[PerryZhang01]

Follows up #2818. Ready for review.

#2818 fixed Kimi-K3's Triton 3.8 import failure (key contains names that are not kernel arguments: NC) as a patch on the vLLM image build, which only reaches CI once vllm-latest is rebuilt. Until then every PR's Kimi-K3 cell still fails, as in the runs of #2785 and #2816 after #2818 merged. This moves the fix into the plugin, where it takes effect with the ATOM install.

Dev record

• Found: the failure is in @triton.autotune at import time, so the kernel cannot be patched after import.
• Decided: register_model wraps triton.autotune and drops non-argument keys only for kernels defined under vllm.models.kimi_k3, the same edit as vllm#60446. Check first: vLLM's model-inspection subprocess runs load_general_plugins before importing the model class, so the wrapper is in place there too.
• Decided: the image patch and its dockerfile steps are removed; docker/ matches main before fix(plugin): green the vLLM 0.31 plugin CI cells; V4-Flash and GLM-5.2 on Model Runner V2 #2818. The PR job's rebuild-on-recipe-change rule from fix(plugin): green the vLLM 0.31 plugin CI cells; V4-Flash and GLM-5.2 on Model Runner V2 #2818 stays, so this PR's job builds an image without the patch.
• Left undone: none. With a vLLM that carries vllm#60446 the wrapper drops nothing and can be removed.

Named result

On the unpatched vllm-latest image (Triton 3.8.0): import vllm.models.kimi_k3 raises NC. After load_general_plugins(), ATOM's K3 model module and kda_backend import, and chunk_kda_fwd_kernel_intra_sub_chunk keys on ['BK', 'BT', 'HV'].

This PR's CI (40996045263) built a vLLM image without the patch; all six cells pass on V2. Kimi-K3 0.956 (threshold 0.94) with the plugin's INFO line for the dropped NC and no NC error; V4-Flash 0.9575, GLM-5.2 0.9272, M3 0.9424, V4.1 0.9287, Qwen3.5 0.8453.

Gates

• tests/test_vllm_kimi_k3_kda_autotune_patch.py: 2 passed (real Triton, no vLLM or GPU).
• black clean; ruff clean on the new files.

🤖 Generated with Claude Code

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 2835 --add-label <label>