[https://nvbugs/6705034][fix] Route zero-token causal-conv input to the channel-major kernel

[nv-guomingz]

Description

A zero-token x reaches causal_conv1d_fwd with the strides its no-op transpose left behind (an empty tensor is always reported contiguous, so the caller's .contiguous() does not restride it). The stride-based layout classification therefore saw it as channel-last, and because 0-byte CUDA allocations can alias each other, the channel-last in-place guard rejected the call with:

RuntimeError: causal_conv1d_fwd: the channel-last path cannot run in-place; `out` must not overlap x

This broke unittest/_torch/modules/kimi_kda/test_kda_prefill_op.py::test_kda_mixer_empty_prefill on DGX_B200 (https://nvbugs/6705034), waived in #18549.

The channel-last kernel also cannot run an empty input anyway (zero token chunks would give an empty launch grid), so this change routes seqlen == 0 to the channel-major kernel, which handles empty input (conv-state write-back only) the way it always has before the channel-last path was introduced in #18081.

Test Coverage

• Unwaives unittest/_torch/modules/kimi_kda/test_kda_prefill_op.py::test_kda_mixer_empty_prefill (DGX_B200), which exercises exactly this zero-token path.

PR Checklist

• PR title and description added
• Test coverage: existing unit test unwaived

🤖 Generated with Claude Code

Dev Engineer Review

• Routes seqlen == 0 inputs to the channel-major causal-convolution kernel.
• Preserves existing channel-last selection for non-empty inputs.
• Avoids empty channel-last launches and supports conv-state write-back.
• Removes the DGX B200 waiver and associated NVBUG reference for test_kda_mixer_empty_prefill.
• No public API changes or unintended configuration changes identified.

QA Engineer Review

• Test-list change: removed test_kda_mixer_empty_prefill from tests/integration/test_lists/waives.txt.
• No test-db or qa/ files were modified.
• The test function was not changed. Its waiver was removed to enable coverage.
• CBTS coverage data is unavailable.

Verdict: needs follow-up

[nv-guomingz]

/bot run

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 72a6018c-5a7f-4475-9894-ef3cc13330f9

📥 Commits

Reviewing files that changed from the base of the PR and between 84a59d6 and 3310718.

📒 Files selected for processing (2)

• cpp/tensorrt_llm/thop/causalConv1dOp.cpp
• tests/integration/test_lists/waives.txt

💤 Files with no reviewable changes (1)

• tests/integration/test_lists/waives.txt

Included review availability: Your plan provides up to 12 included reviews per hour; 11 remain after this review.

---

Walkthrough

The forward kernel selection now routes zero-token inputs to the channel-major kernel. The related DGX_B200 test waiver and NVBUG reference were removed.

Changes

Causal convolution empty-input handling

Layer / File(s)	Summary
Zero-token kernel selection and waiver removal cpp/tensorrt_llm/thop/causalConv1dOp.cpp, tests/integration/test_lists/waives.txt	causalConv1dFwd routes zero-token inputs to the channel-major kernel regardless of strides. The DGX_B200 waiver for test_kda_mixer_empty_prefill was removed.

Estimated code review effort: 1 (Trivial) | ~5 minutes

Merge Risk: ⚪ Minimal · up to 33107

The change routes zero-token causal-convolution inputs to the channel-major kernel and re-enables the targeted empty-prefill test; no actionable merge-blocking risk remains beyond normal checks and review.

Suggested reviewers: bowenfu

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 2 functions across 1 files.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the NVBugs reference, fix type, and primary change: routing zero-token causal-convolution inputs to the channel-major kernel.
Description check	✅ Passed	The description explains the zero-token stride and aliasing failure, the channel-major routing fix, and the relevant test coverage. It is mostly complete, although several checklist items are not expl…
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

Full details: Description check

Explanation

The description explains the zero-token stride and aliasing failure, the channel-major routing fix, and the relevant test coverage. It is mostly complete, although several checklist items are not explicitly addressed.

• Fix all pre-merge checks with AI

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[tensorrt-cicd]

PR_Github #70879 [ run ] triggered by Bot. Commit: 3310718 Link to invocation

[tensorrt-cicd]

PR_Github #70879 [ run ] completed with state FAILURE. Commit: 3310718
/LLM/main/L0_MergeRequest_PR pipeline #58051 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

[nv-guomingz]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #70923 [ run ] triggered by Bot. Commit: 3310718 Link to invocation