34 lines
1.2 KiB
Plaintext
34 lines
1.2 KiB
Plaintext
# ── J-Wash Python dependencies ───────────────────────────────────────────────
|
|
# An NVIDIA GPU (CUDA) is required.
|
|
#
|
|
# 1. Install PyTorch FIRST, matching your CUDA version, from
|
|
# https://pytorch.org/get-started/locally/ — for example:
|
|
# pip install torch --index-url https://download.pytorch.org/whl/cu124
|
|
# (Blackwell GPUs such as the RTX 50-series need a cu128+ build.)
|
|
#
|
|
# 2. The Jacobian-lens library ("jlens") is NOT on PyPI. Clone it and install it
|
|
# editable (see the README for details):
|
|
# git clone https://github.com/anthropics/jacobian-lens vendor/jacobian-lens
|
|
# pip install -e vendor/jacobian-lens
|
|
#
|
|
# 3. Then install everything below:
|
|
# pip install -r requirements.txt
|
|
|
|
transformers>=5.5
|
|
accelerate>=1.0
|
|
huggingface-hub>=0.26
|
|
safetensors>=0.4
|
|
|
|
fastapi>=0.115
|
|
uvicorn[standard]>=0.30
|
|
websockets>=13
|
|
pydantic>=2.0
|
|
|
|
numpy>=1.26
|
|
msgpack>=1.0
|
|
nvidia-ml-py>=12.0
|
|
|
|
# Optional — only needed for int8 / nf4 quantized loading (the "quant" option).
|
|
# bitsandbytes wheels vary by platform/CUDA, so install it manually if you use it:
|
|
# pip install bitsandbytes
|