[OthmanAdi]

A trained Qwen3 code-readout checkpoint currently has to use DecisionFormer's Transformers server because the existing vLLM adapter accepts the bare Qwen3.5/3.8 decision layout. This adds an explicit export to Qwen3ForSequenceClassification and a matching vLLM pooling path.

decisionformer export MODEL OUTPUT --format sequence-classification copies the merged backbone and independently trained answer matrix into standard model.* and score.weight tensors. It preserves the tokenizer, prompt, code order, context limit and calibration metadata. The adapter requests raw classifier logits, slices to the active options, then applies the saved question-type temperature and existing response formatting.

The first conversion scope is merged, unquantized, causal Qwen3 with one code readout and last-token pooling. Exports use float32 weights, which approximately doubles backbone storage relative to 16-bit checkpoints. Existing canonical checkpoints and the Qwen3.5 serving path are retained. Incomplete weight loads are refused, and failed exports are staged and cleaned before publication without overwriting an existing destination.

Validation:

• Full Linux CPU suite: 309 passed, 13 skipped. Ruff and git diff --check pass.
• Actual vLLM CPU 0.32.0, PyTorch 2.13.0 and Transformers 5.20.2: a locally trained tiny Qwen3 checkpoint completed 16 decisions across four requests, including choice, yes/no, score, mixed lengths, JSON and empty states. Maximum active-logit difference was 6.71e-8, probability difference 2.15e-8, with no prompt or choice mismatches against the FP32 reference.
• The separate opt-in actual-engine test also passed. Offline tests cover independent HF loading, learned-head preservation, padding, active-option normalization, per-type temperatures, single/sharded readout files, missing weights and failed writes.
• Four Windows distributed-launcher tests fail before model loading because that PyTorch build lacks libuv support. The same launcher failure reproduces on unchanged main. The full suite above ran in Linux.

The guide includes reproduction commands and the pinned CPU container setup. examples/compare_classifier.py prints reference/exported answers and exits nonzero on drift. DECISIONFORMER_RUN_VLLM=1 enables the actual-engine test.

These checks establish the tested CPU interoperability path. GPU execution, model accuracy and throughput improvements were not measured, and no acceleration claim is made.

[Rohit7sali]

We are running some tests to check the PR, we will update soon.