[AMD] Add ATOM+MTP, dsr1 fp8/fp4

[seungrokj]

Hi @cquil11 @functionstackx

This is an MTP enablement of ATOM with a new docker img.

--use-chat-template is applied.

DS R1 FP4 / FP8 with MTP is enabled.

Regards,
Seungrok

[seungrokj]

/sweep test-config --config-files .github/configs/amd-master.yaml --runner-config .github/configs/runners.yaml --config-keys dsr1-fp8-mi355x-atom-mtp dsr1-fp4-mi355x-atom-mtp

[github-actions]

@seungrokj Kicking off a sweep.

Run: https://github.com/InferenceMAX/InferenceMAX/actions/runs/21861432813
Command: test-config --config-files .github/configs/amd-master.yaml --runner-config .github/configs/runners.yaml --config-keys dsr1-fp8-mi355x-atom-mtp dsr1-fp4-mi355x-atom-mtp
Pinned ref: 044abaf
Approval: not required (trusted collaborator).

[functionstackx]

@cquil11 or @Oseltamivir can u review this plz?

[functionstackx]

@pr-claude review please

[claude]

Claude finished @functionstackx's task in 3m 20s —— View job

---

Review of PR #673

• Gather context and read changed files
• Validate master config and perf-changelog
• Validate container image accessibility
• Validate launch script enroot import
• Validate benchmark script formatting and MTP requirements
• Validate model prefix format
• Check for bugs and inconsistencies between scripts
• Post review summary

---

Summary

No blocking issues found. Two minor warnings posted as inline comments:

1. benchmarks/dsr1_fp8_mi355x_atom_mtp.sh:77 — Trailing backslash on rm -rf line creates an unintentional line continuation. Also, this cleanup block is absent from the FP4 MTP script.
2. .github/configs/amd-master.yaml:65,70,295,298 — Trailing double spaces in some spec-decoding: mtp } entries (cosmetic inconsistency).

Validation checklist

Check	Result
perf-changelog.yaml updated	✅
Container image publicly accessible (rocm/atom:... on Docker Hub)	✅
Launch script has enroot import docker:// pattern	✅
--use-chat-template in MTP benchmark scripts	✅
Server commands use multi-line formatting	✅
Model prefix is valid (dsr1)	✅
SPEC_DECODING pipeline correctly wired through config → sweep → workflow → runner → script	✅
EP_SIZE defaults to 1 when ep not in config	✅

Overall the PR is well-structured and consistent with existing atom configs. LGTM with the minor nits above.

[cquil11]

@seungrokj Sweep completing now, then we will merge.