All pull requests

Pull requests

• Fix Metal matmul on single-head and padded-stride views

  #4033·Cassian433 opened Oct 11, 2026·

  ·
  1

• candle-kernels: fix int overflow in softmax, rmsnorm and layernorm

  #4032·lit0ne opened Oct 11, 2026·

  ·

• Fix Metal try_clone sharing the source buffer

  #4030·muzammil35 opened Oct 9, 2026·

  ·
  1

• Fix double-paren links in quantized Qwen example READMEs

  #4028·pratikgx opened Oct 8, 2026·

  ·

• feat: support all LFM2.5 text models, including the 8B-A1B MoE

  #4026·ykhrustalev opened Oct 7, 2026·

  ·

• chore(deps): update cutile requirement from =0.3.1 to =0.4.0

  dependenciesrust
  #4025·dependabot[bot] opened Oct 5, 2026Bot·

  ·

• fix(onnx): reduce extrema over all dimensions for empty axes

  #4024·hulkbig opened Oct 5, 2026·

  ·
  1

• Add signed step ranges to tensor indexing

  #4022·tianrking opened Oct 4, 2026·

  ·

• Bail in scatter_set/scatter_add_set when indexes share storage with self

  #4019·stanleys12 opened Oct 2, 2026·

  ·

• Fix NaN from the CPU layer_norm kernel when the row mean dominates the spread

  #4017·breken-ai opened Sep 30, 2026·

  ·

• Fix CPU upsample_bilinear2d returning raw storage for same-size strided inputs

  #4016·breken-ai opened Sep 30, 2026·

  ·

• flash-attn: use smaller tiles for short packed BF16 attention on Hopper

  #4015·michaelfeil opened Sep 29, 2026Contributor·

  ·

• PR: expose clear_kv_cache on quantized_qwen3_moe

  #4010·songday opened Sep 28, 2026·

  ·

• chore(deps): update web-sys requirement from =0.3.104 to =0.3.106

  dependenciesrust
  #4009·dependabot[bot] opened Sep 28, 2026Bot·

  ·

• chore(deps): update rand requirement from 0.9.0 to 0.10.3

  dependenciesrust
  #4008·dependabot[bot] opened Sep 28, 2026Bot·

  ·

• Round halfway quant levels to the even integer

  #4007·SashaMIT opened Sep 28, 2026·1 task done·

  ·

• Fix quantized_phi3 rope for partial rotary embeddings

  #4006·taekop opened Sep 28, 2026·

  ·
  1

• FA3: omit unused scratch initialization

  #4005·michaelfeil opened Sep 27, 2026Contributor·

  ·

• fix(flash-attn-v3): honor finite attention windows

  #4002·michaelfeil opened Sep 27, 2026Contributor·

  ·
  1

• Support tuple checkpoints in PyTorch loader

  #3999·knkski opened Sep 24, 2026·

  ·

• ModernBert::load reads a checkpoint without the model. prefix

  #3992·jordiboehme opened Sep 21, 2026·

  ·

• ModernBERT: read hidden_activation from the config instead of hardcoding gelu

  #3991·jordiboehme opened Sep 21, 2026·

  ·

• ModernBERT: build the attention masks in the model's dtype

  #3989·aovestdipaperino opened Sep 20, 2026·

  ·

• Reduce CPU ConvTranspose1d time with and without padding

  #3988·Hosi121 opened Sep 20, 2026·

  ·

• fix: add overflow check to Layout::new and Layout::contiguous_with_of…

  #3987·ksj1230 opened Sep 17, 2026·

  ·
  1