From c3221b2da2fc314b79c2429679e82b32d8601c75 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Sat, 20 Jun 2026 13:26:18 +0900
Subject: [PATCH 01/13] feat: add minimaxm3-fp4-mi355x-atom-mtp benchmark
 script and CI config

Add single-node ATOM MTP (EAGLE3) benchmark for MiniMax-M3-MXFP4 on MI355X
with TP/EP/DP-attention support and matching amd-master.yaml entry.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .github/configs/amd-master.yaml               | 22 +++++
 .../minimaxm3_fp4_mi355x_atom_mtp.sh          | 82 +++++++++++++++++++
 2 files changed, 104 insertions(+)
 create mode 100644 benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 92f8a5609..d2ccbeb2c 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2876,6 +2876,28 @@ minimaxm3-fp4-mi355x-atom:
       - { tp: 4, conc-start: 1, conc-end: 256 }
       - { tp: 8, conc-start: 1, conc-end: 2 }
 
+minimaxm3-fp4-mi355x-atom-mtp:
+  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: 2, conc-start: 128, conc-end: 256 }
+      - { tp: 4, conc-start: 1, conc-end: 256 }
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 2, conc-start: 128, conc-end: 256 }
+      - { tp: 4, conc-start: 1, conc-end: 256 }
+      - { tp: 8, conc-start: 1, conc-end: 2 }
+
 # 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
diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
new file mode 100644
index 000000000..53238959a
--- /dev/null
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
@@ -0,0 +1,82 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL \
+    TP \
+    CONC \
+    ISL \
+    OSL \
+    RANDOM_RANGE_RATIO \
+    RESULT_FILENAME \
+    EP_SIZE \
+    DP_ATTENTION 
+
+if [[ -n "$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: $DP_ATTENTION"
+
+SERVER_LOG=/workspace/server.log
+
+PARALLEL_ARGS=(-tp "$TP") #TP
+if [ "$DP_ATTENTION" = "true" ]; then
+    if [ "$EP_SIZE" -gt 1 ]; then #DP+EP
+        PARALLEL_ARGS=(-tp "$TP" --enable-expert-parallel --enable-dp-attention )
+    else #DP+TP
+        PARALLEL_ARGS=(-tp "$TP" --enable-dp-attention )
+    fi
+fi 
+
+SPEC_ARGS=(--method eagle3 --draft-model Inferact/MiniMax-M3-EAGLE3 --num-speculative-tokens 3 )
+
+# Start GPU monitoring (power, temperature, clocks every second)
+start_gpu_monitor
+MEM_FRAC_STATIC=0.8
+
+set -x
+# (srok), not yet
+#    --kv_cache_dtype fp8 \
+python3 -m atom.entrypoints.openai_server \
+    --model $MODEL \
+    --server-port $PORT \
+    -tp $TP \
+    --max-model-len $MAX_MODEL_LEN $EP \
+    "${PARALLEL_ARGS[@]}" \
+    "${SPEC_ARGS[@]}" \
+    --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 "$RESULT_FILENAME" \
+    --result-dir /workspace/ \
+    --trust-remote-code \
+    --use-chat-template 
+
+# After throughput, run evaluation only if RUN_EVAL is true
+if [ "${RUN_EVAL}" = "true" ]; then
+    run_eval --framework lm-eval --port "$PORT"
+    append_lm_eval_summary
+fi
+
+# Stop GPU monitoring
+stop_gpu_monitor
+set +x

From d6426c4f567306edb80cf5f3eaa3aa8b988af2db Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Sat, 20 Jun 2026 13:29:08 +0900
Subject: [PATCH 02/13] chore: add perf-changelog entry for
 minimaxm3-fp4-mi355x-atom-mtp

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 perf-changelog.yaml | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 46bee4d44..c05963d36 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3964,6 +3964,13 @@
     - "Remove the runtime SupportsEagle3 source patch now included in the pinned nightly"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1843
 
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-mtp
+  description:
+    - "Add minimaxm3-fp4-mi355x-atom-mtp CI recipe: single-node ATOM with EAGLE3 speculative decoding (3 draft tokens) for MiniMax-M3-MXFP4 on MI355X"
+    - "Supports TP/EP/DP-attention combinations; search space mirrors minimaxm3-fp4-mi355x-atom (ISL=1024,8192 OSL=1024 TP2/TP4/TP8)"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1866
+
 - config-keys:
     - kimik2.5-fp4-gb200-dynamo-trt
   description:

From a1b42c4fe02fc1fe8145af79a86e2ab6ce992a86 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Sat, 20 Jun 2026 13:29:52 +0900
Subject: [PATCH 03/13] fix: update minimaxm3-fp4-mi355x-atom-mtp image to
 MiniMax-M3-20260619

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .github/configs/amd-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index d2ccbeb2c..bd13cd743 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2877,7 +2877,7 @@ minimaxm3-fp4-mi355x-atom:
       - { tp: 8, conc-start: 1, conc-end: 2 }
 
 minimaxm3-fp4-mi355x-atom-mtp:
-  image: rocm/atom-dev:M3
+  image: rocm/atom-dev:MiniMax-M3-20260619
   model: amd/MiniMax-M3-MXFP4
   model-prefix: minimaxm3
   runner: mi355x

From ae97ec64e34a4d410f5bbab7365578ff143fccc3 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Sat, 20 Jun 2026 13:36:44 +0900
Subject: [PATCH 04/13] fix: add spec-decoding: mtp to
 minimaxm3-fp4-mi355x-atom-mtp search space; remove stray \$EP arg

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .github/configs/amd-master.yaml                        | 10 +++++-----
 .../fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh     |  2 +-
 2 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index bd13cd743..96e29427a 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2889,14 +2889,14 @@ minimaxm3-fp4-mi355x-atom-mtp:
     - isl: 1024
       osl: 1024
       search-space:
-      - { tp: 2, conc-start: 128, conc-end: 256 }
-      - { tp: 4, conc-start: 1, conc-end: 256 }
+      - { tp: 2, conc-start: 128, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, conc-start: 1, conc-end: 256, spec-decoding: mtp }
     - isl: 8192
       osl: 1024
       search-space:
-      - { tp: 2, conc-start: 128, conc-end: 256 }
-      - { tp: 4, conc-start: 1, conc-end: 256 }
-      - { tp: 8, conc-start: 1, conc-end: 2 }
+      - { tp: 2, conc-start: 128, conc-end: 256, spec-decoding: mtp }
+      - { tp: 4, conc-start: 1, conc-end: 256, spec-decoding: mtp }
+      - { tp: 8, conc-start: 1, conc-end: 2, spec-decoding: mtp }
 
 # 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
diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
index 53238959a..7276bcb4f 100644
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
@@ -43,7 +43,7 @@ python3 -m atom.entrypoints.openai_server \
     --model $MODEL \
     --server-port $PORT \
     -tp $TP \
-    --max-model-len $MAX_MODEL_LEN $EP \
+    --max-model-len $MAX_MODEL_LEN \
     "${PARALLEL_ARGS[@]}" \
     "${SPEC_ARGS[@]}" \
     --block-size 128 \

From 4fcaa1ef3f9b93baed13f486b0d4e2b5930041c3 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Sat, 20 Jun 2026 13:42:47 +0900
Subject: [PATCH 05/13] fix: remove duplicate -tp and --max-model-len from
 minimaxm3_fp4_mi355x_atom_mtp.sh server args

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .../single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh  | 2 --
 1 file changed, 2 deletions(-)

diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
index 7276bcb4f..120a801e2 100644
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
@@ -42,8 +42,6 @@ set -x
 python3 -m atom.entrypoints.openai_server \
     --model $MODEL \
     --server-port $PORT \
-    -tp $TP \
-    --max-model-len $MAX_MODEL_LEN \
     "${PARALLEL_ARGS[@]}" \
     "${SPEC_ARGS[@]}" \
     --block-size 128 \

From d806d96d19b27df6b976c1b9385bc9a6ef6e1664 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Sat, 20 Jun 2026 15:30:12 +0900
Subject: [PATCH 06/13] fix: trim minimaxm3-fp4-mi355x-atom-mtp ISL=8192 search
 space; remove TP8, narrow TP2 to conc=128

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .github/configs/amd-master.yaml | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 96e29427a..5c047cc5a 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2894,9 +2894,8 @@ minimaxm3-fp4-mi355x-atom-mtp:
     - isl: 8192
       osl: 1024
       search-space:
-      - { tp: 2, conc-start: 128, conc-end: 256, spec-decoding: mtp }
+      - { tp: 2, conc-start: 128, conc-end: 128, spec-decoding: mtp }
       - { tp: 4, conc-start: 1, conc-end: 256, spec-decoding: mtp }
-      - { tp: 8, conc-start: 1, conc-end: 2, spec-decoding: mtp }
 
 # 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

From 86308ff5423c3c3e3255794cb1e518fd612125a7 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 23 Jun 2026 10:46:49 +0900
Subject: [PATCH 07/13] fix: bump minimaxm3-fp4-mi355x-atom-mtp image to
 MiniMax-M3-20260622

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .github/configs/amd-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 30d7da3c4..706ca31d2 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2613,7 +2613,7 @@ minimaxm3-fp4-mi355x-atom:
       - { tp: 8, conc-start: 1, conc-end: 2 }
 
 minimaxm3-fp4-mi355x-atom-mtp:
-  image: rocm/atom-dev:MiniMax-M3-20260619
+  image: rocm/atom-dev:MiniMax-M3-20260622
   model: amd/MiniMax-M3-MXFP4
   model-prefix: minimaxm3
   runner: mi355x

From 4458bbe355121e1a6bb89956557d84bbc1524db2 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 23 Jun 2026 12:02:40 +0900
Subject: [PATCH 08/13] fix: disable prefix caching for
 minimaxm3-fp4-mi355x-atom-mtp

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .../single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh   | 1 +
 1 file changed, 1 insertion(+)

diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
index 120a801e2..3b2a28907 100644
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
@@ -47,6 +47,7 @@ python3 -m atom.entrypoints.openai_server \
     --block-size 128 \
     --gpu-memory-utilization $MEM_FRAC_STATIC \
     --trust-remote-code \
+    --no-enable_prefix_caching \
     > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!

From bb638f914e4282ab64c15f7ee1a851d6871a5171 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 23 Jun 2026 12:27:42 +0900
Subject: [PATCH 09/13] fix: update minimaxm3-fp4-mi355x-atom scripts and image
 bump

- minimaxm3_fp4_mi355x_atom.sh: refactor parallel args, set MAX_MODEL_LEN/
  MAX_NUM_BATCHED_TOKENS/MAX_NUM_SEQS, disable prefix caching, conditional
  --use-chat-template, enable AITER_QUICK_REDUCE_QUANTIZATION=INT4
- minimaxm3_fp4_mi355x_atom_mtp.sh: same server arg additions and conditional
  --use-chat-template
- amd-master.yaml: bump minimaxm3-fp4-mi355x-atom image to MiniMax-M3-20260622,
  drop tp8 search-space entry

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .github/configs/amd-master.yaml               |  3 +-
 .../minimaxm3_fp4_mi355x_atom.sh              | 48 +++++++++----------
 .../minimaxm3_fp4_mi355x_atom_mtp.sh          | 10 +++-
 3 files changed, 32 insertions(+), 29 deletions(-)

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 3a61aed78..3a41a9e21 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2591,7 +2591,7 @@ minimaxm3-fp8-mi355x-vllm-mtp:
 # 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
+  image: rocm/atom-dev:MiniMax-M3-20260622
   model: amd/MiniMax-M3-MXFP4
   model-prefix: minimaxm3
   runner: mi355x
@@ -2610,7 +2610,6 @@ minimaxm3-fp4-mi355x-atom:
       search-space:
       - { tp: 2, conc-start: 128, conc-end: 256 }
       - { tp: 4, conc-start: 1, conc-end: 256 }
-      - { tp: 8, conc-start: 1, conc-end: 2 }
 
 minimaxm3-fp4-mi355x-atom-mtp:
   image: rocm/atom-dev:MiniMax-M3-20260622
diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh
index 80f230f28..9bf74b6e6 100644
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh
@@ -11,8 +11,7 @@ check_env_vars \
     RANDOM_RANGE_RATIO \
     RESULT_FILENAME \
     EP_SIZE \
-    DP_ATTENTION \
-    MAX_MODEL_LEN
+    DP_ATTENTION 
 
 if [[ -n "$SLURM_JOB_ID" ]]; then
   echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
@@ -22,41 +21,40 @@ echo "TP: $TP, CONC: $CONC, ISL: $ISL, OSL: $OSL, EP_SIZE: $EP_SIZE, DP_ATTENTIO
 
 SERVER_LOG=/workspace/server.log
 
-export OMP_NUM_THREADS=1
+PARALLEL_ARGS=(-tp "$TP") #TP
+if [ "$DP_ATTENTION" = "true" ]; then
+    if [ "$EP_SIZE" -gt 1 ]; then #DP+EP
+        PARALLEL_ARGS=(-tp "$TP" --enable-expert-parallel --enable-dp-attention )
+    else #DP+TP
+        PARALLEL_ARGS=(-tp "$TP" --enable-dp-attention )
+    fi
+fi 
 
-# 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
+SPEC_ARGS=()
 
 # 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.
+export AITER_QUICK_REDUCE_QUANTIZATION=INT4
+export MAX_MODEL_LEN=32768
+export MAX_NUM_BATCHED_TOKENS=32768
+export MAX_NUM_SEQS=128
+# (srok), not yet
+#    --kv_cache_dtype fp8 \
 python3 -m atom.entrypoints.openai_server \
     --model $MODEL \
     --server-port $PORT \
-    -tp $TP \
-    --max-model-len $MAX_MODEL_LEN $EP \
+    "${PARALLEL_ARGS[@]}" \
+    "${SPEC_ARGS[@]}" \
     --block-size 128 \
     --gpu-memory-utilization $MEM_FRAC_STATIC \
+    --max-model-len $MAX_MODEL_LEN \
+    --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
+    --max-num-seqs $MAX_NUM_SEQS \
     --trust-remote-code \
+    --no-enable_prefix_caching \
     > $SERVER_LOG 2>&1 &
 
 SERVER_PID=$!
@@ -76,7 +74,7 @@ run_benchmark_serving \
     --max-concurrency "$CONC" \
     --result-filename "$RESULT_FILENAME" \
     --result-dir /workspace/ \
-    --trust-remote-code
+    --trust-remote-code $( [[ ${#SPEC_ARGS[@]} -gt 0 ]] && echo "--use-chat-template" )
 
 # After throughput, run evaluation only if RUN_EVAL is true
 if [ "${RUN_EVAL}" = "true" ]; then
diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
index 3b2a28907..0ee78f504 100644
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
@@ -37,6 +37,10 @@ start_gpu_monitor
 MEM_FRAC_STATIC=0.8
 
 set -x
+export AITER_QUICK_REDUCE_QUANTIZATION=INT4
+export MAX_MODEL_LEN=32768
+export MAX_NUM_BATCHED_TOKENS=32768
+export MAX_NUM_SEQS=128
 # (srok), not yet
 #    --kv_cache_dtype fp8 \
 python3 -m atom.entrypoints.openai_server \
@@ -46,6 +50,9 @@ python3 -m atom.entrypoints.openai_server \
     "${SPEC_ARGS[@]}" \
     --block-size 128 \
     --gpu-memory-utilization $MEM_FRAC_STATIC \
+    --max-model-len $MAX_MODEL_LEN \
+    --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
+    --max-num-seqs $MAX_NUM_SEQS \
     --trust-remote-code \
     --no-enable_prefix_caching \
     > $SERVER_LOG 2>&1 &
@@ -67,8 +74,7 @@ run_benchmark_serving \
     --max-concurrency "$CONC" \
     --result-filename "$RESULT_FILENAME" \
     --result-dir /workspace/ \
-    --trust-remote-code \
-    --use-chat-template 
+    --trust-remote-code $( [[ ${#SPEC_ARGS[@]} -gt 0 ]] && echo "--use-chat-template" )
 
 # After throughput, run evaluation only if RUN_EVAL is true
 if [ "${RUN_EVAL}" = "true" ]; then

From a592385a9b4ea81b158605a719b32b930380998d Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 23 Jun 2026 12:30:54 +0900
Subject: [PATCH 10/13] fix: add minimaxm3-fp4-mi355x-atom to perf-changelog
 entry

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 perf-changelog.yaml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 9934e1f94..82201851b 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3985,6 +3985,7 @@
 
 - config-keys:
     - minimaxm3-fp4-mi355x-atom-mtp
+    - minimaxm3-fp4-mi355x-atom
   description:
     - "Add minimaxm3-fp4-mi355x-atom-mtp CI recipe: single-node ATOM with EAGLE3 speculative decoding (3 draft tokens) for MiniMax-M3-MXFP4 on MI355X"
     - "Supports TP/EP/DP-attention combinations; search space mirrors minimaxm3-fp4-mi355x-atom (ISL=1024,8192 OSL=1024 TP2/TP4/TP8)"

From 182decd12ae646ee71f8901fbc3263c7d8cd3562 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 23 Jun 2026 13:04:18 +0900
Subject: [PATCH 11/13] fix: revert minimaxm3-fp4-mi355x-atom
 image/search-space, delete script, update perf-changelog

- amd-master.yaml: revert image to rocm/atom-dev:M3, restore tp8 search-space entry
- perf-changelog.yaml: remove minimaxm3-fp4-mi355x-atom from mtp entry
- delete minimaxm3_fp4_mi355x_atom.sh

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .github/configs/amd-master.yaml               |  3 +-
 .../minimaxm3_fp4_mi355x_atom.sh              | 87 -------------------
 perf-changelog.yaml                           |  1 -
 3 files changed, 2 insertions(+), 89 deletions(-)
 delete mode 100644 benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 3a41a9e21..3a61aed78 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2591,7 +2591,7 @@ minimaxm3-fp8-mi355x-vllm-mtp:
 # 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:MiniMax-M3-20260622
+  image: rocm/atom-dev:M3
   model: amd/MiniMax-M3-MXFP4
   model-prefix: minimaxm3
   runner: mi355x
@@ -2610,6 +2610,7 @@ minimaxm3-fp4-mi355x-atom:
       search-space:
       - { tp: 2, conc-start: 128, conc-end: 256 }
       - { tp: 4, conc-start: 1, conc-end: 256 }
+      - { tp: 8, conc-start: 1, conc-end: 2 }
 
 minimaxm3-fp4-mi355x-atom-mtp:
   image: rocm/atom-dev:MiniMax-M3-20260622
diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh
deleted file mode 100644
index 9bf74b6e6..000000000
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh
+++ /dev/null
@@ -1,87 +0,0 @@
-#!/usr/bin/env bash
-
-source "$(dirname "$0")/../../benchmark_lib.sh"
-
-check_env_vars \
-    MODEL \
-    TP \
-    CONC \
-    ISL \
-    OSL \
-    RANDOM_RANGE_RATIO \
-    RESULT_FILENAME \
-    EP_SIZE \
-    DP_ATTENTION 
-
-if [[ -n "$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: $DP_ATTENTION"
-
-SERVER_LOG=/workspace/server.log
-
-PARALLEL_ARGS=(-tp "$TP") #TP
-if [ "$DP_ATTENTION" = "true" ]; then
-    if [ "$EP_SIZE" -gt 1 ]; then #DP+EP
-        PARALLEL_ARGS=(-tp "$TP" --enable-expert-parallel --enable-dp-attention )
-    else #DP+TP
-        PARALLEL_ARGS=(-tp "$TP" --enable-dp-attention )
-    fi
-fi 
-
-SPEC_ARGS=()
-
-# Start GPU monitoring (power, temperature, clocks every second)
-start_gpu_monitor
-MEM_FRAC_STATIC=0.8
-
-set -x
-export AITER_QUICK_REDUCE_QUANTIZATION=INT4
-export MAX_MODEL_LEN=32768
-export MAX_NUM_BATCHED_TOKENS=32768
-export MAX_NUM_SEQS=128
-# (srok), not yet
-#    --kv_cache_dtype fp8 \
-python3 -m atom.entrypoints.openai_server \
-    --model $MODEL \
-    --server-port $PORT \
-    "${PARALLEL_ARGS[@]}" \
-    "${SPEC_ARGS[@]}" \
-    --block-size 128 \
-    --gpu-memory-utilization $MEM_FRAC_STATIC \
-    --max-model-len $MAX_MODEL_LEN \
-    --max-num-batched-tokens $MAX_NUM_BATCHED_TOKENS \
-    --max-num-seqs $MAX_NUM_SEQS \
-    --trust-remote-code \
-    --no-enable_prefix_caching \
-    > $SERVER_LOG 2>&1 &
-
-SERVER_PID=$!
-
-# 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 $( [[ ${#SPEC_ARGS[@]} -gt 0 ]] && echo "--use-chat-template" )
-
-# After throughput, run evaluation only if RUN_EVAL is true
-if [ "${RUN_EVAL}" = "true" ]; then
-    run_eval --framework lm-eval --port "$PORT"
-    append_lm_eval_summary
-fi
-
-# Stop GPU monitoring
-stop_gpu_monitor
-set +x
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 82201851b..9934e1f94 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -3985,7 +3985,6 @@
 
 - config-keys:
     - minimaxm3-fp4-mi355x-atom-mtp
-    - minimaxm3-fp4-mi355x-atom
   description:
     - "Add minimaxm3-fp4-mi355x-atom-mtp CI recipe: single-node ATOM with EAGLE3 speculative decoding (3 draft tokens) for MiniMax-M3-MXFP4 on MI355X"
     - "Supports TP/EP/DP-attention combinations; search space mirrors minimaxm3-fp4-mi355x-atom (ISL=1024,8192 OSL=1024 TP2/TP4/TP8)"

From e8b2abacfb5127af564b55b72ec349495673566f Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 23 Jun 2026 13:13:49 +0900
Subject: [PATCH 12/13] fix: restore minimaxm3_fp4_mi355x_atom.sh

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .../minimaxm3_fp4_mi355x_atom.sh              | 89 +++++++++++++++++++
 1 file changed, 89 insertions(+)
 create mode 100644 benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh

diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom.sh
new file mode 100644
index 000000000..80f230f28
--- /dev/null
+++ b/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

From c6e8d15770bd349cdb0dc1aae2618d6decaa0ec2 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Tue, 23 Jun 2026 14:48:25 +0900
Subject: [PATCH 13/13] fix: bump MAX_NUM_SEQS to 256 for
 minimaxm3-fp4-mi355x-atom-mtp

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 .../single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh  | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
index 0ee78f504..fc0619b7b 100644
--- a/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
+++ b/benchmarks/single_node/fixed_seq_len/minimaxm3_fp4_mi355x_atom_mtp.sh
@@ -40,7 +40,7 @@ set -x
 export AITER_QUICK_REDUCE_QUANTIZATION=INT4
 export MAX_MODEL_LEN=32768
 export MAX_NUM_BATCHED_TOKENS=32768
-export MAX_NUM_SEQS=128
+export MAX_NUM_SEQS=256
 # (srok), not yet
 #    --kv_cache_dtype fp8 \
 python3 -m atom.entrypoints.openai_server \