[Xunzhuo]

Overview

Add the recipe for vllm-sr/d3-turbo, next to the six Decision 3.0 recipes in recipes/models/vllm-sr/. It depends on #62584 (d3-turbo's prefix_bidir attention), which is still a draft: this PR stays a draft until #62584 merges, after which every nightly includes the support the recipe describes.

Claims

• vllm serve vllm-sr/d3-turbo with no flags (pooling runner, up to 32,768 tokens per question, the checkpoint's validated length) and the opt-in Long context feature (--max-model-len 131072).
• /v1/systemone examples for a text state with several questions, one image, several images and a video, the output shape (one probability per option), and the offline LLM.classify flow with the state block size that prefix_bidir needs.
• The limits (32,768 tokens per question by default, up to 64 questions, up to 255 options), what is not supported yet (Box and Track questions, images inside questions or options, one read of the state per request), and the measurements, all taken on AMD Instinct MI325X. meta.hardware lists only mi325x: verified; NVIDIA is not measured.

Validation

• node scripts/validate-source.mjs --source vllm --dir <this checkout>/recipes from vllm-project/recipes @ 8d3203a5: ✓ vllm recipes ... are valid against the renderer (216 models); build-recipes-api.mjs renders vllm-sr/d3-turbo with vllm serve vllm-sr/d3-turbo --tensor-parallel-size 1 on mi325x, as for the other Decision 3.0 recipes.
• Every command and example in the guide ran as written on one AMD Instinct MI325X (ROCm 7.2), with the ROCm nightly image vllm/vllm-openai-rocm:nightly-rocm72-e68efc685ab1e6f1eda004ced7157e3ca411da2f and [Model] Support Decision 3.0 prefix_bidir attention (vllm-sr/d3-turbo) #62584's Python files on vllm-sr/d3-turbo @ 689b3af1 (v1.1):
  • vllm serve vllm-sr/d3-turbo answered its first request 270 s after launch with empty caches, the 52 GB download included; /v1/models reports max_model_len 32768.
  • The four /v1/systemone requests returned the answers shown in the guide; through the checkpoint's own runtime (AutoModel.from_pretrained(..., trust_remote_code=True), transformers 5.17.0) the same requests gave the same decisions, with a largest probability difference of 0.028.
  • The offline script printed {'No / false': 0.00089, 'Yes / true': 0.99911}; /v1/systemone gave 0.99911 for the same question.
• The parity, latency and throughput numbers in the guide's Measured section are [Model] Support Decision 3.0 prefix_bidir attention (vllm-sr/d3-turbo) #62584's (its Validation section has the tables).
• pre-commit run --files recipes/models/vllm-sr/d3-turbo.yaml: all hooks pass.

Details

• context_length is 131,072, the model card's long-context limit; the checkpoint's position limit is larger, and vLLM's default is the 32,768-token validated length ([Model] Support Decision 3.0 prefix_bidir attention (vllm-sr/d3-turbo) #62584 applies decision_config.json's validated_length).
• The command builder's cURL and Bench buttons send chat completions, which a decision model does not serve, so the guide points to /v1/systemone, as the other Decision 3.0 recipes do.
• The other six Decision 3.0 recipes are unchanged; this recipe lists them in related_recipes.
• Not a duplicate: no recipe or PR for d3-turbo exists in vllm-project/vllm or vllm-project/recipes (searched for d3-turbo).
• AI assistance was used for the recipe, the measurements and this description. The PR stays a draft until the submitter has reviewed every line.

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[mergify]

Recipe preview: https://vllm-recipes-git-preview-vllm-pr-62585-inferact-inc.vercel.app