Enable collective tests

[mmakevic-amd]

Motivation

This PR enables:

• //xla/backends/gpu/collectives:gpu_cliques_test,
• //xla/backends/gpu/collectives:gpu_collectives_test, and
• //xla/backends/gpu/collectives:loopback_collectives_test.

It also extracts ToNcclCount, ToNcclDataType and ToNcclReduction from xla/backends/gpu/collectives/rccl_communicator.cc to the newly added xla/backends/gpu/collectives/rccl_types.cc and adds xla/backends/gpu/collectives/rccl_types_test.cc. This mirrors nccl_types.cc and nccl_types_test.cc.

Technical Details

1. Switched all enabled tests from xla_cc_test to xla_test target
2. Generalized enabled tests to use the provided platform instead of hardcoded CUDA
3. As mentioned above, separated rccl_types in a separate file and added the appropriate test, mirroring CUDA implementation.

Test Plan

/

Test Result

There are five failing subtests in gpu_collectives_test which will be addressed later. Other tests are passing.

[  FAILED  ] GpuCollectivesTest.CreateRegisteredMemory
[  FAILED  ] GpuCollectivesTest.CreateWithMultipleIds
[  FAILED  ] GpuCollectivesTest.SplitCommunicators
[  FAILED  ] GpuCollectivesTest.CreateSymmetricMemory
[  FAILED  ] GpuCollectivesTest.PutAndWaitSignal

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull request overview

This PR enables previously-disabled GPU collectives tests across GPU backends (CUDA/ROCm), and refactors RCCL-specific NCCL-type conversion helpers into a standalone rccl_types library with a new unit test, mirroring the existing NCCL implementation.

Changes:

• Enable gpu_cliques_test, gpu_collectives_test, and loopback_collectives_test by converting targets from xla_cc_test to xla_test and generalizing platform selection away from hardcoded "CUDA".
• Extract ToNcclCount, ToNcclDataType, and ToNcclReduction from rccl_communicator.cc into new rccl_types.{h,cc}.
• Add rccl_types_test.cc to validate RCCL type conversions.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 11 comments.

Show a summary per file

File	Description
xla/backends/gpu/collectives/rccl_types.h	Adds RCCL type conversion API declarations.
xla/backends/gpu/collectives/rccl_types.cc	Implements RCCL type conversion helpers extracted from communicator.
xla/backends/gpu/collectives/rccl_types_test.cc	Adds unit tests for RCCL type conversions.
xla/backends/gpu/collectives/rccl_communicator.cc	Switches communicator code to use extracted type conversion helpers.
xla/backends/gpu/collectives/loopback_collectives_test.cc	Generalizes platform selection via PlatformUtil instead of hardcoded CUDA.
xla/backends/gpu/collectives/gpu_collectives_test.cc	Generalizes platform selection and GPU capability checks across backends.
xla/backends/gpu/collectives/gpu_cliques_test.cc	Generalizes platform selection via PlatformUtil instead of hardcoded CUDA.
xla/backends/gpu/collectives/BUILD	Enables tests via xla_test, adds rccl_types targets, and adjusts deps/tags per backend.

[mmakevic-amd]

Same as above

[hsharsha]

Should we upstream it to openxla/main?

[mmakevic-amd]

Should we upstream it to openxla/main?

I discussed this with @i-chaochen in https://github.com/ROCm/xla-internal/issues/52#issuecomment-4752364931. Since some of these tests are mGPU, and we won't be enabling them in CI for now, he proposed fixing the failing subtests before upstreaming.

[hsharsha]

I mean the fix, We can enable mgpu tests as nightly

[claude]

bug: FP8 reduction silently produces wrong results — The NCCL counterpart uses (cc.IsAtLeastHopper() || is_reduction_op) ? ncclFloat8e5m2 : ncclInt8, deliberately letting NCCL throw errors for FP8 reductions on unsupported hardware. Here, the logic is has_ocp_fp8_support() ? ncclFloat8e5m2 : ncclInt8, ignoring the is_reduction_op flag entirely.

On non-OCP ROCm hardware, FP8 reductions will silently be mapped to ncclInt8, which would produce incorrect results rather than a clean error. This should either mirror the NCCL behavior (using has_ocp_fp8_support() || is_reduction_op) or include a comment explaining why the divergence is intentional.

[mmakevic-amd]

Hi @mfrancepillois, I see you're the one who worked on fp8 support in openxla@bd94197. Could you let me know if this is a legitimate concern? Are reductions supported?

[i-chaochen]

Yes, reduction Ops and all fp8 Ops are supported in rccl since rocm6.4.... https://github.com/ROCm/frameworks-internal/issues/15398

[claude]

Review Summary

This PR enables three ROCm collective tests by converting them to xla_test, generalizing hardcoded "CUDA" platform names to runtime-resolved values, and extracting ToNcclCount/ToNcclDataType/ToNcclReduction into new rccl_types.{h,cc} files mirroring the existing nccl_types structure.

Key finding: The FP8 type mapping in rccl_types.cc diverges from nccl_types.cc by ignoring the is_reduction_op flag — on non-OCP ROCm hardware, FP8 reductions would silently map to ncclInt8 producing incorrect results instead of a clean error.

6 inline comments posted covering: FP8 correctness, fragile transitive includes, unchecked .value() repetition, missing ROCm skip conditions, error construction style, and a stale BUILD dep.

🤖 Generated with Claude Code

[mmakevic-amd]

I mean the fix, We can enable mgpu tests as nightly

@hsharsha, @i-chaochen, here is the draft PR for upstream openxla#44641

The falling subtests in //xla/backends/gpu/collectives:gpu_collectives_test won't affect our CI atm since it is a multi-gpu test. I will look into them once this is merged.

Let me know if it looks good to you, so I can open a PR