[AMD] Add MiniMax-M3-FP8 MI355X ATOMMESH#1865

Open

seungrokj wants to merge 35 commits into

mainfrom

amd/atom_mesh_0619_m3_fp8

Commits on Jun 23, 2026

[AMD] server_atom: improve config print and cleanup

seungrokj
and
claude
committed
update perf-changelog for dsv4-fp4-mi355x-atom-disagg-mtp

seungrokj
and
claude
committed
[AMD] fix DECODE_MTP_SIZE and BENCH_REQUEST_RATE propagation in atom-disagg

seungrokj
and
claude
committed
[AMD] server_atom: pass SPEC_ARGS to prefill server

seungrokj
and
claude
committed
[AMD] amd-master: fix comment for 1P1D TP8+DPA+TBO+MTP1 config

seungrokj
and
claude
committed
[AMD] dsv4_atom-disagg: remove DECODE_MTP_SIZE from check_env_vars

seungrokj
and
claude
committed
[AMD] bench: use --dsv4 flag for DeepSeek-V4-Pro MTP benchmarks

seungrokj
and
claude
committed
[AMD] server_atom: export IS_MTP=true when SPEC_DECODING=mtp for bench.sh

seungrokj
and
claude
committed
[AMD] server_atom: fix hf-overrides JSON quoting

seungrokj
and
claude
committed
fix: inline --hf-overrides to avoid eval word-splitting, remove OPT_ARGS

seungrokj
and
claude
committed
refactor: extract --hf-overrides into HF_OVERRIDES_ARG variable

seungrokj
and
claude
committed
fix: enable --hf-overrides only for DeepSeek-V4-Pro

seungrokj
and
claude
committed
fix: add HF_OVERRIDES_ARG to INFO config print block

seungrokj
and
claude
committed
fix: replace broken-quote array splice with ${ARRAY[*]} in CMD strings

seungrokj
and
claude
committed
fix: remove ${CUDAGRAPH_OPT} from decode CMD

seungrokj
and
claude
committed
feat: add MiniMax-M3 ATOM disagg CI script and server_atom.sh support

seungrokj
and
claude
committed
feat: add minimaxm3-fp4-mi355x-atom-disagg recipe and AITER_QUICK_REDUCE_QUANTIZATION=INT4

seungrokj
and
claude
committed
feat: export AITER_QUICK_REDUCE_QUANTIZATION=INT4 for non-DSv4 models

seungrokj
and
claude
committed
fix: server_atom.sh and minimaxm3 disagg cleanup

seungrokj
and
claude
committed
fix: dsv4_fp4_mi355x_atom-disagg cleanup

seungrokj
and
claude
committed
fix: set BLOCK_SIZE=128 for MiniMax-M3 in minimaxm3_fp4_mi355x_atom-disagg.sh

seungrokj
and
claude
committed
fix: use KV_CACHE_DTYPE=fp8 for MiniMax-M3 disagg (matches atom server default)

seungrokj
and
claude
committed
feat: update minimaxm3-fp4-mi355x-atom-disagg search space and disable --enable-tbo for non-DSv4 models

seungrokj
and
claude
committed
feat: add MiniMax-M3-MXFP4/MXFP8 to models_atom.yaml; set KV_CACHE_DTYPE default to empty for minimaxm3 disagg

seungrokj
and
claude
committed
fix: set mi355x-disagg runner and add dynamic cudagraph sizes for decode node

seungrokj
and
claude
committed
fix: gate ATOM_MOE_GU_ITLV and AITER_BF16_FP8_MOE_BOUND on DeepSeek-V4-Pro only

seungrokj
and
claude
committed
fix: preserve empty KV_CACHE_DTYPE to skip --kv-cache-dtype flag

seungrokj
and
claude
committed
fix: use KV_CACHE_DTYPE=auto for minimaxm3 disagg to skip --kv-cache-dtype flag

seungrokj
and
claude
committed
fix: align minimaxm3 disagg settings with slurm reference script

seungrokj
and
claude
committed
fix: rename minimaxm3-fp4-mi355x-atom-disagg to minimaxm3-fp8 and remove stale perf-changelog entry

seungrokj
and
claude
committed
feat: add minimaxm3_fp8_mi355x_atom-disagg multi-node benchmark script

seungrokj
and
claude
committed
benchmarks: rename minimaxm3 to dsv4 atom-disagg script and generalize config

seungrokj
and
claude
committed
fix: bump minimaxm3-fp8-mi355x-atom-disagg image and pin MAX_MODEL_LEN

seungrokj
and
claude
committed
Merge branch 'main' into amd/atom_mesh_0619_m3_fp8
seungrokj
authored
Merge branch 'main' into amd/atom_mesh_0619_m3_fp8
seungrokj
authored