Skip to content

[ROCm] ROCm-latest (ROCm 10.1) build enablement for Paddle - #59

Open
soumitra-chatterjee-tech wants to merge 3 commits into
ROCm:developfrom
soumitra-chatterjee-tech:rocm-sync-3-enablement
Open

[ROCm] ROCm-latest (ROCm 10.1) build enablement for Paddle#59
soumitra-chatterjee-tech wants to merge 3 commits into
ROCm:developfrom
soumitra-chatterjee-tech:rocm-sync-3-enablement

Conversation

@soumitra-chatterjee-tech

@soumitra-chatterjee-tech soumitra-chatterjee-tech commented Aug 28, 2026

Copy link
Copy Markdown

Makes the fork build and run on ROCm 10.0.0 on AMD Instinct (MI300X/gfx942 and MI350X/MI355X/gfx950), on top of the sync (#57) and overlay (#58).

  • Compile the host C++ side in HIP language mode so rocThrust/CCCL headers resolve
  • thrust_compat force-include for rocThrust internal algorithm cross-references
  • hipBLASLt datatype macros; warp-rnnt and warp-ctc HIP cmake
  • gfx942 + gfx950 offload-arch and CMAKE_HIP_ARCHITECTURES

Validated on MI300X and MI350X/MI355X: 109/110 upstream tests pass per arch; native elementwise/reduction/matmul kernels are bit-exact against the CPU reference.

Stacked on #58.

…m code

Pure upstream PaddlePaddle/Paddle v3.0.0->v3.4.0 adoption (no AMD-specific changes); the AMD/ROCm overlay and ROCm-latest enablement follow as stacked PRs. Co-Authored-By: Claude <claude@anthropic.com>
…3.4)

The fork AMD/ROCm overlay: HIP kernels, ROCm dynload, the ATen/c10 compat layer, and ROCm build wiring, re-layered on the upstream sync. Co-Authored-By: Claude <claude@anthropic.com>
…tinct

Host C++ in HIP language mode for rocThrust/CCCL, thrust_compat force-include, hipBLASLt datatype macros, warp-rnnt/warp-ctc HIP cmake, gfx942/gfx950 offload + CMAKE_HIP_ARCHITECTURES. Co-Authored-By: Claude <claude@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant