Repository navigation

Coll allocator op - #27

Open

devreal wants to merge 32 commits into

mainfrom

coll-allocator-op

Open

Coll allocator op#27
devreal wants to merge 32 commits into
mainfrom
coll-allocator-op

Commits on May 25, 2026
- coll/base: add allocator param to internal collective algorithms
  
  devreal
  and
  claude
  committed
  5bcd28d
- accelerator/base: switch device pool to basic allocator with segment tracking
  
  devreal
  and
  claude
  committed
  d404a04
- coll/base: replace allocator with GPU op session in reduction algorithms
  
  devreal
  and
  claude
  committed
  b912eb4
- op/session: add stop/restart/free pool for GPU reduction sessions
  
  devreal
  and
  claude
  committed
  e4631fe
- op/rocm: fix build integration
  devreal
  committed
  968b60d
- coll/tuned: Allocate sessions and allocators for coll/base calls
  devreal
  committed
  e59d22e
- Cache expensive allocation parts of a GPU session
  devreal
  committed
  be44a1e
- Fix compile issues
  devreal
  committed
  d8ccda7
Commits on May 26, 2026
- Introduce inheritance for op queues
  devreal
  committed
  2f34ecf
Commits on May 30, 2026
- Make op/cuda and op/rocm dso-by-default
  devreal
  committed
  65e65ff
- Fix NVCC include paths
  devreal
  committed
  5733226
Commits on May 31, 2026
- Fix CUDA build, static initializers not supported
  devreal
  committed
  24af1d2
Commits on Jun 8, 2026
- Register launcher callbacks in init_query
  devreal
  committed
  ce49af1
Commits on Jul 13, 2026
- CUDA: compile for native architecture
  devreal
  committed
  45ebf36
- accelerator/cuda: remove spurious return
  devreal
  committed
  f41125f
- Update more implementations to use device sessions
  devreal
  committed
  772dce3
Commits on Jul 14, 2026
- Work around buggy libofi CXI provider
  devreal
  committed
  d4e97fe
Commits on Jul 17, 2026
- Replace basic allocator with list of allocations
  devreal
  committed
  fef1fdc
Commits on Jul 20, 2026
- Device ops: avoid managed memory for command controls
  devreal
  committed
  1f7e3da
- Only launch kernel if needed and reset device state on kernel exit
  devreal
  committed
  df52d5b
Commits on Jul 25, 2026
- Compile for HIP native architecture by default
  devreal
  committed
  a78c551
- Change to op_rocm_kernels.hip
  devreal
  committed
  b39d9f2
Commits on Jul 26, 2026
- coll/tuned: dynamically select reduce target based on size
  devreal
  committed
  3bfb0c3
Commits on Jul 27, 2026
- Move ompi_op_rocm_kernel_fns_init to component_init_query()
  devreal
  committed
  9109b9d
Commits on Jul 29, 2026
- Introduce cooperative groups for device reduction kernels
  devreal
  committed
  16a27e6
- Cooperative groups require separable compilation/linking
  devreal
  committed
  d78e1ba
- Fix rocm linker for rdc build
  devreal
  committed
  de969ef
- More linker fixes
  devreal
  committed
  58d0e01
Commits on Aug 18, 2026
- Make cooperative kernel use optional
  devreal
  committed
  f3e484b
- Protect cuMemPoolGetAccess from invalid memory pools
  devreal
  committed
  330e62b
- Fix querying device attributes to support CUDA 12
  devreal
  committed
  61a3e37
- Add HIPCCFLAGS for rocm CI
  devreal
  committed
  77d46fa

Commits

Commits on May 25, 2026

Commits on May 26, 2026

Commits on May 30, 2026

Commits on May 31, 2026

Commits on Jun 8, 2026

Commits on Jul 13, 2026

Commits on Jul 14, 2026

Commits on Jul 17, 2026

Commits on Jul 20, 2026

Commits on Jul 25, 2026

Commits on Jul 26, 2026

Commits on Jul 27, 2026

Commits on Jul 29, 2026

Commits on Aug 18, 2026