[AMD] Add Qwen3.8-Flash-Next-FP8 nightly validation on ROCm 7.2 (verified on rocm720)

[michaelzhang-ai]

Motivation

Add the AMD nightly accuracy gate for Qwen/Qwen3.8-Flash-Next-FP8 on ROCm 7.2.

The generic model (qwen4_exp / Qwen4ExpConfig) landed in #37500 (2026-09-08), which closed/superseded #36497. This branch is merged onto current main and the skip guard is gone, so the nightly runs GSM8K instead of OK (skipped=1).

Two jobs, matching the two topologies the released FP8 checkpoint needs — gfx942 cannot hold target + MTP draft + graph/KV state on one GPU, gfx950 can:

Job	Runner	Topology
nightly-2-gpu-mi30x-qwen38-flash-next-fp8-rocm720	linux-mi300-2gpu-sglang	TP2 + EP2, mem-fraction-static 0.75
nightly-1-gpu-mi35x-qwen38-flash-next-fp8-rocm720	linux-mi35x-gpu-1	TP1, mem-fraction-static 0.95

Both exercise the AITER decode-graph path with EAGLE speculation (3 / 1 / 4) against a pinned checkpoint revision, gated on GSM8K ≥ 0.94 plus a multimodal smoke check. Direct AITER paged QSA stays disabled so the correctness gate does not depend on an unreleased AITER API.

Modifications

CI only — two files.

• [AMD] Add Qwen3.8 Flash Next nightly validation — cherry-picked from [AMD] Support Qwen3.8-Flash-Next MXFP4 on ROCm #36601 (3003ddf15), authorship preserved. Adds the two workflow jobs to nightly-test-amd-rocm720.yml and test/registered/amd/accuracy/test_qwen38_flash_next_fp8_eval.py. Suites are registered in the test file via register_amd_ci, so no test/run_suite.py change is needed.
• Merge origin/main — picks up support qwen 3.8 flash next #37500 (Qwen4ExpConfig and the QSA/PLE stack).
• Arm the nightly — drop the Qwen4ExpConfig skip guard that existed only while model support was missing.

Net diff against main is those two files.

Remaining follow-up (not this PR)

The suite is fully on the intended AITER MoE path once #36601's 619632b2c ("Honor explicit AITER MoE backend selection") lands. This test uses --moe-runner-backend aiter together with SGLANG_USE_AITER=0. Until that restack, the CLI flag may not independently select AITER MoE.

NVIDIA QSA follow-ups (#36644 FP8 KV, #36556 SM12x) are not ROCm blockers. Issue #36616 remains an image-rebuild item, not a missing code change.

Accuracy Tests

Wiring was verified on the rocm720 flavor before #37500, with job_select narrowed to the MI35x job:

Run	Head	Result
33205733803	before the old skip guard	infrastructure green, failed on KeyError: 'qwen4_exp'
33206455622	skip guard	green (OK (skipped=1))
33209050367	after merging main	green
33215101724	09a9efb0f	green (OK (skipped=1))

Those runs prove checkout, VRAM clear, ROCm 7.2.0 container setup, dependency install, HF cache, suite resolution, and server CLI. They could not score GSM8K because qwen4_exp was not on main.

This update: skip guard removed; current head 31de6319de includes #37500. A new MI35x job_select dispatch is needed to record a real GSM8K number. For reference, #36601 reports GSM8K 96.80% on gfx950 FP8 TP1 and 96.96% on gfx942 FP8 TP2/EP2 for this path.

Not verified in the old runs: the rocm724 matrix flavor, the 2-GPU MI30x job.

Local checks: both jobs are defined, land on the right runners, appear in the job_select options, self-match their own if gate, and are wired into check-all-jobs.needs without dropping any job that is already on main; the two suite names the workflow invokes are exactly the two the test file registers.

Speed Tests and Profiling

Not applicable — CI-only change.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations. (Not applicable; CI-only.)
• 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 #34829340183
Latest PR Test (Extra): ❌ Run #34829339697
Latest PR Test (AMD ROCm 10): ❌ Run #34829340008

[mintlify]

Preview deployment for your docs. Learn more about Mintlify Previews.

Project	Status	Preview	Updated (UTC)
lmsysorg	🟢 Ready	View Preview	Aug 28, 2026, 7:19 PM

💡 Tip: Enable Workflows to automatically generate PRs for you.

[cursor]

Update after #37500 merged (2026-09-08)

• Merged current main into this branch so Qwen4ExpConfig / qwen4_exp is present.
• Dropped the skip guard in test_qwen38_flash_next_fp8_eval.py. The nightly should now run GSM8K instead of OK (skipped=1).
• Marked the PR ready for review.

Still a follow-up, not this diff: #36601 619632b2c (honor explicit --moe-runner-backend aiter with SGLANG_USE_AITER=0). Issue #36616 remains an image-rebuild item.

Previous rocm720 MI35x wiring proof: https://github.com/sgl-project/sglang/actions/runs/33215101724 (skipped because model support was not on main yet). A new job_select dispatch is needed for a real GSM8K number.