Dependency profiles

FastPLMs loads from Hugging Face model repositories with trust_remote_code=True. This repository is a source, test, and dependency workspace. It is not an installable Python distribution.

core.in and features/*.in declare direct dependencies. profiles/*.in combine those declarations for repository environments. Validation commands constrain Torch and Transformers with constraints/validation.txt.

Create a local validation environment with:

uv venv
uv pip install \
  -r requirements/profiles/cpu-validation.in \
  -c requirements/constraints/validation.txt \
  --torch-backend cpu

CUDA container profiles use the same command without --torch-backend cpu. FP8 profiles also pass --overrides requirements/overrides/cuda.txt.

When you need a full transitive lock for a maintained environment, compile the required profile. Do not lock all incompatible features together:

uv pip compile \
  requirements/profiles/candidate.in \
  -c requirements/constraints/validation.txt \
  --universal \
  --generate-hashes \
  -o requirements/locks/candidate.txt

Generated locks are environment-specific test inputs. Public Hugging Face artifacts must expose only direct runtime dependencies.