Skip to content

Latest commit

 

History

History
28 lines (24 loc) · 2.73 KB

readme.md

File metadata and controls

28 lines (24 loc) · 2.73 KB
Raw

This repo is a learning playground for fine-tuning LLMs (currently Llama 3.2 1B / 3B Instruct) and running them locally (with Transformers.js v3 or Web-LLM).

First attempt process (branch: transformers)

  1. Downloaded full Google Chat transcript between my wife and me using Google Takeout
  2. Processed conversation data and created train.yaml pointing to that data
  3. Used torchtune with train.yaml config and full finetune to finetune Llama 3.2 1B Instruct on my data tune run full_finetune_single_device --config train.yaml
  4. Renamed hf_model_0001_2.pt file to pytorch_model.bin so Optimum can find it
  5. Used transformer.js v3's convert.py script (after copying in the quantize.py script and removing duplicate enum to avoid an import issue) to quantize into onnx files python transformers/scripts/convert.py --model_id "Meta-Llama3.2-1B-Instruct-FT" --task "text-generation"
  6. Used transformer.js v3's quantize.py script for all modes (could also do, for example, --modes "q4fp16" "int8") python transformers/scripts/quantize.py --input_folder "models/elisheldon/Meta-Llama3.2-1B-Instruct-FT" --output_folder "models/elisheldon/Meta-Llama3.2-1B-Instruct-FT/onnx"
  7. Published model to https://huggingface.co/elisheldon/Meta-Llama3.2-1B-Instruct-FT which is consumed by this web code

First attempt issues

  1. WebGPU issues. Current WebGPU implementation is mostly broken Only the q4fp16 quantized model can be run on WebGPU without a memory error or garbage output (at least on an M2 Pro), but the quality of the response is poor as compared with running the exact same model on WASM.
  2. WASM speed. Current WASM implementation runs very slowly, although it does successfully respond with reasonable responses.
  3. Small base model. Llama 3.2 1B Instruct was chosen to avoid 2GB protobuf limits in the onnx stack, but fine-tuning results are much better with fine-tuned Llama 3.2 3B Instruct run as GGUF outside of the browser context.

Second attempt process (branch: webllm)

  1. Started with the full finetune of Llama 3.2 1B Instruct from above, prior to any ONNX conversion, optimization or quanitzation
  2. Followed steps here to convert weights to MLC format, then uploaded to HF mlc_llm convert_weight Meta-Llama3.2-1B-Instruct-FT/ --quantization q4f16_1 -o Meta-Llama3.2-1B-Instruct-FT-q4f16_1-MLC mlc_llm gen_config Meta-Llama3.2-1B-Instruct-FT/ --quantization q4f16_1 --conv-template llama-3_1 -o Meta-Llama3.2-1B-Instruct-FT-q4f16_1-MLC
  3. Created basic chat app using webllm and deployed to https://eli-chat.pages.dev/

Second attempt issues

  1. Small base model. This issue persists, as the Llama 3.2 3B Instruct model and repeating the above flow caused memory issues on an M2 Pro.