[AMD/ROCM] Qwen3.5-397B-A17B BF16 MI355X Atom benchmarks

[seungrokj]

Summary

• Add qwen3.5-bf16-mi355x-atom config to amd-master.yaml
• Add benchmark script benchmarks/single_node/qwen3.5_bf16_mi355x_atom.sh using the Atom framework
• Update perf-changelog.yaml with description

Details

• Model: Qwen/Qwen3.5-397B-A17B (BF16)
• Hardware: MI355X (single node)
• Framework: Atom (rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.post)
• KV cache: FP8
• Sweep: TP4 and TP8, ISL/OSL 1k/1k and 8k/1k, concurrency 4–256

🤖 Generated with Claude Code

[github-actions]

Thanks for the contribution! For vLLM & SGLang, please ensure that your recipes is similar to the official vLLM recipes and/or the SGLang cookbook

• https://cookbook.sglang.io/
• https://docs.vllm.ai/projects/recipes/en/latest/

If it is not, please create a PR first before we can merge your PR into the master branch. Let's ensure that the documentation is first class such that the entire ML community can benefit from your hard work! Thank you

• https://github.com/vllm-project/recipes
• https://github.com/sgl-project/sgl-cookbook

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. If re-running failed jobs is attempted, PR authors are responsible for ensuring it passes. See GitHub's docs on re-running failed jobs: https://docs.github.com/en/actions/how-tos/manage-workflow-runs/re-run-workflows-and-jobs#re-running-failed-jobs-in-a-workflow

As a rule of thumb, generally, PR authors should request a review & get a PR approval from the respective companies' CODEOWNERS before requesting a review from core maintainers.

If additional help is needed, PR authors can reach out to core maintainers over Slack.

[claude]

🟡 The new perf-changelog entry for qwen3.5-bf16-mi355x-atom ends with pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/ (empty PR number) at line 1821. This URL resolves to the GitHub PR listing page rather than this specific PR; it should be updated to /pull/1149 to match this PR for traceability.

Extended reasoning...

What the bug is: Line 1821 of perf-changelog.yaml ends with pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/ — the trailing path component is empty. This is the new entry added in this PR for the qwen3.5-bf16-mi355x-atom config.

How it manifests: When a reader clicks the link, they hit https://github.com/SemiAnalysisAI/InferenceX/pull/ which resolves to the GitHub pull request listing page for the repo, not the specific PR that introduced this changelog entry. Traceability from the changelog entry back to the originating PR is lost.

Why existing patterns don't catch this: Other entries in the file use one of two conventions: a real PR number (e.g. /pull/1129, /pull/1148) or an explicit placeholder (/pull/XXX, /pull/XXXX, /pull/TBD). Placeholders are at least visually recognizable as TODOs. An empty trailing component is structurally indistinguishable from a complete URL on quick inspection, but produces a broken link that silently redirects to the listing page.

Impact: Purely cosmetic/documentation. The benchmark execution path doesn't read pr-link, so this has no runtime effect on the new qwen3.5-bf16-mi355x-atom config or any other benchmark. The harm is limited to anyone trying to follow the changelog history back to the introducing PR.

Step-by-step proof:

1. Open perf-changelog.yaml and scroll to the last block (lines 1815-1821).
2. The block has config-keys: [qwen3.5-bf16-mi355x-atom] and a description for a Day-0 Atom benchmark.
3. Line 1821 reads literally: pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/ — note no number after the trailing slash.
4. Compare to the immediately preceding entry (line 1814): pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1129 — has a real number.
5. Visiting the empty-tail URL in a browser redirects to the repo's PR list, not this PR.

How to fix: Change line 1821 to pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1149 (this PR's number). One-character-class fix; no other files affected.