tayaee/Llama-3.1-Tulu-3-8B-AWQ-4bit
022
tayaee/Llama-3.1-Tulu-3-8B-AWQ-4bit
allenai/Llama-3.1-Tulu-3-8B를 AWQ 4bit(W4A16_ASYM, group 128, lm_head 제외)로 양자화한 모델이다.
- Tool:
llmcompressor(AWQModifier(duo_scaling="both")+QuantizationModifier(scheme="W4A16_ASYM")) - Calibration:
wikitext(wikitext-2-raw-v1, 8 samples, seq 512, 학습용 축소 설정) - 실전 권장: 256~512 samples, seq 2048~4096
사용법 (transformers)
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("tayaee/Llama-3.1-Tulu-3-8B-AWQ-4bit")
m = AutoModelForCausalLM.from_pretrained("tayaee/Llama-3.1-Tulu-3-8B-AWQ-4bit", torch_dtype="auto")사용법 (vLLM)
from vllm import LLM, SamplingParams
llm = LLM(model="tayaee/Llama-3.1-Tulu-3-8B-AWQ-4bit", quantization="compressed-tensors")
print(llm.generate(["Hello, my name is"], SamplingParams(max_tokens=60))[0].outputs[0].text)