PR #179930: Use ROCM_VERSION instead of TORCH_HIP_VERSION to gate expandable segments (#179930)

• Repository: pytorch/pytorch
• Author: gnuthor
• State: closed | merged: False
• Link: https://github.com/pytorch/pytorch/pull/179930

Description (problem / solution / changelog)

Summary:

Use ROCM_VERSION (format: major*10000 + minor*100 + patch) instead of TORCH_HIP_VERSION (format: major*100 + minor) to gate expandable segments on ROCm.

The previous gate TORCH_HIP_VERSION < 702 was accidentally disabling expandable segments on ROCm versions 7.0.x (since TORCH_HIP_VERSION for HIP 7.0 = 700 < 702). The corrected gate ROCM_VERSION < 70000 properly limits the disable to ROCm < 7.0.0.

Also adds a test skip for test_graph_rng_after_failed_capture on ROCm when expandable segments are active. This is a pre-existing issue (tracked in pytorch/pytorch#179911) where CUDA graph capture failure recovery doesn't work correctly with expandable segments on ROCm. The skip is conditioned on TEST_WITH_ROCM and self.expandable_segments so the test still runs on ROCm without expandable segments.

Test Plan:

• CI (internal + GitHub Actions)
• Build verification: buck build fbcode//caffe2/c10/cuda:cuda --config caffe2.enable_hip=true passes
• Test skip verified: test_graph_rng_after_failed_capture will be skipped only when ROCm + expandable segments are both active

Reviewed By: mrajpal, joshuuuasu, banitag1

Differential Revision: D100184839

cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang

Changed files

• c10/cuda/CUDAAllocatorConfig.h (modified, +1/-1)
• test/test_cuda.py (modified, +4/-0)