[TRTLLM-15177][chore] Kimi K3: inline MLA module into modeling_kimi_linear

[brnguyen2]

Description

Deferred cleanup carried over from PR #17269 (Kimi K3 / KimiLinear model support), tracked in TRTLLM-15177.

Inline the kimi_k3_mla module into modeling_kimi_linear.py, matching the per-model modeling_xxx.py convention (e.g. DeepSeek-V3). KimiK3MLAAttention and its rope/dtype helpers move directly into the model file and the standalone modules/kimi_k3_mla package is deleted. The moved code is unchanged apart from following the file's existing Linear as TrtllmLinear alias. kimi_kda stays a standalone module (general enough to warrant it).

No functional change: the moved code is identical; only its location changes.

Supersedes #18159

This replaces #18159, which is being closed rather than rebased. #18159 inlined the pre-refactor version of the module and no longer rebases cleanly onto main: after it was opened, upstream substantially rewrote the very module it deletes — the shared-MLA output-gate-hook refactor (#18375, which removed the K3-local backend-selection machinery and moved the gate to a base-class hook) and the attention-domain consolidation (#17968, which relocated the attention.* import paths). Carrying the old inlined copy forward would silently revert that work, so the delta is different enough to warrant a fresh review.

This PR is scoped to the inline only. The other two items from #18159 — removing the communication_method parameter chain and the KimiLinearConfig tuple default — are dropped from this PR: the comm-method chain is now wired through ConfigurableMoE.determine_communication_method on main and deserves its own evaluation. Those can follow as separate one-concern PRs.

Test Coverage

No new code paths are introduced — this is a pure move. Existing Kimi K3 parity suites under tests/unittest/_torch/modeling/ construct KimiK3MLAAttention via the model and exercise the moved code unchanged.

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (no new code paths here; pure move).

• If PR introduces API changes, an appropriate PR label is added (no API change; internal _torch only).

• Any new dependencies have been scanned (none added).

• CODEOWNERS updated if ownership changes (no ownership change).

• Documentation updated as needed.

• Update tava architecture diagram if there is a significant design change (none).

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

[brnguyen2]

/bot run

[tensorrt-cicd]

PR_Github #72195 [ run ] triggered by Bot. Commit: 04086b8 Link to invocation

[tensorrt-cicd]

PR_Github #72195 [ run ] completed with state SUCCESS. Commit: 04086b8
/LLM/main/L0_MergeRequest_PR pipeline #59237 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[brnguyen2]

Closing this out. This is a pure code-move cleanup (inlining KimiK3MLAAttention into modeling_kimi_linear.py) with no functional change, and the kimi_k3_mla module it targets is under active development — each change there (e.g. #18728) turns this into a modify/delete conflict that needs a fresh semantic re-fold, which already forced one respin (this superseded #18159 for the same reason).

The module is still a single-consumer helper and the per-model inline convention (cf. DeepSeek-V3) still applies, so the cleanup remains valid and can be reopened/reworked later if we want to prioritize it. Not worth carrying as a perpetually-conflicting draft in the meantime.