PGLE/FDO profiling broken on all ROCm GPUs — needs XLA pin bump for rocprofiler-sdk v3

[srinivamd]

Summary

JAX PGLE (Profile-Guided Latency Estimation) / FDO profiling does not work on any ROCm GPU. All 5 tests in tests/pgle_test.py fail. This is a tracking issue for the long-term fix.

Immediate workaround PRs (deselect tests):

• #788 — rocm-jaxlib-v0.9.1
• #790 — rocm-jaxlib-v0.9.2

Jira: ROCM-25581 (gfx120X), ROCM-24266 (gfx94X/gfx950)

Root Cause

Two layers:

1. XLA ROCm profiler uses legacy roctracer v2 API

The current XLA pins for both rocm-jaxlib-v0.9.1 (3cc8846c) and rocm-jaxlib-v0.9.2 (d8b2a5f5) use rocm_tracer_v1.cc / device_tracer_rocm.cc with the legacy roctracer/rocprofiler v2 API.

• On gfx120X (RDNA4): rocprofiler v2 support was deliberately removed (merged 2025-09-22). Every HIP API callback is dropped with "No capturing function was found" — 28,746 dropped callbacks in CI logs, producing empty FDO profiles.
• On gfx94X/gfx950 (CDNA): v2 still works for basic tracing, but the ROCm collector doesn't produce CUPTI-equivalent XPlane annotations needed by ConvertXplaneToProfiledInstructionsProto.

2. Upstream migration exists but is not in the XLA pins

openxla/xla#29769 by cj401-amd ported XLA from roctracer v2 to rocprofiler-sdk v3. The PR was closed (copybara), but the code appears to have partially landed in upstream XLA main (files renamed from rocm_tracer_v1.cc to rocm_tracer.cc with rocprofiler-sdk references).

Four follow-up fix PRs for the v3 code were also closed without merging:

• openxla/xla#38683 — VLOG(1) for LOG(INFO)
• openxla/xla#38771 — clean up includes
• openxla/xla#38777 — error checking and status propagation
• openxla/xla#41997 — stop rocprofiler context in Disable() to prevent cross-session event leak

What needs to happen

1. Bump the XLA pin on rocm-jaxlib-v0.9.2 (and future branches) to a commit that includes the rocprofiler-sdk v3 profiling backend
2. Cherry-pick or resubmit the follow-up fix PRs if they haven't landed upstream
3. Verify FDO profile generation produces valid custom kernel cost entries on both gfx94X and gfx120X
4. Address rocprofiler-sdk gfx120X gaps: rocm-systems PR #5271 documented a 75% dispatch miss rate on RDNA4 DefaultSignal path — closed without merge
5. Re-enable PGLE tests by removing --deselect entries from ci/run_pytest_rocm.sh

AMD contributors on the upstream migration

• cj401-amd (chunyjin@amd.com) — original migration PR
• magaonka-amd (magaonka@amd.com) — follow-up fix PRs