All pull requests

Pull requests

• Optimize NF4 single-token GEMV on Blackwell

  #2098·PiotrPotocki opened Oct 8, 2026·

  ·

• Support skip_zeros in the XPU Triton optimizer kernels

  #2097·jiqing-feng opened Oct 8, 2026Contributor·

  ·
  1

• fix: wrong absmax indexing in AVX512 dequantizeBlockwise4bitCpu

  #2096·NanoCode012 opened Oct 7, 2026·

  ·

• MPS: only use the Hub GEMV kernel when K is a multiple of blocksize

  #2094·egeozkoc opened Sep 30, 2026Contributor·

  ·

• test: align post-reload max_error_count with pre-reload check in test_optimizer32bit

  #2093·CS-liujf opened Sep 23, 2026·

  ·

• Fix device parametrization for optimizer benchmarks

  #2092·emme1t opened Sep 22, 2026·

  ·
  1

• fix: pass backend="cpu" when requesting CPU gemm_4bit_forward kernel

  #2090·SIDDARTHAREDDY8 opened Sep 20, 2026·

  ·
  1

• [ROCm] Preserve float bits in SIMT DPP reduction

  #2089·liminfei-amd opened Sep 15, 2026·

  ·
  1

  4

• fix(optim): reject bias_correction and adam_w_mode in LAMB

  #2088·ConnorMoss02 opened Sep 15, 2026·

  ·

• fix(optim): reject nesterov and dampening in SGD and LARS

  #2085·ConnorMoss02 opened Sep 11, 2026·

  ·

• Fix Params4bit.to() leaving CPU-packed weights packed

  #2083·22elix3r opened Sep 10, 2026·

  ·
  1

  3

• Fix PytorchLARS updates when momentum is zero

  #2082·mottopanikeiku opened Sep 10, 2026·

  ·
  1

  1

• [WIP] Prototype original-parameter AdamW8bit state for FSDP1

  #2081·gss10282025 opened Sep 10, 2026·

  ·

• Warn once when Linear4bit computes in float32 on GPU

  #2077·thangdangjp opened Sep 9, 2026·

  ·

• Make AdEMAMix32bit inherit AdEMAMix so it allocates the double state buffer

  #2076·caiotheodoro opened Sep 8, 2026·

  ·
  1

  1

• chore(deps): bump softprops/action-gh-release from 3.0.2 to 3.0.3 in the actions group

  Dependenciesgithub_actions
  #2074·dependabot[bot] opened Sep 7, 2026Bot·

  ·
  1

• [ROCm] Allow explicit GPU architecture override

  #2073·LOGO127 opened Sep 4, 2026·

  ·
  1

• Add AVX2 path for CPU 4-bit (NF4/FP4) blockwise dequantization

  #2072·rachitj27 opened Sep 4, 2026·

  ·

• fix(cuda): chunk 4-bit quant/dequant launches above INT32_MAX elements (#1785)

  #2071·Anai-Guo opened Sep 4, 2026Contributor·

  ·

• fix(optim): fall back to non-paged state on all devices without managed memory

  Cross PlatformLow PriorityOptimizers
  #2069·egeozkoc opened Sep 3, 2026Contributor·

  ·

  · v0.50.3
  3

• Fix CPU Hub kernel backend selection

  x64 CPU
  #2065·SiluPanda opened Aug 30, 2026·

  ·

  · v0.50.3
  1

  8

• fix(optim): mark unused amsgrad/optim_bits as deprecated in 8-bit Adam classes

  Optimizers
  #2063·kritikagarg opened Aug 28, 2026·

  ·
  1

• Fix N-D input handling in default backend int8 ops

  Cross Platform
  #2062·2sumtech opened Aug 27, 2026·

  ·

• docs: correct Adagrad and matrix multiplication wording

  #2058·Excelius-Wang opened Aug 25, 2026·

  ·

• Add Apple MPS NF4 benchmark

  #2057·hamedrabah opened Aug 25, 2026·

  ·
  1