What problem does it solve? Installing test packages like Transformers can silently replace the vendor-compatible PyTorch build that the compiled torch_fl extension depends on, producing ABI errors and invalid test results. This Skill provides a controlled procedure for installing, pinning, and validating test dependencies so environment problems are never mistaken for backend failures. ## Core Features & Use Cases - ABI Protection: Enforces --no-deps installs and explicit torch version pinning so pip never swaps the torch build underneath torch_fl. - Isolation and Pinning: Recommends one environment per (torch_fl build, torch version, transformers version) tuple with pre-install inspection of interpreter and package versions. - Connectivity Recovery: Provides ordered fallback steps for Hugging Face Hub access using local caches, proxies, and the hf-mirror.com endpoint. - Use Case: Before running official Transformers model tests on a vendor accelerator, install transformers with --no-deps, add accelerate and tokenizers explicitly, run the import and flagos device smoke check, then record the full environment beside the test results. ## Quick Start Set up a validated test environment with transformers and accelerate pinned for my torch_fl build, then run the import and device smoke check before executing the official model tests.
