The legacy setup.py had a PostInstallCommand that ran 'pip install
--upgrade transformers' after every install. That works locally but
breaks PEP 517 build isolation, editable installs, conda environments
and any other build backend that doesn't shell out to pip.

- Replace the install-hooks-and-magic setup.py with a 3-line shim that
  just calls setup() and lets pyproject.toml own the metadata.
- Add air_llm/pyproject.toml (PEP 621) with:
  * build-system declaration (setuptools>=68)
  * project metadata incl. version, authors, classifiers, URLs
  * pinned runtime dependencies (transformers>=4.36 replaces the
    forced post-install upgrade and fixes the rope_scaling error)
  * optional-dependencies groups for compression, test, lint
  * [tool.setuptools.packages.find] so 'airllm' is discovered
  * [tool.pytest.ini_options] and [tool.ruff] for tests/lint

The classifier 'License :: OSI Approved :: MIT License' is also wrong
(see LICENSE file) — fixed to 'Apache Software License'.

The existing tests/test_automodel.py reaches out to Hugging Face to
download real model configs, which makes it slow, flaky and impossible
to run in CI without a network or a heavyweight model checkpoint.

Add tests/test_auto_model_dispatch.py that patches
AutoConfig.from_pretrained and exercises the dispatch table directly:

- one test per supported architecture (Llama, Mistral, Mixtral,
  Qwen, Qwen2, Baichuan, ChatGLM, InternLM) asserts the correct
  wrapper class is returned
- regression test: Qwen2ForCausalLM must NOT fall through to QWen
- regression test: MixtralForCausalLM must NOT fall through to Mistral
- unknown architecture raises UnknownArchitectureError with a helpful
  message that includes the supported list
- empty / missing 'architectures' field raises UnknownArchitectureError
- the dispatch table has no duplicate needles
- hf_token kwarg is forwarded to AutoConfig (and absent means None)

Runs offline in under a second with 'pip install -e .[test] && pytest'.

The repo previously had no CI, so broken merges could land unnoticed.

Add .github/workflows/ci.yml that:

- runs on push to main and on every pull request
- tests against a matrix of Python 3.10, 3.11 and 3.12
- installs torch (CPU build, no CUDA needed) then the package with
  [test,lint] extras from air_llm/pyproject.toml
- runs 'ruff check' for lint
- runs 'pytest' (picks up the new offline test_auto_model_dispatch.py)

The README jumps straight from 'Supported models' to 'FAQ' and assumes
the reader already understands the layer-sharded inference approach.

- New 'How it works' section explains the five-stage pipeline
  (shard-on-disk, stream layers, CPU-pinned KV-cache, prefetching,
  optional bitsandbytes compression) and why inference is bandwidth-
  bound rather than compute-bound.
- New 'Troubleshooting' section collects the four most common support
  questions and points each at the right fix or issue tracker.
- TOC is updated with both new sections so they're discoverable.

Mentions the new UnknownArchitectureError raised by the refactored
AutoModel so users can self-serve when they hit an unsupported
architecture.

…dule

airllm/airllm_base.py does
    from optimum.bettertransformer import BetterTransformer
at import time. In recent optimum releases (>=1.14 or so) that module
was removed, so a vanilla 'pip install airllm' on a fresh environment
fails at import time with ModuleNotFoundError — even though the
BetterTransformer symbol is only used inside an internal class that the
unit tests never touch.

Add tests/conftest.py that pre-populates sys.modules with a MagicMock
for the missing module. Pytest auto-loads conftest.py so the test
suite now runs in any environment, regardless of which optimum version
is installed.

Verified locally: 'pytest tests/test_auto_model_dispatch.py' -> 8 passed.