feat: add pytorch_engine_qwen2_5vl_sm120

[kolmogorov-quyet]

Motivation

• Silence spurious CancelledError logs from asyncio.run_coroutine_threadsafe when the future is cancelled.
• Improve FlashAttention kernel meta-tuning for newer GPUs (adds SM 128 path).
• Keep the repository clean with extra .gitignore entries.
• Add a minimal smoke test for the new engine to ensure CI coverage.

Modifications

File	Key changes
.gitignore	Ignore builder/, lmdeploy/lib/, IDE caches, etc.
lmdeploy/pytorch/kernels/cuda/flashattention.py	Added _kernel_meta_sm128(), refactored meta-selection logic (~40 LoC).
lmdeploy/serve/async_engine.py	Replaced lambda f: f.result() with a safe callback: lambda f: None if f.cancelled() else f.result()
tests/test.py	New smoke test: init engine, run one caption generation, assert non-empty output.

Backward Compatibility

No API breaks; existing engines and interfaces continue to work.

Use Cases

• Cleaner logs under heavy load—no more CancelledError spam.
• Better kernel parameters for Ada/Hopper-next GPUs (RTX 50xx, etc.).
• Quick regression guard via the new unit test.

Checklist

• Code formatted / linted
• Unit tests added & pass (pytest -q)
• Documentation updated if necessary
• Ready for review

[windreamer]

Would you please follow the original style to enable Blackwell support when _nv_cap[0] < 13 and put this branch to the end of the if-elif-else block? I think this will help the community better understand the code.

[windreamer]

And you can try to install pre-commit to help you fix lint errors.

Thank you for your efforts!

[kolmogorov-quyet]

Thank you for your helpful review and suggestions 🙏
I've updated the code to follow the original style for better clarity, as you mentioned.
Also, I’ve set up pre-commit and made sure it passes all hooks, to avoid common PEP8 issues.

Please take a look when you have time — really appreciate your support!

[grimoire]

@lvhan028 any advice about the tests?

[lvhan028]

Hi, @kolmogorov-quyet
We appreciate your contribution. Just a quick note—the lmdeploy/tests directory is intended for unit test cases rather than functional testing.
We've already integrated Qwen2.5-VL model testing into lmdeploy's functional test suite, so you can safely remove this test file

[grimoire]

LGTM