vLLM-SR Decision 3.0 (local)

The d3 family (Apache-2.0) runs locally via Hugging Face AutoModel + trust_remote_code=True and model.system_one(). OpenTryOn registers six tiers on decide:

Registry id	Default HF repo
d3-turbo-local	vllm-sr/d3-turbo
d3-flash-local	vllm-sr/d3-flash
d3-mini-local	vllm-sr/d3-mini
d3-nano-local	vllm-sr/d3-nano
d3-lite-local	vllm-sr/d3-lite
d3-edge-local	vllm-sr/d3-edge

Hosted twins (free API): Decision 3.0 (hosted).

Requirements

pip install opentryon[local]
pip install "transformers==5.17.0" torch torchvision pillow opencv-python-headless safetensors accelerate
pip install flash-linear-attention  # optional: faster linear-attention kernels

The shared opentryon[local] pin uses older transformers; use a separate conda env (same pattern as Liquid d1). d3-turbo is ~27B — plan VRAM accordingly; start with d3-edge-local or d3-lite-local on smaller GPUs.

Overrides: D3_MODEL_ID, or D3_MODEL_ID_TURBO, etc.

CLI

Question types include box and track — see the hosted d3 doc.

opentryon decide --model d3-turbo-local \
  --questions attrs.json --image lookbook.jpg --device cuda

opentryon decide --model d3-mini-local \
  --questions q.json --state "SKU copy for wool blazer" --long-context

Python

from tryon.models.d3 import D3LocalAdapter

adapter = D3LocalAdapter(variant="turbo")
out = adapter.decide(
    questions={"formal": {"type": "noul", "instructions": "Formal wear?"}},
    state="Editorial caption: ivory trouser suit",
    image="shot.jpg",
)

MCP

Local models unlock without an API key. Tools: decide_d3_turbo_local, etc.