[TRTLLM-14814][chore] Add SA speculative-decoding eval config for Kimi K3

[brnguyen2]

Description

Suffix-automaton (SA) speculative decoding is supported for the Kimi K3
example, but examples/kimi_k3/ has eval configurations for the default and
block-reuse modes only. There is no example config for the SA path, so an
accuracy run of it has to be assembled by hand. This adds one.

• New examples/kimi_k3/eval_extra_llm_options_sa.yaml: the base
  eval_extra_llm_options.yaml plus the SA-specific keys — max_batch_size: 8,
  a matching CUDA-graph max_batch_size, overlap scheduler off, chunked
  prefill off, and
  speculative_config: {decoding_type: SA, max_draft_len: 2}.
• run_gsm8k_kimi_k3.sbatch: new --sa flag that selects that config and
  enables the speculative-decoding acceptance summary
  (TLLM_EVAL_SPEC_STATS) by default for that mode. --sa and --reuse
  are mutually exclusive; an explicit TLLM_EVAL_SPEC_STATS still wins.
• examples/kimi_k3/README.md: documents --sa and what the SA
  configuration changes.

Configuration/docs only; no library code changes.

The config keys were checked against the current args schema rather than
copied blindly: speculative_config resolves through the decoding_type
discriminated union to SADecodingConfig in
tensorrt_llm/llmapi/llm_args.py, max_draft_len: 2 satisfies its
validator, and global_pool_size is left unset so the
global_pool_size >= max_batch_size check does not apply. Every other key
is already present in the checked-in sibling configs. One deliberate
deviation from the base config: cuda_graph_config.enable_padding is left
at its default (off) rather than true, matching the configuration SA was
evaluated with; this is called out in a comment in the file.

Still missing for examples/kimi_k3/, out of scope here: there is no
example eval configuration for the DFlash speculative-decoding path, which
is scaffolding only at this point.

Test Coverage

No new automated tests: this adds an example YAML config, an sbatch flag,
and README text, none of which are exercised by CI. Verified locally that
the new YAML parses, that its keys and value constraints match the current
LlmArgs/SADecodingConfig schema, and that the modified sbatch passes
bash -n and selects the expected config file per mode. The accuracy claim
it enables (SA is lossless, so GSM8K should match the non-SA run within
noise) is what the config is intended to let users check.

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.

Dev Engineer Review

• Added SA speculative-decoding configuration for the Kimi K3 GSM8K evaluation.
• Added --sa mode with --sa/--reuse mutual exclusion.
• Preserved explicit TLLM_EVAL_SPEC_STATS environment settings.
• Updated README documentation.
• YAML and shell validation passed.
• No library code or public API changes.
• No test-list changes.

QA Engineer Review

No test changes.

[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: b1df7b3b-7d51-45c7-9923-ed664f237eb3

📥 Commits

Reviewing files that changed from the base of the PR and between d16d01f and cae41e3.

📒 Files selected for processing (3)

• examples/kimi_k3/README.md
• examples/kimi_k3/eval_extra_llm_options_sa.yaml
• examples/kimi_k3/run_gsm8k_kimi_k3.sbatch

---

Walkthrough

Adds suffix-automaton speculative decoding support to the Kimi K3 GSM8K evaluation launcher. The change adds SA-specific configuration, mutually exclusive launcher modes, acceptance-statistics control, and usage documentation.

Changes

Suffix-automaton GSM8K evaluation

Layer / File(s)	Summary
SA evaluation configuration examples/kimi_k3/eval_extra_llm_options_sa.yaml, examples/kimi_k3/README.md	Adds SA evaluation settings for parallelism, batching, scheduling, CUDA graphs, KV cache, and draft length. Documents evaluation limitations.
SA launcher integration examples/kimi_k3/run_gsm8k_kimi_k3.sbatch, examples/kimi_k3/README.md	Adds the --sa mode, rejects simultaneous --sa and --reuse, selects the SA configuration, and controls speculative-decoding statistics through TLLM_EVAL_SPEC_STATS. Documents the command and expected acceptance output.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Suggested reviewers: bowenfu

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely identifies the SA speculative-decoding evaluation configuration added for Kimi K3.
Description check	✅ Passed	The description explains the problem, solution, affected files, validation performed, test coverage, and checklist status.
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.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

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

[brnguyen2]

/bot run

[tensorrt-cicd]

PR_Github #64915 [ run ] triggered by Bot. Commit: cae41e3 Link to invocation

[tensorrt-cicd]

PR_Github #64915 [ run ] completed with state SUCCESS. Commit: cae41e3
/LLM/main/L0_MergeRequest_PR pipeline #52754 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]

/bot run

[tensorrt-cicd]

PR_Github #64954 [ run ] triggered by Bot. Commit: cae41e3 Link to invocation