RTX PRO 6000 Heavy evaluation protocol v2¶
Point-in-time record, 2026-07-12. This round repaired the cross-model reasoning evaluation protocol, reran Mistral Small 4 and Nemotron 3 Super on the same RTX PRO 6000, and tested Poolside Laguna XS 2.1 NVFP4 through both vLLM and SGLang. Nemotron 3 Super remains the selected resident Heavy experiment. No router profile or production tier was promoted.
Raw protocol-v2 artifacts are under 2026-07-12-rtx-pro-6000-heavy-eval-v2-evidence/.
What changed in the evaluation protocol¶
The earlier fixed-total-token comparison could confuse a model that spent its
budget in a reasoning channel with a model that answered incorrectly. Protocol
v2 records an equal visible-answer allocation plus explicit reasoning
headroom, selects the model family's real control (reasoning_effort or
enable_thinking), repeats each item three times, and retains the full visible
answer, reasoning metadata, finish reason, per-attempt budget, and failure
class. A question passes when at least two of three attempts pass. The API
still receives one combined completion cap; the two allocations are evidence
intent, not a hard server-side partition.
The deterministic multiple-choice validator now accepts whitespace-tolerant
final markers such as FINAL = D. This fixed a real false negative in the
first diagnostic ARC run without introducing a model judge. The published runs
used the preserved unanchored validator specs. An independent offline check
confirmed that every passing output ended with its expected marker, so the
later anchored-validator repair does not change these scores. The exact
executed specs and their hashes are bound in
run-lineage.json.
Common local topology¶
| Field | Tested value |
|---|---|
| Host / GPU | Primary Node; one RTX PRO 6000 Blackwell 96 GB, sm_120 |
| Resident engines | vLLM nightly 0.23.1rc1.dev531+ga65f93fb2 |
| Served context / admission | 131,072 tokens; five sequences |
| Prefix cache | disabled for independent-prompt comparison |
| Repetitions / stable threshold | three per item; pass rate >= 0.66 |
| Visible answer allocation | 256 tokens |
The prior Heavy challenger run already established that both models complete 5/5 independent requests at concurrency five and pass their model-appropriate preflight. This round concentrates on repaired quality evidence rather than repeating that capacity result.
ARC-Challenge repeated sanity¶
This is a pinned five-row ARC-Challenge slice. It is a small reasoning-budget sanity check, not a capability leaderboard.
| Model / control | Reasoning headroom | Stable items | Passing attempts | Wall time |
|---|---|---|---|---|
Mistral Small 4, effort none |
0 | 2/5 | 6/15 | 27.91 s |
Mistral Small 4, effort high |
1,024 | 4/5 | 9/15 | 95.40 s |
Mistral Small 4, effort high |
2,048 | 5/5 | 15/15 | 97.46 s |
| Nemotron 3 Super, thinking off | 0 | 2/5 | 7/15 | 39.68 s |
| Nemotron 3 Super, thinking on | 1,024 | 5/5 | 15/15 | 69.91 s |
Both models need reasoning headroom for this slice. Nemotron reached 15/15 with half the headroom Mistral needed and completed the comparable perfect run 27.55 seconds sooner.
MMLU-Pro repeated multidomain sanity¶
The second benchmark is a pinned ten-row validation slice from
TIGER-Lab/MMLU-Pro,
covering ten categories and ten answer choices. It is deliberately short and
must not be reported as a full MMLU-Pro score.
| Model / control | Reasoning headroom | Stable items | Passing attempts | Median attempt | Wall time |
|---|---|---|---|---|---|
Mistral Small 4, effort high |
2,048 | 5/10 | 14/30 | 10.41 s | 384.74 s |
| Nemotron 3 Super, thinking on | 1,024 | 8/10 | 23/30 | 7.27 s | 260.91 s |
| Nemotron 3 Super, thinking on | 2,048 | 8/10 | 23/30 | 6.67 s | 317.88 s |
Nemotron's 1,024-headroom recipe is the useful operating point. Doubling its headroom did not improve a single item or attempt and added 56.97 seconds of wall time. Its two unstable items were computer science and engineering; all six attempts exhausted the reasoning budget without a visible answer. This is evidence for a larger budget follow-up on those tasks, not evidence that the answers would necessarily be correct.
Laguna XS 2.1 NVFP4 engine A/B¶
Laguna was selected as a current official challenger because its model card
advertises 33B total / 3B active parameters, a native 262,144-token window,
NVFP4 weights, and launch-day vLLM and SGLang support. The checkpoint revision
was pinned to 07133fb3df1cc3111478e24ee71a823a598c8c2f.
It did not produce trustworthy output on this RTX PRO 6000. The healthy-run
quality claims, the 262K skip-layer stall, and the trtllm_mha rejection are
operator-observed but incomplete: their response bodies or full logs were lost
when containers were recreated. They are not reproducible quality scores.
Retained logs independently substantiate only the 131K startup stall and the
forced-runner assertion:
- vLLM 0.23 nightly loaded at 262K with FP8 KV and reported capacity for 11.81 full windows, but preflight exposed corrupted text and 0/20 tool calls.
- Disabling FP8 KV on all 40 layers changed vLLM to FlashAttention. The retained 131K log stalled during cache profiling; the same 262K behavior is an operator-observed incomplete result.
- Poolside's Laguna-specific SGLang CUDA 13 image loaded the model and
initialized
SWARadixCache, but both the default and explicit checkpoint chat templates produced repetitive/off-topic output; preflight passed only short coding and structured JSON, while the 131K needle was empty and tool fan-out was 0/20. - SGLang's documented Blackwell
trtllm_mhaattention backend rejected sm_120 because that build supports it only on sm_100. Forcing the TRT-LLM MoE runner then crashed during warmup onrouting_method_type is not None.
The machine-readable failure summary is laguna-xs-21-nvfp4-sm120.failure.json. See also the reproduction record, bounded retained log excerpts, and image digests. This rejects the tested recipes, not the model on other Blackwell products. RadixAttention successfully initialized but cannot repair incorrect model execution; it only reuses compatible prefix state after correctness exists.
Source age and decision use¶
| Source | Observed / source date | Age class | Evidence | RTX PRO 6000 relevance | Decision impact |
|---|---|---|---|---|---|
| Poolside Laguna NVFP4 model card | observed 2026-07-12; model updated 2026-07-09 | current | official primary | Blackwell NVFP4 recipe, not workstation-specific | selected candidate and controls |
| vLLM Laguna recipe | observed 2026-07-12; updated 2026-07-02 | current | official primary | says NVFP4 is Blackwell-only and vLLM >=0.22 | established expected vLLM path |
| SGLang Laguna cookbook | observed 2026-07-12 | current | official primary | verified H200/B300/GB300, not sm_120 | supplied engine/image and exposed hardware gap |
| ARC-Challenge dataset | 2023-12-21 revision | historical stable dataset | official dataset | hardware-independent quality prior | supplied the five-item local sanity slice |
| MMLU-Pro dataset | observed 2026-07-12; updated 2026-05-02 | aging | official dataset | hardware-independent quality prior | supplied the ten-item local sanity slice |
URLs and revisions are preserved in source-registry.json.
Decision boundary¶
Nemotron 3 Super is the best currently measured Heavy experiment on this RTX
PRO 6000: it is stable at five admitted sessions, passes the Heavy correctness
gate, wins both repaired quality slices, and does not benefit from more than
1,024 reasoning-headroom tokens on this sample. It is restored healthy at
:39033. The recommendation remains human-gated and does not validate the
model's advertised one-million-token maximum.