Fix the unit tests so they can actually run

[Hotragn]

What

Make the two existing test modules in air_llm/tests/ actually runnable. Right now neither one does.

Why

test_compression.py fails at import. It does:

from airllm import compress_layer_state_dict, uncompress_layer_state_dict

but those functions live in airllm.utils and are not re-exported from the package root (__init__.py exports split_and_save_layers and NotEnoughSpaceException only). So collecting the module raises:

ImportError: cannot import name 'compress_layer_state_dict' from 'airllm'

Because it's a collection error, it takes down the whole test run, not just that module. Separately, the test quantizes with .cuda() tensors, so even once importable it errors rather than skips on a CPU-only machine.

test_automodel.py went stale in the v3.0 rewrite. It still asserts:

'garage-bAInd/Platypus2-7B': 'AirLLMLlama2',
'mistralai/Mistral-7B-Instruct-v0.1': 'AirLLMMistral',
'mistralai/Mixtral-8x7B-v0.1': 'AirLLMMixtral',

but the rewrite routes all standard architectures through the generic AirLLMBaseModel (only ARCH_OVERRIDES archs get dedicated classes), so those three assertions now fail. It also used a relative import (from ..airllm.auto_model import ...) that disagrees with the absolute imports in the other module — the two couldn't be collected by the same run — and it needed network access to resolve every repo id.

Changes

• test_compression.py: import from airllm.utils; skip instead of error when CUDA is unavailable.
• test_automodel.py: test the routing offline by faking the config lookup, asserting current behavior — custom archs (ChatGLM/QWen/Baichuan/InternLM) → dedicated classes; standard archs (Llama, Mistral, Mixtral, Qwen2/3, Phi3, Gemma2, DeepseekV3) and unknown/missing archs → AirLLMBaseModel. The original against-the-Hub check is kept with corrected expectations, gated behind AIRLLM_TEST_NETWORK=1.
• Use absolute imports in both so one invocation collects the whole directory.

No production code is touched, and no new dependency is added (the network/GPU gates use unittest.skipUnless, not pytest markers, so the suite still runs under plain unittest).

Testing

From air_llm/, both runners agree — 3 passed, 2 skipped (the GPU test and the network test):

python -m pytest -v
python -m unittest discover -s tests -t .    # Ran 5 tests ... OK (skipped=2)

Also confirmed the new offline routing assertions genuinely exercise ARCH_OVERRIDES (they fail if the mapping is changed).

Related: I opened #312 proposing a small CI workflow. This PR is useful on its own regardless of that — but it's also the prerequisite that would let CI run the suite green, so I split it out rather than bundling it there.

[Hotragn]

This became more relevant with v3.1.0: the new tests/test_kimi_k3_split.py adds 9 tests, but the suite can't run as a directory because test_compression.py still fails at import, and a collection error aborts the whole run:

$ cd air_llm && python -m pytest tests/
9 tests collected, 1 error
ERROR tests/test_compression.py
!!!!!!!! Interrupted: 1 error during collection !!!!!!!!

(Checked out v3.1.0 / 64a4e4f clean to confirm.) The 9 Kimi tests are collected fine — they're just never executed, because from airllm import compress_layer_state_dict raises ImportError (those live in airllm.utils and aren't re-exported from the package root). So today they only run if you name the file explicitly.

This PR fixes exactly that, and nothing else needs to change: the new Kimi tests already use the absolute-import style (from airllm.utils import split_and_save_layers) this PR standardizes on, so they pass unmodified. After it, python -m pytest from air_llm/ runs the whole directory green — the Kimi tests included.

[Hotragn]

Refreshed against v3.2.0 — the blast radius grew. With test_qwen3_8_split.py added alongside test_kimi_k3_split.py, the collection error now takes down 15 tests instead of 9:

$ cd air_llm && python -m pytest tests/
15 tests collected, 1 error
ERROR tests/test_compression.py
!!!!!!!! Interrupted: 1 error during collection !!!!!!!!

So every test you've written for Kimi K3 and Qwen3.8 is collected and then discarded, because from airllm import compress_layer_state_dict in test_compression.py raises ImportError (those live in airllm.utils, never re-exported from the package root). They only run today if you name each file explicitly.

This PR is still MERGEABLE/clean against v3.2.0 and needs no rebase. It's test-only, touches no production code, adds no dependency, and both of your new test files pass unmodified afterwards — they already use the absolute-import style it standardises on.