All pull requests

Pull requests

• Sliding-window-attention suffix prefill mode

  #2030·ruiite opened Nov 2, 2026·

  ·

• Add fused Metal kernel for DSA sparse prefill

  #2027·vinceclifford opened Nov 2, 2026Contributor·

  ·

• perf(gemma4): add per_layer_config, KV-shared logits_to_keep prefill trimming, and MTP/MoE sanitization

  #2026·casualuser opened Nov 2, 2026·

  ·

• Support transformers_weights in config.json

  #2025·jundot opened Nov 1, 2026Contributor·

  ·

• Fix hardcoded GELU in "gpt_neox" and "gpt_bigcode"

  #2024·Kingjinsight opened Nov 1, 2026·

  ·

• Finalize the prompt cache for unpadded batches

  #2023·joshuaswarren opened Oct 30, 2026·

  ·

• perf(gated_delta): chunk-parallel differentiable training path for Ga…

  #2021·kru2710shna opened Oct 30, 2026·

  ·
  1

• Recover misordered closing brackets in json_tools parser

  #2020·arian262785-cell opened Oct 29, 2026·

  ·

• Add Mistral Large 4 text model

  #2018·jundot opened Oct 29, 2026Contributor·

  ·
  4

• LFM3 cache tests

  #2017·pierreprudh opened Oct 28, 2026·

  ·

• Add LFM3 Support

  #2016·techwritescode opened Oct 27, 2026·

  ·
  2

• Add Raon-1 mini inference support

  #2015·asynclee opened Oct 26, 2026·

  ·

• Fix MiniCPM RoPE theta default

  #2012·liuyunxian-yana opened Oct 26, 2026·

  ·
  1

• fix head dims rotated, only part gets rotated using partial_rotary_factor instead of all of them

  #2011·begcdn opened Oct 26, 2026·

  ·

• fix(gpt_oss): detect transformers [E, H, 2F] layout in sanitize() (fixes #1994)

  #2009·kru2710shna opened Oct 25, 2026·1 task done·

  ·
  1

• Fix quantized KV cache packed storage width

  #2008·sehejjain opened Oct 25, 2026·

  ·

• apply the checkpoint's sliding_window setting when no layer_types list is present

  #2007·begcdn opened Oct 25, 2026·

  ·

• Skip fully masked key tiles in Flint sliding-window prefill

  #2006·tc3oliver opened Oct 24, 2026·

  ·

• Fix Jamba MoE routing weights to match reference

  #2004·Knightkolla opened Oct 23, 2026Contributor·

  ·
  1

• fix(gemma3n): apply KV sharing without cache (#1674)

  #1999·Yu-Haikuo opened Oct 22, 2026·

  ·
  1

• fix: handle nested parameter tags in Qwen3 Coder parser

  #1996·rafiimam opened Oct 21, 2026·

  ·
  1

• Fix nested parameter tags in qwen3_coder tool parser

  #1993·axiom-of-choice opened Oct 19, 2026Contributor·

  ·
  1

  1

• Add NVIDIA Nemotron-Labs Dory model

  new_model
  #1988·pmichalas opened Oct 17, 2026·1 task done·

  ·
  4

• feat(qwen3_5): honour explicit layer_types over full_attention_interval

  #1987·hermitdave opened Oct 17, 2026·

  ·

• Add --loss-chunk-size to reduce LoRA training memory with large vocabularies

  #1983·spicysharkco opened Oct 16, 2026·

  ·