Meta-Llama-3-70B-Instruct-FP8-KV

Model Overview

Meta-Llama-3-70B-Instruct quantized to FP8 weights and activations using per-tensor quantization, ready for inference with vLLM >= 0.5.0. This model checkpoint also includes per-tensor scales for FP8 quantized KV Cache, accessed through the --kv-cache-dtype fp8 argument in vLLM.

from vllm import LLM
model = LLM(model="neuralmagic/Meta-Llama-3-70B-Instruct-FP8-KV", kv_cache_dtype="fp8")
result = model.generate("Hello, my name is")

Usage and Creation

Produced using AutoFP8 with calibration samples from ultrachat.

from datasets import load_dataset
from transformers import AutoTokenizer

from auto_fp8 import AutoFP8ForCausalLM, BaseQuantizeConfig

pretrained_model_dir = "meta-llama/Meta-Llama-3-70B-Instruct"
quantized_model_dir = "Meta-Llama-3-70B-Instruct-FP8-KV"

tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=True)
tokenizer.pad_token = tokenizer.eos_token

ds = load_dataset("mgoin/ultrachat_2k", split="train_sft")
examples = [tokenizer.apply_chat_template(batch["messages"], tokenize=False) for batch in ds]
examples = tokenizer(examples, padding=True, truncation=True, return_tensors="pt").to("cuda")

quantize_config = BaseQuantizeConfig(
    quant_method="fp8",
    activation_scheme="static",
    ignore_patterns=["re:.*lm_head"],
    kv_cache_quant_targets=("k_proj", "v_proj"),
)

model = AutoFP8ForCausalLM.from_pretrained(pretrained_model_dir, quantize_config)
model.quantize(examples)
model.save_quantized(quantized_model_dir)

Evaluation

Open LLM Leaderboard evaluation scores

Model evaluation results obtained via lm-evaluation-harness.

Benchmark	Meta-Llama-3-70B-Instruct	Meta-Llama-3-70B-Instruct-FP8	Meta-Llama-3-70B-Instruct-FP8-KV (this model)
ARC-c 25-shot	72.69	72.61	72.57
HellaSwag 10-shot	85.50	85.41	85.32
MMLU 5-shot	80.18	80.06	79.69
TruthfulQA 0-shot	62.90	62.73	61.92
WinoGrande 5-shot	83.34	83.03	83.66
GSM8K 5-shot	92.49	91.12	90.83
Average Accuracy	79.51	79.16	79.00
Recovery	100%	99.55%	99.36%