test(bench): --no-thinking / --thinking-budget toggles for bench_isl_osl

[rsafier]

What

Adds --no-thinking and --thinking-budget N to bench/bench_isl_osl.py. Both inject chat_template_kwargs (enable_thinking, thinking_budget) into each request so streaming TTFT/TPOT can be measured with reasoning off or at a fixed budget.

Why

On thinking-default models (e.g. Holo, max_thinking_budget=768) reasoning tokens dominate TTFT, which swamps the prefill signal you actually want to measure. These toggles let the bench isolate prefill/decode from reasoning overhead.

Tooling-only; no engine changes.

---

Related PRs — independent, all target main

Now that #229 is merged, these three each cherry-pick cleanly onto current main with zero conflicts and are independent (no stacking) — review/merge in any order:

• feat(loader): native-BF16 dense loader for no-metadata Holo/Ornith checkpoints #250 — native-BF16 dense loader (no-metadata Holo/Ornith dense)
• test(bench): --no-thinking / --thinking-budget toggles for bench_isl_osl #251 — bench_isl_osl thinking toggles
• fix(loader): native-FP8 SSM shadowed on modelopt-mixed-precision 35B #252 — native-FP8-SSM shadowing fix (modelopt 35B)

Follow-ups (not yet opened): Tier-2 FlashInfer-prefill + decode-concurrency reconciliation vs #229 FI (needs a merge, not a clean pick); and closing older stacked PRs folded into #229 (e.g. #220 FlashInfer-GDN AOT) after a rebase-diff confirms no outstanding delta.

[SeedSource]

[BETA] Automated PR review — SeedSource AI pipeline

Advisory only, not a merge gate. AI-generated (internal risk-triage, Atlas serve gate, GLM-5.2 review), attributed per Atlas convention.

Triage

• Risk: low - serve-test-required (triage advisory): False - path-heuristic (authority): False
• Rationale: Benchmark script update for thinking control parameters; no production or quantization changes.

Serve gate

• Skipped - no serving-path files (docs/CLI/test/example). Zero GPU.

GLM-5.2 code review

Correctness: Sound. enable_thinking defaults to True, preserving prior behavior. chat_template_kwargs is the correct vLLM field for thinking controls. Warmup correctly inherits the toggles, ensuring measured runs aren't invalidated by a thinking-on warmup followed by thinking-off measurement.

Risk: Low. Benchmark-only script, no engine code touched. One concern: --no-thinking behavior depends on the server's chat template actually honoring enable_thinking — if the deployed template ignores it, the script will silently report "pure prefill TTFT" while actually measuring time-to-first-answer. No validation or warning if the response contains reasoning tokens when --no-thinking is set.

Quality issues:

• Line 240 print statement is cramped and hard to read; split across lines.
• thinking_budget=None in the signature should be Optional[int] = None for type clarity, though this is a bench script.
• No guard against --thinking-budget being set without --no-thinking being False (i.e., budget with thinking off is contradictory) — minor, server likely ignores it.

Verdict: Approve. Consider adding a sanity check that --no-thinking responses contain no reasoning tokens.

_{Pipeline: build+serve on GB10, GLM-5.2 via NVIDIA NIM, internal risk-triage classifier. External diff treated as untrusted input. Public repo.}