Qwen2.5-14B-Instruct-quantized.w8a8

Model Overview

• Model Architecture: Qwen2
  • Input: Text
  • Output: Text
• Model Optimizations:
  • Activation quantization: INT8
  • Weight quantization: INT8
• Intended Use Cases: Intended for commercial and research use multiple languages. Similarly to Qwen2.5-14B-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).
• Release Date: 12/10/2024
• Version: 1.0
• Model Developers: Neural Magic

Quantized version of Qwen2.5-14B-Instruct. It achieves an average score of 78.15 on the OpenLLM benchmark version 1 and 35.60 on version 2, whereas the unquantized model achieves 77.85 on version 1 and 35.85 on version 2.

Model Optimizations

This model was obtained by quantizing the weights and activations of Qwen2.5-14B-Instruct to INT8 data type. This optimization reduces the number of bits used to represent weights and activations from 16 to 8, reducing GPU memory requirements (by approximately 50%) and increasing matrix-multiply compute throughput (by approximately 2x). Weight quantization also reduces disk size requirements by approximately 50%.

Only weights and activations of the linear operators within transformers blocks are quantized. Weights are quantized with a symmetric static per-channel scheme, where a fixed linear scaling factor is applied between INT8 and floating point representations for each output channel dimension. Activations are quantized with a symmetric dynamic per-token scheme, computing a linear scaling factor at runtime for each token between INT8 and floating point representations.

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-ent/Qwen2.5-14B-Instruct-quantized.w8a8"
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 OpenLLM leaderboard tasks (version 1) with the lm-evaluation-harness (commit 3814Bbd54bc621086e05aa1b030d8d4d5635b25e6) and the vLLM engine, using the following command:

lm_eval \
  --model vllm \
  --model_args pretrained="neuralmagic-ent/Qwen2.5-14B-Instruct-quantized.w8a8",dtype=auto,gpu_memory_utilization=0.9,add_bos_token=True,max_model_len=4096,enable_chunk_prefill=True,tensor_parallel_size=1 \
  --tasks openllm \
  --batch_size auto

Accuracy

Benchmark	Qwen2.5-14B-Instruct	Qwen2.5-14B-Instruct-quantized.w8a8 (this model)	Recovery
OpenLLM v1	MMLU (5-shot)	79.87	79.75	99.9%
	ARC Challenge (25-shot)	68.94	69.20	100.4%
	GSM-8K (5-shot, strict-match)	83.55	85.52	102.36%
	Hellaswag (10-shot)	85.30	85.07	99.7%
	Winogrande (5-shot)	80.51	80.43	99.9%
	TruthfulQA (0-shot, mc2)	68.93	68.94	100.0%
	Average	77.85	78.15	100.4%
OpenLLM v2	MMLU-Pro (5-shot)	48.32		%
	IFEval (0-shot)	81.15		%
	BBH (3-shot)	64.30		%
	Math-lvl-5 (4-shot)	0.00		***
	GPQA (0-shot)	36.74		%
	MuSR (0-shot)	40.20		%
	Average	45.12		%

*** Reference value too low to report meaningful recovery.