feat: Add warmup functionality to reduce search latency

[majiayu000]

Summary

• Add enable_warmup parameter to HNSW and DiskANN embedding servers to pre-load model at startup
• Implement warmup() method on LeannSearcher for manual pre-warming before first search
• Add auto-warmup option during LeannSearcher initialization (enable_warmup=True)
• Model warmup is done via dummy embedding computation, ensuring cache is hot
• Add comprehensive tests for warmup functionality

Test Plan

• Added unit tests in tests/test_warmup.py
• Tests cover warmup method, auto-warmup, server parameter passing
• All Python files pass syntax validation

Fixes #177
Fixes #159

[yichuan-w]

Thanks, this is a known issue for a long time, we will look into that!! cc @andylizf , and can you fix the lint error here?

[majiayu000]

The macOS-13 CI jobs show as 'cancelled' rather than 'failed' - this appears to be a GitHub Actions runner issue, not a code problem. All other builds (macos-14, macos-15, ubuntu) passed successfully.

Could you please re-run the cancelled macOS-13 jobs?

[yichuan-w]

Sure, I will do that later, sorry for the late responses since I was on vacation. And thanks again for your contribution!

[majiayu000]

No worries, thanks for re-running the CI! Let me know if there's anything else that needs to be addressed.

[andylizf]

Thanks for implementing this warmup feature! The functionality looks good and solves a real latency problem.

A few suggestions for potential future improvements (not blocking for this PR):

1. Extract common warmup logic

The warmup code in hnsw_embedding_server.py and diskann_embedding_server.py is duplicated. Consider extracting to a shared module:

# leann/warmup.py
def warmup_embedding_model(model_name: str, embedding_mode: str, provider_options=None) -> float:
    """Pre-load embedding model by computing a dummy embedding."""
    ...

2. Clarify enable_warmup semantics

Currently enable_warmup in LeannSearcher controls two things:

• Server-side model preloading
• Client-side dummy query

Consider separating these concerns in a future refactor.

3. Avoid calling private methods

LeannSearcher.warmup() calls self.backend_impl._ensure_server_running(). Consider exposing a public interface for this.

---

These are minor design notes - the PR is good to merge as-is. Nice work! 🎉

[andylizf]

What do you think? @yichuan-w

[yichuan-w]

yeah, I guess it is good to merge if @majiayu000 can solve the conflict and pass the PR, btw @andylizf , whats the relationship between this and #176