141 lines
4.5 KiB
Python
141 lines
4.5 KiB
Python
# Unsloth should be imported before transformers to ensure all optimizations are applied.
|
|
from unsloth import FastLanguageModel
|
|
|
|
from pathlib import Path
|
|
from threading import Thread
|
|
from transformers import AutoTokenizer, TextIteratorStreamer, pipeline
|
|
from typing import Callable, Iterator, Optional
|
|
from xml_schema_validator import XmlLogitsProcessor
|
|
import os
|
|
import torch
|
|
|
|
from . import LlmEngine
|
|
from .. import util
|
|
|
|
class QwQLlmEngine(LlmEngine):
|
|
|
|
def __init__(
|
|
self,
|
|
model_path: Path,
|
|
temperature: float,
|
|
xml_schema_text: Optional[str] = None,
|
|
):
|
|
"""
|
|
Initialize the QwQ LLM Engine.
|
|
|
|
Args:
|
|
model_path: Local path to the model
|
|
temperature: Sampling temperature
|
|
xml_schema_text: Optional XML schema to validate against
|
|
"""
|
|
self._temperature = temperature
|
|
|
|
# Load tokenizer
|
|
self._tokenizer = AutoTokenizer.from_pretrained(
|
|
model_path,
|
|
)
|
|
|
|
# Load model
|
|
self._model, _returned_tokenizer = FastLanguageModel.from_pretrained(
|
|
model_path,
|
|
load_in_4bit = True, # False for LoRA 16bit
|
|
fast_inference = True, # Enable vLLM fast inference
|
|
gpu_memory_utilization = 0.8, # Reduce if out of memory
|
|
tokenizer = self._tokenizer,
|
|
)
|
|
|
|
# enable unsloth optimizations
|
|
FastLanguageModel.for_inference(self._model)
|
|
|
|
# Create inference pipeline
|
|
self._pipeline = pipeline(
|
|
"text-generation",
|
|
model=self._model,
|
|
tokenizer=self._tokenizer,
|
|
return_full_text=False,
|
|
)
|
|
|
|
if xml_schema_text:
|
|
self._logits_processor = XmlLogitsProcessor(self._tokenizer, xml_schema_text)
|
|
else:
|
|
self._logits_processor = None
|
|
|
|
|
|
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},
|
|
]
|
|
|
|
text = self._tokenizer.apply_chat_template(
|
|
messages,
|
|
tokenize=False,
|
|
add_generation_prompt=False,
|
|
)
|
|
|
|
streamer = TextIteratorStreamer(
|
|
self._tokenizer,
|
|
skip_prompt=True,
|
|
)
|
|
|
|
generation_kwargs = {
|
|
"text_inputs": text,
|
|
"do_sample": True,
|
|
"temperature": self._temperature,
|
|
"streamer": streamer,
|
|
"use_cache": True,
|
|
}
|
|
|
|
if self._logits_processor:
|
|
generation_kwargs["logits_processor"] = [self._logits_processor.copy()]
|
|
|
|
generation_thread = Thread(
|
|
target=self._pipeline,
|
|
kwargs=generation_kwargs
|
|
)
|
|
|
|
generation_thread.start()
|
|
|
|
for text in util.stop_before_value(streamer, self._tokenizer.eos_token):
|
|
yield text
|
|
if should_stop():
|
|
break
|
|
|
|
generation_thread.join()
|
|
|
|
def token_count(self, system_prompt: str, main_context: str) -> int:
|
|
"""
|
|
Count tokens for the given system prompt and main context.
|
|
|
|
Args:
|
|
system_prompt: The system prompt string
|
|
main_context: The main context string
|
|
|
|
Returns:
|
|
int: Total number of tokens
|
|
"""
|
|
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:
|
|
return self._pipeline.model.config.max_position_embeddings
|