[simone-chen]

• Kimi KDA prefill retained the completed convolution input while allocating independent core inputs, and retained FlashKDA output/state/workspace buffers into the next point.
• Release the convolution input and all three views after its timing/logging completes. Clear native buffers in each point’s cleanup, including after failures.
• Preserve native workspace sizing, callback grids, dispatch, input generation, timing settings and partial-row/failure accounting. Oversized native workloads still fail explicitly; this change does not make the whole grid fit on B200.
• Stacked on fix: adapt vLLM 0.31 collector runtime contracts #588 (simonec/vllm031-scale-qualification); no runtime or collection-request pins changed.

Validation

• 58 vLLM/SGLang KDA contract tests passed; Ruff lint and formatting passed.
• Weak-reference lifetime tests execute the collector with both prefill backends and verify continuation after a core failure, unchanged timing parameters and retained partial convolution rows.
• Mutation checks fail when either cleanup is removed: four failures for retained convolution inputs; two FlashKDA failures for retained native buffers.
• GPU collection has not run for this change. Full CI will be requested for the exact PR head.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}

[simone-chen]

/ok to test 1f7451a

[simone-chen]

Superseded by consolidated recurrent/MoE compatibility PR #605, stacked on attention #604. The MoE cleanup is preserved once, including the duplicate in #603. Communication remains #591. Original branches and campaign commit references are preserved; closing without merging.