Docs: vLLM backend needs a vLLM nightly and supports Qwen3.5 only

README.md

@@ -31,7 +31,7 @@ Optional extras, combined as `pip install "decisionformer[serve,qlora]"`:
 |---|---|
 | `serve` | the `/v1/systemone` API server (FastAPI, uvicorn) |
 | `qlora` | bitsandbytes, for QLoRA |
-| `vllm` | the vLLM serving backend (a vLLM build from 2026-10-15 or later) |
+| `vllm` | vLLM, for the vLLM serving backend; Decision 3.0 models need its nightly build (see [Serve as an API](#serve-as-an-api)) |
 | `fla` | fast Gated DeltaNet kernels for Qwen3.5 backbones |
 | `data` | datasets, for Hub datasets (`hf:` inputs) and the example converters |
 
@@ -126,10 +126,13 @@ schemas are at `/docs`. Over-long requests are refused with HTTP 422, never trun
 runs at once; requests that arrive while it is busy are answered together, up to `--max-batch` (16) at a time.
 
 For code-readout checkpoints on Qwen3.5 (the Decision 3.0 layout), vLLM can run the model behind the same API.
-Other checkpoints fail with a clear error that points to the default backend:
+This needs a vLLM nightly build from 2026-10-15 or later (vLLM 0.32.0 on PyPI predates its decision-model support)
+and a CUDA GPU, Ampere or newer for non-causal checkpoints such as `vllm-sr/d3-edge`. Other checkpoints fail with a
+clear error that points to the default backend:
 
 ```bash
-pip install "decisionformer[vllm]"
+pip install "decisionformer[serve]"
+pip install -U vllm --pre --extra-index-url https://wheels.vllm.ai/nightly
 decisionformer serve vllm-sr/d3-edge --backend vllm
 ```
 

pyproject.toml

@@ -38,7 +38,7 @@ Dataset = "https://huggingface.co/datasets/FrontiersMind/deciformer-15k"
 [project.optional-dependencies]
 qlora = ["bitsandbytes>=0.45"]
 serve = ["fastapi>=0.110", "uvicorn>=0.29"]
-vllm = ["fastapi>=0.110", "uvicorn>=0.29", "vllm>=0.12"]  # decision-model support: vLLM nightly from 2026-10-15
+vllm = ["fastapi>=0.110", "uvicorn>=0.29", "vllm>=0.12"]  # Decision 3.0 models need a vLLM nightly from 2026-10-15 (0.32.0 predates it)
 fla = ["flash-linear-attention>=0.5"]  # fast Gated DeltaNet kernels for Qwen3.5/3.8 backbones
 data = ["datasets>=3"]
 dev = ["pytest>=8", "ruff>=0.6", "mypy>=1.11", "fastapi>=0.110", "uvicorn>=0.29", "httpx>=0.27"]

src/decisionformer/cli.py

@@ -279,7 +279,7 @@ def build_parser() -> argparse.ArgumentParser:
     serve.add_argument("--host", default="127.0.0.1")
     serve.add_argument("--port", type=int, default=8000)
     serve.add_argument("--backend", choices=("transformers", "vllm"), default="transformers",
-                       help="vllm: vLLM computes the readout (Qwen3.5/3.8 code-readout checkpoints only)")
+                       help="vllm: vLLM computes the readout (Qwen3.5 code-readout checkpoints; needs a vLLM nightly)")
     serve.add_argument("--max-batch", type=int, default=16,
                        help="requests arriving together are answered in one batch of up to this many")
     serve.add_argument("--api-key-env", default="DECISIONFORMER_API_KEY",

src/decisionformer/vllm_backend.py

@@ -2,11 +2,11 @@
 
 vLLM computes the answer-code logits; decisionformer renders the prompt (byte-identical to the transformers
 backend) and turns logits into answers with the checkpoint's per-type temperatures, so both backends answer the
-same way. vLLM's decision-model support covers code-readout checkpoints whose backbone is a bare Qwen3.5/3.8 model
+same way. vLLM's decision-model support covers code-readout checkpoints whose backbone is a bare Qwen3.5 model
 (`Qwen3_5Model` plus `decision_config.json` and `readout.safetensors`): decisionformer exports on those bases,
 Decision 3.0 (`vllm-sr/d3*`) and AutoJev-format exports such as FrontiersMind/Lumma-fev-27B. Anything else raises
-`UnsupportedBackend`; serve it with the default transformers backend. Needs a vLLM build with decision-model
-support (`pip install decisionformer[vllm]`, vLLM nightly from 2026-10-15 on) and a CUDA or ROCm GPU.
+`UnsupportedBackend`; serve it with the default transformers backend. Needs a vLLM nightly from 2026-10-15
+on (vLLM 0.32.0 on PyPI predates its decision-model support) and a CUDA or ROCm GPU.
 """
 
 from __future__ import annotations
@@ -34,7 +34,7 @@ def check_vllm_support(root: Path, config: dict[str, Any]) -> None:
     """Raise UnsupportedBackend unless vLLM's decision-model path can load this checkpoint as it is."""
     def refuse(reason: str) -> None:
         raise UnsupportedBackend(f"vLLM cannot serve {root}: {reason}. vLLM serves code-readout decision checkpoints "
-                                 "with a bare Qwen3.5/3.8 backbone (Qwen3_5Model). Use the transformers backend "
+                                 "with a bare Qwen3.5 backbone (Qwen3_5Model). Use the transformers backend "
                                  "(the default) for this model.")
 
     if not (root / READOUT).is_file() or (root / HEADS).is_file():