llama.onnx

News

04/?? deploy <llama.onnx, quant table> to aarch64

04/19 remove GPTQ zero point guidance

04/18 export mixed-precision quant table from GPTQ-for-LLaMa

04/11 add 13GB onnx-fp16 models

04/11 add memory pool, support 2GB RAM laptop ⭐

04/10 reduce onnx model size to 26GB

04/10 support temperature add topk logits warp

04/07 add onnxruntime demo

04/05 init project

Features

• Release llama 7B onnx models
• With a 400-lines onnxruntime alpaca demo
  • neither torch nor transformers required
  • support memory pool, works on 2GB laptop/PC (very slow 🐢)

Why do this ?

1. Visualization. graphviz crashed on llama model. LLM visualization tool must support nest or operator folding feature
2. Quatization. LLM often repeat itself, just like fractal. For llama quantization, loading part of decoder backbone would be enough (400MB). It could be quantized partially
3. Embeded device. Small board IO error occurs when dd a big single file
4. Distributed system. Inference LLM on many hybrid (FPGA/NPU/GPGPU) devices would be simple
5. onnx tools. Device manufacturer has support onnx well, there is no reason to neglect it

Usage

Download onnx models here:

Precision	Size	URL
fp32	26GB	huggingface
fp16	13GB	huggingface or 硬件模型库

Here is the graph to call them:

Try onnxruntime demo, no torch required, and the precision has been checked.

$ python3 -m pip install -r requirements.txt
$ python3 demo-single.py ${FP16_ONNX_DIR} "bonjour"
..
# If you only have 4GB memory, use `--poolsize`
$ python3 demo-single.py ${FP16_ONNX_DIR} "bonjour" --poolsize 4
..
Bonjour.

# Try more options
$ python3 demo-single.py --help

Export onnx

STEP1 Convert to HF format

These models converted from alpaca huggingface.

• If you are using LLaMa or llama.cpp, convert it to HF format first. Here are steps:

  # install transformers master
  $ git clone https://github.com/huggingface/transformers
  $ cd transformers && python3 setup.py install
  ..
  $ cd src/transformers
  $ python3 src/transformers/models/llama/convert_llama_weights_to_hf.py  --input_dir ${LLaMa_PATH}  --model_size 7B  --output_dir ${HF_PATH}

• If you are using alpaca-lora, use this script to merge LoRA weights.

• If you are using alpaca, go STEP2.

STEP2 torch.onnx.export

Checkout transformers to this hacking branch, run single inference.

$ python3 tools/export-onnx.py ${PATH_ALPACA_7B}

STEP3 convert to fp16

Use onnxconverter-common.float16

$ cd tools
$ python3 -m pip install -r requirements.txt
$ python3 convert-fp32-to-fp16.py ${FP32_PATH} ${FP16_PATH}

Quantization

Mixed-precision kernel optimization is on the way. Here is a part of guidance.

Notes

1. Any logits_processor or BeamSearch not implemented, so the result would be not good
2. I have compared the output values of onnxruntime-cpu and torch-cuda, and the maximum error is 0.002, not bad
3. The current state is equivalent to these configurations

temperature=0.1
total_tokens=2000
top_p=1.0
top_k=40
repetition_penalty=1.0

Acknowlegements

• llama
• alpaca
• alpaca-lora
• transformers
• peft
• Chinese-LLaMA-Alpaca

License

GPLv3