# Validated production stack for Linux, CUDA 12.8, and PyTorch cu128.
# pip install . does not consume this file automatically; use
# a release GPU bundle manifest or pass -c explicitly when building artifacts.
# Modified FlashAttention and other native project wheels are intentionally
# absent: the hash-pinned bundle manifest is their only install source.
torch==2.10.0
transformers==5.13.0
accelerate==1.14.0
deepspeed==0.19.2
transformer-engine[core_cu12,pytorch]==2.13.0
cuda-bindings==12.9.4
cuda-python==12.9.4
apache-tvm-ffi==0.1.12
flash-linear-attention==0.5.2
tilelang==0.1.13
build==1.5.0
ninja==1.13.0
setuptools-scm==9.2.2
nvidia-cutlass-dsl==4.5.2
nvidia-cutlass-dsl-libs-base==4.5.2
torch-c-dlpack-ext==0.1.5
quack-kernels==0.5.0
nvidia-nccl-cu12==2.30.4
nvidia-nvshmem-cu12==3.3.9
wandb==0.27.1
