Laguna S 2.1 and Laguna XS¶
Current status and review date¶
Decision snapshot
- Product role: Laguna S is a retained rollback-era qualified recipe; Laguna XS is a rejected compatibility path. Neither label reports a current live route.
- Selected or best-qualified configuration: Laguna S 2.1 NVFP4 with thinking disabled at 262,144 tokens; retained TP=1 and exclusive TP=2 evidence describe different measured lanes.
- Measured hardware: one RTX PRO 6000 for the historical TP=1 lane and two equal RTX PRO 6000 cards over PCIe for the exclusive TP=2 refresh.
- Evidence: Laguna S has
functional,capacity, andqualityevidence; Laguna XS hascompatibility-onlyandhistorical-invalidevidence. - Decision: Laguna S was a dated
rollback; the TP=2 refresh isno-promotion; Laguna XS isrejectedunder the tested recipes. - Important limitation: thinking-enabled Laguna S can exhaust its visible-answer budget, and some Laguna XS response bodies/full logs were not retained.
- Review dates: retained evidence through 2026-08-01; dossier-format review 2026-08-31.
Open the retained recipe registry or jump to the decision, known limitations, or dated evidence.
Review narrative¶
2026-07-12 — Laguna XS engine A/B rejection¶
Laguna XS was attempted through vLLM and SGLang on one RTX PRO 6000. The vLLM
FP8-KV lane exposed corrupted text and 0/20 tools; the non-FP8 fallback stalled
during cache profiling. SGLang loaded but produced repetitive or off-topic
output, missed the long-context and tool gates, and its documented
trtllm_mha backend rejected sm_120. Some operator-observed healthy-run
responses and full logs were lost when containers were recreated, so they are
not reproducible quality claims.
Outcome: rejected for the tested configurations, not a judgment about the
model on every Blackwell product.
2026-07-26 — Laguna S single-card qualification and promotion¶
Laguna S used vLLM 0.23.1rc1.dev1327+gf25953cc5, NVFP4 weights, FP8 KV,
262,144 context, and thinking disabled. Repeated protocol-v3 quality passed,
including 32K, 128K, and 240K retrieval, tools, session recall, unified diff,
and timeout triage. Thinking-enabled testing also showed an intermittent
failure that consumed the complete 4,352-token completion allowance and
returned no visible answer, so the qualified contract forced thinking off.
The dated campaign then recorded a human-approved Heavy promotion with
GPT-OSS Puzzle retained as rollback.
Outcome: the exact TP=1 disabled-thinking recipe qualified and was promoted in that dated campaign; this dossier does not claim it remains live.
2026-08-01 — Exclusive TP=2 refresh¶
The same exact Laguna S checkpoint ran on two equal PRO 6000 cards with pinned vLLM 0.25.1, NVFP4/FP8 KV, 262,144 context, one admitted request, no DFlash, and thinking disabled. Smoke, JSON, 30K retrieval, tools, extended tools, intelligence 6/6, session 3/3, and tools 3/3 passed. At 32K it measured 1.97-second TTFT, 15,134 effective prefill tok/s, and 70.9 tok/s decode. At a 231,457-token prompt it measured 31.85-second TTFT, 7,252 effective prefill tok/s, and 66.0 tok/s decode.
Outcome: qualified as a TP=2 no-promotion comparison lane; no production
alias or router profile changed.
Immutable identity¶
Laguna S 2.1¶
- Model:
poolside/Laguna-S-2.1-NVFP4revision07614121b31898586430f189d27a25a0be310843. - TP=1 served name:
laguna-s-2.1-nvfp4. - TP=2 served name:
laguna-s-2.1-nvfp4-tp2. - TP=1 runtime: vLLM
0.23.1rc1.dev1327+gf25953cc5, image referencevllm/vllm-openai:nightly-f25953cc59f9b4ba9b04b16228d2b86dcfbcbdb1. - TP=1 registry digest: Not retained; the engine-revision tag is not a registry digest.
- TP=2 runtime: vLLM 0.25.1, image
sha256:e4f88a835143cd22aee2397a26ec6bb80b3a4a6fe0c882bcbc63822904766089.
Laguna XS 2.1¶
- Model:
poolside/Laguna-XS-2.1-NVFP4revision07133fb3df1cc3111478e24ee71a823a598c8c2fin the retained July 12 evaluation. - Qualification: no qualified served recipe resulted.
- Retained runtimes: vLLM image
sha256:907377dddef392f6b679d9c071e1c33c3935b4dc993b61d0352e391a5319ff3eand SGLang imagesha256:0261d11a9bf5ef8041d68d55f0bac3f3375330e21353245e23ef461ffbe57da5.
Tested hardware and topology¶
Laguna S TP=1¶
- Measured: one RTX PRO 6000 Blackwell 96 GB on an isolated endpoint.
- Execution mode: single-card, five or fewer admitted requests depending on the workload; the promotion contract served thinking disabled.
Laguna S TP=2¶
- Measured: two equal RTX PRO 6000 Blackwell Max-Q cards.
- Execution mode: exclusive TP=2 over PCIe without NVLink, concurrency one.
- Comparability boundary: runtime, topology, admission, and workload differ from the TP=1 lane, so this is not a topology-only speed A/B.
Laguna XS compatibility probes¶
- Measured: one RTX PRO 6000, isolated candidate endpoints.
- Boundary: the failures apply to the exact tested vLLM/SGLang recipes.
Engine, quantization, KV, context, and concurrency recipe¶
Laguna S TP=1 rollback-era recipe¶
- Engine and image: vLLM
0.23.1rc1.dev1327+gf25953cc5, commit-tagged nightly image above. - Weights and KV: compressed-tensors NVFP4 and FP8 KV.
- Contract: 262,144 tokens; request-level
chat_template_kwargs.enable_thinking=false. - Recipe: single-card registry entry.
Laguna S TP=2 refresh¶
- Engine and image: pinned vLLM 0.25.1 image digest above.
- Weights and KV: NVFP4 and FP8 KV.
- Contract: TP=2, 262,144 tokens, one admitted request, no DFlash, thinking disabled.
- Recipe: TP=2 campaign registry.
Laguna XS rejected probes¶
- vLLM: FP8-KV and non-FP8 fallback probes only.
- SGLang: Laguna-specific CUDA 13 image with default and explicit template probes.
- Recipe: retained experimental Compose.
Evidence by measurement class¶
Laguna S TP=1¶
- Status:
functional,capacity, andquality. - Measured: retrieval passed at 32K, 128K, and 240K targets; repeated tools/session/intelligence checks passed. Short-output capacity completed 10/10 at c1 and 40/40 at c8, with 75.46 and 83.24 aggregate tok/s.
- Evidence: qualification finding and its linked raw artifacts.
Laguna S TP=2¶
- Status:
functional,capacity, andquality;no-promotion. - Measured: functional and repeated gates passed; 32K measured 1.97-second TTFT, 15,134 prefill tok/s, and 70.9 tok/s decode; 231,457 prompt tokens measured 31.85-second TTFT, 7,252 prefill tok/s, and 66.0 tok/s decode.
- Evidence: campaign finding, 32K artifact, and 240K artifact.
Laguna XS¶
- Status:
compatibility-onlyandhistorical-invalid; tested recipes rejected. - Measured: the retained negative evidence includes corrupted text, 0/20 tools, startup stalls, empty long-context output, and backend rejection.
- Limit: full response bodies/logs for some operator-observed attempts were not retained.
- Evidence: engine A/B finding and machine-readable failure.
Decision and promotion state¶
Retained or selected¶
- Laguna S TP=1: dated
rollback-era qualified recipe. Do not interpret this as a current live assignment. - Laguna S TP=2:
no-promotion; retained as a qualified comparison lane.
Rejected¶
- Laguna XS:
rejectedunder the tested vLLM and SGLang configurations; reconsideration requires fresh immutable identity and complete evidence.
Failures and gotchas¶
Reasoning and evidence limits¶
- Laguna S thinking: the qualified contract disables thinking because the enabled path intermittently exhausted the completion budget without a visible answer.
- Laguna XS retention gap: some response bodies and complete logs were not retained, so those attempts remain historical-invalid rather than quality results.
Runtime and topology limits¶
- Laguna XS FP8 KV: produced corrupted output and 0/20 tools.
- Laguna XS fallback: non-FP8 vLLM stalled; SGLang output remained
repetitive/off-topic and its documented
trtllm_mhapath rejected sm_120. - Cross-lane comparison: TP=1 and TP=2 results use different runtimes, admission, and workloads.