[AMD] Add dsv4-fp4-mi355x-atom-disagg-mtp config

[Oseltamivir]

Summary

Adds a DeepSeek-V4-Pro FP4 MI355X ATOM disaggregated MTP benchmark, dsv4-fp4-mi355x-atom-disagg-mtp (spec-decoding=mtp, num-speculative-tokens=3 via DECODE_MTP_SIZE).

This is config-only — no new benchmark script:

• Mirrors the dsv4-fp4-mi355x-atom-disagg topology: 2P1D DPA+TP8 (conc 256–2048) and 1P1D TP8 (conc 4–128).
• Reuses benchmarks/multi_node/dsv4_fp4_mi355x_atom-disagg.sh; the multinode launcher builds the script name without a _mtp suffix, so the same script serves STP and MTP.
• MTP is driven by the DeepSeek-V4-Pro mtp_flags already present in benchmarks/multi_node/amd_utils/models_atom.yaml (--method mtp --num-speculative-tokens) plus DECODE_MTP_SIZE in the decode additional-settings; server_atom.sh composes the spec args.
• Results are disambiguated from the base config by the spec-mtp segment in RESULT_FILENAME (no collision).

Image

Bumped to rocm/atom-dev:nightly_202607201550 (2026-07-20) from the base config's stale 2026-06-10 nightly. Verified on-node that this image carries the dsv4 MTP path (deepseek_v4 -> deepseek_v4_mtp -> DeepseekV4MTPModel) and the atomesh disagg router (/usr/local/bin/atomesh).

Draft depth

DECODE_MTP_SIZE=3 on both arms, matching the single-node dsv4_fp4_mi355x_atom_mtp.sh recipe. The Atom MTP head runs autoregressively N times, so N=3 is valid even though dsv4 ships one nextn layer. A shallower high-concurrency depth (per the dsr1-fp4-mi355x-sglang-disagg-mtp concurrency taper) may follow once results land.

Validation

• generate_sweep_configs.py test-config — passes Pydantic, generates 2 MTP arms, max-model-len 9472 auto-injected.
• pytest utils/matrix_logic/ — 224 passed.
• Not yet run on real GPUs (MI355X cluster was saturated); intended to run through the normal run-sweep path.

Note: perf-changelog.yaml pr-link will be updated to this PR's URL in a follow-up commit.

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档