INCModel4/Phi-4-multimodal-instruct-W4A16-AutoRound

Model Details

This model is an int4 (W4A16, group_size 128, symmetric) model of microsoft/Phi-4-multimodal-instruct generated by intel/auto-round with 200 tuning iterations. Please follow the license of the original model.

• Quantized (W4A16): text decoder linear projections
• Kept in BF16: LoRA adapters, token embeddings, lm_head, vision and audio components, normalization layers

How to Use

Environment

uv pip install vllm==0.33.0
uv pip install transformers==5.20.2

vLLM Usage

vllm serve INCModel4/Phi-4-multimodal-instruct-W4A16-AutoRound \
    --kv-cache-dtype auto \
    --dtype bfloat16 \
    --trust-remote-code

Evaluation

Benchmark	BF16	W4A16	Recovery
PIQA	77.58	77.20	99.51%
HellaSwag	51.73	50.64	97.88%
MMLU	64.76	63.48	98.02%
GSM8K	63.61	64.97	102.15%
Average	64.42	64.07	99.46%

lm_eval

lm_eval --model local-completions \
  --model_args model=INCModel4/Phi-4-multimodal-instruct-W4A16-AutoRound,tokenizer=microsoft/Phi-4-multimodal-instruct,tokenizer_backend=huggingface,base_url=http://127.0.0.1:8000/v1/completions,tokenized_requests=True,num_concurrent=32,max_retries=50,timeout=1800,max_length=131073,max_gen_toks=2048,trust_remote_code=True,add_bos_token=False \
  --tasks gsm8k --batch_size 32 --output_path results/gsm8k --log_samples --seed 42 \
  --num_fewshot 5 --apply_chat_template --fewshot_as_multiturn --gen_kwargs max_gen_toks=2048

lm_eval --model local-completions \
  --model_args model=INCModel4/Phi-4-multimodal-instruct-W4A16-AutoRound,tokenizer=microsoft/Phi-4-multimodal-instruct,tokenizer_backend=huggingface,base_url=http://127.0.0.1:8000/v1/completions,tokenized_requests=True,num_concurrent=32,max_retries=50,timeout=1800,max_length=131073,max_gen_toks=2048,trust_remote_code=True,add_bos_token=False \
  --tasks piqa,mmlu,hellaswag --batch_size 32 --output_path results/piqa_mmlu_hellaswag --log_samples --seed 42

Generate the Model

curl -fsSL https://huggingface.co/INCModel4/Phi-4-multimodal-instruct-W4A16-AutoRound/resolve/main/auto-round-block-list-fallback.patch \
    | patch -p1 -d "$(python -c 'import auto_round; from pathlib import Path; print(Path(auto_round.__file__).resolve().parent.parent)')"
auto-round \
    --model_name microsoft/Phi-4-multimodal-instruct \
    --scheme W4A16 \
    --iters 200 \
    --nsamples 128 \
    --ignore_layers lora_A,lora_B,embed_tokens_extend,lm_head \
    --format auto_round \
    --output_dir Phi-4-multimodal-instruct-W4A16-AutoRound \
    --seqlen 2048 \
    --group_size 128 \
    --to_quant_block_names model.layers

Ethical Considerations and Limitations

The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs. Therefore, before deploying any applications of the model, developers should perform safety testing.

Caveats and Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. Here are a couple of useful links to learn more about Intel's AI software:

• Intel Neural Compressor
• AutoRound

Disclaimer

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.

Cite

@article{cheng2023optimize,
  title={Optimize weight rounding via signed gradient descent for the quantization of llms},
  author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi},
  journal={arXiv preprint arXiv:2309.05516},
  year={2023}
}

arxiv github

Signed-off-by: Sun, Xuehao xuehao.sun@intel.com