[AMD] Add MiniMax-M3-FP4 MI355X ATOM EAGLE3 / non-EAGLE3 update 0623#1917
Open
seungrokj wants to merge 18 commits into
Open
[AMD] Add MiniMax-M3-FP4 MI355X ATOM EAGLE3 / non-EAGLE3 update 0623#1917seungrokj wants to merge 18 commits into
seungrokj wants to merge 18 commits into
Claude / Claude Code Review
completed
Jun 24, 2026 in 8m 45s
Code review found 1 important issue
Found 5 candidates, confirmed 2. See review comments for details.
Details
| Severity | Count |
|---|---|
| 🔴 Important | 1 |
| 🟡 Nit | 0 |
| 🟣 Pre-existing | 0 |
| Severity | File:Line | Issue |
|---|---|---|
| 🔴 Important | benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh:33 |
SPEC_ARGS swapped between fp4 atom MTP and non-MTP scripts |
Annotations
Check failure on line 33 in benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh
claude / Claude Code Review
SPEC_ARGS swapped between fp4 atom MTP and non-MTP scripts
The fp4 atom non-MTP and MTP scripts have their `SPEC_ARGS` arrays swapped relative to their fp8 siblings (and relative to their own names): `minimaxm3_fp4_mi355x_atom.sh` line 33 sets the EAGLE3 draft-model args while `minimaxm3_fp4_mi355x_atom_mtp.sh` line 33 sets `SPEC_ARGS=()`. Result: the new `minimaxm3-fp4-mi355x-atom-mtp` recipe (the headline feature of this PR) will silently run **without** speculative decoding, while the existing non-MTP recipe will run **with** EAGLE3 + chat template.
Loading