Issues

• benchmark_e2e.py on Windows: timed runs can be power-throttled (decode 1.9–2.0× slower)

  platform:windows
  microsoft/onnxruntime-genai#2914 · Filter by author john-rocky opened on Nov 9, 2026

• Model builder writes state_window even when it is 0, so its models do not load on ONNX Runtime 1.28 (the version windowsml[with-ort] installs)

  microsoft/onnxruntime-genai#2902 · Filter by author john-rocky opened on Nov 7, 2026

• Contribution offer: Qwen3-ASR in the model builder and runtime (#1968)

  microsoft/onnxruntime-genai#2885 · Filter by author john-rocky opened on Nov 6, 2026
  1

• Model builder emits CPU GroupQueryAttention with fused rotary for head sizes the CPU kernel refuses

  ep:DML
  microsoft/onnxruntime-genai#2833 · Filter by author JohnnyTeutonic opened on Oct 31, 2026

• Model builder: Mistral sliding_window dropped (no layer_types, every layer full attention), regression since 0.15.2

  microsoft/onnxruntime-genai#2832 · Filter by author JohnnyTeutonic opened on Oct 31, 2026
  1

• Model builder: Gemma 3 sliding layers inherit the global layers RoPE linear scaling (rescale_factors)

  microsoft/onnxruntime-genai#2831 · Filter by author JohnnyTeutonic opened on Oct 31, 2026

• Feature request: save and restore a generator's KV cache state (export/import past key values)

  microsoft/onnxruntime-genai#2830 · Filter by author dittops opened on Oct 31, 2026

• Guidance: which appended tokens reach the grammar depends on the previous call

  microsoft/onnxruntime-genai#2817 · Filter by author ykhrustalev opened on Oct 30, 2026
  1

• Model builder: CPU fp32 models with head_size % 16 != 0 build but fail at the first token (fused rotary in GroupQueryAttention)

  ep:DML
  microsoft/onnxruntime-genai#2791 · Filter by author JohnnyTeutonic opened on Oct 27, 2026

• Model builder: Mistral's sliding window is no longer applied (local_window_size -1 on every layer)

  microsoft/onnxruntime-genai#2790 · Filter by author JohnnyTeutonic opened on Oct 27, 2026

• Model builder: Gemma 3's sliding-window layers get the global layers' linear RoPE scaling

  microsoft/onnxruntime-genai#2789 · Filter by author JohnnyTeutonic opened on Oct 27, 2026
  1

• Model builder: Gemma 3 fails with KeyError 'rope_type' under transformers 5

  microsoft/onnxruntime-genai#2788 · Filter by author JohnnyTeutonic opened on Oct 27, 2026

• Support Liquid AI d1 decision models

  microsoft/onnxruntime-genai#2743 · Filter by author ykhrustalev opened on Oct 16, 2026
  2

• [Model Support] poolside Laguna 2.1 (LagunaForCausalLM / laguna)

  microsoft/onnxruntime-genai#2704 · Filter by author haozhiii opened on Oct 10, 2026

• [Model Support] StepFun Step-3.5-Flash (Step3p5ForCausalLM / step3p5)

  microsoft/onnxruntime-genai#2703 · Filter by author zoe226 opened on Oct 10, 2026

• [Model Support] NVIDIA Nemotron 3 Super 120B A12B (NemotronHForCausalLM / nemotron_h)

  microsoft/onnxruntime-genai#2702 · Filter by author zoe226 opened on Oct 10, 2026

• Allow hybrid prefixes to rebuild after their last checkpoint is evicted

  microsoft/onnxruntime-genai#2661 · Filter by author bmehta001 opened on Oct 1, 2026
  1

• Design: Whisper-compatible timestamp decoding primitives for ORT GenAI

  microsoft/onnxruntime-genai#2659 · Filter by author joelst opened on Sep 30, 2026

• Allow comments in genai_config.json

  enhancement
  microsoft/onnxruntime-genai#2653 · Filter by author baijumeswani opened on Sep 30, 2026

• Engine::Run hangs on the second uncached request with DFlash2 + CUDA graphs + max_scheduled_tokens=512

  bugep:CUDA
  microsoft/onnxruntime-genai#2640 · Filter by author tianleiwu opened on Sep 28, 2026
  1

• Feature request: grammar fast-forward support in Engine

  microsoft/onnxruntime-genai#2636 · Filter by author ShahVandit opened on Sep 28, 2026
  1

• OnnxRuntimeGenAIChatClient duplicates the last token when JSON-schema guidance completes the generation

  microsoft/onnxruntime-genai#2633 · Filter by author juanluelguerre opened on Sep 26, 2026

• Support Qwen3.8-Flash-Next (Qwen4ExpForConditionalGeneration) in Model Builder

  microsoft/onnxruntime-genai#2598 · Filter by author zoe226 opened on Sep 20, 2026
  2

• PagedAttention: CUDA illegal memory access on a full-context request when max_batch_size>=2 and max_scheduled_tokens=2048

  ep:CUDA
  microsoft/onnxruntime-genai#2583 · Filter by author tianleiwu opened on Sep 16, 2026
  1

  

• Any plans for a WebAssembly build?

  microsoft/onnxruntime-genai#2536 · Filter by author dariyooo opened on Sep 5, 2026