Skip to content

[AMD] Add MiniMax-M3-FP8 MI355X ATOM EAGLE3 / non-EAGLE3 update 0623#1916

Open
seungrokj wants to merge 10 commits into
mainfrom
amd/m3_atom_single_fp8_0623
Open

[AMD] Add MiniMax-M3-FP8 MI355X ATOM EAGLE3 / non-EAGLE3 update 0623#1916
seungrokj wants to merge 10 commits into
mainfrom
amd/m3_atom_single_fp8_0623

[AMD] add perf-changelog entry for minimaxm3-fp8-mi355x-atom

a245590
Select commit
Loading
Failed to load commit list.
Claude / Claude Code Review completed Jun 24, 2026 in 8m 24s

Code review found 3 potential issues

Found 5 candidates, confirmed 3. See review comments for details.

Details

Severity Count
🔴 Important 0
🟡 Nit 3
🟣 Pre-existing 0
Severity File:Line Issue
🟡 Nit perf-changelog.yaml:4004-4010 Missing perf-changelog entry for mtp recipe updates
🟡 Nit benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi355x_atom.sh:77 Dead --use-chat-template conditional copied from mtp variant
🟡 Nit perf-changelog.yaml:4009 perf-changelog pr-link points to wrong PR

Annotations

Check warning on line 4010 in perf-changelog.yaml

See this annotation in the file changed.

@claude claude / Claude Code Review

Missing perf-changelog entry for mtp recipe updates

The existing changelog entry for `minimaxm3-fp8-mi355x-atom-mtp` (around line 4011) is left untouched at PR #1867 / image `...-20260619`, even though this PR materially changes that recipe: image bump to `...-20260623`, `AITER_QUICK_REDUCE_QUANTIZATION=INT4` commented out, `ATOM_M3_SPARSE_USE_ASM_PA=1` added, and most notably `--kv_cache_dtype fp8` is now enabled (previously commented out as '(srok), not yet'). Per the established convention in this file — e.g. the `minimaxm3-fp8-mi300x-vllm-mtp

Check warning on line 77 in benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi355x_atom.sh

See this annotation in the file changed.

@claude claude / Claude Code Review

Dead --use-chat-template conditional copied from mtp variant

Dead `--use-chat-template` conditional copied from the mtp sibling. On line 33 `SPEC_ARGS=()` is initialized empty and never mutated anywhere in this non-mtp file, so the trailing `$( [[ ${#SPEC_ARGS[@]} -gt 0 ]] && echo "--use-chat-template" )` on line 77 always expands to nothing — pure dead code. Nit: drop the conditional (and the unused `SPEC_ARGS=()` / `"${SPEC_ARGS[@]}"` plumbing if you want to go further) so this script stops mimicking the mtp variant's spec-decoding shape.

Check warning on line 4009 in perf-changelog.yaml

See this annotation in the file changed.

@claude claude / Claude Code Review

perf-changelog pr-link points to wrong PR

The `pr-link` for this new `minimaxm3-fp8-mi355x-atom` perf-changelog entry points to `/pull/1929`, but this is PR #1916 — readers following the changelog will land on the wrong (or non-existent) PR. Update to `https://github.com/SemiAnalysisAI/InferenceX/pull/1916`.