Llama.turbine

This is an experimental pytorch model that can load from llama.cpp gguf files, run eagerly and compile to Turbine/IREE. We are using it to evaluate different approaches for interfacing with the llama.cpp ecosystem.

Setup instructions

Prerequisite

Setup SHARK-Turbine and it's python required environment.

Prepping model

python python/turbine_llamacpp/model_downloader.py --hf_model_name="openlm-research/open_llama_3b"

by this point you should see a directory named downloaded_open_llama_3b in you working directory. This is typically the downloaded_<name_of_your_model>.

Setup Llama.cpp

git clone https://github.com/ggerganov/llama.cpp
pip install gguf

Converting HF model to GGUF.

# Convert HF to GGUF and quantize to q8.
python llama.cpp/convert.py downloaded_open_llama_3b --outfile ggml-model-q8_0.gguf --outtype q8_0

Running Llama.turbine

Running on pytorch

python python/turbine_llamacpp/model.py  --gguf_path=/path/to/ggml-model-q8_0.gguf

Generating MLIR

python python/turbine_llamacpp/compile.py  --gguf_path=python/turbine_llamacpp/ggml-model-q8_0.gguf