test(qwen): probe chunked-prefill validation gap

[chaofengw-nv]

Background

This Draft PR is an intentional mutation experiment. The Qwen long-context E2E case validates prefill receipts across 513 chunks, while semantic parity is covered only by a shorter case. This PR tests whether premerge can reject a semantic corruption that begins after the second prefill chunk.

Exit Criteria

• Record the exact-head TRTMC Internal CI / Automated premerge gate result.
• If premerge passes, treat that as evidence of a false-green validation gap.
• If premerge fails on semantic correctness, treat that as evidence that an existing gate covers the mutation.
• Do not merge this PR under any outcome.

Implementation

• Deliberately replay the second 64-token input chunk for the third and every later prefill launch.
• Preserve the prompt-token count, launch count, maximum chunk size, KV-cache capacity, and generated-token count observed by the existing runtime receipt checks.

Change categories

• Model or runtime behavior
• Public API
• ABI
• Bundle or artifact format
• Dependencies
• Documentation only
• CI or developer tooling

Validation

Commands and Results

• clang-format --dry-run --Werror families/qwen/runtime/pipeline.cpp: passed.
• pytest families/qwen/tests/test_e2e.py -q: 3 passed, 12 skipped; GPU E2E is intentionally left to exact-head premerge CI.

Hardware, Environment, and Revisions

• Repository head: aa75b06e022d1fbe00f209affd4207c0911af72e.
• Local validation: Linux CPU environment with the repository Python environment and system clang-format.
• Target E2E: Qwen/Qwen3-0.6B at the manifest-pinned checkpoint revision, BF16 native KV cache.

Not Run / Remaining Gaps

• No target-GPU E2E was run locally on this PR head; the experiment requires exact-head premerge CI to determine whether the intentional semantic corruption is detected.

Contributor Self-Review

• I have completed a self-review of this change.

Notes For Future Readers

• Start review at QwenTextGenerationPipeline::run_prefill_batched.
• This change intentionally corrupts model behavior and must never be merged. Close the PR after the CI observation is recorded.

Risk level

• Low
• Medium
• High

High risk by design: the mutation silently substitutes stale prompt tokens after the second prefill chunk. The Draft state and explicit do-not-merge exit criterion are required safeguards.

[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

---

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

[chaofengw-nv]

Probe result

Confirmed false green. Do not merge this PR.

• Tested head: aa75b06e022d1fbe00f209affd4207c0911af72e
• Deliberate defect: prompts requiring at least three 64-token prefill chunks replay the second chunk from the third chunk onward, corrupting long-prompt semantics while preserving chunk/KV telemetry.
• Exact-head result: TRTMC Internal CI / Automated premerge gate = PASS.

This demonstrates that the current premerge coverage can accept a Qwen chunked-prefill semantic regression. The PR remains Draft and must be closed, not merged, after the evidence is captured.

[github-actions]

This is an automated Internal CI result; no review from an individual maintainer is requested.

TRTMC Protected CI result
=========================

Status: FAILED
Pull request: #1324
Head commit: 453245c91faa9386dd1c80dff78f3574ce42b4a8
Reason: Automated internal CI failed; details withheld

Protected failure details are not transferred to the public repository.

Open the public Source Actions run from the automated status link above.

[chaofengw-nv]

Final negative-control result

The experiment is complete. Do not merge this PR.

• Current head: 453245c91faa9386dd1c80dff78f3574ce42b4a8
• Public Stable Community CI: PASS
• Public TRTMC Internal CI / Automated premerge gate: FAIL, as intended after adding the third-chunk semantic oracle
• Clean test-only PR test(qwen): require third-chunk semantic parity #1337 at 5f47cc4b1c002d97ceb1f6fc2610bec92a2b59af: both public statuses PASS

This PR contains deliberate runtime corruption and is being closed as preserved negative-control evidence, without merge.