[mickqian]

Summary

• wire MoonViT through the existing encoder-DP sharding helper and preserve the TP=1 projector-facing tensor contract
• make MoonViT layers tensor-parallel when encoder DP is disabled, and avoid TP-collective CUDA-graph capture for DP-sharded image work
• eliminate per-layer GPU synchronization for packed attention metadata and bound the serving-time positional interpolation cache
• resolve the language model once before prefill graph setup and cache supported image-processor kwargs

Before / after behavior

This is primarily a Kimi-VL correctness/enablement change, not a valid before/after throughput benchmark: before, Kimi-VL TP=1 encoder-DP returns a list where the projector requires a tensor and fails on serving; after, the output is concatenated and the server completes image requests. No cross-engine number is presented as a PR performance result.

Validation

• pre-commit on all changed files
• registered Kimi-VL encoder/DP, interpolation-cache eviction, ViT capture-context, and HF image-processor tests
• the Kimi-VL entry was intentionally removed from the shared test_encoder_dp matrix.

---

CI States

Latest PR Test (Base): ✅ Run #29237951501
Latest PR Test (Extra): ❌ Run #29237951515

[gemini-code-assist]

Code Review

This pull request implements tensor parallelism (TP) and data parallelism (DP) support for the Kimi-VL model and its MoonViT encoder. It also introduces several performance optimizations, including caching interpolated positional embeddings, passing precomputed max_seqlen to avoid GPU synchronization, and caching accepted kwargs in image processor patches to bypass slow exception paths. The review feedback highlights a critical issue where the new positional embedding cache grows without bound during inference, potentially causing a GPU memory leak and Out-Of-Memory (OOM) crashes. It is recommended to limit the cache size and evict old entries to ensure stability in long-running serving environments.

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.

[mickqian]

/tag-and-rerun-ci