[AMD][CI] Merge the four MI35x DeepSeek-V3.2 nightly jobs into two to save runtime

[michaelzhang-ai]

Motivation

MI35x DeepSeek-V3.2 coverage is split across four jobs — accuracy and perf, basic and MTP — while the MI30x V3.2 jobs in the same workflow do the same work in two, each running accuracy then perf. Both steps in a pair serve the same weights, so the split buys a second container setup and a second cold weight load per config and nothing else.

Modifications

nightly-accuracy-8-gpu-mi35x-deepseek-v32-rocm720      + nightly-perf-8-gpu-mi35x-deepseek-v32-basic-rocm720
  -> nightly-8-gpu-mi35x-deepseek-v32-rocm720
nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp-rocm720  + nightly-perf-8-gpu-mi35x-deepseek-v32-mtp-rocm720
  -> nightly-8-gpu-mi35x-deepseek-v32-mtp-rocm720

New names follow the MI30x convention with an mi35x infix. Suites, launch flags, step timeouts and --timeout-per-file are carried over verbatim; job_select and check-all-jobs are updated to the new names. 45 jobs -> 43. No test file or suite touched.

The two merged jobs end up with exactly the gating shape of the MI30x V3.2 pair in this same workflow — accuracy blocking, perf non-blocking, no if: guard — so MI35x and MI30x V3.2 now behave identically.

• Perf is skipped when accuracy fails, by default step behaviour. The split jobs ran the two independently, so a failing night still produced a perf table; benchmarking a configuration that just missed its GSM8K floor is not worth ~27 min of an 8-GPU node.

Accuracy Tests

No test, suite, flag, threshold or timeout changes.

Both merged jobs passed in run 32907097246 (rocm720, continue_on_error=false), each running its accuracy step and then its perf step. The dispatch resolved both new job names with no old accuracy-/perf- name present, checking the renames against real Actions resolution. That run predates #36396, but these two job bodies are byte-identical to what it executed.

pre-commit passes, including check for duplicate workflow job names and validate registered test CI registries. Every job body diffed against main: only the two merged jobs and check-all-jobs differ, so #36396's additions and removals are preserved.

Speed Tests and Profiling

Four split jobs from the 8/24 and 8/26 scheduled runs versus the two merged jobs from run 32907097246. All rocm720, linux-mi35x-gpu-8, GPU-hours = wall time x 8.

	jobs	total	setup+install	test	GPU-h
split (mean of 2 nights)	4	134.5 min	26.3	108.0	17.93
merged	2	114.0 min	12.3	101.6	15.20
delta	-2	-20.5 min (-15%)	-14.0	-6.4	-2.73

Across both image flavors that is ~5.5 GPU-h/night, and 8 scheduled MI35x 8-GPU job runs become 4 — likely the bigger win given queue depths on that pool. Of the 20.5 min, the 14.0 min of eliminated setup is mechanical and reliable; the 6.4 min of test time is one sample against two and within the baseline's own 4.9 min spread, so the defensible floor is ~3.7 GPU-h/night from setup alone.

On a night where accuracy fails, skipping perf reclaims a further ~27 min of node time per affected config, which the split jobs spent benchmarking a broken configuration.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests
• Update documentation
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ✅ Run #33122934517
Latest PR Test (Extra): ❌ Run #33122934627
Latest PR Test (AMD ROCm 7.2): ➖ No AMD PR run found for this commit.