All pull requests

Pull requests

• Enable co-resident programs for the AMDGPU EP

  #2803·Aditya Lohia (aditya-dl) opened Oct 28, 2026Contributor·

  ·
  3

• Fix sustained decode throughput accounting in the Python E2E benchmark

  #2802·BBaditi opened Oct 28, 2026·

  ·

• [Do not merge] Qwen-3.8 Flash Next runtime on v0.17.0

  #2801·Wang, Zhenze (zz002) opened Oct 28, 2026·

  ·
  1

• Load models from named native file handles across platforms

  #2800·Mustapha Jaber (mustjab) opened Oct 28, 2026Contributor·

  ·
  4

• Refactor ASR postprocessor APIs and add generic processing dispatch

  #2799·Mason Corey (themason2011) opened Oct 27, 2026·

  ·
  1

• feat: add LFM3 support to the model builder and runtime

  #2798·Yuri Khrustalev (ykhrustalev) opened Oct 27, 2026Contributor·

  ·
  6

• [Speech] [Timestamp] Refactor opaque PostProcessor result access

  #2797·Mason Corey (themason2011) opened Oct 27, 2026·

  ·
  2

• Isolate native test lifetimes and private Apple telemetry delegates

  #2796·bmehta001 opened Oct 27, 2026Contributor·

  ·
  1

• [Do not merge] Keep LoRA graph inputs on device across pipeline stages

  #2795·Chang Liu (cliu1003) opened Oct 27, 2026Contributor·

  ·
  1

• [Do not merge] Feed sliding-window sequence lengths on AMDGPU

  #2794·Chang Liu (cliu1003) opened Oct 27, 2026Contributor·

  ·
  1

• Restore Mistral sliding window metadata in model builder

  #2793·lsmlhi_25 (FOWEPJF255) opened Oct 27, 2026·

  ·
  1

  2

• [Do not merge] Pin onnxruntime-extensions to main plus #1136

  #2792·Wang, Zhenze (zz002) opened Oct 27, 2026·

  ·

• enables webgpu ep benchmarking for genai engine

  #2787·Prathik Rao (prathikr) opened Oct 26, 2026Contributor·

  ·
  4

• [Qwen38] Fix Windows S=8 staged PLE lookup, add adaptive context length budget, and optimize long-context DFlash2 prefill

  #2786·Tianlei Wu (tianleiwu) opened Oct 26, 2026Contributor·

  ·

• [CUDA] Support batch_size > 1 and ragged batching in qwen38_flash_graph Engine

  #2785·Tianlei Wu (tianleiwu) opened Oct 26, 2026Contributor·

  ·

• [CUDA] Guard QMoE offload warmup, skip out-of-window DFlash2 prefill, and pin windowed decode graph columns

  #2784·Tianlei Wu (tianleiwu) opened Oct 26, 2026Contributor·3 tasks done·

  ·

• Move to scalesets for CICD

  #2783·eserscor opened Oct 26, 2026Contributor·

  ·
  1

• IndexerCache: fall back to a growing cache when the graph has no past_sequence_length

  #2782·Wang, Zhenze (zz002) opened Oct 26, 2026·

  ·
  1

• Skip the Qwen3.5 MTP head when the checkpoint has no mtp.* weights

  #2781·Eduardo Worrel (eduardoworrel) opened Oct 26, 2026·

  ·
  12

• Support text-only Qwen3_5ForCausalLM checkpoints in the model builder

  #2780·Eduardo Worrel (eduardoworrel) opened Oct 26, 2026·

  ·
  4

• Support fixed-shape Gemma4 vision preprocessing

  #2779·ananyaa-ms opened Oct 26, 2026·

  ·
  2

• [Flint] Fix BF16 KV-cache rewind and beam-reorder dispatch

  #2778·Rui Ren (rui-ren) opened Oct 25, 2026Contributor·

  ·
  1

• [CUDA] Support S=8 (7-draft DFlash2) verification with 2-bank compact state_update replay

  #2777·Tianlei Wu (tianleiwu) opened Oct 24, 2026Contributor·

  ·

• Add procedural Qwen 3.8 Flash builder

  #2776·Akshay Sonawane (apsonawane) opened Oct 23, 2026Contributor·

  ·
  9

• [WebGPU] Batch fixed-state copies at phase boundaries

  #2775·aciddelgado opened Oct 23, 2026Contributor·

  ·