Repository navigation

vulkan: allow proxy collectives without external semaphore FD functions - #23

Open

AndrewTilson wants to merge 15 commits into

pwilkin:vulkan-tp-p2pfrom

AndrewTilson:review/proxy-init-fix

Commits on Jun 26, 2026
- ggml-backend-meta: fix SPLIT_MODE_TENSOR segfault on multi_buffer KV-cache views
  pwilkin
  committed
  325ccea
- vulkan: GPU-pipelined multi-GPU AllReduce for -sm tensor
  pwilkin
  committed
  5dc559c
- vulkan: portable CPU-proxy fallback for cross-driver -sm tensor AllReduce
  pwilkin
  committed
  68c289b
- fix Windows build
  pwilkin
  committed
  a448deb
Commits on Jun 27, 2026
- vulkan: generalize -sm tensor AllReduce to >2 devices
  pwilkin
  committed
  362fdd2
- ggml-backend : resolve multi-buffer wrappers to sub-buffers in meta backend
  pwilkin
  committed
  e578ca2
- vulkan: add O(n) ring AllReduce for -sm tensor (GGML_VK_COMM_RING)
  pwilkin
  committed
  01ea44e
- vulkan: add CPU-proxy bridge support to the ring AllReduce
  
  pwilkin
  and
  claude
  committed
  f192a63
Commits on Jun 28, 2026
- vulkan: prototype device-to-device peer buffers for -sm tensor (GGML_VK_COMM_D2D)
  
  pwilkin
  and
  claude
  committed
  aa52642
- vulkan: F16 staging for the ring AllReduce
  
  pwilkin
  and
  claude
  committed
  831bc20
- vulkan: strip comments from the -sm tensor comm code
  
  pwilkin
  and
  claude
  committed
  5a829a9
- vulkan: drop D2D, make the ring the default, remove the OFF toggle
  
  pwilkin
  and
  claude
  committed
  bd1d318
- vulkan: remove the all-to-all pipeline and fp32 staging from the comm
  
  pwilkin
  and
  claude
  committed
  75e7750
Commits on Jun 29, 2026
- vulkan: fix -sm tensor decode crawl, comm init robustness, add opt-in tree all-reduce
  pwilkin
  committed
  13ac901
Commits on Sep 9, 2026
- vulkan: allow proxy collectives without external semaphore FD functions
  AndrewTilson
  committed
  e5d3e1c

Navigation Menu

Repository navigation

vulkan: allow proxy collectives without external semaphore FD functions - #23

vulkan: allow proxy collectives without external semaphore FD functions#23
AndrewTilson wants to merge 15 commits into
pwilkin:vulkan-tp-p2pfrom
AndrewTilson:review/proxy-init-fix

Commits on Jun 26, 2026

ggml-backend-meta: fix SPLIT_MODE_TENSOR segfault on multi_buffer KV-cache views

vulkan: GPU-pipelined multi-GPU AllReduce for -sm tensor

vulkan: portable CPU-proxy fallback for cross-driver -sm tensor AllReduce

fix Windows build

Commits on Jun 27, 2026

vulkan: generalize -sm tensor AllReduce to >2 devices

ggml-backend : resolve multi-buffer wrappers to sub-buffers in meta backend

vulkan: add O(n) ring AllReduce for -sm tensor (GGML_VK_COMM_RING)

vulkan: add CPU-proxy bridge support to the ring AllReduce

Commits on Jun 28, 2026

vulkan: prototype device-to-device peer buffers for -sm tensor (GGML_VK_COMM_D2D)

vulkan: F16 staging for the ring AllReduce

vulkan: strip comments from the -sm tensor comm code

vulkan: drop D2D, make the ring the default, remove the OFF toggle

vulkan: remove the all-to-all pipeline and fp32 staging from the comm

Commits on Jun 29, 2026

vulkan: fix -sm tensor decode crawl, comm init robustness, add opt-in tree all-reduce

Commits on Sep 9, 2026

vulkan: allow proxy collectives without external semaphore FD functions