[CI] Parity: classify newly observed ROCm skips

[ethanwee1]

Summary

• classify Triton kernel, Triton heuristic, and fused-attention tests as PT2.0 - Inductor
• classify backend skips, including TVM coverage, as PT2.0 - Dynamo
• classify the new hipSOLVER DnXsytrs requires ROCm >= 7.14 guard as hipSolver/Magma
• support both dotted and slash-separated test file names

Dashboard analysis

Reviewed the AI analysis generated on 2026-08-04. It confirms that the inductor.test_triton_heuristics and TVM backend movements are classifier artifacts addressed by this PR. I also checked the underlying analyzed mi350 report rather than relying only on the AI recommendations: all 62 incoming unclassified skips are now covered—60 hipSOLVER version guards and two inductor.test_fused_attention tests.

The analysis also reports large historical convolution, distributed, FSDP, and Misc category movements, but explicitly identifies them as unrelated collection/classifier churn and does not provide test-level evidence for a new category rule. This PR does not add speculative broad rules for those movements.

This PR was split from the earlier parity automation stack and contains only skip-reason classification changes.

Test plan

• Run Python syntax, lint, and whitespace checks
• Run the classifier against the analyzed mi350 parity CSV
• Confirm coverage rises from 62 unclassified skips to 0 (100% of 2,877 target tests classified)
• Verify this PR changes only auto_classify_skip_reasons.py

[ethanwee1]

Validation runs

Dispatched parity.yml on the fix branch (ethanwee1/pytorch) for SHA 050ce9e1226b95d6b5938bdd5e245491f7219fd8 — the exact SHA that was failing before the fix — across all ROCm arches:

arch	run	result	CUDA shards detected
mi350	30567121048	✅ success	default=14, distributed=10 (25 CUDA jobs)
mi200	30569040561	✅ success	default=14, distributed=10 (25 CUDA jobs)
mi300	30569038602	⚠️ flagged incomplete	default=14, distributed=10 (25 CUDA jobs)

The CUDA baseline now downloads correctly on every arch — no more TEST KEY ... DOES NOT EXIST IN JOBS / no artifacts found on the CUDA side.

The mi300 run is flagged failed only by the intentional graceful-degradation step: for this specific SHA the ROCm mi300 test-report artifacts (default/distributed/inductor) are absent in S3/GHA, so a partial report was produced and the job self-flagged. That is a pre-existing data-availability path unrelated to this change (the CUDA fix itself worked identically there).

[ethanwee1]

Rebased onto latest develop (0376854). Ready for Jenkins re-run.

[ethanwee1]

Tested — now detects shard counts for all architectures

Second commit extends dynamic shard detection to every ROCm arch (mi350/mi300/mi200/…) plus the CUDA trunk baseline. The standalone ROCm arch workflows run their shards inside a reusable workflow (workflow_call), so those shards don't appear in the jobs API — derive_shard_count now also consults the commit check-runs API before the config fallback.

Validation runs on the PR branch code (ethanwee1/pytorch, SHA 21f65597):

arch	default	distributed	inductor	CUDA	run	report
mi350	8	3	2	14/10	✅ 30654399051	366,939 rows
mi200	10	3	4	14/10	✅ 30654402900	360,145 rows
mi300	8 (config had 6)	3	4 (config had 2)	14/10	⚠️ 30654401126	shards correct; artifacts in trunk-rocm-sandbox (separate follow-up)

mi300 proves the detection catches an actual upstream reshard (6→8, 2→4) that the hardcoded config missed. Its run only fails on the distinct "which run hosts the reusable-workflow artifacts" issue — not a shard-count problem — which is being tracked separately.

[ethanwee1]

Added a third commit: auto-detect ROCm job-name prefixes (resolve_job_prefix). The per-arch prefix in parity_job_config.json (e.g. linux-noble-rocm-py3.12-mi300, the nightly gfx942→mi350 rename) is now a self-healing hint — detected from the resolved run's jobs API + commit check-runs when it drifts, kept as-is when still present. CUDA stays configured (trunk carries multiple deliberate CUDA families). Live no-regression mi350 run: 30655560233 ✅ (366,948 rows). Merged to the deployment target ethanwee1/pytorch main as #5.

[jithunnair-amd]

Follow-up fix for resolve_job_prefix scoping: ethanwee1#6 (stacked on this branch). Scopes check-runs to workflow runs, filters ROCm-only prefixes, and adds arch verification.

[ethanwee1]

Ported the two remaining deployment-fork fixes into this PR:

• Busted job URLs (_wf_run_ids.json per-job run-id map; summarize builds each shard's URL from it). Matches fork #7.
• Artifact-hosting run (resolve_artifact_download: ROCm arch shards that run under trunk-rocm-sandbox are downloaded from the run that actually ran them, filtered by exact job ids). Matches fork #8.

This PR now carries the full set: CUDA shard detection, all-arch ROCm shard detection (check-runs), job-name prefix auto-detection, busted-job-URL fix, and artifact-hosting-run resolution.