Add CI tests for AMD GPU accelerator detection (real-hardware fixtures) - #271
Add CI tests for AMD GPU accelerator detection (real-hardware fixtures)#271hmeiland wants to merge 2 commits into
Conversation
Sorry, something went wrong.
Adds workflow-based tests for both AMD detection methods in accelpath, using output captured from real hardware (AMD Radeon Pro V710, gfx1101) so no AMD GPU is required in CI: - Method 1 (KFD sysfs): make the topology root overridable via $EESSI_KFD_TOPOLOGY_ROOT so captured nodes/*/properties fixtures can be injected. Covers the gfx_target_version -> gfxMAJMINSTEP decode and the CPU-node (gfx_target_version 0) skip. - Method 2 (amd-smi): fake amd-smi stub on $PATH replaying real 'amd-smi static --asic' output, mirroring the existing nvidia-smi tests. New workflow tests_archdetect_amd_gpu.yml runs kfd_sysfs (v710, cpu_only) and amd_smi (v710) matrices. NVIDIA detection and the no-accelerator exit-2 path remain unchanged. This work was created with the help of AI (GitHub Copilot / an AI coding assistant).
|
The main reason for not merging #205 and #214 way back when was that we weren't ready to expose the AMD GPU software yet (i.e. it was only accessible by explicitely setting the EESSI_ACCELERATOR_TARGET_OVERRIDE=accel/amd/gfx90a). We'd like at least the core software to be available, and functional. I guess we should have a discussion on what that core software is - we are up to the level of HIP/RCCL now. Maybe we consider that enough and we expose it? Will discuss in our support meeting. |
Sorry, something went wrong.
|
Discussed in the support meeting. We think the time is right to actually expose this, i.e. add the archdetect support. Only challenge now is to find the time to review this PR :) |
Sorry, something went wrong.
|
Just as a note: from the brief look I gave it, this replaces #205, but NOT #214. I.e. if this PR is approved, we should close #205 as 'superseded' by this PR, but we should probably take #214 out of draft mode and still assess it. Note that #214 in the opening post also states we probably need to update workflows/tests_eessi_extend_module.yml. That is NOT resolved by the current PR, so that gap remains. |
Sorry, something went wrong.
on it... |
Sorry, something went wrong.
|
Superseded by #273. That PR consolidates AMD GPU accelerator detection in init/eessi_archdetect.sh plus the CI tests (KFD-topology fixtures for MI250X/MI300X/V710 + no-device cases), is rebased on current main, and has all checks passing. Please review #273 instead; closing this one shortly to avoid two competing implementations of the same detection. |
Sorry, something went wrong.
Summary
Adds workflow-based CI tests for AMD GPU accelerator detection in eessi_archdetect.sh, building on the detection logic proposed in #205. All tests use output captured from real hardware (an AMD Radeon Pro V710 / NAVI32, gfx1101), so no AMD GPU is required in CI.
This addresses the AMD-CI gap tracked in #214, complementing the detection work in #205.
What's covered
accelpath()'s AMD path (from #205) has two detection methods; this PR tests both:
Changes
Verification
Verified locally against the #205 branch, and the gfx1101 result was confirmed on the live V710 (both KFD gfx_target_version=110001 and amd-smi TARGET_GRAPHICS_VERSION: gfx1101 decode to accel/amd/gfx1101):
Notes