Repository navigation

[ET-VK] Cooperative-matrix kernels for 4-bit linear on Intel Xe2 and NVIDIA - #24000

Draft

xuyanwen2012 wants to merge 11 commits into

pytorch:mainfrom

sarc-acl:pr/wmma-families

Draft

[ET-VK] Cooperative-matrix kernels for 4-bit linear on Intel Xe2 and NVIDIA#24000
xuyanwen2012 wants to merge 11 commits into
pytorch:mainfrom
sarc-acl:pr/wmma-families

Commits on Oct 19, 2026
- [ET-VK] Full subgroups, Vulkan memory model and family-key properties in vk_api
  
  xuyanwen2012
  and
  claude
  committed
  7396952
- [ET-VK] Cooperative-matrix family key and kernel rows for 4-bit linear
  
  xuyanwen2012
  and
  claude
  committed
  0582c1c
- [ET-VK] q4gsw linear: cooperative-matrix family kernels
  
  xuyanwen2012
  and
  claude
  committed
  3b8c77f
- [ET-VK] dq8ca_q4gsw linear: cooperative-matrix family kernels
  
  xuyanwen2012
  and
  claude
  committed
  68fcab5
- [ET-VK] Tests for the cooperative-matrix family kernels
  
  xuyanwen2012
  and
  claude
  committed
  c20abb6
- [ET-VK] q4gsw family kernels: partial last M tile on texture IO
  
  xuyanwen2012
  and
  claude
  committed
  d537e47
- [ET-VK] Row-major dq8ca family kernel: no reads past M rows
  
  xuyanwen2012
  and
  claude
  committed
  2c79704
- [ET-VK] Family kernel selection: NVIDIA rows on any aligned shape, Xe2 8da4w threshold 20e6
  
  xuyanwen2012
  and
  claude
  committed
  3345345
- [ET-VK] Fence the build commands in the custom_ops CMakeLists comment
  
  xuyanwen2012
  and
  claude
  committed
  1d8fc04
Commits on Oct 22, 2026
- [ET-VK] Enable vulkanMemoryModelDeviceScope with the memory model
  
  xuyanwen2012
  and
  claude
  committed
  dd34a66
- [ET-VK] Bind the Xe2 q4gsw texture3d input in the general layout
  
  xuyanwen2012
  and
  claude
  committed
  d5edd60