[Misc] Various simplifications and typing fixes (#5368)
This commit is contained in:
@@ -3,6 +3,7 @@ from typing import Any, Dict, List, Optional, Tuple
|
||||
|
||||
import torch
|
||||
|
||||
from vllm.config import SpeculativeConfig
|
||||
from vllm.distributed.communication_op import broadcast_tensor_dict
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.rejection_sampler import RejectionSampler
|
||||
@@ -30,7 +31,7 @@ def create_spec_worker(*args, **kwargs) -> "SpecDecodeWorker":
|
||||
WorkerWrapper. It constructs a SpecDecodeWorker from the speculative config.
|
||||
"""
|
||||
assert "speculative_config" in kwargs
|
||||
speculative_config = kwargs.get("speculative_config")
|
||||
speculative_config: SpeculativeConfig = kwargs.get("speculative_config")
|
||||
assert speculative_config is not None
|
||||
|
||||
target_worker = Worker(*args, **kwargs)
|
||||
@@ -109,12 +110,11 @@ class SpecDecodeWorker(LoraNotSupportedWorkerBase):
|
||||
logger.info("Configuring SpecDecodeWorker with proposer=%s",
|
||||
type(proposer_worker))
|
||||
|
||||
return SpecDecodeWorker(
|
||||
proposer_worker,
|
||||
scorer_worker,
|
||||
disable_by_batch_size=disable_by_batch_size,
|
||||
rejection_sampler=RejectionSampler(
|
||||
disable_bonus_tokens=disable_bonus_tokens, ))
|
||||
return SpecDecodeWorker(proposer_worker,
|
||||
scorer_worker,
|
||||
disable_by_batch_size=disable_by_batch_size,
|
||||
rejection_sampler=RejectionSampler(
|
||||
disable_bonus_tokens=disable_bonus_tokens))
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user