robotlegs-qwen3.5-4b-onnx

Robotlegs Qwen3.5 4B ONNX bundles

These are immutable, ready-to-load text-generation bundles for Robotlegs. They were exported offline from Qwen/Qwen3.5-4B revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a with ONNX Runtime GenAI 0.13.1. Each provider directory contains its exact source and tool provenance.

• cpu/int4-rtn-block-32-acc-level-4 targets the CPU execution provider.
• cuda/int4-rtn-block-32 targets the CUDA execution provider.

The graph adapter preserves text input_ids, maps the external two-dimensional text position contract onto Qwen3.5's internal three-axis positions, expands an unsupported RMS-normalization contrib operator into standard ONNX operations, and installs a text-only ChatML template with thinking disabled. Robotlegs does not perform any conversion, quantization, optimization, or graph rewriting at application startup.

The CPU bundle passed a complete ONNX Runtime GenAI load and constrained generation smoke test. CUDA graph execution must be accepted on a CUDA-capable host; an export performed without GPU access is not execution evidence.