ft_req.txt

# ==============================================================================
# Requirements for v3 -> v4 NADRA fine-tuning on the HBL GPU server
# ==============================================================================
#
# HOW TO GET THESE ONTO THE SERVER (no internet on the GPU box):
#
#   On your internet-connected laptop, with the SAME Python version and OS
#   family (Windows) as the target server:
#
#       pip download -r requirements.txt -d wheelhouse
#
#   Then copy the entire `wheelhouse` folder to the GPU server and install
#   fully offline with:
#
#       pip install --no-index --find-links=wheelhouse -r requirements.txt
#
#   IMPORTANT: run `pip download` on a machine with the same OS (Windows)
#   and Python minor version (e.g. 3.10.x) as the server, or the downloaded
#   wheels won't match and pip will silently try to hit the network on the
#   server and fail.
#
# BITSANDBYTES / WINDOWS NOTE:
#   Current stable bitsandbytes (0.45.x+) publishes official Windows CUDA
#   wheels directly on PyPI -- a plain `pip download bitsandbytes` on the
#   laptop should work, no need for third-party Windows forks.
#   Compatibility matrix to double-check against your server's CUDA Toolkit
#   before committing to a long run:
#     - CUDA Toolkit 11.7-12.6 supported on Windows
#     - Requires MSVC 19.38+ (VS2022) runtime present
#     - NF4/4-bit quantization needs NVIDIA Maxwell or newer -- your A16
#       (Ampere, compute capability 8.6) is well within spec.
#   Test 4-bit model loading in isolation on the server FIRST (small script,
#   just load + quantize, no training) before starting an unattended run --
#   CUDA/driver mismatches surface as first-op runtime errors, not install
#   errors, so a clean `pip install` does not guarantee it will actually work.
#
# The training script defaults to optim="adamw_torch" (no bitsandbytes
# optimizer kernels needed at all). Only bitsandbytes' 4-bit LOADING path
# is required by default. Set USE_BNB_PAGED_OPTIMIZER = True in the script
# only after separately confirming the 8-bit optimizer kernels work on this
# server -- it's a distinct risk surface from 4-bit quantized loading.
#
# ==============================================================================

torch>=2.1,<2.5
transformers>=4.44,<4.50
accelerate>=0.33,<0.35
peft>=0.12,<0.14
bitsandbytes>=0.45,<0.50
datasets>=2.20,<2.22
numpy>=1.24,<2.0
pandas>=2.0,<2.3
safetensors>=0.4.3