perf: avoid redundant Kimi vision projection copies

[mickqian]

Summary

Kimi K2.5/K2.7's non-DP image path currently does:

1. concatenate per-image MoonViT outputs;
2. run the multimodal projector;
3. split the projected tensor back into per-image views;
4. immediately concatenate those views at the call site.

The language-model embedding path only consumes the packed tensor. This PR keeps the packed result and also accepts an already-packed tensor, removing the redundant split/cat and its extra allocation/copy.

Performance

NVIDIA RTX 5090 (SM120), CUDA 13.0, BF16, 1152-wide MoonViT outputs with a lightweight flattening projector. This isolates the packing overhead and is not an end-to-end serving claim.

Image output chunks	Before p50	After p50	Speedup
1	0.0242 ms	0.0140 ms	1.73x
2	0.0222 ms	0.0126 ms	1.76x
4	0.0231 ms	0.0126 ms	1.84x
8	0.0274 ms	0.0134 ms	2.05x

The benchmark uses the same projector call in both paths; only the post-projector split/cat is changed.

Correctness / validation

• Added CPU coverage for variable-resolution per-image outputs and already-packed outputs.
• Pre-commit hooks, AST validation, registration validation, and git diff --check pass.
• The local macOS environment cannot collect the model test because its preinstalled huggingface_hub==1.6.0 rejects an existing SGLang strict dataclass definition; this is unrelated to the patch and is covered by the repository CI environment.

---

CI States

Latest PR Test (Base): 🚫 Run #29550245282
Latest PR Test (Extra): ❌ Run #29550245261

[gemini-code-assist]

Code Review

This pull request refactors mm_projection_auto in kimi_k25.py to return a single flattened feature tensor directly, avoiding redundant split and concatenate operations. It also adds unit tests to verify this behavior. The review feedback points out a correctness bug where a 3D tensor could be returned instead of a 2D tensor if mm_projector is None, and suggests a performance optimization to avoid calling torch.cat when there is only a single image in the sequence.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.