README.md

metadata

tags:
  - int8
  - vllm
  - llm-compressor
language:
  - en
pipeline_tag: text-generation
license: apache-2.0
base_model:
  - Qwen/Qwen2.5-32B

Qwen2.5-32B-quantized.w8a16

Model Overview

• Model Architecture: Qwen2
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Weight quantization: INT8
• Intended Use Cases: Similarly to Qwen2.5-32B, this is a base language model.
• Out-of-scope: Use in any manner that violates applicable laws or regulations (including trade compliance laws).
• Release Date: 10/09/2024
• Version: 1.0
• Model Developers: Neural Magic

Quantized version of Qwen2.5-32B. It achieves an OpenLLMv1 score of 75.4, compared to 75.3 for Qwen2.5-32B.

Model Optimizations

This model was obtained by quantizing the weights of Qwen2.5-32B to INT8 data type. 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 of the linear operators within transformers blocks are quantized. Symmetric per-channel quantization is applied, in which a linear scaling per output dimension maps the INT8 and floating point representations of the quantized weights. The GPTQ algorithm is applied for quantization, as implemented in the llm-compressor library.

Deployment

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/Qwen2.5-32B-quantized.w8a16"
number_gpus = 1
max_model_len = 8192

sampling_params = SamplingParams(temperature=0.7, top_p=0.8, max_tokens=256)

tokenizer = AutoTokenizer.from_pretrained(model_id)

prompt = "Give me a short introduction to large language model."

llm = LLM(model=model_id, tensor_parallel_size=number_gpus, max_model_len=max_model_len)

outputs = llm.generate(prompt, sampling_params)

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

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

Evaluation

The model was evaluated on the OpenLLMv1 benchmark, composed of MMLU, ARC-Challenge, GSM-8K, Hellaswag, Winogrande and TruthfulQA. Evaluation was conducted using lm-evaluation-harness and the vLLM engine.

Accuracy

Category	Benchmark	Qwen2.5-32B	Qwen2.5-32B-quantized.w8a16 (this model)	Recovery
OpenLLM v1
	MMLU (5-shot)	83.25	83.19	99.9%
	ARC Challenge (25-shot)	66.30	66.04	99.6%
	GSM-8k (5-shot, strict-match)	78.09	78.62	100.7%
	Hellaswag (10-shot)	85.08	85.14	100.1%
	Winogrande (5-shot)	81.29	81.61	100.4%
	TruthfulQA (0-shot, mc2)	57.76	57.78	100.0%
	Average	75.30	75.40	100.1%

Reproduction

The results were obtained using the following command:

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic/Qwen2.5-32B-quantized.w8a16",dtype=auto,max_model_len=4096,add_bos_token=True,tensor_parallel_size=1 \
  --tasks openllm \
  --batch_size auto