Repository navigation

refactor(EPv2): replace per-module device staging with shared buffer - #703

Merged

kawhil-amd merged 12 commits into

kawhil-amd:refactor/ep-staging-shared-buffer

Oct 12, 2026

Merged

refactor(EPv2): replace per-module device staging with shared buffer#703
kawhil-amd merged 12 commits into
ROCm:mainfrom
kawhil-amd:refactor/ep-staging-shared-buffer

Commits on Oct 12, 2026
- refactor(EPv2): replace per-module __device__ staging with shared buffer
  
  kawhil-amd
  and
  claude
  committed
  fea3c3c
- fix(EPv2): move dispatch_staging init before _build_kernels
  
  kawhil-amd
  and
  claude
  committed
  47d2aec
- fix(EPv2): fail on the host when staging_base is unbound
  
  kawhil-amd
  and
  claude
  committed
  bcb7339
- fix(EPv2): bump EpArgsFieldCount after rebase onto main
  
  kawhil-amd
  and
  claude
  committed
  92ae111
- fix(EPv2): reject a misaligned staging_base on the host
  
  kawhil-amd
  and
  claude
  committed
  8a76165
- fix(EPv2): release the dispatch staging in _close_backend
  
  kawhil-amd
  and
  claude
  committed
  e6ec5e7
- test(EPv2): pin ep_staging_bytes against the device layout
  
  kawhil-amd
  and
  claude
  committed
  cbeaf83
- fix(EPv2): resolve max_recv in ep_staging_bytes the way EpMaxRecv does
  
  kawhil-amd
  and
  claude
  committed
  57895ba
- fix(jit/v2): route KernelSpec::Launch through Derived::LaunchRaw
  
  kawhil-amd
  and
  claude
  committed
  bdb6670
- refactor(EPv2): drop the dead zero case and refresh the staging comments
  
  kawhil-amd
  and
  claude
  committed
  f069364
- refactor(EPv2): drop the unused scale-slot count, keep what it documented
  
  kawhil-amd
  and
  claude
  committed
  780c8c4
- style: apply black and clang-format over the files this PR touches
  
  kawhil-amd
  and
  claude
  committed
  814868b