# Versions of our final run (NVIDIA A100 40 GB, CUDA 12.8, Python 3.12).
torch==2.8.0
transformers==5.18.0
accelerate==1.15.0
peft==0.21.1
huggingface_hub==1.33.0
numpy==2.1.3
# Qwen3.5 linear-attention (Gated DeltaNet) kernels; 0.5+ needs a newer Triton than torch 2.8 ships
flash-linear-attention==0.4.0
# Our run also had causal-conv1d==1.7.0 (fast short-convolution kernel). It needs torch at build time, so it cannot
# be installed from this file; for identical kernels install it afterwards (optional, transformers falls back to torch):
#   pip install causal-conv1d==1.7.0 --no-build-isolation