All pull requests

Pull requests

• Retain useful byte-based command batches on roomy g13s GPUs

  #54·tijs opened Nov 6, 2026·

  ·
  1

• Integrate upstream Metal kernel fixes (reductions, scan, indexing, conv, FFT, norms, large grids)

  #53·rcfa opened Nov 5, 2026·

  ·

• Integrate upstream Metal runtime fixes (fence deadlock, residency error, metallib lookup, SDPA fallback)

  #52·rcfa opened Nov 5, 2026·

  ·

• Integrate upstream io fixes (short reads, .npy shape checks, GGUF size overflow, atomic save)

  #51·rcfa opened Nov 5, 2026·

  ·

• Integrate upstream fixes to transforms and ops (vmap, VJP/JVP, compile, conv, stream arguments)

  #50·rcfa opened Nov 5, 2026·

  ·

• Integrate upstream fixes to array views, strides and contiguity flags

  #49·rcfa opened Nov 5, 2026·

  ·

• Integrate upstream CPU kernel fixes (reductions, binary ops, conv, FP8, fp16 masks, float64 math)

  #48·rcfa opened Nov 5, 2026·

  ·

• Build integer arrays from floats with saturation

  #47·rcfa opened Nov 5, 2026·

  ·

• Convert complex64 to integers with saturation on the CPU

  #46·rcfa opened Nov 5, 2026·

  ·

• Convert floats to integers with saturation on the CPU

  #45·rcfa opened Nov 5, 2026·

  ·

• Fix unaligned uint32 stores in the CPU RandomBits kernel

  #44·rcfa opened Nov 3, 2026·

  ·

• Wrap integer sums and products outside the Simd operators on the CPU

  #43·rcfa opened Nov 3, 2026·

  ·

• Do signed integer negation, abs and pow in unsigned types on the CPU

  #42·rcfa opened Nov 3, 2026·

  ·

• Take upstream's left_shift / right_shift fix (#4694) and test the counts outside the type width in C++

  #41·rcfa opened Nov 3, 2026·

  ·

• Build with clang < 16: capture and typename in the CPU backend

  #40·rcfa opened Nov 3, 2026·

  ·

• Do signed integer add, subtract and multiply in unsigned types on the CPU

  #39·rcfa opened Nov 3, 2026·

  ·

• Move past FLOAT64 metadata values by their size in the file

  #38·rcfa opened Nov 3, 2026·

  ·

• Read GGUF metadata without assuming alignment

  #37·rcfa opened Nov 3, 2026·

  ·

• Sum long float rows in blocks on the CPU

  #34·rcfa opened Oct 30, 2026·

  ·

• Synchronize streams before the allocator statistics baseline

  #33·rcfa opened Oct 30, 2026·

  ·

• Apply the nvfp4 global scales in the CPU qqmm

  #31·rcfa opened Oct 28, 2026·

  ·

• Guard the CPU quantized matmuls against M, N or K of 0

  #30·rcfa opened Oct 28, 2026·

  ·

• Compute each CPU arange element from start, not a running sum

  #29·rcfa opened Oct 28, 2026·

  ·

• Give min / -1 its arm64 result in the CPU integer division ops

  #28·rcfa opened Oct 28, 2026·

  ·

• Pass lda and ldb through to the bf16 and fp16 CPU matmuls

  #26·rcfa opened Oct 21, 2026·

  ·