All pull requests

Pull requests

• Launch fast.cross_entropy in slices past the Metal grid limit

  #4776·ArupKantiDas opened Oct 26, 2026Contributor·

  ·

• Fix Metal and CUDA affine quantization for all-negative groups

  #4775·lgyStoic opened Oct 26, 2026·

  ·
  1

• Use pytest to run CPU tests in parallel

  #4774·zcbenz opened Oct 26, 2026Member·

  ·

• Keep the existing file when save, save_safetensors or save_gguf fails

  #4773·aleroot opened Oct 26, 2026Contributor·

  ·

• [Metal] Fix fill loop in layer_norm

  #4772·luciusacarr opened Oct 26, 2026Contributor·

  ·
  1

  1

• Fix SliceUpdate vmap with unbatched update

  #4770·mahimn01 opened Oct 26, 2026Contributor·

  ·

• Fix premature rounding in addmm

  #4769·tetsuo opened Oct 25, 2026·

  ·

• Check scatter and gather indices are in bounds

  #4768·turingH opened Oct 25, 2026·

  ·

• [jaccl] Detect dead RDMA links instead of crashing

  #4767·ipazarbasi opened Oct 25, 2026·

  ·

• Route of per-channel gate in gated_delta_update fixed

  #4766·Exorust opened Oct 25, 2026·

  ·

• Fix gguf tensor to f16 size overflow

  #4764·turingH opened Oct 25, 2026·

  ·

• Fix GPU concatenate for outputs larger than 2^31 elements

  #4762·citruslab opened Oct 24, 2026·3 tasks done·

  ·

• [CUDA] Expose compilation targets via device_info

  #4761·dhiltgen opened Oct 24, 2026Contributor·

  ·
  3

• Fix GPU concatenate for outputs larger than 2^31 elements

  #4758·tivris opened Oct 24, 2026Contributor·

  ·

• Add optional source IP binding for ring connections

  #4757·ronaldmannak opened Oct 24, 2026Contributor·

  ·

• Fix FP8 NaN preservation in to_fp8 and E8M0 scale dequantization

  #4756·FaizPalwala opened Oct 23, 2026·

  ·
  1

• Bigger MoE tiles on M1 GPUs

  await response
  #4754·a3rosol opened Oct 23, 2026·

  ·
  1

• Use gemv_wide for fp32 matmuls with 2-3 rows

  await verification
  #4742·utsav1033 opened Oct 22, 2026·

  ·

• feat(nn): add freqs and partial_rotary_factor support to RoPE

  low priority
  #4732·shadyuwugurl opened Oct 20, 2026·

  ·

• Add optional int8 path for quantized matmul with affine weights

  await verification
  #4722·magnuslundstedt opened Oct 18, 2026Contributor·

  ·
  1

• Fix sharded-to-all bias and integer split positions on quantized layers

  await verificationlow priority
  #4715·ronaldmannak opened Oct 17, 2026Contributor·

  ·

• Debug time graph verification

  #4711·zcbenz opened Oct 16, 2026Member·

  ·

• Fix fftn size one axis aliasing a temporary

  #4705·StockerMC opened Oct 15, 2026·

  ·
  1

  3

• Make GPU scheduler know the spin is waiting for something

  low priority
  #4703·aleroot opened Oct 15, 2026Contributor·

  ·
  3

• Fix fast.cross_entropy VJP on non-contiguous Metal inputs

  #4702·StockerMC opened Oct 15, 2026·

  ·
  1

  1