All pull requests

Pull requests

• cuDNN: bottom-right causal alignment and packed sequences for attention

  #3438·CarloLucibello opened Oct 21, 2026Contributor·

  ·
  2

• cuSOLVER: Fall back to 64-bit implementations when legacy workspace sizes overflow

  #3437·michel2323 opened Oct 20, 2026Member·

  ·

• Lower scalar and vector LLVM math intrinsics through device libraries

  #3433·maleadt opened Oct 20, 2026Member·

  ·
  1

• [DO NOT MERGE] Diagnose cuSOLVERMg failures on the MultiGPU agent

  #3428·maleadt opened Oct 19, 2026Member·

  ·

• Reject kernel constants that are not isbits, Symbols or types

  #3427·maleadt opened Oct 19, 2026Member·

  ·
  1

• Narrow index arithmetic with CUDA-owned specialization policy

  kernelsperformance
  #3389·maleadt opened Oct 13, 2026Member·

  ·
  11

• [release-6.2] Port to LLVM.jl 10 (GPUCompiler 1.23.1), release 6.2.3

  installationkernels
  #3381·vchuravy opened Oct 11, 2026Member·

  ·
  1

• KA: round prefer_blocks workgroup sizes to warp multiples

  kernelsperformance
  #3373·vchuravy opened Oct 10, 2026Member·

  ·
  2

  

• Specialize kernels on how their array arguments alias

  kernelsperformance
  #3372·vchuravy opened Oct 9, 2026Member·

  ·

• Enzyme: forward rule for CuArray(::DataRef) must wrap the shadow DataRef

  extensions
  #3371·vchuravy opened Oct 9, 2026Member·

  ·
  

• Keep kernel launches type-stable with index-typed device arrays

  kernelsperformance
  #3294·maleadt opened Sep 24, 2026Member·4/4·

  ·
  2

• Add an index type parameter to CuDeviceArray

  kernelsperformance
  #3293·maleadt opened Sep 24, 2026Member·3/4·

  ·
  2

• Add hostcalls for device-to-host services

  kernelsruntime
  #3243·maleadt opened Aug 20, 2026Member·

  ·
  2

• Try enabling fwd broadcating Enzyme test

  ciextensionsupstream
  #3214·kshyatt opened Jul 24, 2026Member·

  ·
  5

• Implement unsafe_wrap! for in-place re-wrapping of non-owning CuArrays

  arraysspeculative
  #3194·ameligrana opened Jul 12, 2026·

  ·
  1

• Expose directed rounding for Float64 WMMA tensor cores

  kernels
  #3143·orkolorko opened May 16, 2026Contributor·4 tasks·

  ·
  2

• Use Republic.jl to forward subpackage APIs

  installation
  #3103·AntonOresten opened Apr 15, 2026Contributor·

  ·
  2

• CUPTI Profiler Host, Range Profiler, and PM Sampling APIs

  runtime
  #3059·gbaraldi opened Mar 23, 2026Member·1 of 2 tasks·

  ·
  3

• Add direct sparse gpu constructors for COO, CSC, CSR and BSR matrices

  libraries
  #3022·nHackel opened Jan 23, 2026·

  ·
  1

• Add AnyCuDeviceArray variations and CuScalar

  arrayskernelsspeculative
  #2849·moukle opened Aug 22, 2025·

  ·
  4

• make CUDA randn work with Zygote

  extensionsneeds changes
  #2581·bgctw opened Dec 9, 2024·

  ·
  6

• WIP: Native I/O.

  kernelsspeculative
  #2485·maleadt opened Sep 5, 2024Member·

  ·

• High Level Wrapper for Fused Matmul + Bias + Activation

  librariesperformance
  #2360·avik-pal opened May 4, 2024·

  ·

• Use TaskLocalValues

  runtimespeculative
  #2075·vchuravy opened Sep 8, 2023Member·

  ·
  1

• Support FFT adjoint plans and test

  libraries
  #2073·gaurav-arya opened Sep 4, 2023·

  ·