petals
πΈ Run LLMs at home, BitTorrent-style. Fine-tuning and inference up to 10x faster than offloading
νμΌ νμκΈ°
μ΅μ’ λ²μ λ€μ΄λ‘λ (.zip)- check-style.yaml
- push-docker-image.yaml
- run-tests.yaml
- benchmark_forward.py
- benchmark_inference.py
- benchmark_training.py
- prompt-tuning-personachat.ipynb
- prompt-tuning-sst2.ipynb
- __init__.py
- run_dht.py
- run_prod_server.sh
- run_server.py
- __init__.py
- sequence_info.py
- sequence_manager.py
- spending_policy.py
- __init__.py
- config.py
- from_pretrained.py
- inference_session.py
- lm_head.py
- ptune.py
- remote_forward_backward.py
- remote_generation.py
- remote_sequential.py
- sequential_autograd.py
- __init__.py
- block.py
- config.py
- model.py
- __init__.py
- block.py
- config.py
- model.py
- __init__.py
- block.py
- config.py
- model.py
- speculative_model.py
- __init__.py
- block.py
- config.py
- model.py
- __init__.py
- __init__.py
- backend.py
- block_functions.py
- block_selection.py
- block_utils.py
- from_pretrained.py
- handler.py
- memory_cache.py
- reachability.py
- server.py
- task_pool.py
- task_prioritizer.py
- throughput.py
- __init__.py
- asyncio.py
- auto_config.py
- convert_block.py
- cuda_graphs.py
- dht.py
- disk_cache.py
- hf_auth.py
- logging.py
- misc.py
- packaging.py
- peft.py
- ping.py
- random.py
- version.py
- __init__.py
- constants.py
- data_structures.py
- dht_utils.py
- bootstrap.id
- conftest.py
- server2.id
- test_aux_functions.py
- test_block_exact_match.py
- test_cache.py
- test_chained_calls.py
- test_dtype.py
- test_full_model.py
- test_optimized_layers.py
- test_peft.py
- test_priority_pool.py
- test_remote_sequential.py
- test_sequence_manager.py
- test_server_stats.py
- test_speculative_generation.py
- test_tensor_parallel.py
- test_utils.py
- .gitignore
- Dockerfile
- LICENSE
- pyproject.toml
- README.md
- setup.cfg
# μ€μΉ κ°μ΄λ
git clone https://github.com/bigscience-workshop/petals
κΉνλΈμμ νλ‘μ νΈ μ½λ μ 체λ₯Ό λ΄ μ»΄ν¨ν°λ‘ λ΄λ €λ°μ΅λλ€.
cd petals
λ°©κΈ λ΄λ €λ°μ νλ‘μ νΈ ν΄λ μμΌλ‘ μ΄λν©λλ€.
2. 곡μ μ€μΉ μ€ν¬λ¦½νΈ
μ¬μ μΆμ²pip install git+https://github.com/bigscience-workshop/petals
PyPIμ λ°°ν¬λ ν¨ν€μ§λ₯Ό λ°λ‘ μ€μΉν©λλ€. μμ€ ν΄λ‘ μ΄ νμ μμ΅λλ€.
brew install python
Homebrewλ‘ μ΄λ―Έ λΉλλ ν¨ν€μ§λ₯Ό λ°λ‘ μ€μΉν©λλ€. μμ€ λΉλκ° νμ μμ΅λλ€.
μ΄ λ ν¬μ READMEμ μ ν μ€μ λͺ λ Ήμ΄λ₯Ό κ·Έλλ‘ κ°μ Έμμ΅λλ€.
3. Docker
μ¬μ- Git GitHubμμ νλ‘μ νΈ μ½λλ₯Ό λ΄λ €λ°μΌλ €λ©΄ νμν©λλ€.
- Docker Desktop 컨ν μ΄λλ₯Ό λΉλνκ³ μ€ννλ €λ©΄ νμν©λλ€. μ€μΉ ν μ€νν΄μ λ°±κ·ΈλΌμ΄λμ μΌλμΈμ.
sudo docker run -p 31330:31330 --ipc host --gpus all --volume petals-cache:/cache --rm \
λΉλλ μ΄λ―Έμ§λ₯Ό μ€μ 컨ν μ΄λλ‘ μ€νν©λλ€.
μ΄ λ ν¬μ READMEμ μ ν μ€μ λͺ λ Ήμ΄λ₯Ό κ·Έλλ‘ κ°μ Έμμ΅λλ€.
4. Python
μ¬μpip install git+https://github.com/bigscience-workshop/petals
PyPIμ λ°°ν¬λ ν¨ν€μ§λ₯Ό λ°λ‘ μ€μΉν©λλ€. μμ€ ν΄λ‘ μ΄ νμ μμ΅λλ€.
python -m petals.cli.run_server meta-llama/Meta-Llama-3.1-405B-Instruct
νμ΄μ¬ μ€ν¬λ¦½νΈ(λλ λͺ¨λ)λ₯Ό μ€νν©λλ€.
python -m petals.cli.run_server --port 31330 meta-llama/Meta-Llama-3.1-405B-Instruct
νμ΄μ¬ μ€ν¬λ¦½νΈ(λλ λͺ¨λ)λ₯Ό μ€νν©λλ€.
python3 -m pip install git+https://github.com/bigscience-workshop/petals
requirements.txt λ±μ λͺ μλ νμ΄μ¬ λΌμ΄λΈλ¬λ¦¬λ₯Ό μ€μΉν©λλ€.
python3 -m petals.cli.run_server meta-llama/Meta-Llama-3.1-405B-Instruct
νμ΄μ¬ μ€ν¬λ¦½νΈ(λλ λͺ¨λ)λ₯Ό μ€νν©λλ€.
μ΄ λ ν¬μ READMEμ μ ν μ€μ λͺ λ Ήμ΄λ₯Ό κ·Έλλ‘ κ°μ Έμμ΅λλ€.
