Repository navigation

Fix GPU IPC and collect - #758

Merged

jpsamaroo merged 27 commits into

masterfrom

jps/fix-ipc-ownership

Oct 13, 2026

Merged

Fix GPU IPC and collect#758
jpsamaroo merged 27 commits into
masterfrom
jps/fix-ipc-ownership

Commits on Oct 12, 2026
- Preserve GPU ownership across Distributed worker transfers
  jpsamaroo
  committed
  46a8955
- CUDAExt: Don't take device ownership for aliasing or re-pin host buffers
  
  jpsamaroo
  and
  claude
  committed
  7910770
- datadeps: Stage cross-device remainder copies through the host
  
  jpsamaroo
  and
  claude
  committed
  6cfad12
- DArray: Bring dense GPU tiles to host before collect's cat
  
  jpsamaroo
  and
  claude
  committed
  59b37e9
- AGENTS.md: Record the multi-GPU-per-process lesson
  
  jpsamaroo
  and
  claude
  committed
  b9543d4
- DArray: Wait on the tile's stream before collect's host copy
  
  jpsamaroo
  and
  claude
  committed
  2f13963
- AGENTS.md: Record that reading GPU results needs Dagger's stream
  
  jpsamaroo
  and
  claude
  committed
  82e43d2
- CUDAExt: Release HtoD host registrations when the copy completes
  
  jpsamaroo
  and
  claude
  committed
  fd0be6b
- ROCExt: Release HtoD host registrations when the copy completes
  
  jpsamaroo
  and
  claude
  committed
  a34e32a
- CUDAExt, ROCExt: Unregister pinned host buffers without blocking in a finalizer
  
  jpsamaroo
  and
  claude
  committed
  269e216
- Read device addresses for aliasing without taking stream ownership
  
  jpsamaroo
  and
  claude
  committed
  1f8100c
- GPU exts: Move device-resident chunks and strided device views to/from host
  
  jpsamaroo
  and
  claude
  committed
  312bdff
- CUDAExt, ROCExt: Copy between GPUs of one process with overlapped peer copies
  
  jpsamaroo
  and
  claude
  committed
  a4d56bf
- ROCExt: Upload from pageable host memory instead of registering it
  
  jpsamaroo
  and
  claude
  committed
  9f648de
- Identify copy tasks so GPU stream tracking can execute transfers
  jpsamaroo
  committed
  7ce30ac
- Honor test arguments and dependency resolution in both test entry points
  jpsamaroo
  committed
  ab34cb7
- Give ROCm CI a spare execution thread for host-call polling
  jpsamaroo
  committed
  830e6a6
- Check MPI rank ownership before GPU event tracking unwraps chunks
  jpsamaroo
  committed
  9cacd89
- Record the native HIP failure behind local halo validation
  jpsamaroo
  committed
  4ea7e52
- Flush suite boundaries so GPU CI failures retain progress
  jpsamaroo
  committed
  9a0c6af
- Use the legacy Level Zero adapter for oneMKL CI interop
  jpsamaroo
  committed
  6edfee8
- Keep Intel MPI transfers on host staging until descriptor passing exists
  jpsamaroo
  committed
  ad74089
- Restore Intel MPI CI after validating host-staged transfers
  jpsamaroo
  committed
  96e1710
- Distinguish passing GPU assertions from upstream finalizer errors
  jpsamaroo
  committed
  e5a95c8
- Select external AMD codegen on every ROCm CI worker
  jpsamaroo
  committed
  9a17c2b
- Keep rebased investigation lessons uniquely numbered
  jpsamaroo
  committed
  cdebe66
- Keep Intel MPI benchmarks disabled after the larger GEMM crashes
  jpsamaroo
  committed
  edc18fa

Commits

Commits on Oct 12, 2026