[https://nvbugs/6451032][fix] Reserve extra dflash context slot for dummy requests

[amukkara]

Summary by CodeRabbit

• Bug Fixes

  • Improved speculative decoding handling for unknown or temporary requests by assigning them to a dedicated scratch slot, preventing incorrect reuse of the first active slot.
  • Increased internal slot capacity to support safer request mapping during warmup, padding, and context processing.

• Tests

  • Removed waivers for DFlash Qwen3.5 integration tests, allowing both parameterized scenarios to run normally.

Description

Problem: DFlash keeps a per-request context K/V cache indexed by a slot. With cuda graph padding enabled, the padding requests defaulted to slot 0 corrupting whichever real request happened to occupy slot 0. This caused non-deterministic acceptance-length collapse (one request in a batch dropping from ~3.0 to ~0.55) and made test_dflash_qwen3_5_4b flaky.

Solution: Reserve one extra scratch slot for dummy requests. _ctx_len and the K/V buffers are now sized max_batch + 1, with _dummy_slot = max_batch. Real requests still only draw slots 0..max_batch-1.

Test Coverage

unittest/_torch/speculative/hw_agnostic/test_dflash.py::test_dflash_qwen3_5_4b
Increase acceptance length threshold in dflash unit tests to catch future regressions early.

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

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

• Test cases are provided for new code paths (see test instructions)

• If PR introduces API changes, an appropriate PR label is added - either api-compatible or api-breaking. For api-breaking, include BREAKING in the PR title.

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

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

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

GitHub Bot Help

To see a list of available CI bot commands, please comment /bot help.

[coderabbitai]

📝 Walkthrough

Walkthrough

DFlash speculative decoding now uses an additional dummy slot for unknown request IDs, including warmup and CUDA graph padding. Context-length and K/V buffers accommodate the extra slot, and two DFlash integration test waivers are removed.

Changes

DFlash dummy-slot handling

Layer / File(s)	Summary
Reserve dummy-slot buffers tensorrt_llm/_torch/speculative/dflash.py	DFlash reserves an extra slot at max_batch and allocates context-length and K/V buffers for max_batch + 1 slots.
Route unknown requests to dummy slot tensorrt_llm/_torch/speculative/dflash.py, tests/integration/test_lists/waives.txt	Request mappings during preparation and post-prefill reconstruction use the dummy slot for missing IDs, and two DFlash test waiver entries are removed.

Estimated code review effort: 3 (Moderate) | ~15 minutes

Suggested reviewers: zhaoyangwang-nvidia

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
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.
Title check	✅ Passed	The title follows the required ticket/type format and clearly describes the DFlash dummy-slot fix.
Description check	✅ Passed	The description includes the problem, solution, test coverage, and checklist completion, so it is mostly complete.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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