# The versions the service was measured and released with. Every line is a pin on purpose: the
# image is rebuilt often and the model is not, and a torch or transformers that moved under it
# would change the numbers without changing the checkpoint. Bump them together, then rerun
# tests/live_service.py against the container before shipping.
#
# torch from PyPI carries CUDA 12.8 and runs on Blackwell (sm_120) and older; the container needs
# only the host driver through nvidia-container-toolkit.
torch==2.10.0
transformers==5.3.0
huggingface-hub==1.7.1
accelerate==1.13.0
safetensors==0.7.0
tokenizers==0.22.2
jinja2>=3.1
einops==0.8.2
# The linear-attention kernels the hybrid trunk wants on CUDA; without them p50 triples.
# Off the GPU the service hides them, so the same image serves a CPU too.
fla-core==0.5.2
flash-linear-attention==0.5.2
fastapi==0.140.0
uvicorn==0.51.0
requests==2.32.5
