[emme1t]

Llama.from_pretrained() currently lists the remote repository before resolving a cached model, so HF_HUB_OFFLINE=1 fails even when the GGUF is already cached. In offline mode, list the local Hub snapshot or a populated local_dir, then reuse the existing filename/glob matching and download resolution. Online repository discovery stays unchanged.

Fixes #1801.

Validation on Windows x64, Python 3.12.13, with the pinned llama.cpp CPU build:

• 19 targeted tests pass, including 11 new cases for cached files, nested paths, globs, shards, missing/ambiguous matches, local directories and online discovery. The final new test file gives 10 failed / 1 passed on the original code.
• The 11 new cases also pass with huggingface-hub==0.23.0; the main run uses 2.4.0.
• Real offline inference with ggml-org/models / tinyllamas/stories260K.gguf succeeds for both the literal filename and tinyllamas/stories260*.gguf. Both generate 8 tokens, , there was a little girl, from Once upon a time.
• Repository-wide Ruff check and format check pass. The full multi-model test suite was not run locally.

The native build used LLVM-MinGW with Ninja, GGML_NATIVE=OFF, GGML_OPENMP=OFF, and _WIN32_WINNT=0x0A00. Offline matching considers files already present on disk. A requested local_dir must contain the model, following Hugging Face Hub's offline behavior.