diff --git a/recipes/glm-5-3-flash-nvfp4-on-2x-dgx-spark-atlas-rc4-2-tp2-dflash2-1m/recipe.yaml b/recipes/glm-5-3-flash-nvfp4-on-2x-dgx-spark-atlas-rc4-2-tp2-dflash2-1m/recipe.yaml index ff13c0f1d..eecf64f4d 100644 --- a/recipes/glm-5-3-flash-nvfp4-on-2x-dgx-spark-atlas-rc4-2-tp2-dflash2-1m/recipe.yaml +++ b/recipes/glm-5-3-flash-nvfp4-on-2x-dgx-spark-atlas-rc4-2-tp2-dflash2-1m/recipe.yaml @@ -187,3 +187,4 @@ provenance: - 'https://github.com/Enntity/sparkglm/blob/main/bench/RIGMARK.md' - 'https://github.com/othexmr/GLM-5.3-Flash-NVFP4-2x-4x-DGX-Sparks-RiNGSiDE' notes: "Submitted 2026-10-13 from lar-scout ingest card t_6e8bdfa6. Repo @ 3c7d67146929f350ed75fcdfd5e5349db843b483 (verified via git ls-remote + clone; HEAD = RC4.2 merge PR #47). ENGINE CORRECTION: the serving engine is ATLAS (Enntity/atlas fork, sparkglm/atlas-20261012-rc4.2-layered @ dfb9c26a, tree 89e392a3, install/atlas-source.json; image ghcr.io/enntity/atlas-sparkglm built from that commit + the install/ tree), NOT vLLM — the repo's vLLM lane is retired at tag vllm-final ('Atlas is now the only engine'); the RiNGSiDE vLLM TP2 column is the external reference, not this recipe. Launch argv VERBATIM from install/profiles/4x1m.json server_argv @ 3c7d671 (serve.py resolves ${model}/${drafter} and appends --rank/--master-addr/--master-port/--bind=127.0.0.1/--port=8893/--model-name=glm-5.3-flash-atlas); profile env (~220 keys) transcribed verbatim into the same file. Hardware: hw.card family stem dgx-spark + device_count 2 resolves against seed slug dgx-spark-x2 (migration 0066 pre-POST, 0064/0065 convention) — first 2-node Spark recipe on the site. Weights: nvidia/GLM-5.3-Flash-NVFP4 @ 423acf37 (start.sh pin; HF HEAD since da920bb0) + drafter incoai/GLM-5.3-Flash-DFlash2 @ 7d74cdd8 (start.sh pin; bf582e4e = drafter repo's current HEAD). Licenses: repo AGPL-3.0-only; weights MIT (HF tags); drafter CC BY-NC-ND 4.0 NON-COMMERCIAL (card said CC BY-NC — ND is the published term). Perf rows TRANSCRIBED from the repo's raw receipts results/2026-10-12-rc4/raw/ (SHA256SUMS bundle): RigMark cells per bench/RIGMARK.md on the 8x128k profile, single-stream rows on production 4x1m; RiNGSiDE comparison figures are the repo's reference numbers, not re-measured. Other repo receipts: greedy c1==c4 4/4 (rc4_on.json), logprob hash 246cd1331b41, arithmetic 491/500 + needle 24/24, strict-96-key JSON 39 s @ 59 tok/s. Drivers 580.173.02 / 580.178.04 (carveout-validated, docs/INSTALL-MANUAL.md). No dupe: zero 2-Spark recipes and zero nvidia-NVFP4 GLM lanes on the site at ingest." +max_context_len: 1048576