[AMD] Add DeepSeek-V4.1-Flash MI35x nightly accuracy test

[michaelzhang-ai]

Motivation

DeepSeek-V4.1-Flash has no AMD test registered, so nothing in CI exercises the MI35x FP4 path that the cookbook recipe publishes for it.

Modifications

• New test test/registered/amd/accuracy/mi35x/test_deepseek_v41_flash_eval_mi35x.py: a GSM8K 8-shot completion eval (1319 questions) with deepseek-ai/DeepSeek-V4.1-Flash.

  • Server args are the MI350X High-Throughput cell of docs/src/snippets/configs/deepseek-ai/deepseek-v4_1.jsx, as dsv4.1-amd: serve DeepSeek-V4.1 on gfx950 #41308 left it: --tp 4 --ep-size 4 --mem-fraction-static 0.8 --speculative-algorithm DSPARK --speculative-dspark-block-size 5 --cuda-graph-max-bs-decode 256 --max-running-requests 256 --enable-decoder-swa-bounded-replay. The attention and MoE backends are left to resolve on HIP.
  • Env is the recipe's AITER_FLYDSL_FORCE_REDUCE=1 AITER_BF16_FP8_MOE_BOUND=0. SGLANG_USE_AITER and SGLANG_MOE_PADDING come from the ROCm image. AITER_BF16_FP8_MOE_BOUND=0 stays until [FlyDSL] [Bugfix] Route a16w4 SiLU MoE to CK-Tile and apply swiglu_limit ROCm/aiter#5802 is in the AITER pin.
  • It uses the completion harness, like the DeepSeek-V4-Flash MI35x test, so V4.1's thinking mode can't route the answer into reasoning_content.

• New nightly job nightly-4-gpu-mi35x-deepseek-v41-flash in nightly-test-amd.yml (dropdown option, job, check-all-jobs.needs), suite nightly-amd-4-gpu-mi35x-deepseek-v41-flash. It runs on linux-mi35x-gpu-8 and uses four GPUs, like the MiniMax TP4 jobs.

• The suite is added to NIGHTLY_SUITES in test/run_suite.py.

• The DeepSeek-V3.2 MI35x nightly jobs are removed. This means nightly-8-gpu-mi35x-deepseek-v32 and nightly-8-gpu-mi35x-deepseek-v32-mtp: their job definitions, their job_select options and their check-all-jobs.needs entries. They no longer run on the schedule or by manual dispatch. This offsets the MI35x runner time the new job adds. DeepSeek-V3.2 has been replaced by V4, and both jobs passed on all three ROCm images for the last 8 scheduled nightlies (48/48), at about 100 runner-minutes per night. The accuracy and perf tests stay in the tree, and the MI30x V3.2 jobs are unchanged.

This is MI35x only because the FP4 experts need gfx95x. The recipe publishes no MI30x cell.

Status and threshold

• The gfx950 series this needs (dsv4.1-amd: KV cache layouts, FP4 indexer, compressor and router kernels #41019, dsv4.1-amd: gfx950 sparse decode attention and sorted top-k #41020, dsv4.1-amd: fused mHC boundary and all-reduce + mHC post kernels #41021, dsv4.1-amd: serve DeepSeek-V4.1 on gfx950 #41308) is all on main, and this branch is rebased on top of it.

• Threshold 0.89, from measured runs. The 4-GPU DeepSeek-V4.1-Flash job ran manually on all three ROCm images (MI35x, all 1319 questions):

  ROCm	Run	GSM8K	Invalid	Test (server start + eval)	Eval only	Whole job
  rocm10	36797490320	0.911	0.001	18.1 min	62 s	24 min
  rocm724	36868324806	0.917	0.000	9.3 min	62 s	16 min
  rocm720	36868333055	0.920	0.001	8.6 min	52 s	15 min

  The threshold sits 0.02 under the lowest score (rocm10), replacing the provisional 0.86. For comparison, dsv4.1-amd: serve DeepSeek-V4.1 on gfx950 #41308 measured 0.870–0.890 with the same cell on 200 questions.

• The rocm10 run took about 9 minutes longer to start the server, likely because the weights weren't yet cached on that runner. est_time=1200 covers that case. Container setup and dependency install add about 6–7 minutes to every job. So the new job adds about 15–24 minutes per ROCm image per night on linux-mi35x-gpu-8, and uses four of its eight GPUs.

• A Low-Latency (DSpark, decode cap 128) variant and a perf step are left for follow-ups.

Checklist

• Dispatch nightly-test-amd.yml with job_select=nightly-4-gpu-mi35x-deepseek-v41-flash on this branch for rocm10, rocm724 and rocm720, and set the threshold from the results (0.911 / 0.917 / 0.920 → threshold 0.89)

---

CI States

Latest PR Test (Base): ❌ Run #36871380250
Latest PR Test (Extra): ❌ Run #36871379298
Latest PR Test (AMD ROCm 10): ❌ Run #36871380180