[Klaud Cold] Add minimaxm3-fp4-mi355x-atom (upstream branch for full-sweep validation) (#1813)

* minimaxm3-fp4-mi355x-atom: day-zero MiniMax-M3 MXFP4 MI355X atom recipe

Smoke-tested on MI355X (mia1-p01-g07): TP4 conc-1 1k1k served and benched
clean (mean TPOT 6.8ms). KV cache left at default dtype — amd/MiniMax-M3-MXFP4
has no calibrated FP8 KV scales, so --kv_cache_dtype fp8 asserts in the MSA
fused_qknorm kernel.

* minimaxm3-fp4-mi355x-atom: route amd/MiniMax-M3* weights to NFS cache

* minimaxm3-fp4-mi355x-atom: fill perf-changelog pr-link

* minimaxm3-fp4-mi355x-atom: use matrix MAX_MODEL_LEN (isl+osl+256)

* trigger full-sweep validation

* perf-changelog: point minimaxm3-fp4-mi355x-atom pr-link to upstream PR #1813

---------

Co-authored-by: shekhar <shekhar.pandey@amd.com>
Co-authored-by: seungrokj <144636725+seungrokj@users.noreply.github.com>

Author: 3 people

.github/configs/amd-master.yaml

@@ -2851,6 +2851,28 @@ minimaxm3-fp8-mi355x-vllm-mtp:
       - { tp: 4, conc-start: 1, conc-end: 64, spec-decoding: mtp }
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 128, conc-end: 256, spec-decoding: mtp }
 
+# MiniMax-M3 MXFP4 MI355X atom recipe:
+# https://github.com/ROCm/ATOM/blob/5d42d49f9e4292e5b61475917e92e7ec1b1dacb7/recipes/MiniMax-M3.md
+# block size 128 is mandatory for MSA. TP4 on a single gfx950 node, per the recipe.
+minimaxm3-fp4-mi355x-atom:
+  image: rocm/atom-dev:M3
+  model: amd/MiniMax-M3-MXFP4
+  model-prefix: minimaxm3
+  runner: mi355x
+  precision: fp4
+  framework: atom
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 1024
+      osl: 1024
+      search-space:
+      - { tp: 4, conc-start: 1, conc-end: 128 }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 4, conc-start: 1, conc-end: 128 }
+
 # MiniMax-M3 MXFP8 MI300X day-zero recipe. Reuse the dedicated ROCm image and
 # MI355X serving shape, but retain the default BF16 KV cache because this
 # checkpoint lacks calibrated ROCm FP8 attention scales. Use the TP8-only H100

benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh

@@ -0,0 +1,89 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME \
+    EP_SIZE \
+    DP_ATTENTION \
+    MAX_MODEL_LEN
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTION: $DP_ATTENTION"
+
+SERVER_LOG=/workspace/server.log
+
+export OMP_NUM_THREADS=1
+
+# Use the matrix-supplied MAX_MODEL_LEN (isl + osl + 256). Eval-only jobs need a
+# larger window for the eval prompts, so override it from the eval context.
+if [ "${EVAL_ONLY}" = "true" ]; then
+    setup_eval_context
+    MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN"
+fi
+
+if [ "$EP_SIZE" -gt 1 ]; then
+  EP=" --enable-expert-parallel"
+else
+  EP=" "
+fi
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+MEM_FRAC_STATIC=0.8
+
+set -x
+
+# Flags follow the ATOM MiniMax-M3 MXFP4 recipe (FP4 on 4xMI355 section):
+# https://github.com/ROCm/ATOM/blob/5d42d49f9e4292e5b61475917e92e7ec1b1dacb7/recipes/MiniMax-M3.md
+# --block-size 128 is mandatory for MiniMax MSA. KV cache is left at the default
+# dtype: amd/MiniMax-M3-MXFP4 ships no calibrated FP8 KV scales, so
+# --kv_cache_dtype fp8 trips an assertion (k_scale is None) in the MSA
+# fused_qknorm kernel during init.
+python3 -m atom.entrypoints.openai_server \
+    --model $MODEL \
+    --server-port $PORT \
+    -tp $TP \
+    --max-model-len $MAX_MODEL_LEN $EP \
+    --block-size 128 \
+    --gpu-memory-utilization $MEM_FRAC_STATIC \
+    --trust-remote-code \
+    > $SERVER_LOG 2>&1 &
+
+SERVER_PID=$!
+
+# Wait for server to be ready
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+export PYTHONDONTWRITEBYTECODE=1
+run_benchmark_serving \
+    --model "$MODEL" \
+    --port "$PORT" \
+    --backend vllm \
+    --input-len "$ISL" \
+    --output-len "$OSL" \
+    --random-range-ratio "$RANDOM_RANGE_RATIO" \
+    --num-prompts "$((CONC * 10))" \
+    --max-concurrency "$CONC" \
+    --result-filename "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x

perf-changelog.yaml

@@ -3918,6 +3918,13 @@
     - "This issue is now fixed in the latest TRTLLM release."
     - "Also update all configs for DSR1 TRTLLM FP8 to reflect latest released image usage"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1767
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom
+  description:
+    - "Add day-zero MiniMax-M3 MXFP4 (amd/MiniMax-M3-MXFP4) single-node atom benchmark on MI355X, following the ROCm/ATOM MiniMax-M3 recipe (TP4, block size 128 for MSA, default KV cache dtype)."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1813
+
 
 - config-keys:
     - glm5-fp4-gb300-dynamo-trt

runners/launch_mi355x-amds.sh

@@ -242,8 +242,9 @@ else
     fi
 
     # MiniMax-M3 weights are not staged on the node-local /var/lib NVMe cache;
-    # they are pre-downloaded once to the NFS share instead.
-    if [[ "$MODEL" == MiniMaxAI/MiniMax-M3* ]]; then
+    # they are pre-downloaded once to the NFS share instead. Covers both the
+    # MiniMaxAI MXFP8 checkpoint and the amd MXFP4 atom checkpoint.
+    if [[ "$MODEL" == MiniMaxAI/MiniMax-M3* || "$MODEL" == amd/MiniMax-M3* ]]; then
         export HF_HUB_CACHE_MOUNT="/it-share/hf-hub-cache/"
     fi