Qwen2-72B-Instruct-FP8

Model Overview

• Model Architecture: Qwen2
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: FP8
  • Activation quantization: FP8
• Intended Use Cases: Intended for commercial and research use in English. Similarly to Meta-Llama-3-8B-Instruct, this models is intended for assistant-like chat.
• Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws). Use in languages other than English.
• Release Date: 6/8/2024
• Version: 1.0
• License(s): tongyi-qianwen
• Model Developers: Neural Magic

Quantized version of Qwen2-72B-Instruct. It achieves an average score of 80.34 on the OpenLLM benchmark (version 1), whereas the unquantized model achieves 79.97.

Model Optimizations

This model was obtained by quantizing the weights and activations of Qwen2-72B-Instruct to FP8 data type, ready for inference with vLLM >= 0.5.0. This optimization reduces the number of bits per parameter from 16 to 8, reducing the disk size and GPU memory requirements by approximately 50%.

Only the weights and activations of the linear operators within transformers blocks are quantized. Symmetric per-tensor quantization is applied, in which a single linear scaling maps the FP8 representations of the quantized weights and activations. AutoFP8 is used for quantization with 512 sequences of UltraChat.

Deployment

Use with vLLM

This model can be deployed efficiently using the vLLM backend, as shown in the example below.

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "neuralmagic/Qwen2-72B-Instruct-FP8"

sampling_params = SamplingParams(temperature=0.6, top_p=0.9, max_tokens=256)

tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

prompts = tokenizer.apply_chat_template(messages, tokenize=False)

llm = LLM(model=model_id)

outputs = llm.generate(prompts, sampling_params)

generated_text = outputs[0].outputs[0].text
print(generated_text)

vLLM aslo supports OpenAI-compatible serving. See the documentation for more details.

Creation

This model was created by applying AutoFP8 with calibration samples from ultrachat, as presented in the code snipet below. Although AutoFP8 was used for this particular model, Neural Magic is transitioning to using llm-compressor which supports several quantization schemes and models not supported by AutoFP8.

from datasets import load_dataset
from transformers import AutoTokenizer

from auto_fp8 import AutoFP8ForCausalLM, BaseQuantizeConfig

pretrained_model_dir = "Qwen/Qwen2-72B-Instruct"
quantized_model_dir = "Qwen2-72B-Instruct-FP8"

tokenizer = AutoTokenizer.from_pretrained(pretrained_model_dir, use_fast=True, model_max_length=4096)
tokenizer.pad_token = tokenizer.eos_token

ds = load_dataset("mgoin/ultrachat_2k", split="train_sft").select(range(512))
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"],
)

model = AutoFP8ForCausalLM.from_pretrained(
    pretrained_model_dir, quantize_config=quantize_config
)

model.quantize(examples)
model.save_quantized(quantized_model_dir)

Evaluation

The model was evaluated on the OpenLLM leaderboard tasks (version 1) with the lm-evaluation-harness (commit 383bbd54bc621086e05aa1b030d8d4d5635b25e6) and the vLLM engine, using the following command:

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/Qwen2-72B-Instruct-FP8",dtype=auto,gpu_memory_utilization=0.4,add_bos_token=True,max_model_len=4096 \
  --tasks openllm \
  --batch_size auto

Accuracy

Open LLM Leaderboard evaluation scores

Benchmark	Qwen2-72B-Instruct	Qwen2-72B-Instruct-FP8(this model)	Recovery
MMLU (5-shot)	83.97	84.06	100.1%
ARC Challenge (25-shot)	71.58	72.09	100.7%
GSM-8K (5-shot, strict-match)	87.56	88.93	101.5%
Hellaswag (10-shot)	86.94	86.83	99.87%
Winogrande (5-shot)	82.79	83.18	100.4%
TruthfulQA (0-shot)	66.98	66.95	99.95%
Average	79.97	80.34	100.4%