88 lines
3.0 KiB
Python
88 lines
3.0 KiB
Python
from huggingface_hub import InferenceClient
|
|
from transformers import AutoTokenizer, AutoConfig
|
|
from typing import Iterator, Optional, Callable
|
|
|
|
from . import LlmEngine
|
|
|
|
class HfLlmEngine(LlmEngine):
|
|
"""
|
|
LLM Engine implementation using HuggingFace's InferenceClient.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
model: str,
|
|
temperature: float,
|
|
api_token: Optional[str],
|
|
):
|
|
"""
|
|
Initialize the HuggingFace Inference API LLM Engine.
|
|
|
|
Args:
|
|
model: HuggingFace model ID to use
|
|
temperature: Sampling temperature
|
|
api_token: HuggingFace API token
|
|
"""
|
|
self._model = model
|
|
self._temperature = temperature
|
|
|
|
self._tokenizer = AutoTokenizer.from_pretrained(model, token=api_token)
|
|
self._config = AutoConfig.from_pretrained(model, token=api_token)
|
|
self._client = InferenceClient(token=api_token)
|
|
|
|
def infer(self, system_prompt: str, main_context: str, continuation_text: str, should_stop: Callable[[], bool] = lambda: False) -> Iterator[str]:
|
|
"""
|
|
Run inference using the system prompt and main context.
|
|
|
|
Args:
|
|
system_prompt: The system prompt string
|
|
main_context: The main context string after templating
|
|
continuation_text: Part of the response that is already generated
|
|
should_stop: Callback that returns True when inference should stop
|
|
|
|
Returns:
|
|
Iterator[str]: An iterator that yields the generated text.
|
|
"""
|
|
messages = [
|
|
{"role": "system", "content": system_prompt},
|
|
{"role": "user", "content": main_context},
|
|
{"role": "assistant", "content": continuation_text},
|
|
]
|
|
|
|
stream = self._client.chat_completion(
|
|
model=self._model,
|
|
messages=messages,
|
|
temperature=self._temperature,
|
|
add_generation_prompt=False,
|
|
stream=True
|
|
)
|
|
|
|
try:
|
|
for response in stream:
|
|
if should_stop():
|
|
stream.close()
|
|
break
|
|
if content := response.choices[0].delta.content:
|
|
yield content
|
|
finally:
|
|
stream.close()
|
|
|
|
def token_count(self, system_prompt: str, main_context: str) -> int:
|
|
messages = [
|
|
{"role": "system", "content": system_prompt},
|
|
{"role": "user", "content": main_context}
|
|
]
|
|
prompt = self._tokenizer.apply_chat_template(
|
|
messages, tokenize=False, add_generation_prompt=True
|
|
)
|
|
return len(self._tokenizer.encode(prompt))
|
|
|
|
def token_limit(self) -> int:
|
|
"""
|
|
Get the model's context window size.
|
|
|
|
Returns:
|
|
int: Maximum number of tokens the model can process
|
|
"""
|
|
return self._config.max_position_embeddings
|