All pull requests

Pull requests

• 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·

  ·

• WebGPU: test auxiliary int64 Cast session configuration

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

  ·

• [Qwen38] Add IQ4_NL and Q8_0 GGUF PLE n-gram table runtime support

  model:qwen-3.8
  #2913·Tianlei Wu (tianleiwu) opened Nov 9, 2026Contributor·

  ·
  3

• Fix telemetry shutdown and unload lifetimes and Apple system dependencies

  #2912·bmehta001 opened Nov 8, 2026Contributor·

  ·
  4

• Test: consolidate Bolt

  #2911·Rui Ren (rui-ren) opened Nov 8, 2026Contributor·

  ·

• fix: omit state_window attribute when the window is disabled

  #2910·KVR Vishnu (VISHNU0906) opened Nov 8, 2026·

  ·
  1

  2

• Add native speculative block decoding runtime

  model:qwen-3.8
  #2909·Sayan Shaw (sayanshaw24) opened Nov 8, 2026Collaborator·

  ·

• [Qwen38] Consolidate Win32 GGUF mmap, unified context-tier adaptation, and offload runtime profiles (stacked on #2876, #2864, #2897)

  qmoe_offload
  #2907·Tianlei Wu (tianleiwu) opened Nov 8, 2026Contributor·

  ·

• Cover hybrid prefix rebuilding after checkpoint eviction

  #2904·bmehta001 opened Nov 7, 2026Contributor·

  ·
  1

  1

  

• Enable opt-in variable-width Engine decode graph capture

  #2903·bmehta001 opened Nov 7, 2026Contributor·

  ·
  1

  

• Add Qwen3-ASR to the model builder and runtime

  #2901·Daisuke Majima (MLBoy) (john-rocky) opened Nov 7, 2026·

  ·
  2

• Forward GenAI CUDA streams to plugin provider options

  #2900·Sunghoon Choi (hanbitmyths) opened Nov 7, 2026Collaborator·

  ·
  1

• Preserve Qwen-VL batch padding and per-row vision positions

  #2899·Sunghoon Choi (hanbitmyths) opened Nov 7, 2026Collaborator·

  ·
  4

• Add Ling-3.0-flash-VL export and Engine image inputs

  #2898·Baiju Meswani (baijumeswani) opened Nov 7, 2026Collaborator·

  ·

• Add Qwen3.8-Flash-Next GGUF package builder, IQ4_NL/Q8_0 PLE n-gram table support, and precision options

  model:qwen-3.8
  #2897·Tianlei Wu (tianleiwu) opened Nov 6, 2026Contributor·

  ·
  19

• Recreate Qwen38 decode state on Engine owner thread

  model:qwen-3.8
  #2896·Sayan Shaw (sayanshaw24) opened Nov 6, 2026Collaborator·

  ·
  8

• WebGPU: forward Dawn cache during allocator initialization

  #2893·Sushanth Rajasankar (sushraja-msft) opened Nov 6, 2026Contributor·

  ·
  5

• [3/3] Restore CUDA graph replay for native Bolt partial-acceptance rounds

  model: nemotron-bolt
  #2890·Rui Ren (rui-ren) opened Nov 6, 2026Contributor·

  ·
  1

• Support runtime-selected chunk-local logits

  #2889·Sunghoon Choi (hanbitmyths) opened Nov 6, 2026Collaborator·

  ·
  48

• fix: refuse images or audio in a later SetInputs call

  #2888·Yuri Khrustalev (ykhrustalev) opened Nov 6, 2026Contributor·

  ·
  4

• Model builder: read the LayerNorm epsilon from configs that don't use rms_norm_eps

  #2887·Qiming Bao (14H034160212) opened Nov 6, 2026·

  ·
  2

• Model builder: escape % in the dflash2_precision help so --help works

  #2886·Daisuke Majima (MLBoy) (john-rocky) opened Nov 6, 2026·

  ·
  1

• [Model Builder] Gemma 3: support Transformers 5 rope_parameters and unscaled local RoPE caches

  #2884·Qiming Bao (14H034160212) opened Nov 6, 2026·

  ·
  2

  2