mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-02-07 19:51:50 +08:00
34 lines
1.1 KiB
Python
34 lines
1.1 KiB
Python
from abc import ABC, abstractmethod
|
|
from typing import List, Optional
|
|
|
|
from ..pyexecutor.llm_request import LlmRequest
|
|
from ..pyexecutor.resource_manager import ResourceManager
|
|
from ..pyexecutor.scheduler import ScheduledRequests
|
|
|
|
|
|
class Drafter(ABC):
|
|
"""Abstract base class for all drafter implementations."""
|
|
|
|
def __init__(self, max_concurrency: Optional[int] = None) -> None:
|
|
self.max_concurrency = max_concurrency
|
|
|
|
@abstractmethod
|
|
def prepare_draft_tokens(
|
|
self,
|
|
scheduled_requests: ScheduledRequests,
|
|
resource_manager: Optional[ResourceManager] = None,
|
|
) -> None:
|
|
"""
|
|
Prepare the drafter tokens for the forward computation this step.
|
|
|
|
Args:
|
|
scheduled_requests: The scheduled requests for this iteration
|
|
"""
|
|
raise NotImplementedError
|
|
|
|
def should_use_spec_decode(self, requests: List[LlmRequest]) -> bool:
|
|
"""Check if spec decode should be used for the current iteration."""
|
|
if self.max_concurrency is not None:
|
|
return len(requests) <= self.max_concurrency
|
|
return True
|