All pull requests

Pull requests

• quantized: add iq4_nl, CPU and CUDA

  #4083·flaneur2020 opened Nov 1, 2026·

  ·

• Add backward pass for conv_transpose1d

  #4082·bhaskargurram-ai opened Oct 31, 2026Contributor·

  ·
  1

• Fix dead GLM4, Helium and SmolLM3 links in READMEs

  #4080·sarankumar1325 opened Oct 31, 2026·

  ·

• Add backward pass for conv_transpose1d, and fix a CPU matmul bug it exposed

  #4079·FentonA opened Oct 30, 2026·

  ·
  1

• Make value_states contiguous before the KV cache append in qwen

  #4077·andyjsbell opened Oct 29, 2026·

  ·

• Fix build: wrap __cpuid_count in unsafe (E0133)

  #4076·bnmchone opened Oct 29, 2026·

  ·
  1

• fix(kernels): stop mmvq_gguf reading past the weights at odd row counts

  #4075·cryptomilk opened Oct 29, 2026·

  ·

• Return -inf/inf from log_sum_exp when the max is infinite

  #4073·b84330808 opened Oct 28, 2026·

  ·

• fix(onnx): reduce over all axes when ReduceMax/ReduceMin get empty axes

  #4072·michele-xilo opened Oct 28, 2026·

  ·
  1

• Order NaN consistently in the CPU argsort

  #4071·b84330808 opened Oct 28, 2026·

  ·

• Add LFM2-VL (LiquidAI vision-language models)

  #4069·Jacqkues opened Oct 28, 2026Contributor·

  ·

• Cache CPU thread-count resolution

  #4067·WatNeru opened Oct 26, 2026·

  ·
  1

• fix(conv): prevent unsigned subtraction underflow in conv_transpose1d for l_in=1

  #4065·YamaGucciDev opened Oct 25, 2026·

  ·
  1

• Fix the Metal sdpa causal mask when there are fewer queries than keys

  #4062·dhkts1 opened Oct 23, 2026·

  ·

• feat(core): mmap-backed GGUF loading for quantized tensors

  #4061·lu-zero opened Oct 23, 2026·

  ·

• Add a gelu_tanh op with f32-internal forward and backward kernels to candle-nn

  #4060·DrJesseGlass opened Oct 22, 2026Contributor·

  ·

• Add backward passes to the fused rms_norm, layer_norm and softmax_last_dim ops

  #4059·DrJesseGlass opened Oct 22, 2026Contributor·

  ·

• Keep the CPU RMSNorm scale in f32

  #4057·xgbtxy opened Oct 21, 2026·

  ·
  4

• Metal: 64-token tile for the quantized matrix-matrix kernel

  #4056·dhkts1 opened Oct 21, 2026·

  ·

• Metal: QTensor::quantize_onto no longer leaves a zeroed buffer in the pool

  #4055·dhkts1 opened Oct 21, 2026·

  ·

• Metal quantized matmul: read the input from its own offset

  #4054·dhkts1 opened Oct 21, 2026·

  ·

• fix(kernels): only define the double atomicAdd shim in device code (CUDA 13)

  #4053·freezed-corpse-143 opened Oct 21, 2026·

  ·

• perf(cpu): use argminmax for argmin, argmax, min and max over the last dim

  #4051·jvdd opened Oct 20, 2026·

  ·

• fix(onnx): treat Shape end as exclusive and clamp start/end

  #4050·RaphaelFakhri opened Oct 20, 2026·

  ·
  1

• Fix flip on non-contiguous tensors

  #4048·spicysharkco opened Oct 20, 2026·

  ·