Skip to content

SOURCE_ARCHIVES_NOTES.md

Latest commit

 

History

History
executable file
·
53 lines (42 loc) · 2.89 KB

SOURCE_ARCHIVES_NOTES.md

File metadata and controls

executable file
·
53 lines (42 loc) · 2.89 KB
Raw

Source archival notes — Laguna derivation evidence (retrieved 2026-07-23)

Anchor for all "before/after" judgments: r0b0tlab/laguna-s-2.1-nvfp4-sm121-vllm created 2026-07-22T02:21:00Z; first release v0.25.1-gb10-k7 published 2026-07-22T02:38:49Z; checkpoint-update release 2026-07-23T00:53:06Z (GitHub API, retrieved 2026-07-23).

X posts — Wayback save FAILED (login wall); saved copies below (search-index text, retrieved 2026-07-23)

@ivanfioravanti — https://x.com/ivanfioravanti/status/2079655857856434299

"Laguna S 2.1 performance on a single DGX Spark. Prefill 600-800 tok/s decode is around 15 tok/s on prose and 22-24 on code. Without speculation, decode sits at 13-14 tok/s on every engine we tried on the GB10, which is the memory-bandwidth ceiling for this model." Post date: not visible in index; corroborates the NVIDIA-forum numbers.

@sudoingX — https://x.com/sudoingX/status/2066935217194250356

"the results are in. two 128gb boxes on my desk, the nvidia dgx spark and the amd strix halo... ran them head to head on the exact same model..." (Snapshot id 2066935217194250356 — Spark vs Strix Halo head-to-head.)

@sudoingX — https://x.com/sudoingX/status/2050517565097824303

"a week with the dgx spark, here is what is on it and what i have measured so far... nvidia gb10 sm_121, 124 gb unified lpddr5x at 273 gb/s, cuda 13.0..."

@darvasch (DFlash 3x-on-code / slower-on-agent-reasoning claim)

NOT LOCATED by search on 2026-07-23. Claim survives only as the paraphrase in COMMUNITY_RESEARCH.md. Treat as UNVERIFIED; do not cite publicly.

@stevibe ("Spark 19.44, PRO6000 108, 4x5090 145")

NOT LOCATED by search on 2026-07-23. Treat as UNVERIFIED; do not cite publicly. (Similar cross-hardware comparisons exist from other authors.)

NVIDIA Developer Forums thread (Discourse JSON, retrieved 2026-07-23)

Thread: "Laguna S 2.1 Config & Benchmarks" https://forums.developer.nvidia.com/t/laguna-s-2-1-config-benchmarks/377663 Created: 2026-07-21T18:32:17Z (user serapis) — model-release-day thread.

Key dated receipts (all BEFORE r0b0tlab's 2026-07-22T02:21Z creation):

  • vr8vr8, 2026-07-21T18:57:30Z: "go for num_speculative_tokens = 7. I managed to get speed around 40-50 tokens/s" ← K=7 is community-first.
  • vr8vr8, 2026-07-21T19:14:18Z: c=2 + K=7 results on the eugr-derived recipe (bilikaz/spark-vllm-docker branch feat/laguna-s-2.1-nvfp4).
  • clawdiusmaximus, 2026-07-21T22:17:23Z: with the HF card's 15 spec tokens, "draft acceptance was consistently 0-15%... 6-15 were always at 0.0" ← the reduce-K-below-15 insight, community-first.
  • serapis (OP), 2026-07-21T18:32:17Z: DFlash acceptance ~8% on release config.

Laguna S 2.1 model release date

Poolside released Laguna S 2.1 on 2026-07-21 (MarkTechPost coverage dated 2026-07-21; forum thread same day). The entire tuning conversation—including r0b0tlab's repo—is release-week work.