[AMD/ROCm] qwen3.5-fp8-mi355x-atom, Bump image to rocm/atom:rocm7.2.3_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom20260511

[seungrokj]

Summary

• Bump ATOM image for qwen3.5-fp8-mi355x-atom from rocm/atom:rocm7.2.2_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.2.post to rocm/atom-dev:nightly_202605111702
• TP=4 shows +3.2% to +16.3% throughput improvement across 1k1k and 8k1k workloads (concurrency 4-256)
• Validated via ATOM upstream nightly benchmark run #25686894636 (2026-05-11)

Throughput Comparison (tput/GPU, tok/s)

ISL	OSL	Conc	GPUs	InferenceX	ATOM Upstream	Diff %
1024	1024	4	4	208.99	221.58	+6.0%
1024	1024	8	4	373.92	407.45	+9.0%
1024	1024	16	4	599.13	627.41	+4.7%
1024	1024	32	4	955.48	998.01	+4.5%
1024	1024	64	4	1368.61	1429.45	+4.4%
1024	1024	128	4	1929.98	1992.52	+3.2%
1024	1024	256	4	2861.12	2936.57	+2.6%
8192	1024	4	4	811.98	944.35	+16.3%
8192	1024	8	4	1496.50	1631.33	+9.0%
8192	1024	16	4	2276.05	2417.82	+6.2%
8192	1024	32	4	3388.72	3619.13	+6.8%
8192	1024	64	4	4498.52	4775.50	+6.2%
8192	1024	128	4	5321.91	5943.18	+11.7%
8192	1024	256	4	6925.46	7330.06	+5.8%

Changes

• .github/configs/amd-master.yaml: Update image tag for qwen3.5-fp8-mi355x-atom
• perf-changelog.yaml: Add changelog entry

Test plan

• Verify benchmark runs successfully with new image on MI355X

🤖 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://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node 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/sglang/tree/main/docs_new

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.

[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://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node 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/sglang/tree/main/docs_new

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.

[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://docs.sglang.io/cookbook/intro
• https://recipes.vllm.ai/

If it is not, please create a PR first before we can merge your single node 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/sglang/tree/main/docs_new

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]

LGTM, straightforward ATOM image tag bump for the qwen3.5-fp8-mi355x-atom benchmark config.

Extended reasoning...

Overview

This PR bumps the ATOM container image for the qwen3.5-fp8-mi355x-atom benchmark config from atom0.1.2.post to the nightly atom20260511 build (ROCm 7.2.3). The only changes are a one-line image tag update in .github/configs/amd-master.yaml and a corresponding entry in perf-changelog.yaml.

Security risks

None. This is a benchmark configuration file update — no auth, crypto, permissions, or runtime code is affected. The image is published under the existing trusted rocm/atom namespace already used by many other configs in this file.

Level of scrutiny

Low. This is a mechanical config tweak following an established pattern repeated throughout amd-master.yaml (most other configs reference the same rocm/atom registry). It only affects the single qwen3.5-fp8-mi355x-atom benchmark entry and is self-contained.

Other factors

No bugs were found by the bug-hunting system. The changelog entry follows the same format used by other entries in perf-changelog.yaml. The PR description includes a test plan and notes throughput improvements observed at TP=4.