[AMD][CI] Run only the ROCm 7.2.4 image on the nightly schedule

[michaelzhang-ai]

Motivation

#35603 put both rocm724 and rocm720 on the one 17:30 nightly. Four complete nights ran that way (2026-08-20 … 08-23, 47 job pairs a night) before #36171 changed the image-pull path. This PR makes the schedule run rocm724 only, on cost and redundancy grounds.

The 24-hour overruns on those four nights were not caused by the second leg. They were the shared local registry at 10.44.14.109:5000. On 08-22 a single MI30x job resolved its image at 00:27:35 and finished docker pull from the local registry at 03:33:50 — one successful pull, no fallback, 186 minutes. MI35x jobs never went down that path and held a 4.1 min median throughout, even though their job count doubled too, and the untouched ROCm 7.0 nightly on the same MI30x pools took the same 10x hit without changing its job count at all. With #36171 bypassing the local registry, tonight's run has both legs, an MI30x Setup docker median back at 14.9 min (max 18.8), and 86 of 92 jobs done in about 7.4 h. So both flavors do fit in a night now; the wall-clock argument for touching this workflow is gone.

What remains is cost and redundancy.

Of 188 job pairs over the four nights, 166 reached the same completed verdict, 17 disagreed and 5 involved a cancellation. Only five job-nights had rocm720 fail while rocm724 passed, spread over four distinct jobs, and none reproduced on another night:

job	night
nightly-8-gpu-mi35x-deepseek-r1-hicache	08-20
nightly-8-gpu-mi35x-qwen35-triton-dcp	08-20
nightly-8-gpu-minimax-m27	08-22
nightly-accuracy-8-gpu-mi35x	08-22
nightly-4-gpu-mi35x-minimax-m25	08-23

No durable 7.2.0-specific regression appeared. The unique signal ran the other way: nightly-8-gpu-grok1-int4 failed on 724 three nights out of four while 720 stayed green all four, and grok2, minimax-m3 and glm5-mxfp4 each had a 724-only failure. That is expected, since rocm724 is Ubuntu 24.04 / Python 3.12 / torch 2.11 / triton 3.7 against 720's 22.04 / 3.10 / 2.9.1, and it argues for keeping 724 rather than both. Tonight repeats the pattern: 9 failures on 724, 8 on 720.

On cost, tonight's healthy run puts each leg at roughly 205–212 GPU-hours, so the second leg is about 210 GPU-hours a night.

Modifications

• nightly-test-amd-rocm720.yml: the rocm_version matrix resolves to ["rocm724"] when no input is given, so the schedule and push triggers run one leg. all still expands to both.
• Same file: the workflow_dispatch default moves from all to rocm724, with rocm720 and all still in the dropdown. The workflow_call default was already rocm724; its stale comment is corrected.
• amd-ci-job-monitor.yml: rocm_flavors drops to rocm724, which that script is documented to keep in sync with the matrix.

Verified that every trigger resolves as intended (schedule / push → ["rocm724"], dispatch default → ["rocm724"], dispatch rocm720 → ["rocm720"], dispatch or caller all → both), that no existing caller passes all (amd-aiter-scout, bot-bump-sglang-version and release-branch-cut all take the rocm724 default), and that check_workflow_job_names.py, actionlint and the full pre-commit set pass.

Open question for AMD CI owners

This is a coverage-vs-cost call, not a fix for an outage, so it is reasonable to reject it. rocm720 images are still published nightly by release-docker-amd-rocm720-nightly.yml and still built for stable releases by release-docker-amd.yml, so after this change they have no scheduled coverage. Three ways to go:

1. merge this and accept that the published 7.2.0 images are untested;
2. merge this plus a weekly 7.2.0 pass over a smoke subset (needs github.event.schedule in the concurrency group so the two crons do not cancel each other);
3. drop this PR and keep both legs nightly, now that the registry fix has restored the ~7 h runtime, and revisit if the ~210 GPU-hours a night becomes a problem.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests
• Update documentation
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #32796016068
Latest PR Test (Extra): ❌ Run #32796015960
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.