Skip to content

Add DeepReinforce Ornith-1.0 to cookbook - #29404

Merged
zijiexia merged 2 commits into
sgl-project:mainfrom
averyjones4:ornith-dev
Jul 7, 2026
Merged

zijiexia merged 2 commits into
sgl-project:mainfrom
averyjones4:ornith-dev

Conversation

@averyjones4

@averyjones4 averyjones4 commented Jun 26, 2026 •

Copy link
Copy Markdown
Contributor

Summary

  • Add a DeepReinforce Ornith-1.0 cookbook page
  • Add deployment configs for Ornith-1.0 397B, 397B-FP8, 35B, 35B-FP8, and 9B
  • Add DeepReinforce overview card and logo asset
  • Local preview checked at /cookbook/autoregressive/DeepReinforce/Ornith-1.0

CI States

Latest PR Test (Base): ✅ Run #28901846551
Latest PR Test (Extra): ❌ Run #28901846313

@github-actions github-actions Bot added the documentation Improvements or additions to documentation label Jun 26, 2026

@gemini-code-assist gemini-code-assist Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request introduces a new cookbook page and configuration for deploying the DeepReinforce Ornith-1.0 model family using SGLang. It includes documentation, navigation updates, and deployment configurations for various model sizes (397B, 35B, 9B) and formats. Additionally, the _deployment.jsx snippet was updated to conditionally render the playground link based on the presence of playgroundFeatures in the configuration. There are no review comments, and I have no feedback to provide.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

@zijiexia zijiexia self-assigned this Jun 26, 2026
@zijiexia

zijiexia commented Jun 29, 2026 •

Copy link
Copy Markdown
Collaborator

Thanks for the contribution! Nice clean config and a thorough model table. mint validate passes and the HF links all resolve. A few things need to change before this can merge — three are blocking.

🔴 Blocking

  1. Port must be 30000, not 8000. The PORT/CURL_PORT placeholder defaults, the curl block, and the Python examples (base_url="http://localhost:8000/v1") all use 8000. Every other cookbook config (deepseek-v4, glm-5.2, laguna-m1, lfm2.5, unlimited-ocr) uses 30000 — please standardize.

  2. Logo asset doesn't meet the spec. cards/logos/ expects a 940×525 RGBA transparent PNG named <vendor>.png. The submitted deepreinforce-card.webp is a lossy VP8/YUV webp with no alpha channel (not transparent) and the wrong extension/name. Please export deepreinforce.png (940×525, transparent, icon-only) and git add -f it, then update the <Card img=...> in intro.mdx.

⚠️ Please address

  1. verified: true on all 9 cells is too strong. The config comment defines "verified" as "matches the model-card recipe," but in the cookbook verified means a recipe that was actually run/signed off. With zero benchmarks on the page and the FP8 READMEs still pointing at the non-FP8 repos (per your own comment), please don't mark cells verified unless they were actually run — especially 397B BF16 on a single H200 node.

  2. Flag order. --host/--port should come last (the Playground insert anchors assume this). Currently they precede --context-length / --mem-fraction-static / the parsers.

  3. FP8 should be the quant axis, not a variant. Modeling 397b-fp8/35b-fp8 as separate variants (with a single model quant) doubles the variant list and diverges from sibling configs (Laguna uses quant: bf16/fp8/nvfp4). Consider variants 397b/35b/9b × quant model/fp8.

  4. 397B BF16 memory headroom is tight. ~794 GB weights vs ~958 GB usable (8×141 GB × 0.85) leaves only ~17% for KV at --context-length 262144, below the usual 20–30% target — --mem-fraction-static 0.85 is aggressive here and risks OOM. Worth re-checking against an actual run.

Looks good otherwise — strategy id default matches convention, nav/card placement and frontmatter are correct, and validation is green. Happy to re-review once the blocking items are resolved.

@averyjones4
averyjones4 force-pushed the ornith-dev branch 2 times, most recently from 7ae5d6c to 5c6ad6d Compare July 1, 2026 00:41
@averyjones4

averyjones4 commented Jul 1, 2026 •

Copy link
Copy Markdown
Contributor Author

Thanks for the detailed review. I have addressed the comments in the latest push:

  1. Standardized the serving port to 30000 across the defaults, curl command, and Python examples.

  2. Replaced the logo asset with cards/logos/deepreinforce.png, exported as a 940×525 RGBA transparent PNG, and updated the Card img reference in intro.mdx.

  3. Removed the verified: false fields from the Ornith cells, also removed the Verified / Not Verified badge for this page. (And we confirm that the model of 397b(bf16) can be deployed on a single h200 node(8 gpus)).

  4. Moved --host and --port to the end of each launch command.

  5. Reworked the axes to use model size as the variant and quantization as BF16 / FP8, instead of modeling FP8 as a separate variant.

  6. Reduced --mem-fraction-static from 0.85 to 0.8 for the 397B cells. This leaves more KV/runtime headroom for the 397B BF16 H200 recipe.

…nced strategy

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@zijiexia
zijiexia merged commit 631213c into sgl-project:main Jul 7, 2026
82 of 86 checks passed
Chronostasys pushed a commit to MindLab-Research/sglang that referenced this pull request Aug 24, 2026
Co-authored-by: averyjones4 <averyjones4@users.noreply.github.com>
Co-authored-by: Zijie Xia <zijie.xia@radixark.ai>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

documentation Improvements or additions to documentation

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants