All pull requests

Pull requests

• Reject speculative fallback in DFlash graph-restore tests

  #2957·bmehta001 opened Nov 12, 2026Contributor·

  ·
  1

• Enable paired prefix reuse for full-attention DFlash

  #2956·bmehta001 opened Nov 12, 2026Contributor·2/2·

  ·

• Share initializers between the multimodal embedding and decoder sessions

  #2953·Pierluigi Urru (stwykd) opened Nov 12, 2026·

  ·
  6

• [CUDA] Improve performance of DFlash2 model packages in ORT GenAI

  #2949·kunal-vaishnavi opened Nov 11, 2026Contributor·

  ·
  27

• Xiao.legacy decoder builders

  #2946·xiaoh-amd (xiaoh-amd) opened Nov 11, 2026·

  ·
  1

• Add recompute preemption to the continuous-batching engine

  #2945·Gopalakrishnan Nallasamy (GopalakrishnanN) opened Nov 11, 2026Contributor·

  ·
  3

• Enable bounded prefix caching for native Bolt

  #2942·bmehta001 opened Nov 11, 2026Contributor·1/2·

  ·
  25

  

• Keep Qwen per-image vision views on their source device

  #2941·bmehta001 opened Nov 11, 2026Contributor·

  ·
  2

• Model builder: make Whisper's cross-attention KV caches per decoder layer

  #2940·Daisuke Majima (MLBoy) (john-rocky) opened Nov 11, 2026·

  ·
  3

• [Model Builder] Gemma: add the RMSNorm weight offset in FP32 and use the exact embedding scale

  #2939·Qiming Bao (14H034160212) opened Nov 11, 2026Contributor·

  ·
  8

• Support R=16 target prefill in Nemotron-BOLT

  #2936·BBaditi opened Nov 10, 2026·

  ·

• Report per-model support for external prefill chunking

  #2934·Sayan Shaw (sayanshaw24) opened Nov 10, 2026Collaborator·

  ·
  3

• feat: add LFM3-Audio speech input to the builder and runtime

  #2933·Yuri Khrustalev (ykhrustalev) opened Nov 10, 2026Contributor·

  ·
  21

• fix: build LFM3 from transformers' native model classes

  #2932·Yuri Khrustalev (ykhrustalev) opened Nov 10, 2026Contributor·

  ·
  10

• [Speech] Return ASR timestamp intervals as integer milliseconds

  #2930·Mason Corey (themason2011) opened Nov 9, 2026Contributor·

  ·
  4

• Add a reproducible coding performance benchmark

  #2928·Baiju Meswani (baijumeswani) opened Nov 9, 2026Collaborator·

  ·
  6

• Add an opt-in Qwen WebGPU export profile

  #2927·Baiju Meswani (baijumeswani) opened Nov 9, 2026Collaborator·

  ·
  2

• Emit signed rotary and video assets for packed Ling VL

  #2925·Baiju Meswani (baijumeswani) opened Nov 9, 2026Collaborator·

  ·

• Preserve Bolt verification arithmetic and enable one-token ordinary decode

  #2923·Rui Ren (rui-ren) opened Nov 9, 2026Contributor·3/3·

  ·
  9

• [CUDA] Complete Bolt target-only decoding and shared-weight review fixes

  #2921·Rui Ren (rui-ren) opened Nov 9, 2026Contributor·2/3·

  ·
  9

• Guard fixed-request DFlash2 padding behind compatible ORT policy

  #2920·Rui Ren (rui-ren) opened Nov 9, 2026Contributor·1/3·

  ·
  9

• Emit video processor assets for packed Qwen-family models

  #2918·Baiju Meswani (baijumeswani) opened Nov 9, 2026Collaborator·2/2·

  ·

• Add reusable Engine video inputs and model-owned vision preparation

  #2917·Baiju Meswani (baijumeswani) opened Nov 9, 2026Collaborator·1/2·

  ·

• Preserve Bolt target arithmetic across speculative verification boundaries

  #2916·Rui Ren (rui-ren) opened Nov 9, 2026Contributor·

  ·
  4

• WebGPU: test auxiliary int64 Cast session configuration

  #2915·Sushanth Rajasankar (sushraja-msft) opened Nov 9, 2026Contributor·

  ·