaling factors. This should generally be supplied, when '
'KV cache dtype is FP8. Otherwise, KV cache scaling factors '
'default to 1.0, which may cause accuracy issues. '
'FP8_E5M2 (without scaling) is only supported on cuda version'
'greater than 11.8. On ROCm (AMD GPU), FP8_E4M3 is instead '
'supported for common inference criteria.')
parser.add_argument('--max-model-len',
type=int,
default=EngineArgs.max_model_len,
help='Model context length. If unspecified, will '
'be automatically derived from the model config.')
parser.add_argument(
'--guided-decoding-backend',
type=str,
default='outlines',
choices=['outlines', 'lm-format-enforcer'],
help='Which engine will be used for guided decoding'
' (JSON schema / regex etc) by default. Currently support '
'https://github.com/outlines-dev/outlines and '
'https://github.com/noamgat/lm-format-enforcer.'
' Can be overridden per request via guided_decoding_backend'
' parameter.')
# Parallel arguments
parser.add_argument('--worker-use-ray',
action='store_true',
help='Use Ray for distributed serving, will be '
'automatically set when using more than 1 GPU.')
parser.add_argument('--pipeline-parallel-size',
'-pp',
type=int,
default=EngineArgs.pipeline_parallel_size,
help='Number of pipeline stages.')
parser.add_argument('--tensor-parallel-size',
'-tp',
type=int,
default=EngineArgs.tensor_parallel_size,
help='Number of tensor parallel replicas.')
parser.add_argument(
'--max-parallel-loading-workers',
type=int,
default=EngineArgs.max_parallel_loading_workers,
help='Load model sequentially in multiple batches, '
'to avoid RAM OOM when using tensor '
'parallel and large models.')
parser.add_argument(
'--ray-workers-use-nsight',
action='store_true',
help='If specified, use nsight to profile Ray workers.')
# KV cache arguments
parser.add_argument('--block-size',
type=int,
default=EngineArgs.block_size,
choices=[8, 16, 32],
help='Token block size for contiguous chunks of '
'tokens.')
parser.add_argument('--enable-prefix-caching',
action='store_true',
help='Enables automatic prefix caching.')
parser.add_argument('--use-v2-block-manager',
action='store_true',
help='Use BlockSpaceMangerV2.')
parser.add_argument(
'--num-lookahead-slots',
type=int,
default=EngineArgs.num_lookahead_slots,
help='Experimental scheduling config necessary for '
'speculative decoding. This will be replaced by '
'speculative config in the future; it is present '
'to enable correctness tests until then.')
parser.add_argument('--seed',
type=int,
default=EngineArgs.seed,
help='Random seed for operations.')
parser.add_argument('--swap-space',
type=int,
default=EngineArgs.swap_space,
help='CPU swap space size (GiB) per GPU.')
parser.add_argument(
'--gpu-memory-utilization',
type=float,
default=EngineArgs.gpu_memory_utilization,
help='The fraction of GPU memory to be used for the model '
'executor, which can range from 0 to 1. For example, a value of '
'0.5 would imply 50%% GPU memory utilization. If unspecified, '
'will use the default value of 0.9.')
parser.add_argument(
'--num-gpu-blocks-override',
type=int,
default=None,
help='If specified, ignore GPU profiling result and use this number'
'of GPU blocks. Used for testing preemption.')
parser.add_argument('--max-num-batched-tokens',
type=int,
default=EngineArgs.max_num_batched_tokens,
help='Maximum number of batched tokens per '
'iteration.')
parser.add_argument('--max-num-seqs',
type=int,
default=EngineArgs.max_num_seqs,
help='Maximum number of sequences per iteration.')
parser.add_argument(
'--max-logprobs',
type=int,
default=EngineArgs.max_logprobs,
help=('Max number of log probs to return logprobs is specified in'
' SamplingParams.'))
parser.add_argument('--disable-log-stats',
action='store_true',
help='Disable logging statistics.')
# Quantization settings.
parser.add_argument('--quantization',
'-q',
type=nullable_str,
choices=[*QUANTIZATION_METHODS, None],
default=EngineArgs.quantization,
help='Method used to quantize the weights. If '
'None, we first check the `quantization_config` '
'attribute in the model config file. If that is '
'None, we assume the model weights are not '
'quantized and use `dtype` to determine the data '
'type of the weights.')
parser.add_argument('--enforce-eager',
action='store_true',
help='Always use eager-mode PyTorch. If False, '
'will use eager mode and CUDA graph in hybrid '
'for maximal performance and flexibility.')
parser.add_argument('--max-context-len-to-capture',
type=int,
default=EngineArgs.max_context_len_to_capture,
help='Maximum context length covered by CUDA '
'graphs. When a sequence has context length '
'larger than this, we fall back to eager mode. '
'(DEPRECATED. Use --max-seq_len-to-capture instead'
')')
parser.add_argument('--max-seq_len-to-capture',
type=int,
default=EngineArgs.max_seq_len_to_capture,
help='Maximum sequence length covered by CUDA '
'graphs. When a sequence has context length '
'larger than this, we fall back to eager mode.')
parser.add_argument('--disable-custom-all-reduce',
action='store_true',
default=EngineArgs.disable_custom_all_reduce,
help='See ParallelConfig.')
parser.add_argument('--tokenizer-pool-size',
type=int,
default=EngineArgs.tokenizer_pool_size,
help='Size of tokenizer pool to use for '
'asynchronous tokenization. If 0, will '
'use synchronous tokenization.')
parser.add_argument('--tokenizer-pool-type',
type=str,
default=EngineArgs.tokenizer_pool_type,
help='Type of tokenizer pool to use for '
'asynchronous tokenization. Ignored '
'if tokenizer_pool_size is 0.')
parser.add_argument('--tokenizer-pool-extra-config',
type=nullable_str,
default=EngineArgs.tokenizer_pool_extra_config,
help='Extra config for tokenizer pool. '
'This should be a JSON string that will be '
'parsed into a dictionary. Ignored if '
'tokenizer_pool_size is 0.')
# LoRA related configs
parser.add_argument('--enable-lora',
action='store_true',
help='If True, enable handling of LoRA adapters.')
parser.add_argument('--max-loras',
type=int,
default=EngineArgs.max_loras,
help='Max number of LoRAs in a single batch.')
parser.add_argument('--max-lora-rank',
type=int,
default=EngineArgs.max_lora_rank,
help='Max LoRA rank.')
parser.add_argument(
'--lora-extra-vocab-size',
type=int,
default=EngineArgs.lora_extra_vocab_size,
help=('Maximum size of extra vocabulary that can be '
'present in a LoRA adapter (added to the base '
'model vocabulary).'))
parser.add_argument(
'--lora-dtype',
type=str,
default=EngineArgs.lora_dtype,
choices=['auto', 'float16', 'bfloat16', 'float32'],
help=('Data type for LoRA. If auto, will default to '
'base model dtype.'))
parser.add_argument(
'--max-cpu-loras',
type=int,
default=EngineArgs.max_cpu_loras,
help=('Maximum number of LoRAs to store in CPU memory. '
'Must be >= than max_num_seqs. '
'Defaults to max_num_seqs.'))
parser.add_argument(
'--fully-sharded-loras',
action='store_true',
help=('By default, only half of the LoRA computation is '
'sharded with tensor parallelism. '
'Enabling this will use the fully sharded layers. '
'At high sequence length, max rank or '
'tensor parallel size, this is likely faster.'))
parser.add_argument("--device",
type=str,
default=EngineArgs.device,
choices=["auto", "cuda", "neuron", "cpu", "musa"],
help='Device type for vLLM execution.')
# Related to Vision-language models such as llava
parser.add_argument(
'--image-input-type',
type=nullable_str,
default=None,
choices=[
t.name.lower() for t in VisionLanguageConfig.ImageInputType
],
help=('The image input type passed into vLLM. '
'Should be one of "pixel_values" or "image_features".'))
parser.add_argument('--image-token-id',
type=int,
default=None,
help=('Input id for image token.'))
parser.add_argument(
'--image-input-shape',
type=nullable_str,
default=None,
help=('The biggest image input shape (worst for memory footprint) '
'given an input type. Only used for vLLM\'s profile_run.'))
parser.add_argument(
'--image-feature-size',
type=int,
default=None,
help=('The image feature size along the context dimension.'))
parser.add_argument(
'--scheduler-delay-factor',
type=float,
default=EngineArgs.scheduler_delay_factor,
help='Apply a delay (of delay factor multiplied by previous'
'prompt latency) before scheduling next prompt.')
parser.add_argument(
'--enable-chunked-prefill',
action='store_true',
help='If set, the prefill requests can be chunked based on the '
'max_num_batched_tokens.')
parser.add_argument(
'--speculative-model',
type=nullable_str,
default=EngineArgs.speculative_model,
help=
'The name of the draft model to be used in speculative decoding.')
parser.add_argument(
'--num-speculative-tokens',
type=int,
default=EngineArgs.num_speculative_tokens,
help='The number of speculative tokens to sample from '
'the draft model in speculative decoding.')
parser.add_argument(
'--speculative-max-model-len',
type=int,
default=EngineArgs.speculative_max_model_len,
help='The maximum sequence length supported by the '
'draft model. Sequences over this length will skip '
'speculation.')
parser.add_argument(
'--ngram-prompt-lookup-max',
type=int,
default=EngineArgs.ngram_prompt_lookup_max,
help='Max size of window for ngram prompt lookup in speculative '
'decoding.')
parser.add_argument(
'--ngram-prompt-lookup-min',
type=int,
default=EngineArgs.ngram_prompt_lookup_min,
help='Min size of window for ngram prompt lookup in speculative '
'decoding.')
parser.add_argument('--model-loader-extra-config',
type=nullable_str,
default=EngineArgs.model_loader_extra_config,
help='Extra config for model loader. '
'This will be passed to the model loader '
'corresponding to the chosen load_format. '
'This should be a JSON string that will be '
'parsed into a dictionary.')
parser.add_argument(
"--served-model-name",
nargs="+",
type=str,
default=None,
help="The model name(s) used in the API. If multiple "
"names are provided, the server will respond to any "
"of the provided names. The model name in the model "
"field of a response will be the first name in this "
"list. If not specified, the model name will be the "
"same as the `--model` argument. Noted that this name(s)"
"will also be used in `model_name` tag content of "
"prometheus metrics, if multiple names provided, metrics"
"tag will take the first one.")
return parser
@classmethod
def from_cli_args(cls, args: argparse.Namespace) -> 'EngineArgs':
# Get the list of attributes of this dataclass.
attrs = [attr.name for attr in dataclasses.fields(cls)]
# Set the attributes from the parsed arguments.
engine_args = cls(**{attr: getattr(args, attr) for attr in attrs})
return engine_args
def create_engine_config(self, ) -> EngineConfig:
device_config = DeviceConfig(self.device)
model_config = ModelConfig(
self.model, self.tokenizer, self.tokenizer_mode,
self.trust_remote_code, self.dtype, self.seed, self.revision,
self.code_revision, self.tokenizer_revision, self.max_model_len,
self.quantization, self.quantization_param_path,
self.enforce_eager, self.max_context_len_to_capture,
self.max_seq_len_to_capture, self.max_logprobs,
self.skip_tokenizer_init, self.served_model_name)
cache_config = CacheConfig(self.block_size,
self.gpu_memory_utilization,
self.swap_space, self.kv_cache_dtype,
self.num_gpu_blocks_override,
model_config.get_sliding_window(),
self.enable_prefix_caching)
parallel_config = ParallelConfig(
self.pipeline_parallel_size, self.tensor_parallel_size,
self.worker_use_ray, self.max_parallel_loading_workers,
self.disable_custom_all_reduce,
TokenizerPoolConfig.create_config(
self.tokenizer_pool_size,
self.tokenizer_pool_type,
self.tokenizer_pool_extra_config,
), self.ray_workers_use_nsight)
speculative_config = SpeculativeConfig.maybe_create_spec_config(
target_model_config=model_config,
target_parallel_config=parallel_config,
target_dtype=self.dtype,
speculative_model=self.speculative_model,
num_speculative_tokens=self.num_speculative_tokens,
speculative_max_model_len=self.speculative_max_model_len,
enable_chunked_prefill=self.enable_chunked_prefill,
use_v2_block_manager=self.use_v2_block_manager,
ngram_prompt_lookup_max=self.ngram_prompt_lookup_max,
ngram_prompt_lookup_min=self.ngram_prompt_lookup_min,
)
scheduler_config = SchedulerConfig(
self.max_num_batched_tokens,
self.max_num_seqs,
model_config.max_model_len,
self.use_v2_block_manager,
num_lookahead_slots=(self.num_lookahead_slots
if speculative_config is None else
speculative_config.num_lookahead_slots),
delay_factor=self.scheduler_delay_factor,
enable_chunked_prefill=self.enable_chunked_prefill,
)
lora_config = LoRAConfig(
max_lora_rank=self.max_lora_rank,
max_loras=self.max_loras,
fully_sharded_loras=self.fully_sharded_loras,
lora_extra_vocab_size=self.lora_extra_vocab_size,
lora_dtype=self.lora_dtype,
max_cpu_loras=self.max_cpu_loras if self.max_cpu_loras
and self.max_cpu_loras > 0 else None) if self.enable_lora else None
load_config = LoadConfig(
load_format=self.load_format,
download_dir=self.download_dir,
model_loader_extra_config=self.model_loader_extra_config,
)
if self.image_input_type:
if (not self.image_token_id or not self.image_input_shape
or not self.image_feature_size):
raise ValueError(
'Specify `image_token_id`, `image_input_shape` and '
'`image_feature_size` together with `image_input_type`.')
vision_language_config = VisionLanguageConfig(
image_input_type=VisionLanguageConfig.
get_image_input_enum_type(self.image_input_type),
image_token_id=self.image_token_id,
image_input_shape=str_to_int_tuple(self.image_input_shape),
image_feature_size=self.image_feature_size,
)
else:
vision_language_config = None
decoding_config = DecodingConfig(
guided_decoding_backend=self.guided_decoding_backend)
return EngineConfig(model_config=model_config,
cache_config=cache_config,
parallel_config=parallel_config,
scheduler_config=scheduler_config,
device_config=device_config,
lora_config=lora_config,
vision_language_config=vision_language_config,
speculative_config=speculative_config,
load_config=load_config,
decoding_config=decoding_config)
@dataclass
class AsyncEngineArgs(EngineArgs):
"""Arguments for asynchronous vLLM engine."""
engine_use_ray: bool = False
disable_log_requests: bool = False
max_log_len: Optional[int] = None
@staticmethod
def add_cli_args(parser: argparse.ArgumentParser,
async_args_only: bool = False) -> argparse.ArgumentParser:
if not async_args_only:
parser = EngineArgs.add_cli_args(parser)
parser.add_argument('--engine-use-ray',
action='store_true',
help='Use Ray to start the LLM engine in a '
'separate process as the server process.')
parser.add_argument('--disable-log-requests',
action='store_true',
help='Disable logging requests.')
parser.add_argument('--max-log-len',
type=int,
default=None,
help='Max number of prompt characters or prompt '
'ID numbers being printed in log.'
'\n\nDefault: Unlimited')
return parser
# These functions are used by sphinx to build the documentation
def _engine_args_parser():
return EngineArgs.add_cli_args(argparse.ArgumentParser())
def _async_engine_args_parser():
return AsyncEngineArgs.add_cli_args(argparse.ArgumentParser(),
async_args_only=True)

View File

@@ -0,0 +1,737 @@
import asyncio
import time
from functools import partial
from typing import (Any, AsyncIterator, Callable, Dict, Iterable, List,
Optional, Set, Tuple, Type, Union)
from transformers import PreTrainedTokenizer
import vllm.envs as envs
from vllm.config import DecodingConfig, ModelConfig
from vllm.core.scheduler import SchedulerOutputs
from vllm.engine.arg_utils import AsyncEngineArgs
from vllm.engine.llm_engine import LLMEngine
from vllm.executor.ray_utils import initialize_ray_cluster, ray
from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
from vllm.outputs import RequestOutput
from vllm.sampling_params import SamplingParams
from vllm.sequence import ExecuteModelRequest, MultiModalData, SamplerOutput
from vllm.usage.usage_lib import UsageContext
logger = init_logger(__name__)
ENGINE_ITERATION_TIMEOUT_S = envs.VLLM_ENGINE_ITERATION_TIMEOUT_S
class AsyncEngineDeadError(RuntimeError):
pass
def _raise_exception_on_finish(
task: asyncio.Task, error_callback: Callable[[Exception],
None]) -> None:
msg = ("Task finished unexpectedly. This should never happen! "
"Please open an issue on Github.")
exception = None
try:
task.result()
# NOTE: This will be thrown if task exits normally (which it should not)
raise AsyncEngineDeadError(msg)
except Exception as e:
exception = e
logger.error("Engine background task failed", exc_info=e)
error_callback(exception)
raise AsyncEngineDeadError(
msg + " See stack trace above for the actual cause.") from e
class AsyncStream:
"""A stream of RequestOutputs for a request that can be
iterated over asynchronously."""
def __init__(self, request_id: str) -> None:
self.request_id = request_id
self._queue: asyncio.Queue = asyncio.Queue()
self._finished = False
def put(self, item: Union[RequestOutput, Exception]) -> None:
if self._finished:
return
self._queue.put_nowait(item)
def finish(self) -> None:
self._queue.put_nowait(StopAsyncIteration())
self._finished = True
@property
def finished(self) -> bool:
return self._finished
def __aiter__(self):
return self
async def __anext__(self) -> RequestOutput:
result = await self._queue.get()
if isinstance(result, Exception):
raise result
return result
class RequestTracker:
"""Synchronous abstraction for tracking requests."""
def __init__(self) -> None:
self._request_streams: Dict[str, AsyncStream] = {}
self._finished_requests: asyncio.Queue[str] = asyncio.Queue()
self._new_requests: asyncio.Queue[Tuple[AsyncStream,
dict]] = asyncio.Queue()
self.new_requests_event = asyncio.Event()
def __contains__(self, item):
return item in self._request_streams
def __len__(self) -> int:
return len(self._request_streams)
def propagate_exception(self,
exc: Exception,
request_id: Optional[str] = None) -> None:
"""Propagate an exception to request streams
(all if request_id is None)."""
if request_id is not None:
self._request_streams[request_id].put(exc)
self.abort_request(request_id)
else:
for rid, stream in self._request_streams.items():
stream.put(exc)
self.abort_request(rid)
def process_request_output(self,
request_output: RequestOutput,
*,
verbose: bool = False) -> None:
"""Process a request output from the engine."""
request_id = request_output.request_id
self._request_streams[request_id].put(request_output)
if request_output.finished:
if verbose:
logger.info("Finished request %s.", request_id)
self.abort_request(request_id)
def process_exception(self,
request_id: str,
exception: Exception,
*,
verbose: bool = False) -> None:
"""Propagate an exception from the engine."""
self._request_streams[request_id].put(exception)
if verbose:
logger.info("Finished request %s.", request_id)
self.abort_request(request_id)
def add_request(self, request_id: str,
**engine_add_request_kwargs) -> AsyncStream:
"""Add a request to be sent to the engine on the next background
loop iteration."""
if request_id in self._request_streams:
raise KeyError(f"Request {request_id} already exists.")
stream = AsyncStream(request_id)
self._new_requests.put_nowait((stream, {
"request_id": request_id,
**engine_add_request_kwargs
}))
self.new_requests_event.set()
return stream
def abort_request(self, request_id: str, *, verbose: bool = False) -> None:
"""Abort a request during next background loop iteration."""
if verbose:
logger.info("Aborted request %s.", request_id)
self._finished_requests.put_nowait(request_id)
if request_id not in self._request_streams or self._request_streams[
request_id].finished:
# The request has already finished or been aborted.
return
self._request_streams[request_id].finish()
def get_new_and_finished_requests(self) -> Tuple[List[Dict], Set[str]]:
"""Get the new requests and finished requests to be
sent to the engine."""
new_requests: List[Dict] = []
finished_requests: Set[str] = set()
while not self._finished_requests.empty():
request_id = self._finished_requests.get_nowait()
finished_requests.add(request_id)
self._request_streams.pop(request_id, None)
while not self._new_requests.empty():
stream, new_request = self._new_requests.get_nowait()
if stream.request_id in finished_requests:
# The request has already been aborted.
stream.finish()
continue
self._request_streams[stream.request_id] = stream
new_requests.append(new_request)
return new_requests, finished_requests
async def wait_for_new_requests(self):
if not self.has_new_requests():
await self.new_requests_event.wait()
self.new_requests_event.clear()
def has_new_requests(self):
return not self._new_requests.empty()
class _AsyncLLMEngine(LLMEngine):
"""Extension of LLMEngine to add async methods."""
async def step_async(self) -> List[RequestOutput]:
"""Performs one decoding iteration and returns newly generated results.
The workers are ran asynchronously if possible.
This function performs one decoding iteration of the engine. It first
schedules the sequences to be executed in the next iteration and the
token blocks to be swapped in/out/copy. Then, it executes the model
and updates the scheduler with the model outputs. Finally, it decodes
the sequences and returns the newly generated results.
"""
seq_group_metadata_list, scheduler_outputs = self.scheduler.schedule()
if not scheduler_outputs.is_empty():
# Execute the model.
execute_model_req = ExecuteModelRequest(
seq_group_metadata_list=seq_group_metadata_list,
blocks_to_swap_in=scheduler_outputs.blocks_to_swap_in,
blocks_to_swap_out=scheduler_outputs.blocks_to_swap_out,
blocks_to_copy=scheduler_outputs.blocks_to_copy,
num_lookahead_slots=scheduler_outputs.num_lookahead_slots,
running_queue_size=scheduler_outputs.running_queue_size,
)
output = await self.model_executor.execute_model_async(
execute_model_req)
else:
output = []
request_outputs = self._process_model_outputs(
output, scheduler_outputs.scheduled_seq_groups,
scheduler_outputs.ignored_seq_groups, seq_group_metadata_list)
# Log stats.
self.do_log_stats(scheduler_outputs, output)
return request_outputs
async def encode_request_async(
self,
request_id: str, # pylint: disable=unused-argument
prompt: Optional[str],
prompt_token_ids: Optional[List[int]] = None,
lora_request: Optional[LoRARequest] = None,
):
if prompt_token_ids is None:
assert prompt is not None
prompt_token_ids = await self.tokenizer.encode_async(
request_id=request_id,
prompt=prompt,
lora_request=lora_request)
return prompt_token_ids
async def add_request_async(
self,
request_id: str,
prompt: Optional[str],
sampling_params: SamplingParams,
prompt_token_ids: Optional[List[int]] = None,
arrival_time: Optional[float] = None,
lora_request: Optional[LoRARequest] = None,
multi_modal_data: Optional[MultiModalData] = None,
) -> None:
if lora_request is not None and not self.lora_config:
raise ValueError(f"Got lora_request {lora_request} but LoRA is "
"not enabled!")
if arrival_time is None:
arrival_time = time.time()
prompt_token_ids = await self.encode_request_async(
request_id=request_id,
prompt=prompt,
prompt_token_ids=prompt_token_ids,
lora_request=lora_request)
return self.add_request(request_id,
prompt=prompt,
prompt_token_ids=prompt_token_ids,
sampling_params=sampling_params,
arrival_time=arrival_time,
lora_request=lora_request,
multi_modal_data=multi_modal_data)
async def check_health_async(self) -> None:
self.model_executor.check_health()
class AsyncLLMEngine:
"""An asynchronous wrapper for LLMEngine.
This class is used to wrap the LLMEngine class to make it asynchronous. It
uses asyncio to create a background loop that keeps processing incoming
requests. The LLMEngine is kicked by the generate method when there
are requests in the waiting queue. The generate method yields the outputs
from the LLMEngine to the caller.
NOTE: For the comprehensive list of arguments, see `LLMEngine`.
Args:
worker_use_ray: Whether to use Ray for model workers. Required for
distributed execution. Should be the same as
`parallel_config.worker_use_ray`.
engine_use_ray: Whether to make LLMEngine a Ray actor. If so, the
async frontend will be executed in a separate process as the
model workers.
log_requests: Whether to log the requests.
max_log_len: Maximum number of prompt characters or prompt ID numbers
being printed in log.
start_engine_loop: If True, the background task to run the engine
will be automatically started in the generate call.
*args: Arguments for LLMEngine.
*kwargs: Arguments for LLMEngine.
"""
_engine_class: Type[_AsyncLLMEngine] = _AsyncLLMEngine
def __init__(self,
worker_use_ray: bool,
engine_use_ray: bool,
*args,
log_requests: bool = True,
max_log_len: Optional[int] = None,
start_engine_loop: bool = True,
**kwargs) -> None:
self.worker_use_ray = worker_use_ray
self.engine_use_ray = engine_use_ray
self.log_requests = log_requests
self.max_log_len = max_log_len
self.engine = self._init_engine(*args, **kwargs)
self.background_loop: Optional[asyncio.Future] = None
# We need to keep a reference to unshielded
# task as well to prevent it from being garbage
# collected
self._background_loop_unshielded: Optional[asyncio.Task[Any]] = None
self.start_engine_loop = start_engine_loop
self._errored_with: Optional[BaseException] = None
# Lazy initialized fields
self._request_tracker: RequestTracker
@classmethod
def from_engine_args(
cls,
engine_args: AsyncEngineArgs,
start_engine_loop: bool = True,
usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
) -> "AsyncLLMEngine":
"""Creates an async LLM engine from the engine arguments."""
# Create the engine configs.
engine_config = engine_args.create_engine_config()
if engine_config.device_config.device_type == "neuron":
from vllm.executor.neuron_executor import NeuronExecutorAsync
executor_class = NeuronExecutorAsync
elif engine_config.device_config.device_type == "cpu":
assert not engine_config.parallel_config.worker_use_ray, (
"Ray is not supported with the CPU backend.")
from vllm.executor.cpu_executor import CPUExecutorAsync
executor_class = CPUExecutorAsync
elif engine_config.parallel_config.worker_use_ray:
initialize_ray_cluster(engine_config.parallel_config)
from vllm.executor.ray_gpu_executor import RayGPUExecutorAsync
executor_class = RayGPUExecutorAsync
else:
assert engine_config.parallel_config.world_size == 1, (
"Ray is required if parallel_config.world_size > 1.")
from vllm.executor.gpu_executor import GPUExecutorAsync
executor_class = GPUExecutorAsync
# Create the async LLM engine.
engine = cls(
engine_config.parallel_config.worker_use_ray,
engine_args.engine_use_ray,
**engine_config.to_dict(),
executor_class=executor_class,
log_requests=not engine_args.disable_log_requests,
log_stats=not engine_args.disable_log_stats,
max_log_len=engine_args.max_log_len,
start_engine_loop=start_engine_loop,
usage_context=usage_context,
)
return engine
@property
def is_running(self) -> bool:
return (self.background_loop is not None
and self._background_loop_unshielded is not None
and not self._background_loop_unshielded.done())
@property
def is_stopped(self) -> bool:
return self.errored or (self.background_loop is not None and
self._background_loop_unshielded is not None
and self._background_loop_unshielded.done())
@property
def errored(self) -> bool:
return self._errored_with is not None
def set_errored(self, exc: Exception) -> None:
self._errored_with = exc
def _error_callback(self, exc: Exception) -> None:
self.set_errored(exc)
self._request_tracker.propagate_exception(exc)
async def get_tokenizer(self) -> "PreTrainedTokenizer":
if self.engine_use_ray:
return await self.engine.get_tokenizer.remote() # type: ignore
else:
return self.engine.get_tokenizer()
def start_background_loop(self) -> None:
"""Start the background loop."""
if self.errored:
raise AsyncEngineDeadError(
"Background loop has errored already.") from self._errored_with
if self.is_running:
raise RuntimeError("Background loop is already running.")
# Initialize the RequestTracker here so it uses the right event loop.
self._request_tracker = RequestTracker()
self._background_loop_unshielded = asyncio.get_event_loop(
).create_task(self.run_engine_loop())
self._background_loop_unshielded.add_done_callback(
partial(_raise_exception_on_finish,
error_callback=self._error_callback))
self.background_loop = asyncio.shield(self._background_loop_unshielded)
def _init_engine(self, *args,
**kwargs) -> Union[_AsyncLLMEngine, "ray.ObjectRef"]:
if not self.engine_use_ray:
engine_class = self._engine_class
elif self.worker_use_ray:
engine_class = ray.remote(num_cpus=0)(self._engine_class).remote
else:
# FIXME(woosuk): This is a bit hacky. Be careful when changing the
# order of the arguments.
cache_config = kwargs["cache_config"]
parallel_config = kwargs["parallel_config"]
if parallel_config.tensor_parallel_size == 1:
num_gpus = cache_config.gpu_memory_utilization
else:
num_gpus = 1
engine_class = ray.remote(num_gpus=num_gpus)(
self._engine_class).remote
return engine_class(*args, **kwargs)
async def engine_step(self) -> bool:
"""Kick the engine to process the waiting requests.
Returns True if there are in-progress requests."""
new_requests, finished_requests = (
self._request_tracker.get_new_and_finished_requests())
for new_request in new_requests:
# Add the request into the vLLM engine's waiting queue.
# TODO: Maybe add add_request_batch to reduce Ray overhead
try:
if self.engine_use_ray:
await self.engine.add_request.remote( # type: ignore
**new_request)
else:
await self.engine.add_request_async(**new_request)
except ValueError as e:
# TODO: use a vLLM specific error for failed validation
self._request_tracker.process_exception(
new_request["request_id"],
e,
verbose=self.log_requests,
)
if finished_requests:
await self._engine_abort(finished_requests)
if self.engine_use_ray:
request_outputs = await self.engine.step.remote() # type: ignore
else:
request_outputs = await self.engine.step_async()
# Put the outputs into the corresponding streams.
for request_output in request_outputs:
self._request_tracker.process_request_output(
request_output, verbose=self.log_requests)
return len(request_outputs) > 0
async def _engine_abort(self, request_ids: Iterable[str]):
if self.engine_use_ray:
await self.engine.abort_request.remote(request_ids) # type: ignore
else:
self.engine.abort_request(request_ids)
async def run_engine_loop(self):
has_requests_in_progress = False
while True:
if not has_requests_in_progress:
logger.debug("Waiting for new requests...")
await self._request_tracker.wait_for_new_requests()
logger.debug("Got new requests!")
# Abort if iteration takes too long due to unrecoverable errors
# (eg. NCCL timeouts).
try:
has_requests_in_progress = await asyncio.wait_for(
self.engine_step(), ENGINE_ITERATION_TIMEOUT_S)
except asyncio.TimeoutError as exc:
logger.error(
"Engine iteration timed out. This should never happen!")
self.set_errored(exc)
raise
await asyncio.sleep(0)
async def add_request(
self,
request_id: str,
prompt: Optional[str],
sampling_params: SamplingParams,
prompt_token_ids: Optional[List[int]] = None,
arrival_time: Optional[float] = None,
lora_request: Optional[LoRARequest] = None,
multi_modal_data: Optional[MultiModalData] = None,
) -> AsyncStream:
if self.log_requests:
shortened_prompt = prompt
shortened_token_ids = prompt_token_ids
if self.max_log_len is not None:
if shortened_prompt is not None:
shortened_prompt = shortened_prompt[:self.max_log_len]
if shortened_token_ids is not None:
shortened_token_ids = shortened_token_ids[:self.
max_log_len]
logger.info(
"Received request %s: prompt: %r, "
"sampling_params: %s, prompt_token_ids: %s, "
"lora_request: %s.", request_id, shortened_prompt,
sampling_params, shortened_token_ids, lora_request)
if not self.is_running:
if self.start_engine_loop:
self.start_background_loop()
else:
raise AsyncEngineDeadError(
"Background loop is not running. If it was running, "
"inspect the output to find the stacktrace of the "
"error that caused the background loop to stop "
"(AsyncEngineDeadError).")
if arrival_time is None:
arrival_time = time.time()
if self.engine_use_ray:
prompt_token_ids = await (
self.engine.encode_request_async.remote( # type: ignore
request_id=request_id,
prompt=prompt,
prompt_token_ids=prompt_token_ids,
lora_request=lora_request))
else:
prompt_token_ids = await self.engine.encode_request_async(
request_id=request_id,
prompt=prompt,
prompt_token_ids=prompt_token_ids,
lora_request=lora_request)
stream = self._request_tracker.add_request(
request_id,
prompt=prompt,
sampling_params=sampling_params,
prompt_token_ids=prompt_token_ids,
arrival_time=arrival_time,
lora_request=lora_request,
multi_modal_data=multi_modal_data,
)
return stream
async def generate(
self,
prompt: Optional[str],
sampling_params: SamplingParams,
request_id: str,
prompt_token_ids: Optional[List[int]] = None,
lora_request: Optional[LoRARequest] = None,
multi_modal_data: Optional[MultiModalData] = None
) -> AsyncIterator[RequestOutput]:
"""Generate outputs for a request.
Generate outputs for a request. This method is a coroutine. It adds the
request into the waiting queue of the LLMEngine and streams the outputs
from the LLMEngine to the caller.
Args:
prompt: The prompt string. Can be None if prompt_token_ids is
provided.
sampling_params: The sampling parameters of the request.
request_id: The unique id of the request.
prompt_token_ids: The token IDs of the prompt. If None, we
use the tokenizer to convert the prompts to token IDs.
lora_request: LoRA request to use for generation, if any.
multi_modal_data: Multi modal data per request.
Yields:
The output `RequestOutput` objects from the LLMEngine for the
request.
Details:
- If the engine is not running, start the background loop,
which iteratively invokes
:meth:`~vllm.engine.async_llm_engine.AsyncLLMEngine.engine_step`
to process the waiting requests.
- Add the request to the engine's `RequestTracker`.
On the next background loop, this request will be sent to
the underlying engine.
Also, a corresponding `AsyncStream` will be created.
- Wait for the request outputs from `AsyncStream` and yield them.
Example:
>>> # Please refer to entrypoints/api_server.py for
>>> # the complete example.
>>>
>>> # initialize the engine and the example input
>>> engine = AsyncLLMEngine.from_engine_args(engine_args)
>>> example_input = {
>>> "prompt": "What is LLM?",
>>> "stream": False, # assume the non-streaming case
>>> "temperature": 0.0,
>>> "request_id": 0,
>>> }
>>>
>>> # start the generation
>>> results_generator = engine.generate(
>>> example_input["prompt"],
>>> SamplingParams(temperature=example_input["temperature"]),
>>> example_input["request_id"])
>>>
>>> # get the results
>>> final_output = None
>>> async for request_output in results_generator:
>>> if await request.is_disconnected():
>>> # Abort the request if the client disconnects.
>>> await engine.abort(request_id)
>>> # Return or raise an error
>>> ...
>>> final_output = request_output
>>>
>>> # Process and return the final output
>>> ...
"""
# Preprocess the request.
arrival_time = time.time()
try:
stream = await self.add_request(
request_id,
prompt,
sampling_params,
prompt_token_ids=prompt_token_ids,
arrival_time=arrival_time,
lora_request=lora_request,
multi_modal_data=multi_modal_data,
)
async for request_output in stream:
yield request_output
except (Exception, asyncio.CancelledError) as e:
# If there is an exception or coroutine is cancelled, abort the
# request.
self._abort(request_id)
raise e
async def abort(self, request_id: str) -> None:
"""Abort a request.
Abort a submitted request. If the request is finished or not found,
this method will be a no-op.
Args:
request_id: The unique id of the request.
"""
if not self.is_running:
raise AsyncEngineDeadError(
"Background loop is not running. If it was running, "
"inspect the output to find the stacktrace of the "
"error that caused the background loop to stop "
"(AsyncEngineDeadError).")
return self._abort(request_id)
def _abort(self, request_id: str) -> None:
"""Abort a request.
Abort a submitted request. If the request is finished or not found,
this method will be a no-op.
Args:
request_id: The unique id of the request.
"""
self._request_tracker.abort_request(request_id,
verbose=self.log_requests)
async def get_model_config(self) -> ModelConfig:
"""Get the model configuration of the vLLM engine."""
if self.engine_use_ray:
return await self.engine.get_model_config.remote() # type: ignore
else:
return self.engine.get_model_config()
async def get_decoding_config(self) -> DecodingConfig:
"""Get the decoding configuration of the vLLM engine."""
if self.engine_use_ray:
return await self.engine.get_decoding_config.remote( # type: ignore
)
else:
return self.engine.get_decoding_config()
async def do_log_stats(
self,
scheduler_outputs: Optional[SchedulerOutputs] = None,
model_output: Optional[List[SamplerOutput]] = None) -> None:
if self.engine_use_ray:
await self.engine.do_log_stats.remote( # type: ignore
scheduler_outputs, model_output)
else:
self.engine.do_log_stats()
async def check_health(self) -> None:
"""Raises an error if engine is unhealthy."""
t = time.perf_counter()
logger.debug("Starting health check...")
if self.is_stopped:
raise AsyncEngineDeadError("Background loop is stopped.")
if self.engine_use_ray:
try:
await self.engine.check_health.remote() # type: ignore
except ray.exceptions.RayActorError as e:
raise RuntimeError("Engine is dead.") from e
else:
await self.engine.check_health_async()
logger.debug("Health check took %fs", time.perf_counter() - t)

784
vllm/engine/llm_engine.py Normal file
View File

@@ -0,0 +1,784 @@
import time
from typing import Iterable, List, Optional, Type, Union
from transformers import GenerationConfig, PreTrainedTokenizer
import vllm
from vllm.config import (CacheConfig, DecodingConfig, DeviceConfig, LoadConfig,
LoRAConfig, ModelConfig, ParallelConfig,
SchedulerConfig, SpeculativeConfig,
VisionLanguageConfig)
from vllm.core.scheduler import (ScheduledSequenceGroup, Scheduler,
SchedulerOutputs)
from vllm.engine.arg_utils import EngineArgs
from vllm.engine.metrics import StatLogger, Stats
from vllm.engine.output_processor.interfaces import (
SequenceGroupOutputProcessor)
from vllm.engine.output_processor.stop_checker import StopChecker
from vllm.engine.output_processor.util import create_output_by_sequence_group
from vllm.executor.executor_base import ExecutorBase
from vllm.executor.ray_utils import initialize_ray_cluster
from vllm.logger import init_logger
from vllm.lora.request import LoRARequest
from vllm.outputs import RequestOutput
from vllm.sampling_params import SamplingParams
from vllm.sequence import (ExecuteModelRequest, MultiModalData, SamplerOutput,
Sequence, SequenceGroup, SequenceGroupMetadata,
SequenceStatus)
from vllm.transformers_utils.detokenizer import Detokenizer
from vllm.transformers_utils.tokenizer_group import (BaseTokenizerGroup,
get_tokenizer_group)
from vllm.usage.usage_lib import (UsageContext, is_usage_stats_enabled,
usage_message)
from vllm.utils import Counter
logger = init_logger(__name__)
_LOCAL_LOGGING_INTERVAL_SEC = 5
def _load_generation_config_dict(model_config: ModelConfig):
try:
return GenerationConfig.from_pretrained(
model_config.model,
revision=model_config.revision,
).to_diff_dict()
except OSError:
# Not found.
return {}
class LLMEngine:
"""An LLM engine that receives requests and generates texts.
This is the main class for the vLLM engine. It receives requests
from clients and generates texts from the LLM. It includes a tokenizer, a
language model (possibly distributed across multiple GPUs), and GPU memory
space allocated for intermediate states (aka KV cache). This class utilizes
iteration-level scheduling and efficient memory management to maximize the
serving throughput.
The `LLM` class wraps this class for offline batched inference and the
`AsyncLLMEngine` class wraps this class for online serving.
NOTE: The config arguments are derived from the `EngineArgs` class. For the
comprehensive list of arguments, see `EngineArgs`.
Args:
model_config: The configuration related to the LLM model.
cache_config: The configuration related to the KV cache memory
management.
parallel_config: The configuration related to distributed execution.
scheduler_config: The configuration related to the request scheduler.
device_config: The configuration related to the device.
lora_config (Optional): The configuration related to serving multi-LoRA.
vision_language_config (Optional): The configuration related to vision
language models.
speculative_config (Optional): The configuration related to speculative
decoding.
executor_class: The model executor class for managing distributed
execution.
log_stats: Whether to log statistics.
usage_context: Specified entry point, used for usage info collection
"""
def __init__(
self,
model_config: ModelConfig,
cache_config: CacheConfig,
parallel_config: ParallelConfig,
scheduler_config: SchedulerConfig,
device_config: DeviceConfig,
load_config: LoadConfig,
lora_config: Optional[LoRAConfig],
vision_language_config: Optional[VisionLanguageConfig],
speculative_config: Optional[SpeculativeConfig],
decoding_config: Optional[DecodingConfig],
executor_class: Type[ExecutorBase],
log_stats: bool,
usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
) -> None:
logger.info(
"Initializing an LLM engine (v%s) with config: "
"model=%r, speculative_config=%r, tokenizer=%r, "
"skip_tokenizer_init=%s, tokenizer_mode=%s, revision=%s, "
"tokenizer_revision=%s, trust_remote_code=%s, dtype=%s, "
"max_seq_len=%d, download_dir=%r, load_format=%s, "
"tensor_parallel_size=%d, disable_custom_all_reduce=%s, "
"quantization=%s, enforce_eager=%s, kv_cache_dtype=%s, "
"quantization_param_path=%s, device_config=%s, "
"decoding_config=%r, seed=%d, served_model_name=%s)",
vllm.__version__,
model_config.model,
speculative_config,
model_config.tokenizer,
model_config.skip_tokenizer_init,
model_config.tokenizer_mode,
model_config.revision,
model_config.tokenizer_revision,
model_config.trust_remote_code,
model_config.dtype,
model_config.max_model_len,
load_config.download_dir,
load_config.load_format,
parallel_config.tensor_parallel_size,
parallel_config.disable_custom_all_reduce,
model_config.quantization,
model_config.enforce_eager,
cache_config.cache_dtype,
model_config.quantization_param_path,
device_config.device,
decoding_config,
model_config.seed,
model_config.served_model_name,
)
# TODO(woosuk): Print more configs in debug mode.
self.model_config = model_config
self.cache_config = cache_config
self.lora_config = lora_config
self.vision_language_config = vision_language_config
self.parallel_config = parallel_config
self.scheduler_config = scheduler_config
self.device_config = device_config
self.speculative_config = speculative_config
self.load_config = load_config
self.decoding_config = decoding_config or DecodingConfig()
self.log_stats = log_stats
if not self.model_config.skip_tokenizer_init:
self.tokenizer: BaseTokenizerGroup
self._init_tokenizer()
self.detokenizer = Detokenizer(self.tokenizer)
else:
self.detokenizer = None
self.tokenizer = None
self.seq_counter = Counter()
self.generation_config_fields = _load_generation_config_dict(
model_config)
self.model_executor = executor_class(
model_config=model_config,
cache_config=cache_config,
parallel_config=parallel_config,
scheduler_config=scheduler_config,
device_config=device_config,
lora_config=lora_config,
vision_language_config=vision_language_config,
speculative_config=speculative_config,
load_config=load_config,
)
self._initialize_kv_caches()
# If usage stat is enabled, collect relevant info.
if is_usage_stats_enabled():
from vllm.model_executor.model_loader import (
get_architecture_class_name)
usage_message.report_usage(
get_architecture_class_name(model_config),
usage_context,
extra_kvs={
# Common configuration
"dtype":
str(model_config.dtype),
"tensor_parallel_size":
parallel_config.tensor_parallel_size,
"block_size":
cache_config.block_size,
"gpu_memory_utilization":
cache_config.gpu_memory_utilization,
# Quantization
"quantization":
model_config.quantization,
"kv_cache_dtype":
cache_config.cache_dtype,
# Feature flags
"enable_lora":
bool(lora_config),
"enable_prefix_caching":
cache_config.enable_prefix_caching,
"enforce_eager":
model_config.enforce_eager,
"disable_custom_all_reduce":
parallel_config.disable_custom_all_reduce,
})
if self.tokenizer:
# Ping the tokenizer to ensure liveness if it runs in a
# different process.
self.tokenizer.ping()
# Create the scheduler.
# NOTE: the cache_config here have been updated with the numbers of
# GPU and CPU blocks, which are profiled in the distributed executor.
self.scheduler = Scheduler(scheduler_config, cache_config, lora_config)
# Metric Logging.
if self.log_stats:
self.stat_logger = StatLogger(
local_interval=_LOCAL_LOGGING_INTERVAL_SEC,
labels=dict(model_name=model_config.served_model_name),
max_model_len=self.model_config.max_model_len)
self.stat_logger.info("cache_config", self.cache_config)
# Create sequence output processor, e.g. for beam search or
# speculative decoding.
self.output_processor = (
SequenceGroupOutputProcessor.create_output_processor(
self.scheduler_config,
self.detokenizer,
self.scheduler,
self.seq_counter,
self.get_tokenizer_for_seq,
stop_checker=StopChecker(
self.scheduler_config.max_model_len,
self.get_tokenizer_for_seq,
),
))
def _initialize_kv_caches(self) -> None:
"""Initialize the KV cache in the worker(s).
The workers will determine the number of blocks in both the GPU cache
and the swap CPU cache.
"""
num_gpu_blocks, num_cpu_blocks = (
self.model_executor.determine_num_available_blocks())
if self.cache_config.num_gpu_blocks_override is not None:
num_gpu_blocks_override = self.cache_config.num_gpu_blocks_override
logger.info(
"Overriding num_gpu_blocks=%d with "
"num_gpu_blocks_override=%d", num_gpu_blocks,
num_gpu_blocks_override)
num_gpu_blocks = num_gpu_blocks_override
self.cache_config.num_gpu_blocks = num_gpu_blocks
self.cache_config.num_cpu_blocks = num_cpu_blocks
self.model_executor.initialize_cache(num_gpu_blocks, num_cpu_blocks)
@classmethod
def from_engine_args(
cls,
engine_args: EngineArgs,
usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
) -> "LLMEngine":
"""Creates an LLM engine from the engine arguments."""
# Create the engine configs.
engine_config = engine_args.create_engine_config()
# Initialize the cluster and specify the executor class.
if engine_config.device_config.device_type == "neuron":
from vllm.executor.neuron_executor import NeuronExecutor
executor_class = NeuronExecutor
elif engine_config.device_config.device_type == "cpu":
from vllm.executor.cpu_executor import CPUExecutor
executor_class = CPUExecutor
elif engine_config.parallel_config.worker_use_ray:
initialize_ray_cluster(engine_config.parallel_config)
from vllm.executor.ray_gpu_executor import RayGPUExecutor
executor_class = RayGPUExecutor
else:
assert engine_config.parallel_config.world_size == 1, (
"Ray is required if parallel_config.world_size > 1.")
from vllm.executor.gpu_executor import GPUExecutor
executor_class = GPUExecutor
# Create the LLM engine.
engine = cls(
**engine_config.to_dict(),
executor_class=executor_class,
log_stats=not engine_args.disable_log_stats,
usage_context=usage_context,
)
return engine
def __reduce__(self):
# This is to ensure that the LLMEngine is not referenced in
# the closure used to initialize Ray worker actors
raise RuntimeError("LLMEngine should not be pickled!")
def __del__(self):
# Shutdown model executor when engine is garbage collected
# Use getattr since __init__ can fail before the field is set
if model_executor := getattr(self, "model_executor", None):
model_executor.shutdown()
def get_tokenizer(self) -> "PreTrainedTokenizer":
return self.tokenizer.get_lora_tokenizer(None)
def get_tokenizer_for_seq(self,
sequence: Sequence) -> "PreTrainedTokenizer":
return self.tokenizer.get_lora_tokenizer(sequence.lora_request)
def _init_tokenizer(self, **tokenizer_init_kwargs):
init_kwargs = dict(
tokenizer_id=self.model_config.tokenizer,
enable_lora=bool(self.lora_config),
max_num_seqs=self.scheduler_config.max_num_seqs,
max_input_length=None,
tokenizer_mode=self.model_config.tokenizer_mode,
trust_remote_code=self.model_config.trust_remote_code,
revision=self.model_config.tokenizer_revision)
init_kwargs.update(tokenizer_init_kwargs)
self.tokenizer = get_tokenizer_group(
self.parallel_config.tokenizer_pool_config, **init_kwargs)
def _verify_args(self) -> None:
self.model_config.verify_with_parallel_config(self.parallel_config)
self.cache_config.verify_with_parallel_config(self.parallel_config)
if self.lora_config:
self.lora_config.verify_with_model_config(self.model_config)
self.lora_config.verify_with_scheduler_config(
self.scheduler_config)
def encode_request(
self,
request_id: str, # pylint: disable=unused-argument
prompt: Optional[str],
prompt_token_ids: Optional[List[int]] = None,
lora_request: Optional[LoRARequest] = None,
):
if prompt_token_ids is None:
assert prompt is not None
prompt_token_ids = self.tokenizer.encode(request_id=request_id,
prompt=prompt,
lora_request=lora_request)
return prompt_token_ids
def add_request(
self,
request_id: str,
prompt: Optional[str],
sampling_params: SamplingParams,
prompt_token_ids: Optional[List[int]] = None,
arrival_time: Optional[float] = None,
lora_request: Optional[LoRARequest] = None,
multi_modal_data: Optional[MultiModalData] = None,
) -> None:
"""Add a request to the engine's request pool.
The request is added to the request pool and will be processed by the
scheduler as `engine.step()` is called. The exact scheduling policy is
determined by the scheduler.
Args:
request_id: The unique ID of the request.
prompt: The prompt string. Can be None if prompt_token_ids is
provided.
sampling_params: The sampling parameters for text generation.
prompt_token_ids: The token IDs of the prompt. If None, we
use the tokenizer to convert the prompts to token IDs.
arrival_time: The arrival time of the request. If None, we use
the current monotonic time.
multi_modal_data: Multi modal data per request.
Details:
- Set arrival_time to the current time if it is None.
- Set prompt_token_ids to the encoded prompt if it is None.
- Create `best_of` number of :class:`~vllm.Sequence` objects.
- Create a :class:`~vllm.SequenceGroup` object
from the list of :class:`~vllm.Sequence`.
- Add the :class:`~vllm.SequenceGroup` object to the scheduler.
Example:
>>> # initialize engine
>>> engine = LLMEngine.from_engine_args(engine_args)
>>> # set request arguments
>>> example_prompt = "Who is the president of the United States?"
>>> sampling_params = SamplingParams(temperature=0.0)
>>> request_id = 0
>>>
>>> # add the request to the engine
>>> engine.add_request(
>>> str(request_id),
>>> example_prompt,
>>> SamplingParams(temperature=0.0))
>>> # continue the request processing
>>> ...
"""
if lora_request is not None and not self.lora_config:
raise ValueError(f"Got lora_request {lora_request} but LoRA is "
"not enabled!")
max_logprobs = self.get_model_config().max_logprobs
if (sampling_params.logprobs
and sampling_params.logprobs > max_logprobs) or (
sampling_params.prompt_logprobs
and sampling_params.prompt_logprobs > max_logprobs):
raise ValueError(f"Cannot request more than "
f"{max_logprobs} logprobs.")
if arrival_time is None:
arrival_time = time.time()
prompt_token_ids = self.encode_request(
request_id=request_id,
prompt=prompt,
prompt_token_ids=prompt_token_ids,
lora_request=lora_request)
# Create the sequences.
block_size = self.cache_config.block_size
seq_id = next(self.seq_counter)
eos_token_id = None
if self.tokenizer:
eos_token_id = self.tokenizer.get_lora_tokenizer(
lora_request).eos_token_id
else:
logger.warning("Use None for EOS token id because tokenizer is "
"not initialized")
seq = Sequence(seq_id, prompt, prompt_token_ids, block_size,
eos_token_id, lora_request)
# Defensive copy of SamplingParams, which are used by the sampler,
# this doesn't deep-copy LogitsProcessor objects
sampling_params = sampling_params.clone()
# Add the eos token id into the sampling_params to support min_tokens
# processing
if seq.eos_token_id is not None:
sampling_params.all_stop_token_ids.add(seq.eos_token_id)
sampling_params.update_from_generation_config(
self.generation_config_fields)
# Create the sequence group.
seq_group = SequenceGroup(request_id, [seq], sampling_params,
arrival_time, lora_request, multi_modal_data)
# Add the sequence group to the scheduler.
self.scheduler.add_seq_group(seq_group)
def abort_request(self, request_id: Union[str, Iterable[str]]) -> None:
"""Aborts a request(s) with the given ID.
Args:
request_id: The ID(s) of the request to abort.
Details:
- Refer to the
:meth:`~vllm.core.scheduler.Scheduler.abort_seq_group`
from class :class:`~vllm.core.scheduler.Scheduler`.
Example:
>>> # initialize engine and add a request with request_id
>>> request_id = str(0)
>>> # abort the request
>>> engine.abort_request(request_id)
"""
self.scheduler.abort_seq_group(request_id)
def get_model_config(self) -> ModelConfig:
"""Gets the model configuration."""
return self.model_config
def get_decoding_config(self) -> DecodingConfig:
"""Gets the decoding configuration."""
return self.decoding_config
def get_num_unfinished_requests(self) -> int:
"""Gets the number of unfinished requests."""
return self.scheduler.get_num_unfinished_seq_groups()
def has_unfinished_requests(self) -> bool:
"""Returns True if there are unfinished requests."""
return self.scheduler.has_unfinished_seqs()
def _process_model_outputs(
self,
output: List[SamplerOutput],
scheduled_seq_groups: List[ScheduledSequenceGroup],
ignored_seq_groups: List[SequenceGroup],
seq_group_metadata_list: List[SequenceGroupMetadata],
) -> List[RequestOutput]:
"""Apply the model output to the sequences in the scheduled seq groups.
Returns RequestOutputs that can be returned to the client.
"""
now = time.time()
# Organize outputs by [sequence group][step] instead of
# [step][sequence group].
output_by_sequence_group = create_output_by_sequence_group(
sampler_outputs=output, num_seq_groups=len(scheduled_seq_groups))
# Update the scheduled sequence groups with the model outputs.
for scheduled_seq_group, outputs, seq_group_meta in zip(
scheduled_seq_groups, output_by_sequence_group,
seq_group_metadata_list):
seq_group = scheduled_seq_group.seq_group
seq_group.update_num_computed_tokens(
scheduled_seq_group.token_chunk_size)
self.output_processor.process_prompt_logprob(seq_group, outputs)
if seq_group_meta.do_sample:
self.output_processor.process_outputs(seq_group, outputs)
# Free the finished sequence groups.
self.scheduler.free_finished_seq_groups()
# Create the outputs.
request_outputs: List[RequestOutput] = []
for scheduled_seq_group in scheduled_seq_groups:
seq_group = scheduled_seq_group.seq_group
seq_group.maybe_set_first_token_time(now)
request_output = RequestOutput.from_seq_group(seq_group)
request_outputs.append(request_output)
for seq_group in ignored_seq_groups:
request_output = RequestOutput.from_seq_group(seq_group)
request_outputs.append(request_output)
return request_outputs
def step(self) -> List[RequestOutput]:
"""Performs one decoding iteration and returns newly generated results.
.. figure:: https://i.imgur.com/sv2HssD.png
:alt: Overview of the step function
:align: center
Overview of the step function.
Details:
- Step 1: Schedules the sequences to be executed in the next
iteration and the token blocks to be swapped in/out/copy.
- Depending on the scheduling policy,
sequences may be `preempted/reordered`.
- A Sequence Group (SG) refer to a group of sequences
that are generated from the same prompt.
- Step 2: Calls the distributed executor to execute the model.
- Step 3: Processes the model output. This mainly includes:
- Decodes the relevant outputs.
- Updates the scheduled sequence groups with model outputs
based on its `sampling parameters` (`use_beam_search` or not).
- Frees the finished sequence groups.
- Finally, it creates and returns the newly generated results.
Example:
>>> # Please see the example/ folder for more detailed examples.
>>>
>>> # initialize engine and request arguments
>>> engine = LLMEngine.from_engine_args(engine_args)
>>> example_inputs = [(0, "What is LLM?",
>>> SamplingParams(temperature=0.0))]
>>>
>>> # Start the engine with an event loop
>>> while True:
>>> if example_inputs:
>>> req_id, prompt, sampling_params = example_inputs.pop(0)
>>> engine.add_request(str(req_id), prompt, sampling_params)
>>>
>>> # continue the request processing
>>> request_outputs = engine.step()
>>> for request_output in request_outputs:
>>> if request_output.finished:
>>> # return or show the request output
>>>
>>> if not (engine.has_unfinished_requests() or example_inputs):
>>> break
"""
seq_group_metadata_list, scheduler_outputs = self.scheduler.schedule()
if not scheduler_outputs.is_empty():
execute_model_req = ExecuteModelRequest(
seq_group_metadata_list=seq_group_metadata_list,
blocks_to_swap_in=scheduler_outputs.blocks_to_swap_in,
blocks_to_swap_out=scheduler_outputs.blocks_to_swap_out,
blocks_to_copy=scheduler_outputs.blocks_to_copy,
num_lookahead_slots=scheduler_outputs.num_lookahead_slots,
running_queue_size=scheduler_outputs.running_queue_size,
)
output = self.model_executor.execute_model(
execute_model_req=execute_model_req)
else:
output = []
request_outputs = self._process_model_outputs(
output, scheduler_outputs.scheduled_seq_groups,
scheduler_outputs.ignored_seq_groups, seq_group_metadata_list)
# Log stats.
self.do_log_stats(scheduler_outputs, output)
return request_outputs
def do_log_stats(
self,
scheduler_outputs: Optional[SchedulerOutputs] = None,
model_output: Optional[List[SamplerOutput]] = None) -> None:
"""Forced log when no requests active."""
if self.log_stats:
self.stat_logger.log(
self._get_stats(scheduler_outputs, model_output))
def _get_stats(
self,
scheduler_outputs: Optional[SchedulerOutputs],
model_output: Optional[List[SamplerOutput]] = None) -> Stats:
"""Get Stats to be Logged to Prometheus.
Args:
scheduler_outputs: Optional, used to populate metrics related to
the scheduled batch,
model_output: Optional, used to emit speculative decoding metrics
which are created by the workers.
"""
now = time.time()
# System State
# Scheduler State
num_running_sys = len(self.scheduler.running)
num_swapped_sys = len(self.scheduler.swapped)
num_waiting_sys = len(self.scheduler.waiting)
# KV Cache Usage in %
num_total_gpu = self.cache_config.num_gpu_blocks
num_free_gpu = self.scheduler.block_manager.get_num_free_gpu_blocks()
gpu_cache_usage_sys = 1.0 - (num_free_gpu / num_total_gpu)
num_total_cpu = self.cache_config.num_cpu_blocks
cpu_cache_usage_sys = 0.
if num_total_cpu > 0:
num_free_cpu = self.scheduler.block_manager.get_num_free_cpu_blocks(
)
cpu_cache_usage_sys = 1.0 - (num_free_cpu / num_total_cpu)
# Iteration stats
num_prompt_tokens_iter = 0
num_generation_tokens_iter = 0
time_to_first_tokens_iter: List[float] = []
time_per_output_tokens_iter: List[float] = []
# Request stats
# Latency
time_e2e_requests: List[float] = []
# Metadata
num_prompt_tokens_requests: List[int] = []
num_generation_tokens_requests: List[int] = []
best_of_requests: List[int] = []
n_requests: List[int] = []
finished_reason_requests: List[str] = []
# NOTE: This loop assumes prefill seq_groups are before
# decode seq_groups in scheduled_seq_groups.
if scheduler_outputs is not None:
num_generation_tokens_from_prefill_groups = 0.
# NOTE: if scheduler_outputs.num_prefill_groups > 0 and
# the len of scheduler_outputs.scheduled_seq_groups is !=
# scheduler_outputs.num_prefill_groups, this means that
# chunked prefills have been detected.
for idx, scheduled_seq_group in enumerate(
scheduler_outputs.scheduled_seq_groups):
group_was_prefill = idx < scheduler_outputs.num_prefill_groups
seq_group = scheduled_seq_group.seq_group
# NOTE: a seq_group that completed all of its prefill tokens
# in the last iteration will have seq_group.is_prefill() = False
# with group_was_prefill = True
if group_was_prefill:
# Number of prompt tokens.
num_prompt_tokens_iter += (
scheduled_seq_group.token_chunk_size)
# If the seq_group just finished the prefill state
# get TTFT.
if not seq_group.is_prefill():
latency = seq_group.get_last_latency(now)
time_to_first_tokens_iter.append(latency)
# One generation token per finished prefill.
num_generation_tokens_from_prefill_groups += (
seq_group.num_seqs())
else:
# TPOTs.
latency = seq_group.get_last_latency(now)
time_per_output_tokens_iter.append(latency)
# Because of chunked prefill, we can have a single sequence
# group that does multiple prompt_runs. To prevent logging
# the same metadata more than once per request, we standardize
# on logging request level information for finished requests,
# which can only happen once.
if seq_group.is_finished():
# Latency timings
time_e2e_requests.append(now -
seq_group.metrics.arrival_time)
# Metadata
num_prompt_tokens_requests.append(
len(seq_group.prompt_token_ids))
num_generation_tokens_requests.extend([
seq.get_output_len()
for seq in seq_group.get_finished_seqs()
])
best_of_requests.append(seq_group.sampling_params.best_of)
n_requests.append(seq_group.sampling_params.n)
finished_reason_requests.extend([
SequenceStatus.get_finished_reason(seq.status)
for seq in seq_group.get_finished_seqs()
])
# Number of generation tokens.
# num_batched_tokens equals the number of prompt_tokens plus the
# number of decode_tokens in a single iteration. So,
# num_generation_tokens = num_batched_tokens - num_prompt_tokens
# + num_generation_tokens_from_prefill_groups (since we generate
# one token on prefills on iters where the prefill finishes).
num_generation_tokens_iter = (
scheduler_outputs.num_batched_tokens - num_prompt_tokens_iter +
num_generation_tokens_from_prefill_groups)
# Spec decode, if enabled, emits specialized metrics from the worker in
# sampler output.
if model_output and (model_output[0].spec_decode_worker_metrics
is not None):
spec_decode_metrics = model_output[0].spec_decode_worker_metrics
else:
spec_decode_metrics = None
return Stats(
now=now,
# System stats
# Scheduler State
num_running_sys=num_running_sys,
num_swapped_sys=num_swapped_sys,
num_waiting_sys=num_waiting_sys,
# KV Cache Usage in %
gpu_cache_usage_sys=gpu_cache_usage_sys,
cpu_cache_usage_sys=cpu_cache_usage_sys,
# Iteration stats
num_prompt_tokens_iter=num_prompt_tokens_iter,
num_generation_tokens_iter=num_generation_tokens_iter,
time_to_first_tokens_iter=time_to_first_tokens_iter,
time_per_output_tokens_iter=time_per_output_tokens_iter,
spec_decode_metrics=spec_decode_metrics,
# Request stats
# Latency
time_e2e_requests=time_e2e_requests,
# Metadata
num_prompt_tokens_requests=num_prompt_tokens_requests,
num_generation_tokens_requests=num_generation_tokens_requests,
best_of_requests=best_of_requests,
n_requests=n_requests,
finished_reason_requests=finished_reason_requests,
)
def add_lora(self, lora_request: LoRARequest) -> bool:
return self.model_executor.add_lora(lora_request)
def remove_lora(self, lora_id: int) -> bool:
return self.model_executor.remove_lora(lora_id)
def list_loras(self) -> List[int]:
return self.model_executor.list_loras()
def check_health(self) -> None:
self.model_executor.check_health()

368
vllm/engine/metrics.py Normal file
View File

@@ -0,0 +1,368 @@
import time
from dataclasses import dataclass
from typing import TYPE_CHECKING
from typing import Counter as CollectionsCounter
from typing import Dict, List, Optional, Protocol, Union
import numpy as np
from prometheus_client import (REGISTRY, Counter, Gauge, Histogram, Info,
disable_created_metrics)
from vllm.logger import init_logger
if TYPE_CHECKING:
from vllm.spec_decode.metrics import SpecDecodeWorkerMetrics
logger = init_logger(__name__)
disable_created_metrics()
# The begin-* and end* here are used by the documentation generator
# to extract the metrics definitions.
# begin-metrics-definitions
class Metrics:
labelname_finish_reason = "finished_reason"
def __init__(self, labelnames: List[str], max_model_len: int):
# Unregister any existing vLLM collectors
for collector in list(REGISTRY._collector_to_names):
if hasattr(collector, "_name") and "vllm" in collector._name:
REGISTRY.unregister(collector)
# Config Information
self.info_cache_config = Info(
name='vllm:cache_config',
documentation='information of cache_config')
# System stats
# Scheduler State
self.gauge_scheduler_running = Gauge(
name="vllm:num_requests_running",
documentation="Number of requests currently running on GPU.",
labelnames=labelnames)
self.gauge_scheduler_waiting = Gauge(
name="vllm:num_requests_waiting",
documentation="Number of requests waiting to be processed.",
labelnames=labelnames)
self.gauge_scheduler_swapped = Gauge(
name="vllm:num_requests_swapped",
documentation="Number of requests swapped to CPU.",
labelnames=labelnames)
# KV Cache Usage in %
self.gauge_gpu_cache_usage = Gauge(
name="vllm:gpu_cache_usage_perc",
documentation="GPU KV-cache usage. 1 means 100 percent usage.",
labelnames=labelnames)
self.gauge_cpu_cache_usage = Gauge(
name="vllm:cpu_cache_usage_perc",
documentation="CPU KV-cache usage. 1 means 100 percent usage.",
labelnames=labelnames)
# Iteration stats
self.counter_prompt_tokens = Counter(
name="vllm:prompt_tokens_total",
documentation="Number of prefill tokens processed.",
labelnames=labelnames)
self.counter_generation_tokens = Counter(
name="vllm:generation_tokens_total",
documentation="Number of generation tokens processed.",
labelnames=labelnames)
self.histogram_time_to_first_token = Histogram(
name="vllm:time_to_first_token_seconds",
documentation="Histogram of time to first token in seconds.",
labelnames=labelnames,
buckets=[
0.001, 0.005, 0.01, 0.02, 0.04, 0.06, 0.08, 0.1, 0.25, 0.5,
0.75, 1.0, 2.5, 5.0, 7.5, 10.0
])
self.histogram_time_per_output_token = Histogram(
name="vllm:time_per_output_token_seconds",
documentation="Histogram of time per output token in seconds.",
labelnames=labelnames,
buckets=[
0.01, 0.025, 0.05, 0.075, 0.1, 0.15, 0.2, 0.3, 0.4, 0.5, 0.75,
1.0, 2.5
])
# Request stats
# Latency
self.histogram_e2e_time_request = Histogram(
name="vllm:e2e_request_latency_seconds",
documentation="Histogram of end to end request latency in seconds.",
labelnames=labelnames,
buckets=[1.0, 2.5, 5.0, 10.0, 15.0, 20.0, 30.0, 40.0, 50.0, 60.0])
# Metadata
self.histogram_num_prompt_tokens_request = Histogram(
name="vllm:request_prompt_tokens",
documentation="Number of prefill tokens processed.",
labelnames=labelnames,
buckets=build_1_2_5_buckets(max_model_len),
)
self.histogram_num_generation_tokens_request = Histogram(
name="vllm:request_generation_tokens",
documentation="Number of generation tokens processed.",
labelnames=labelnames,
buckets=build_1_2_5_buckets(max_model_len),
)
self.histogram_best_of_request = Histogram(
name="vllm:request_params_best_of",
documentation="Histogram of the best_of request parameter.",
labelnames=labelnames,
buckets=[1, 2, 5, 10, 20],
)
self.histogram_n_request = Histogram(
name="vllm:request_params_n",
documentation="Histogram of the n request parameter.",
labelnames=labelnames,
buckets=[1, 2, 5, 10, 20],
)
self.counter_request_success = Counter(
name="vllm:request_success_total",
documentation="Count of successfully processed requests.",
labelnames=labelnames + [Metrics.labelname_finish_reason])
# Deprecated in favor of vllm:prompt_tokens_total
self.gauge_avg_prompt_throughput = Gauge(
name="vllm:avg_prompt_throughput_toks_per_s",
documentation="Average prefill throughput in tokens/s.",
labelnames=labelnames,
)
# Deprecated in favor of vllm:generation_tokens_total
self.gauge_avg_generation_throughput = Gauge(
name="vllm:avg_generation_throughput_toks_per_s",
documentation="Average generation throughput in tokens/s.",
labelnames=labelnames,
)
# end-metrics-definitions
def build_1_2_5_buckets(max_value: int):
"""
Builds a list of buckets with increasing powers of 10 multiplied by
mantissa values (1, 2, 5) until the value exceeds the specified maximum.
Example:
>>> build_1_2_5_buckets(100)
[1, 2, 5, 10, 20, 50, 100]
"""
mantissa_lst = [1, 2, 5]
exponent = 0
buckets = []
while True:
for m in mantissa_lst:
value = m * 10**exponent
if value <= max_value:
buckets.append(value)
else:
return buckets
exponent += 1
@dataclass
class Stats:
"""Created by LLMEngine for use by StatLogger."""
now: float
# System stats (should have _sys suffix)
# Scheduler State
num_running_sys: int
num_waiting_sys: int
num_swapped_sys: int
# KV Cache Usage in %
gpu_cache_usage_sys: float
cpu_cache_usage_sys: float
# Iteration stats (should have _iter suffix)
num_prompt_tokens_iter: int
num_generation_tokens_iter: int
time_to_first_tokens_iter: List[float]
time_per_output_tokens_iter: List[float]
# Request stats (should have _requests suffix)
# Latency
time_e2e_requests: List[float]
# Metadata
num_prompt_tokens_requests: List[int]
num_generation_tokens_requests: List[int]
best_of_requests: List[int]
n_requests: List[int]
finished_reason_requests: List[str]
spec_decode_metrics: Optional["SpecDecodeWorkerMetrics"] = None
class SupportsMetricsInfo(Protocol):
def metrics_info(self) -> Dict[str, str]:
...
class StatLogger:
"""StatLogger is used LLMEngine to log to Promethus and Stdout."""
def __init__(self, local_interval: float, labels: Dict[str, str],
max_model_len: int) -> None:
# Metadata for logging locally.
self.last_local_log = time.time()
self.local_interval = local_interval
# Tracked stats over current local logging interval.
self.num_prompt_tokens: List[int] = []
self.num_generation_tokens: List[int] = []
# Prometheus metrics
self.labels = labels
self.metrics = Metrics(labelnames=list(labels.keys()),
max_model_len=max_model_len)
def info(self, type: str, obj: SupportsMetricsInfo) -> None:
if type == "cache_config":
self.metrics.info_cache_config.info(obj.metrics_info())
def _get_throughput(self, tracked_stats: List[int], now: float) -> float:
return float(np.sum(tracked_stats) / (now - self.last_local_log))
def _local_interval_elapsed(self, now: float) -> bool:
elapsed_time = now - self.last_local_log
return elapsed_time > self.local_interval
def _log_prometheus(self, stats: Stats) -> None:
# System state data
self._log_gauge(self.metrics.gauge_scheduler_running,
stats.num_running_sys)
self._log_gauge(self.metrics.gauge_scheduler_swapped,
stats.num_swapped_sys)
self._log_gauge(self.metrics.gauge_scheduler_waiting,
stats.num_waiting_sys)
self._log_gauge(self.metrics.gauge_gpu_cache_usage,
stats.gpu_cache_usage_sys)
self._log_gauge(self.metrics.gauge_cpu_cache_usage,
stats.cpu_cache_usage_sys)
# Iteration level data
self._log_counter(self.metrics.counter_prompt_tokens,
stats.num_prompt_tokens_iter)
self._log_counter(self.metrics.counter_generation_tokens,
stats.num_generation_tokens_iter)
self._log_histogram(self.metrics.histogram_time_to_first_token,
stats.time_to_first_tokens_iter)
self._log_histogram(self.metrics.histogram_time_per_output_token,
stats.time_per_output_tokens_iter)
# Request level data
# Latency
self._log_histogram(self.metrics.histogram_e2e_time_request,
stats.time_e2e_requests)
# Metadata
finished_reason_counter = CollectionsCounter(
stats.finished_reason_requests)
self._log_counter_labels(self.metrics.counter_request_success,
finished_reason_counter,
Metrics.labelname_finish_reason)
self._log_histogram(self.metrics.histogram_num_prompt_tokens_request,
stats.num_prompt_tokens_requests)
self._log_histogram(
self.metrics.histogram_num_generation_tokens_request,
stats.num_generation_tokens_requests)
self._log_histogram(self.metrics.histogram_n_request, stats.n_requests)
self._log_histogram(self.metrics.histogram_best_of_request,
stats.best_of_requests)
def _log_gauge(self, gauge: Gauge, data: Union[int, float]) -> None:
# Convenience function for logging to gauge.
gauge.labels(**self.labels).set(data)
def _log_counter(self, counter: Counter, data: Union[int, float]) -> None:
# Convenience function for logging to counter.
counter.labels(**self.labels).inc(data)
def _log_counter_labels(self, counter: Counter, data: CollectionsCounter,
label_key: str) -> None:
# Convenience function for collection counter of labels.
for label, count in data.items():
counter.labels(**{**self.labels, label_key: label}).inc(count)
def _log_histogram(self, histogram: Histogram,
data: Union[List[int], List[float]]) -> None:
# Convenience function for logging list to histogram.
for datum in data:
histogram.labels(**self.labels).observe(datum)
def _log_prometheus_interval(self, prompt_throughput: float,
generation_throughput: float) -> None:
# Logs metrics to prometheus that are computed every logging_interval.
# Support legacy gauge metrics that make throughput calculations on
# the vLLM side. Moving forward, we should use counters like
# counter_prompt_tokens, counter_generation_tokens
# Which log raw data and calculate summaries using rate() on the
# grafana/prometheus side. See
# https://github.com/vllm-project/vllm/pull/2316#discussion_r1464204666
self.metrics.gauge_avg_prompt_throughput.labels(
**self.labels).set(prompt_throughput)
self.metrics.gauge_avg_generation_throughput.labels(
**self.labels).set(generation_throughput)
def log(self, stats: Stats) -> None:
"""Called by LLMEngine.
Logs to prometheus and tracked stats every iteration.
Logs to Stdout every self.local_interval seconds."""
# Log to prometheus.
self._log_prometheus(stats)
# Save tracked stats for token counters.
self.num_prompt_tokens.append(stats.num_prompt_tokens_iter)
self.num_generation_tokens.append(stats.num_generation_tokens_iter)
# Log locally every local_interval seconds.
if self._local_interval_elapsed(stats.now):
# Compute summary metrics for tracked stats (and log them
# to promethus if applicable).
prompt_throughput = self._get_throughput(self.num_prompt_tokens,
now=stats.now)
generation_throughput = self._get_throughput(
self.num_generation_tokens, now=stats.now)
self._log_prometheus_interval(
prompt_throughput=prompt_throughput,
generation_throughput=generation_throughput)
# Log to stdout.
logger.info(
"Avg prompt throughput: %.1f tokens/s, "
"Avg generation throughput: %.1f tokens/s, "
"Running: %d reqs, Swapped: %d reqs, "
"Pending: %d reqs, GPU KV cache usage: %.1f%%, "
"CPU KV cache usage: %.1f%%",
prompt_throughput,
generation_throughput,
stats.num_running_sys,
stats.num_swapped_sys,
stats.num_waiting_sys,
stats.gpu_cache_usage_sys * 100,
stats.cpu_cache_usage_sys * 100,
)
# Reset tracked stats for next interval.
self.num_prompt_tokens = []
self.num_generation_tokens = []
self.last_local_log = stats.now
if stats.spec_decode_metrics is not None:
logger.info(
self._format_spec_decode_metrics_str(
stats.spec_decode_metrics))
def _format_spec_decode_metrics_str(
self, metrics: "SpecDecodeWorkerMetrics") -> str:
return ("Speculative metrics: "
f"Draft acceptance rate: {metrics.draft_acceptance_rate:.3f}, "
f"System efficiency: {metrics.system_efficiency:.3f}, "
f"Number of speculative tokens: {metrics.num_spec_tokens}, "
f"Number of accepted tokens: {metrics.accepted_tokens}, "
f"Number of draft tokens tokens: {metrics.draft_tokens}, "
f"Number of emitted tokens tokens: {metrics.emitted_tokens}.")

View File

View File

@@ -0,0 +1,76 @@
from abc import ABC, abstractmethod
from typing import Callable, List
from transformers import PreTrainedTokenizer
from vllm.config import SchedulerConfig
from vllm.core.scheduler import Scheduler
from vllm.engine.output_processor.stop_checker import StopChecker
from vllm.sequence import Sequence, SequenceGroup, SequenceGroupOutput
from vllm.transformers_utils.detokenizer import Detokenizer
from vllm.utils import Counter
class SequenceGroupOutputProcessor(ABC):
"""Interface for logic that processes new token ids in sequence groups,
managing detokenization, stop checking, and freeing/forking sequences with
the scheduler.
This is highly coupled with the LLMEngine and should be seen as an extension
of it. The logic is separated to simplify the LLMEngine class and allow
separate implementations for single-step decoding (which supports beam
search sequence forking) and multi-step decoding (which does not support
beam search, but does support speculative decoding).
"""
@staticmethod
def create_output_processor(
scheduler_config: SchedulerConfig,
detokenizer: Detokenizer,
scheduler: Scheduler,
seq_counter: Counter,
get_tokenizer_for_seq: Callable[[Sequence], PreTrainedTokenizer],
stop_checker: "StopChecker",
):
"""Create an output processor.
This returns a single-step output processor if num_lookahead_slots is
zero, else returns a multi-step output processor.
"""
if scheduler_config.num_lookahead_slots == 0:
# Importing here to avoid cycle.
from vllm.engine.output_processor.single_step import (
SingleStepOutputProcessor)
return SingleStepOutputProcessor(
scheduler_config,
detokenizer,
scheduler,
seq_counter,
stop_checker,
)
else:
# Importing here to avoid cycle.
from vllm.engine.output_processor.multi_step import (
MultiStepOutputProcessor)
return MultiStepOutputProcessor(
detokenizer,
scheduler,
seq_counter,
get_tokenizer_for_seq,
stop_checker,
)
@abstractmethod
def process_outputs(self, sequence_group: SequenceGroup,
outputs: List[SequenceGroupOutput]) -> None:
"""Process new token ids for the sequence group. Handles logic such as
detokenization, stop checking, and freeing/forking sequences in the
scheduler.
"""
pass
@abstractmethod
def process_prompt_logprob(self, seq_group: SequenceGroup,
outputs: List[SequenceGroupOutput]) -> None:
"""Update prompt logprobs received from outputs to seq_group."""
pass

View File

@@ -0,0 +1,142 @@
import functools
from typing import Callable, List
from transformers import PreTrainedTokenizer
from vllm.core.scheduler import Scheduler
from vllm.engine.output_processor.interfaces import (
SequenceGroupOutputProcessor)
from vllm.engine.output_processor.stop_checker import StopChecker
from vllm.logger import init_logger
from vllm.sampling_params import SamplingParams
from vllm.sequence import (Sequence, SequenceGroup, SequenceGroupOutput,
SequenceOutput, SequenceStatus)
from vllm.transformers_utils.detokenizer import Detokenizer
from vllm.utils import Counter
logger = init_logger(__name__)
class MultiStepOutputProcessor(SequenceGroupOutputProcessor):
"""SequenceGroupOutputProcessor which handles logic related to
detokenization and stopping conditions. It specializes to "multi-step
decoding", where vLLM's worker may generate multiple tokens per invocation.
This is currently mutually exclusive with advanced sampling techniques like
beam search, which motivates the separation of this logic from the single
step output processor.
This class is responsible for things such as correctly appending all new
token ids to their sequence, detokenizing new token ids, truncating new
output tokens after an eos token, and correctly handling the case where the
number of new output tokens per sequence differs in a single batch.
"""
def __init__(
self,
detokenizer: Detokenizer,
scheduler: Scheduler,
seq_counter: Counter,
get_tokenizer_for_seq: Callable[[Sequence], PreTrainedTokenizer],
stop_checker: StopChecker,
):
self.detokenizer = detokenizer
self.scheduler = scheduler
self.seq_counter = seq_counter
self.get_tokenizer_for_seq = get_tokenizer_for_seq
self.stop_checker = stop_checker
def process_prompt_logprob(self, seq_group: SequenceGroup,
outputs: List[SequenceGroupOutput]) -> None:
# TODO(sang): Prompt logprob currently not implemented in multi step
# workers.
self._log_prompt_logprob_unsupported_warning_once()
@staticmethod
@functools.lru_cache()
def _log_prompt_logprob_unsupported_warning_once():
logger.warning(
"Prompt logprob is not supported by multi step workers. "
"(e.g., speculative decode uses multi step workers).")
def process_outputs(self, sequence_group: SequenceGroup,
outputs: List[SequenceGroupOutput]) -> None:
"""Append new tokens in the outputs to sequences in the sequence group.
This only supports sequence groups of size 1. It supports greater than
one new token per sequence.
This applies logic like stop condition checking and detokenization,
including freeing finished sequences. It also handles cases where there
are tokens emitted after the EOS token.
"""
seqs = sequence_group.get_seqs(status=SequenceStatus.RUNNING)
assert seqs, "expected running sequences"
assert len(seqs) == 1, (
"Beam search not supported in multi-step decoding.")
seq = seqs[0]
# Since there's only one sequence per sequence group, we can take the
# first sample.
samples = [outputs[step].samples[0] for step in range(len(outputs))]
# -1 means the output token is not valid (eg. due to spec decode
# rejecting tokens).
valid_samples = [
sample for sample in samples if sample.output_token != -1
]
assert valid_samples
self._process_seq_outputs(seq, valid_samples,
sequence_group.sampling_params)
def _process_seq_outputs(self, seq: Sequence,
valid_samples: List[SequenceOutput],
sampling_params: SamplingParams) -> None:
output_token_ids = [sample.output_token for sample in valid_samples]
output_logprobs = [sample.logprobs for sample in valid_samples]
# Truncate to max_tokens if necessary.
remaining_tokens = sampling_params.max_tokens - (seq.get_output_len() +
len(output_token_ids))
if remaining_tokens < 0:
valid_samples = valid_samples[:remaining_tokens]
output_token_ids = output_token_ids[:remaining_tokens]
# Truncate any tokens after EOS. This is required as spec decode
# generates a fixed number of tokens without evaluating stopping
# conditions within the block. This can cause an eos token to be
# unintentionally ignored.
if not sampling_params.ignore_eos:
eos_token_id = self.get_tokenizer_for_seq(seq).eos_token_id
# Avoiding .index calls as exception throwing in the happy path
# is expensive.
for i in range(len(output_token_ids)):
if output_token_ids[i] == eos_token_id:
output_token_ids = output_token_ids[:i + 1]
valid_samples = valid_samples[:i + 1]
break
# Incrementally append tokens to the sequence, as if we had only one new
# token.
for output_token_id, output_logprob in zip(output_token_ids,
output_logprobs):
seq.append_token_id(
token_id=output_token_id,
logprobs=output_logprob,
)
new_char_count = 0
if sampling_params.detokenize:
new_char_count = self.detokenizer.decode_sequence_inplace(
seq, sampling_params)
self.stop_checker.maybe_stop_sequence(
seq,
new_char_count=new_char_count,
sampling_params=sampling_params)
if seq.is_finished():
break
if seq.is_finished():
self.scheduler.free_seq(seq)

View File

@@ -0,0 +1,284 @@
from typing import Dict, List, Tuple, Union
from vllm.config import SchedulerConfig
from vllm.core.scheduler import Scheduler
from vllm.engine.output_processor.interfaces import (
SequenceGroupOutputProcessor)
from vllm.engine.output_processor.stop_checker import StopChecker
from vllm.logger import init_logger
from vllm.sampling_params import SamplingParams
from vllm.sequence import (Sequence, SequenceGroup, SequenceGroupOutput,
SequenceOutput, SequenceStatus)
from vllm.transformers_utils.detokenizer import Detokenizer
from vllm.utils import Counter
logger = init_logger(__name__)
class SingleStepOutputProcessor(SequenceGroupOutputProcessor):
"""SequenceGroupOutputProcessor which handles "output processing" logic,
which happens after the model returns generated token ids and before
scheduling of the next batch. Output processing logic includes
detokenization, and determining if a sequence is finished (e.g. via max len
or eos token).
The SingleStepOutputProcessor is specialized to the case where the model
emits at most a single token per invocation, which precludes configurations
such as speculative decoding or multi-step decoding. This enables beam
search sampling, which requires forking/finishing/freeing sequences in a way
that is currently difficult to schedule multiple steps ahead of time.
"""
def __init__(
self,
scheduler_config: SchedulerConfig,
detokenizer: Detokenizer,
scheduler: Scheduler,
seq_counter: Counter,
stop_checker: StopChecker,
):
self.scheduler_config = scheduler_config
self.detokenizer = detokenizer
self.scheduler = scheduler
self.seq_counter = seq_counter
self.stop_checker = stop_checker
def process_outputs(self, sequence_group: SequenceGroup,
outputs: List[SequenceGroupOutput]) -> None:
"""Append all new tokens to sequences in the sequence group. Fork any
surviving beam candidates; free any unsurviving ones.
Invokes detokenizer to detokenize new tokens, and also marks sequences
as finished if they meet stop conditions.
"""
assert (len(outputs) == 1
), f"{type(self)} does not support multiple outputs per step"
return self._process_sequence_group_outputs(sequence_group, outputs[0])
def process_prompt_logprob(self, seq_group: SequenceGroup,
outputs: List[SequenceGroupOutput]) -> None:
assert len(outputs) == 1, ("Single step should only has 1 output.")
output = outputs[0]
prompt_logprobs = output.prompt_logprobs
if (prompt_logprobs is not None
and seq_group.sampling_params.detokenize and self.detokenizer):
self.detokenizer.decode_prompt_logprobs_inplace(
seq_group, prompt_logprobs)
if not seq_group.prompt_logprobs:
# The first prompt token's logprob is None because it doesn't
# have tokens that are precedent.
seq_group.prompt_logprobs = [None]
seq_group.prompt_logprobs.extend(prompt_logprobs)
def _process_sequence_group_outputs(self, seq_group: SequenceGroup,
outputs: SequenceGroupOutput) -> None:
# Process samples
samples = outputs.samples
parent_seqs = seq_group.get_seqs(status=SequenceStatus.RUNNING)
existing_finished_seqs = seq_group.get_finished_seqs()
parent_child_dict: Dict[int, List[SequenceOutput]] = {
parent_seq.seq_id: []
for parent_seq in parent_seqs
}
for sample in samples:
parent_child_dict[sample.parent_seq_id].append(sample)
# List of (child, parent)
child_seqs: List[Tuple[Sequence, Sequence]] = []
# Process the child samples for each parent sequence
for parent in parent_seqs:
child_samples: List[SequenceOutput] = parent_child_dict[
parent.seq_id]
if len(child_samples) == 0:
# This parent sequence has no children samples. Remove
# the parent sequence from the sequence group since it will
# not be used in the future iterations.
parent.status = SequenceStatus.FINISHED_ABORTED
seq_group.remove(parent.seq_id)
self.scheduler.free_seq(parent)
continue
# Fork the parent sequence if there are multiple child samples.
for child_sample in child_samples[:-1]:
new_child_seq_id: int = next(self.seq_counter)
child = parent.fork(new_child_seq_id)
child.append_token_id(child_sample.output_token,
child_sample.logprobs)
child_seqs.append((child, parent))
# Continue the parent sequence for the last child sample.
# We reuse the parent sequence here to reduce redundant memory
# copies, especially when using non-beam search sampling methods.
last_child_sample = child_samples[-1]
parent.append_token_id(last_child_sample.output_token,
last_child_sample.logprobs)
child_seqs.append((parent, parent))
for seq, _ in child_seqs:
if seq_group.sampling_params.detokenize and self.detokenizer:
new_char_count = self.detokenizer.decode_sequence_inplace(
seq, seq_group.sampling_params)
else:
new_char_count = 0
self.stop_checker.maybe_stop_sequence(seq, new_char_count,
seq_group.sampling_params)
# Non-beam search case
if not seq_group.sampling_params.use_beam_search:
# For newly created child sequences, add them to the sequence group
# and fork them in block manager if they are not finished.
for seq, parent in child_seqs:
if seq is not parent:
seq_group.add(seq)
if not seq.is_finished():
self.scheduler.fork_seq(parent, seq)
# Free the finished and selected parent sequences' memory in block
# manager. Keep them in the sequence group as candidate output.
# NOTE: we need to fork the new sequences before freeing the
# old sequences.
for seq, parent in child_seqs:
if seq is parent and seq.is_finished():
self.scheduler.free_seq(seq)
return
# Beam search case
# Select the child sequences to keep in the sequence group.
selected_child_seqs = []
unselected_child_seqs = []
beam_width = seq_group.sampling_params.best_of
length_penalty = seq_group.sampling_params.length_penalty
# Select the newly finished sequences with the highest scores
# to replace existing finished sequences.
# Tuple of (seq, parent, is_new)
existing_finished_seqs = [(seq, None, False)
for seq in existing_finished_seqs]
new_finished_seqs = [(seq, parent, True) for seq, parent in child_seqs
if seq.is_finished()]
all_finished_seqs = existing_finished_seqs + new_finished_seqs
# Sort the finished sequences by their scores.
all_finished_seqs.sort(key=lambda x: x[0].get_beam_search_score(
length_penalty=length_penalty, eos_token_id=x[0].eos_token_id),
reverse=True)
for seq, parent, is_new in all_finished_seqs[:beam_width]:
if is_new:
# A newly generated child sequence finishes and has a high
# score, so we will add it into the sequence group.
selected_child_seqs.append((seq, parent))
for seq, parent, is_new in all_finished_seqs[beam_width:]:
if is_new:
# A newly generated child sequence finishes but has a low
# score, so we will not add it into the sequence group.
# Additionally, if this sequence is a continuation of a
# parent sequence, we will need remove the parent sequence
# from the sequence group.
unselected_child_seqs.append((seq, parent))
else:
# An existing finished sequence has a low score, so we will
# remove it from the sequence group.
seq_group.remove(seq.seq_id)
# select the top beam_width sequences from the running
# sequences for the next iteration to continue the beam
# search.
running_child_seqs = [(seq, parent) for seq, parent in child_seqs
if not seq.is_finished()]
# Sort the running sequences by their scores.
running_child_seqs.sort(key=lambda x: x[0].get_beam_search_score(
length_penalty=length_penalty, eos_token_id=x[0].eos_token_id),
reverse=True)
# Check if we can stop the beam search.
if len(running_child_seqs) == 0:
# No running sequences, stop the beam search.
stop_beam_search = True
elif len(all_finished_seqs) < beam_width:
# Not enough finished sequences, continue the beam search.
stop_beam_search = False
else:
# Check the early stopping criteria
best_running_seq = running_child_seqs[0][0]
current_worst_seq = all_finished_seqs[beam_width - 1][0]
stop_beam_search = self._check_beam_search_early_stopping(
seq_group.sampling_params.early_stopping,
seq_group.sampling_params, best_running_seq, current_worst_seq)
if stop_beam_search:
# Stop the beam search and remove all the running sequences from
# the sequence group.
unselected_child_seqs.extend(running_child_seqs)
else:
# Continue the beam search and select the top beam_width sequences
# to continue the beam search.
selected_child_seqs.extend(running_child_seqs[:beam_width])
# The remaining running sequences will not be used in the next
# iteration. Again, if these sequences are continuations of
# parent sequences, we will need to remove the parent sequences
# from the sequence group.
unselected_child_seqs.extend(running_child_seqs[beam_width:])
# For newly created child sequences, add them to the sequence group
# and fork them in block manager if they are not finished.
for seq, parent in selected_child_seqs:
if seq is not parent:
seq_group.add(seq)
if not seq.is_finished():
self.scheduler.fork_seq(parent, seq)
# Free the finished and selected parent sequences' memory in block
# manager. Keep them in the sequence group as candidate output.
for seq, parent in selected_child_seqs:
if seq is parent and seq.is_finished():
self.scheduler.free_seq(seq)
# Remove the unselected parent sequences from the sequence group and
# free their memory in block manager.
for seq, parent in unselected_child_seqs:
if seq is parent:
# Remove the parent sequence if it is not selected for next
# iteration
seq_group.remove(seq.seq_id)
self.scheduler.free_seq(seq)
def _check_beam_search_early_stopping(
self,
early_stopping: Union[bool, str],
sampling_params: SamplingParams,
best_running_seq: Sequence,
current_worst_seq: Sequence,
) -> bool:
assert sampling_params.use_beam_search
length_penalty = sampling_params.length_penalty
if early_stopping is True:
return True
current_worst_score = current_worst_seq.get_beam_search_score(
length_penalty=length_penalty,
eos_token_id=current_worst_seq.eos_token_id)
if early_stopping is False:
highest_attainable_score = best_running_seq.get_beam_search_score(
length_penalty=length_penalty,
eos_token_id=best_running_seq.eos_token_id)
else:
assert early_stopping == "never"
if length_penalty > 0.0:
# If length_penalty > 0.0, beam search will prefer longer
# sequences. The highest attainable score calculation is
# based on the longest possible sequence length in this case.
max_possible_length = max(
best_running_seq.get_prompt_len() +
sampling_params.max_tokens,
self.scheduler_config.max_model_len)
highest_attainable_score = (
best_running_seq.get_beam_search_score(
length_penalty=length_penalty,
eos_token_id=best_running_seq.eos_token_id,
seq_len=max_possible_length))
else:
# Otherwise, beam search will prefer shorter sequences. The
# highest attainable score calculation is based on the current
# sequence length.
highest_attainable_score = (
best_running_seq.get_beam_search_score(
length_penalty=length_penalty,
eos_token_id=best_running_seq.eos_token_id))
return current_worst_score >= highest_attainable_score

View File

@@ -0,0 +1,101 @@
from typing import Callable, Optional
from transformers import PreTrainedTokenizer
from vllm.sampling_params import SamplingParams
from vllm.sequence import Sequence, SequenceStatus
class StopChecker:
"""LLMEngine helper class which separates out the logic involving stop
checking. This checks things such as: whether the eos token was emitted,
whether the max_tokens has been consumed, whether a stop string has been
emitted, or if we have exceeded the max model len.
"""
def __init__(self, max_model_len: int,
get_tokenizer_for_seq: Callable[[Sequence],
PreTrainedTokenizer]):
self.max_model_len = max_model_len
self.get_tokenizer_for_seq = get_tokenizer_for_seq
def maybe_stop_sequence(self, seq: Sequence, new_char_count: int,
sampling_params: SamplingParams) -> None:
"""Stop the finished sequences.
new_char_count is the number of chars added to the
sequence's output text for the newly generated token
"""
# Check if the minimum number of tokens has been generated yet;
# skip the stop string/token checks if not
if seq.get_output_len() < sampling_params.min_tokens:
return
# Check if the sequence has generated the EOS token.
if ((not sampling_params.ignore_eos)
and seq.get_last_token_id() == seq.eos_token_id):
seq.status = SequenceStatus.FINISHED_STOPPED
return
# Check if a stop token was encountered.
# This assumes a single token produced per step.
last_token_id = seq.get_last_token_id()
if last_token_id in sampling_params.stop_token_ids:
if new_char_count and (
not sampling_params.include_stop_str_in_output):
# Remove last token
seq.output_text = seq.output_text[:-new_char_count]
seq.status = SequenceStatus.FINISHED_STOPPED
seq.stop_reason = last_token_id
return
# Check if any stop strings are matched.
stop_str = self._check_stop_strings(seq, new_char_count,
sampling_params)
if stop_str is not None:
seq.status = SequenceStatus.FINISHED_STOPPED
seq.stop_reason = stop_str
return
# Check if the sequence has reached max_model_len.
if seq.get_len() > self.max_model_len:
seq.status = SequenceStatus.FINISHED_LENGTH_CAPPED
return
# Check if the sequence has reached max_tokens.
if seq.get_output_len() == sampling_params.max_tokens:
seq.status = SequenceStatus.FINISHED_LENGTH_CAPPED
return
@staticmethod
def _check_stop_strings(seq: Sequence, new_char_count: int,
sampling_params: SamplingParams) -> Optional[str]:
"""Check if any stop strings are matched and truncate sequence
output text accordingly.
Returns the stop string if matched or else None.
"""
if not new_char_count:
return None
for stop_str in sampling_params.stop:
stop_string_len = len(stop_str)
# Avoid searching already-searched text.
stop_index = seq.output_text.find(
stop_str, -new_char_count - stop_string_len)
if stop_index == -1:
continue
if sampling_params.include_stop_str_in_output:
# Truncate to end of stop string.
stop_index += stop_string_len
if stop_index >= len(seq.output_text):
# No truncation required.
return stop_str
# Truncate the output text to either the beginning
# or end of the stop string.
seq.output_text = seq.output_text[:stop_index]
return stop_str
return None

View File

@@ -0,0 +1,19 @@
from typing import List
from vllm.sequence import SamplerOutput, SequenceGroupOutput
def create_output_by_sequence_group(
sampler_outputs: List[SamplerOutput],
num_seq_groups: int) -> List[List[SequenceGroupOutput]]:
"""Helper method which transforms a 2d list organized by
[step][sequence group] into [sequence group][step].
"""
output_by_sequence_group: List[List[SamplerOutput]] = [
[] for _ in range(num_seq_groups)
]
for step in sampler_outputs:
for i, sequence_group_output in enumerate(step):
output_by_sequence_group[i].append(sequence_group_output)
return output_by_sequence_group