From dc74d0a484b0d0f8e3538415173f8d8706308cf5 Mon Sep 17 00:00:00 2001
From: Duyi-Wang <duyi.wang@amd.com>
Date: Wed, 24 Jun 2026 05:07:33 +0000
Subject: [PATCH 01/10] [AMD] Add MiniMax-M3-MXFP4 MI355X vLLM disagg recipe

Disaggregated (prefill/decode) vLLM recipe for amd/MiniMax-M3-MXFP4 on MI355X
over the MoRI-IO KV connector.

Recipe:
- benchmarks/multi_node/minimaxm3_fp4_mi355x_vllm-disagg.sh: launcher.
- models_vllm.yaml: MiniMax-M3-MXFP4 entry. block-size 128 (MSA), TRITON_ATTN,
  --language-model-only, AITER MoE, minimax_m3 parsers. No --kv-cache-dtype fp8
  (the checkpoint ships no calibrated FP8 KV scales).
- amd-master.yaml: minimaxm3-fp4-mi355x-vllm-disagg config, 8k1k, two layouts
  (1P1D TP4 and 2P1D TP4), conc 1..512.

Supporting fixes to the shared vllm-disagg path:
- server_vllm.sh: count prefill/decode GPUs from the per-worker TP size
  (PREFILL_TP_SIZE*xP / DECODE_TP_SIZE*yD) instead of GPUS_PER_NODE*xP. With
  TP < node GPU count (e.g. TP4 on an 8-GPU node) the old expression
  over-counted, corrupting PREFILL_GPUS/DECODE_GPUS and halving tput_per_gpu.
- env.sh / job.slurm: set the MoRI-IO RDMA QP knobs (MORI_IO_QP_MAX_SEND_WR
  etc.) for the vllm-disagg path. They were only set in the SGLang branch, so
  vllm-disagg ran at the default send-queue depth and stalled at high
  concurrency ("SQ full").
---
 .github/configs/amd-master.yaml               |  50 +++++++
 benchmarks/multi_node/amd_utils/env.sh        | 122 +++++++++---------
 .../multi_node/amd_utils/models_vllm.yaml     |   6 +
 .../multi_node/amd_utils/server_vllm.sh       |   2 +-
 .../minimaxm3_fp4_mi355x_vllm-disagg.sh       |  78 +++++++++++
 5 files changed, 198 insertions(+), 60 deletions(-)
 create mode 100755 benchmarks/multi_node/minimaxm3_fp4_mi355x_vllm-disagg.sh

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index a3f352adaf..4fafc96206 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2587,6 +2587,56 @@ 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 vLLM disaggregated (prefill/decode) config.
+minimaxm3-fp4-mi355x-vllm-disagg:
+  image: rocm/sgl-dev:vllm-0.23.1-rocm723-mi35x-mori-0623-m3
+  model: amd/MiniMax-M3-MXFP4
+  model-prefix: minimaxm3
+  runner: mi355x-disagg
+  precision: fp4
+  framework: vllm-disagg
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 1P TP4 + 1D TP4 (2 nodes total), conc sweep 1..512 (single job, looped)
+      - spec-decoding: "none"
+        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128, 256, 512 ]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "PREFILL_NODES=1"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "DECODE_NODES=1"
+      # 2P TP4 + 1D TP4 (3 nodes total), conc sweep 1..512 (single job, looped)
+      - spec-decoding: "none"
+        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128, 256, 512 ]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "PREFILL_NODES=2"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "DECODE_NODES=1"
+
 # 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.
diff --git a/benchmarks/multi_node/amd_utils/env.sh b/benchmarks/multi_node/amd_utils/env.sh
index a24347114b..a05182bf16 100755
--- a/benchmarks/multi_node/amd_utils/env.sh
+++ b/benchmarks/multi_node/amd_utils/env.sh
@@ -45,6 +45,68 @@ set +x
 
 export NCCL_IB_HCA=${NCCL_IB_HCA:-$IBDEVICES}
 
+# =============================================================================
+# MoRI-specific environment
+# =============================================================================
+# Shared by the vLLM MoRIIOConnector and the SGLang/MoRI KV-transfer path.
+
+export MORI_IO_SQ_BACKOFF_TIMEOUT_US=50000
+export MORI_IO_QP_MAX_SEND_WR=16384
+export MORI_IO_QP_MAX_CQE=32768
+export MORI_IO_QP_MAX_SGE=2
+export MORI_IO_TC_DISABLE=0
+
+# QoS/DSCP configuration
+# Priority order: 1) Set by runner, 2) Detect via nicctl, 3) Detect from hostname
+if [[ -n "$MORI_RDMA_TC" ]]; then
+    echo "[INFO] Using MORI_RDMA_TC=$MORI_RDMA_TC (set by runner or environment)"
+elif command -v nicctl &> /dev/null; then
+    ND_PRIO=$(nicctl show qos  2>/dev/null | awk '/PFC no-drop priorities/ {print $NF; exit}')
+    ND_DSCP=$(nicctl show qos 2>/dev/null| awk -v p="$ND_PRIO" '
+$1 == "DSCP" && $2 == ":" && $NF == p {
+    print $3; exit
+}')
+
+    if [[ -n "$ND_DSCP" ]] && [[ -n "$ND_PRIO" ]]; then
+        TC=$(( 4 * ND_DSCP ))
+        export MORI_RDMA_SL=$ND_PRIO
+        export MORI_IO_SL=$ND_PRIO
+        export MORI_RDMA_TC=$TC
+        export MORI_IO_TC=$TC
+        echo "[INFO] Detected QoS config from nicctl: MORI_RDMA_TC=$MORI_RDMA_TC, MORI_RDMA_SL=$MORI_RDMA_SL, MORI_IO_TC=$MORI_IO_TC, MORI_IO_SL=$MORI_IO_SL"
+    else
+        echo "[WARN] nicctl available but QoS data unavailable; trying hostname detection."
+        # Fall back to hostname-based detection
+        NODENAME=$(hostname -s)
+        if [[ $NODENAME == GPU* ]] || [[ $NODENAME == smci355-ccs-aus* ]]; then
+            export MORI_RDMA_TC=96
+            export MORI_IO_TC=96
+            echo "[INFO] Auto-detected MORI_RDMA_TC=$MORI_RDMA_TC from hostname $NODENAME"
+        elif [[ $NODENAME == mia1* ]]; then
+            export MORI_RDMA_TC=104
+            export MORI_IO_TC=104
+            echo "[INFO] Auto-detected MORI_RDMA_TC=$MORI_RDMA_TC from hostname $NODENAME"
+        else
+            echo "[INFO] Unable to detect MORI_RDMA_TC from hostname. Skipping RDMA QoS configuration."
+        fi
+    fi
+else
+    # nicctl not available, try hostname-based detection
+    NODENAME=$(hostname -s)
+    if [[ $NODENAME == GPU* ]] || [[ $NODENAME == smci355-ccs-aus* ]]; then
+        export MORI_RDMA_TC=96
+        export MORI_IO_TC=96
+        echo "[INFO] Auto-detected MORI_RDMA_TC=$MORI_RDMA_TC from hostname $NODENAME"
+    elif [[ $NODENAME == mia1* ]]; then
+        export MORI_RDMA_TC=104
+        export MORI_IO_TC=104
+        echo "[INFO] Auto-detected MORI_RDMA_TC=$MORI_RDMA_TC from hostname $NODENAME"
+    else
+        echo "[INFO] nicctl not found and unable to detect from hostname. Skipping RDMA QoS configuration."
+        echo "       This is normal for clusters without QoS or outside Docker containers."
+    fi
+fi
+
 # =============================================================================
 # Engine-specific environment
 # =============================================================================
@@ -120,7 +182,7 @@ $1 == "DSCP" && $2 == ":" && $NF == p {
 
 else
     # =========================================================================
-    # SGLang/MoRI-specific environment
+    # SGLang-specific environment
     # =========================================================================
 
     export SGLANG_USE_AITER=1
@@ -131,13 +193,6 @@ else
     export MORI_COMBINE_DTYPE_DECODE=fp8
     export SGLANG_MORI_QP_PER_TRANSFER=4
     export SGLANG_MORI_NUM_WORKERS=4
-    export MORI_IO_SQ_BACKOFF_TIMEOUT_US=50000
-
-    export MORI_IO_QP_MAX_SEND_WR=16384
-    export MORI_IO_QP_MAX_CQE=32768
-    export MORI_IO_QP_MAX_SGE=4
-
-    export MORI_IO_TC_DISABLE=0
 
     export SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=3600
     export SGLANG_DISAGGREGATION_WAITING_TIMEOUT=3600
@@ -179,57 +234,6 @@ else
     # 1 mirrors router logs to stdout via tee (useful for live debugging).
     export SGLANG_ROUTER_STDOUT_LOGS="${SGLANG_ROUTER_STDOUT_LOGS:-0}"
 
-    # QoS/DSCP configuration
-    # Priority order: 1) Set by runner, 2) Detect via nicctl, 3) Detect from hostname
-    if [[ -n "$MORI_RDMA_TC" ]]; then
-        echo "[INFO] Using MORI_RDMA_TC=$MORI_RDMA_TC (set by runner or environment)"
-    elif command -v nicctl &> /dev/null; then
-        ND_PRIO=$(nicctl show qos  2>/dev/null | awk '/PFC no-drop priorities/ {print $NF; exit}')
-        ND_DSCP=$(nicctl show qos 2>/dev/null| awk -v p="$ND_PRIO" '
-$1 == "DSCP" && $2 == ":" && $NF == p {
-    print $3; exit
-}')
-
-        if [[ -n "$ND_DSCP" ]] && [[ -n "$ND_PRIO" ]]; then
-            TC=$(( 4 * ND_DSCP ))
-            export MORI_RDMA_SL=$ND_PRIO
-            export MORI_IO_SL=$ND_PRIO
-            export MORI_RDMA_TC=$TC
-            export MORI_IO_TC=$TC
-            echo "[INFO] Detected QoS config from nicctl: MORI_RDMA_TC=$MORI_RDMA_TC, MORI_RDMA_SL=$MORI_RDMA_SL, MORI_IO_TC=$MORI_IO_TC, MORI_IO_SL=$MORI_IO_SL"
-        else
-            echo "[WARN] nicctl available but QoS data unavailable; trying hostname detection."
-            # Fall back to hostname-based detection
-            NODENAME=$(hostname -s)
-            if [[ $NODENAME == GPU* ]] || [[ $NODENAME == smci355-ccs-aus* ]]; then
-                export MORI_RDMA_TC=96
-                export MORI_IO_TC=96
-                echo "[INFO] Auto-detected MORI_RDMA_TC=$MORI_RDMA_TC from hostname $NODENAME"
-            elif [[ $NODENAME == mia1* ]]; then
-                export MORI_RDMA_TC=104
-                export MORI_IO_TC=104
-                echo "[INFO] Auto-detected MORI_RDMA_TC=$MORI_RDMA_TC from hostname $NODENAME"
-            else
-                echo "[INFO] Unable to detect MORI_RDMA_TC from hostname. Skipping RDMA QoS configuration."
-            fi
-        fi
-    else
-        # nicctl not available, try hostname-based detection
-        NODENAME=$(hostname -s)
-        if [[ $NODENAME == GPU* ]] || [[ $NODENAME == smci355-ccs-aus* ]]; then
-            export MORI_RDMA_TC=96
-            export MORI_IO_TC=96
-            echo "[INFO] Auto-detected MORI_RDMA_TC=$MORI_RDMA_TC from hostname $NODENAME"
-        elif [[ $NODENAME == mia1* ]]; then
-            export MORI_RDMA_TC=104
-            export MORI_IO_TC=104
-            echo "[INFO] Auto-detected MORI_RDMA_TC=$MORI_RDMA_TC from hostname $NODENAME"
-        else
-            echo "[INFO] nicctl not found and unable to detect from hostname. Skipping RDMA QoS configuration."
-            echo "       This is normal for clusters without QoS or outside Docker containers."
-        fi
-    fi
-
     # FIXME: WA for latest upstream 0305 image
     export PYTHONPATH=/sgl-workspace/aiter:${PYTHONPATH}
 
diff --git a/benchmarks/multi_node/amd_utils/models_vllm.yaml b/benchmarks/multi_node/amd_utils/models_vllm.yaml
index a566fe449b..ddfa5356eb 100644
--- a/benchmarks/multi_node/amd_utils/models_vllm.yaml
+++ b/benchmarks/multi_node/amd_utils/models_vllm.yaml
@@ -38,6 +38,12 @@ MiniMax-M2.5:
   env: "VLLM_USE_V1=1 VLLM_ROCM_USE_AITER=1 VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 VLLM_ENGINE_READY_TIMEOUT_S=3600 VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1"
   hf_dir: "models--MiniMaxAI--MiniMax-M2.5"
 
+MiniMax-M3-MXFP4:
+  prefill_flags: "--tensor-parallel-size 8 --max-num-batched-tokens 16384 --max-num-seqs 512 --block-size 128 --language-model-only --attention-backend TRITON_ATTN --moe-backend aiter --no-enable-prefix-caching --gpu-memory-utilization 0.90 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice"
+  decode_flags: "--tensor-parallel-size 8 --max-num-batched-tokens 16384 --max-num-seqs 512 --block-size 128 --language-model-only --attention-backend TRITON_ATTN --moe-backend aiter --no-enable-prefix-caching --gpu-memory-utilization 0.90 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice"
+  env: "VLLM_USE_V1=1 VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_MOE=1 VLLM_USE_BREAKABLE_CUDAGRAPH=0 VLLM_ENGINE_READY_TIMEOUT_S=3600"
+  hf_dir: "models--amd--MiniMax-M3-MXFP4"
+
 gpt-oss-120b:
   prefill_flags: "--tensor-parallel-size 8"
   decode_flags: "--tensor-parallel-size 8"
diff --git a/benchmarks/multi_node/amd_utils/server_vllm.sh b/benchmarks/multi_node/amd_utils/server_vllm.sh
index f02b1cd563..c3a6dbf14a 100755
--- a/benchmarks/multi_node/amd_utils/server_vllm.sh
+++ b/benchmarks/multi_node/amd_utils/server_vllm.sh
@@ -302,7 +302,7 @@ if [ "$NODE_RANK" -eq 0 ]; then
     cd $WS_PATH
 
     export ROUTER_PORT=$ROUTER_PORT
-    BENCH_CMD="bash $WS_PATH/bench.sh ${xP} ${yD} $((GPUS_PER_NODE*xP)) $((GPUS_PER_NODE*yD)) \
+    BENCH_CMD="bash $WS_PATH/bench.sh ${xP} ${yD} $((PREFILL_TP_SIZE*xP)) $((DECODE_TP_SIZE*yD)) \
         $MODEL_DIR $MODEL_NAME /run_logs/slurm_job-${SLURM_JOB_ID} ${BENCH_INPUT_LEN} \
         ${BENCH_OUTPUT_LEN} \"${BENCH_MAX_CONCURRENCY}\" ${BENCH_REQUEST_RATE} \
         ${BENCH_RANDOM_RANGE_RATIO} ${BENCH_NUM_PROMPTS_MULTIPLIER}"
diff --git a/benchmarks/multi_node/minimaxm3_fp4_mi355x_vllm-disagg.sh b/benchmarks/multi_node/minimaxm3_fp4_mi355x_vllm-disagg.sh
new file mode 100755
index 0000000000..2658b8615c
--- /dev/null
+++ b/benchmarks/multi_node/minimaxm3_fp4_mi355x_vllm-disagg.sh
@@ -0,0 +1,78 @@
+#!/usr/bin/env bash
+
+source "$(dirname "$0")/../benchmark_lib.sh"
+
+check_env_vars \
+    CONC_LIST \
+    ISL \
+    OSL \
+    IMAGE \
+    SPEC_DECODING \
+    MODEL_PATH \
+    PREFILL_NUM_WORKERS \
+    PREFILL_TP \
+    PREFILL_EP \
+    PREFILL_DP_ATTN \
+    DECODE_NUM_WORKERS \
+    DECODE_TP \
+    DECODE_EP \
+    DECODE_DP_ATTN \
+    PREFILL_NODES \
+    DECODE_NODES \
+    RANDOM_RANGE_RATIO \
+    FRAMEWORK
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+  echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+set -x
+
+cd "$GITHUB_WORKSPACE/benchmarks/multi_node/amd_utils" || exit 1
+
+export TIME_LIMIT="08:00:00"
+export MODEL_PATH=$MODEL_PATH
+export MODEL_NAME=$MODEL_NAME
+export CONTAINER_IMAGE=$IMAGE
+
+if [[ "${PREFILL_EP:-1}" -eq 1 ]]; then
+    export PREFILL_ENABLE_EP=false
+else
+    export PREFILL_ENABLE_EP=true
+fi
+
+if [[ "$PREFILL_DP_ATTN" == "true" ]]; then
+    export PREFILL_ENABLE_DP=true
+else
+    export PREFILL_ENABLE_DP=false
+fi
+
+if [[ "${DECODE_EP:-1}" -eq 1 ]]; then
+    export DECODE_ENABLE_EP=false
+else
+    export DECODE_ENABLE_EP=true
+fi
+
+if [[ "$DECODE_DP_ATTN" == "true" ]]; then
+    export DECODE_ENABLE_DP=true
+else
+    export DECODE_ENABLE_DP=false
+fi
+
+JOB_ID=$(bash ./submit.sh $PREFILL_NODES \
+    $PREFILL_NUM_WORKERS \
+    $DECODE_NODES \
+    $DECODE_NUM_WORKERS \
+    $ISL $OSL "${CONC_LIST// /x}" inf \
+    ${PREFILL_ENABLE_EP} ${PREFILL_ENABLE_DP} \
+    ${DECODE_ENABLE_EP} ${DECODE_ENABLE_DP} \
+    ${PREFILL_TP} ${DECODE_TP} \
+    ${RANDOM_RANGE_RATIO} \
+    "${NODE_LIST:-}")
+
+if [[ $? -ne 0 ]]; then
+    echo "Failed to submit job" >&2
+    exit 1
+fi
+
+echo "$JOB_ID"

From 7b3ed239d73c6aa835fae4f9b7e4ec2b2f0cbdfd Mon Sep 17 00:00:00 2001
From: Duyi-Wang <duyi.wang@amd.com>
Date: Wed, 24 Jun 2026 06:30:59 +0000
Subject: [PATCH 02/10] [AMD] perf-changelog: add MiniMax-M3-MXFP4 MI355X vLLM
 disagg entry

---
 perf-changelog.yaml | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 9c45a352ad..887abf5e53 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4146,7 +4146,7 @@
     - "Image: lmsysorg/sglang:nightly-dev-cu13-20260528-0abe6a85"
     - "8 topologies: low-latency 1p1d-tp8-tp8 + 1p6d-dep8-tp8; mid-curve 1p1d through 6p1d-dep8-dep16."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1676
-  
+
 - config-keys:
     - minimaxm3-fp8-b300-dynamo-vllm
   description:
@@ -4162,7 +4162,7 @@
     - "server_atom.sh: fix _MAX_CONC assignment before cudagraph size check; gate ATOM_MOE_GU_ITLV/AITER_BF16_FP8_MOE_BOUND on DeepSeek-V4-Pro only"
     - "Search space: ISL=8192 and ISL=1024, 1P1D TP4, conc 1-512"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1865
-  
+
 - config-keys:
     - minimaxm3-fp8-mi355x-vllm-disagg
   description:
@@ -4183,10 +4183,16 @@
     - "server_atom.sh: fix _MAX_CONC assignment before cudagraph size check; gate ATOM_MOE_GU_ITLV/AITER_BF16_FP8_MOE_BOUND on DeepSeek-V4-Pro only"
     - "Search space: ISL=8192 and ISL=1024, 1P1D TP4, conc 1-512"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1927
-  
+
 - config-keys:
     - dsv4-fp4-b200-dynamo-vllm
   description:
     - "Update the DeepSeek-V4-Pro B200 disaggregated Dynamo-vLLM benchmark to the vllm/vllm-openai:v0.23.0 image"
     - "Lower max-num-batched-tokens to 16384 and gpu-memory-utilization to 0.9 on the high-throughput and max-throughput recipes to avoid OOM"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1899
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-vllm-disagg
+  description:
+    - "Initial submission: MiniMax-M3 MXFP4 disagg (prefill/decode) on MI355X with vLLM over the MoRI-IO KV connector (8k/1k)."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1914

From 6592e7f3b4ebd6cead2e684b82ab853c192e1ff0 Mon Sep 17 00:00:00 2001
From: Duyi-Wang <duyi.wang@amd.com>
Date: Wed, 24 Jun 2026 07:03:51 +0000
Subject: [PATCH 03/10] [AMD] server_vllm.sh: default PREFILL/DECODE_TP_SIZE to
 a full node

Mirror server_sglang.sh / server_atom.sh so the bench.sh GPU count never
resolves to 0 if submit.sh did not export the per-worker TP size.
---
 benchmarks/multi_node/amd_utils/server_vllm.sh | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/benchmarks/multi_node/amd_utils/server_vllm.sh b/benchmarks/multi_node/amd_utils/server_vllm.sh
index c3a6dbf14a..f19ce8560b 100755
--- a/benchmarks/multi_node/amd_utils/server_vllm.sh
+++ b/benchmarks/multi_node/amd_utils/server_vllm.sh
@@ -39,6 +39,9 @@ BENCH_MAX_CONCURRENCY="${BENCH_MAX_CONCURRENCY:-512}"
 DRY_RUN="${DRY_RUN:-0}"
 GPUS_PER_NODE="${GPUS_PER_NODE:-8}"
 
+PREFILL_TP_SIZE="${PREFILL_TP_SIZE:-$GPUS_PER_NODE}"
+DECODE_TP_SIZE="${DECODE_TP_SIZE:-$GPUS_PER_NODE}"
+
 ROUTER_PORT="${ROUTER_PORT:-30000}"
 SERVER_PORT="${SERVER_PORT:-2584}"
 ENGINE_ID="${ENGINE_ID:-${MODEL_NAME}-pd-run}"

From a29be9b833babebe005f46ab5a9663487cc57c77 Mon Sep 17 00:00:00 2001
From: Duyi-Wang <duyi.wang@amd.com>
Date: Wed, 24 Jun 2026 09:28:41 +0000
Subject: [PATCH 04/10] [AMD] job.slurm: inject MoRI-IO QP knobs via docker -e
 for vllm-disagg

env.sh exports the MORI_IO_* QP knobs but they do not propagate to the vLLM
worker processes, so inject them into the container base env via docker -e.
MORI_IO_TC_DISABLE is intentionally omitted: the TC value is detected per-node
in env.sh and cannot reach the workers, so enabling TC steering without a TC
value would just fall back to TC=0; leave MoRI-IO at its library default instead.
---
 benchmarks/multi_node/amd_utils/job.slurm | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/benchmarks/multi_node/amd_utils/job.slurm b/benchmarks/multi_node/amd_utils/job.slurm
index 895f1ac2fb..3c851ad5ab 100755
--- a/benchmarks/multi_node/amd_utils/job.slurm
+++ b/benchmarks/multi_node/amd_utils/job.slurm
@@ -404,6 +404,10 @@ if [[ "$ENGINE" == "vllm-disagg" ]]; then
         -e HSA_ENABLE_SDMA=1
         -e PROXY_STREAM_IDLE_TIMEOUT=\${PROXY_STREAM_IDLE_TIMEOUT:-300}
         -e PYTHONPYCACHEPREFIX=/tmp/pycache
+        -e MORI_IO_SQ_BACKOFF_TIMEOUT_US=\${MORI_IO_SQ_BACKOFF_TIMEOUT_US:-50000}
+        -e MORI_IO_QP_MAX_SEND_WR=\${MORI_IO_QP_MAX_SEND_WR:-16384}
+        -e MORI_IO_QP_MAX_CQE=\${MORI_IO_QP_MAX_CQE:-32768}
+        -e MORI_IO_QP_MAX_SGE=\${MORI_IO_QP_MAX_SGE:-2}
     )
 elif [[ "$ENGINE" == "atom-disagg" ]]; then
     DOCKER_ENV_ENGINE=(

From 68c3d63945892f371e5b7f894cdbd91cee2b808a Mon Sep 17 00:00:00 2001
From: Duyi-Wang <duyi.wang@amd.com>
Date: Wed, 24 Jun 2026 12:18:41 +0000
Subject: [PATCH 05/10] [AMD] minimaxm3-fp4 disagg: switch to mori-0624-2
 image, read_mode, 1P1D only

- amd-master.yaml: bump image to rocm/sgl-dev:vllm-0.23.1-rocm723-mi35x-mori-0624-2;
  drop the 2P1D layout and cap the 1P1D conc sweep at 256.
- server_vllm.sh: set read_mode:true on the MoRIIOConnector kv-transfer-config.
- job.slurm: drop the MORI_IO_* docker -e injection (handled by the image / env.sh).
---
 .github/configs/amd-master.yaml           | 23 +++--------------------
 benchmarks/multi_node/amd_utils/job.slurm |  4 ----
 2 files changed, 3 insertions(+), 24 deletions(-)

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 4fafc96206..cdd91d83e7 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2589,7 +2589,7 @@ minimaxm3-fp8-mi355x-vllm-mtp:
 
 # MiniMax-M3 MXFP4 MI355X vLLM disaggregated (prefill/decode) config.
 minimaxm3-fp4-mi355x-vllm-disagg:
-  image: rocm/sgl-dev:vllm-0.23.1-rocm723-mi35x-mori-0623-m3
+  image: rocm/sgl-dev:vllm-0.23.1-rocm723-mi35x-mori-0624-2
   model: amd/MiniMax-M3-MXFP4
   model-prefix: minimaxm3
   runner: mi355x-disagg
@@ -2602,9 +2602,9 @@ minimaxm3-fp4-mi355x-vllm-disagg:
     - isl: 8192
       osl: 1024
       search-space:
-      # 1P TP4 + 1D TP4 (2 nodes total), conc sweep 1..512 (single job, looped)
+      # 1P TP4 + 1D TP4 (2 nodes total), conc sweep 1..256 (single job, looped)
       - spec-decoding: "none"
-        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128, 256, 512 ]
+        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128, 256 ]
         prefill:
           num-worker: 1
           tp: 4
@@ -2619,23 +2619,6 @@ minimaxm3-fp4-mi355x-vllm-disagg:
           dp-attn: false
           additional-settings:
           - "DECODE_NODES=1"
-      # 2P TP4 + 1D TP4 (3 nodes total), conc sweep 1..512 (single job, looped)
-      - spec-decoding: "none"
-        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128, 256, 512 ]
-        prefill:
-          num-worker: 2
-          tp: 4
-          ep: 1
-          dp-attn: false
-          additional-settings:
-          - "PREFILL_NODES=2"
-        decode:
-          num-worker: 1
-          tp: 4
-          ep: 1
-          dp-attn: false
-          additional-settings:
-          - "DECODE_NODES=1"
 
 # MiniMax-M3 MXFP4 MI355X atom recipe:
 # https://github.com/ROCm/ATOM/blob/5d42d49f9e4292e5b61475917e92e7ec1b1dacb7/recipes/MiniMax-M3.md
diff --git a/benchmarks/multi_node/amd_utils/job.slurm b/benchmarks/multi_node/amd_utils/job.slurm
index 3c851ad5ab..895f1ac2fb 100755
--- a/benchmarks/multi_node/amd_utils/job.slurm
+++ b/benchmarks/multi_node/amd_utils/job.slurm
@@ -404,10 +404,6 @@ if [[ "$ENGINE" == "vllm-disagg" ]]; then
         -e HSA_ENABLE_SDMA=1
         -e PROXY_STREAM_IDLE_TIMEOUT=\${PROXY_STREAM_IDLE_TIMEOUT:-300}
         -e PYTHONPYCACHEPREFIX=/tmp/pycache
-        -e MORI_IO_SQ_BACKOFF_TIMEOUT_US=\${MORI_IO_SQ_BACKOFF_TIMEOUT_US:-50000}
-        -e MORI_IO_QP_MAX_SEND_WR=\${MORI_IO_QP_MAX_SEND_WR:-16384}
-        -e MORI_IO_QP_MAX_CQE=\${MORI_IO_QP_MAX_CQE:-32768}
-        -e MORI_IO_QP_MAX_SGE=\${MORI_IO_QP_MAX_SGE:-2}
     )
 elif [[ "$ENGINE" == "atom-disagg" ]]; then
     DOCKER_ENV_ENGINE=(

From 3d5a50797524b59bfd5f7e999e5275dbf6339762 Mon Sep 17 00:00:00 2001
From: Duyi-Wang <duyi.wang@amd.com>
Date: Wed, 24 Jun 2026 14:00:43 +0000
Subject: [PATCH 06/10] [AMD] minimaxm3-fp4: prefill-only INT4 quick-reduce,
 max-num-batched-tokens 32768

- models_vllm.yaml: add prefill_env (VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4,
  VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB=2048); bump prefill+decode
  --max-num-batched-tokens 16384 -> 32768.
- server_vllm.sh: add a prefill_env / PREFILL_MODEL_ENVS channel (mirrors the
  existing decode_env path) so prefill-only env reaches every prefill rank.
---
 benchmarks/multi_node/amd_utils/models_vllm.yaml |  5 +++--
 benchmarks/multi_node/amd_utils/server_vllm.sh   | 12 ++++++++++++
 2 files changed, 15 insertions(+), 2 deletions(-)

diff --git a/benchmarks/multi_node/amd_utils/models_vllm.yaml b/benchmarks/multi_node/amd_utils/models_vllm.yaml
index ddfa5356eb..646cdba2ac 100644
--- a/benchmarks/multi_node/amd_utils/models_vllm.yaml
+++ b/benchmarks/multi_node/amd_utils/models_vllm.yaml
@@ -39,9 +39,10 @@ MiniMax-M2.5:
   hf_dir: "models--MiniMaxAI--MiniMax-M2.5"
 
 MiniMax-M3-MXFP4:
-  prefill_flags: "--tensor-parallel-size 8 --max-num-batched-tokens 16384 --max-num-seqs 512 --block-size 128 --language-model-only --attention-backend TRITON_ATTN --moe-backend aiter --no-enable-prefix-caching --gpu-memory-utilization 0.90 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice"
-  decode_flags: "--tensor-parallel-size 8 --max-num-batched-tokens 16384 --max-num-seqs 512 --block-size 128 --language-model-only --attention-backend TRITON_ATTN --moe-backend aiter --no-enable-prefix-caching --gpu-memory-utilization 0.90 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice"
+  prefill_flags: "--tensor-parallel-size 8 --max-num-batched-tokens 32768 --max-num-seqs 512 --block-size 128 --language-model-only --attention-backend TRITON_ATTN --moe-backend aiter --no-enable-prefix-caching --gpu-memory-utilization 0.90 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice"
+  decode_flags: "--tensor-parallel-size 8 --max-num-batched-tokens 32768 --max-num-seqs 512 --block-size 128 --language-model-only --attention-backend TRITON_ATTN --moe-backend aiter --no-enable-prefix-caching --gpu-memory-utilization 0.90 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice"
   env: "VLLM_USE_V1=1 VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_MOE=1 VLLM_USE_BREAKABLE_CUDAGRAPH=0 VLLM_ENGINE_READY_TIMEOUT_S=3600"
+  prefill_env: "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB=2048"
   hf_dir: "models--amd--MiniMax-M3-MXFP4"
 
 gpt-oss-120b:
diff --git a/benchmarks/multi_node/amd_utils/server_vllm.sh b/benchmarks/multi_node/amd_utils/server_vllm.sh
index f19ce8560b..55154cd015 100755
--- a/benchmarks/multi_node/amd_utils/server_vllm.sh
+++ b/benchmarks/multi_node/amd_utils/server_vllm.sh
@@ -145,10 +145,12 @@ pf = bash_escape(m.get('prefill_flags', '--tensor-parallel-size 8'))
 df = bash_escape(m.get('decode_flags', '--tensor-parallel-size 8'))
 ev = bash_escape(m.get('env', ''))
 dev = bash_escape(m.get('decode_env', ''))
+pev = bash_escape(m.get('prefill_env', ''))
 print(f'PREFILL_SERVER_CONFIG=\"{pf}\"')
 print(f'DECODE_SERVER_CONFIG=\"{df}\"')
 print(f'MODEL_ENVS=\"{ev}\"')
 print(f'DECODE_MODEL_ENVS=\"{dev}\"')
+print(f'PREFILL_MODEL_ENVS=\"{pev}\"')
 ")"
 
 echo "Loaded model configuration for: $MODEL_NAME"
@@ -251,6 +253,11 @@ if [ "$NODE_RANK" -eq 0 ]; then
 
     setup_vllm_env
 
+    for env_pair in ${PREFILL_MODEL_ENVS}; do
+        export "$env_pair"
+        echo "[PREFILL_ENV] $env_pair"
+    done
+
     # Router is started as an external container by job.slurm (VLLM_ROUTER_IMAGE)
     echo "Using external vllm-router container (started by job.slurm on this node)"
 
@@ -420,6 +427,11 @@ elif [ "$NODE_RANK" -gt 0 ] && [ "$NODE_RANK" -lt "$xP" ]; then
 
     setup_vllm_env
 
+    for env_pair in ${PREFILL_MODEL_ENVS}; do
+        export "$env_pair"
+        echo "[PREFILL_ENV] $env_pair"
+    done
+
     SERVED_MODEL="${MODEL_NAME}"
     PREFILL_CMD="vllm serve ${MODEL_PATH} \
         --served-model-name ${SERVED_MODEL} \

From 2eb5331e1593953b1ed892813fab4faa0444d7b0 Mon Sep 17 00:00:00 2001
From: Duyi-Wang <duyi.wang@amd.com>
Date: Wed, 24 Jun 2026 14:05:37 +0000
Subject: [PATCH 07/10] [AMD] minimaxm3-fp4 disagg: cap 1P1D conc sweep at 128

---
 .github/configs/amd-master.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index cdd91d83e7..698f2d02d3 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2602,9 +2602,9 @@ minimaxm3-fp4-mi355x-vllm-disagg:
     - isl: 8192
       osl: 1024
       search-space:
-      # 1P TP4 + 1D TP4 (2 nodes total), conc sweep 1..256 (single job, looped)
+      # 1P TP4 + 1D TP4 (2 nodes total), conc sweep 1..128 (single job, looped)
       - spec-decoding: "none"
-        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128, 256 ]
+        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128 ]
         prefill:
           num-worker: 1
           tp: 4

From 5dc09e4ed63d9e9263dee012a4bd5ac96630126b Mon Sep 17 00:00:00 2001
From: Duyi-Wang <duyi.wang@amd.com>
Date: Thu, 25 Jun 2026 05:45:29 +0000
Subject: [PATCH 08/10] [AMD] minimaxm3-fp4: drop prefill_env (INT4
 quick-reduce) and its plumbing

Remove the prefill-only VLLM_ROCM_QUICK_REDUCE_* env from the MiniMax-M3-MXFP4
model entry, and the now-unused prefill_env / PREFILL_MODEL_ENVS channel in
server_vllm.sh.
---
 benchmarks/multi_node/amd_utils/models_vllm.yaml |  1 -
 benchmarks/multi_node/amd_utils/server_vllm.sh   | 12 ------------
 2 files changed, 13 deletions(-)

diff --git a/benchmarks/multi_node/amd_utils/models_vllm.yaml b/benchmarks/multi_node/amd_utils/models_vllm.yaml
index 646cdba2ac..13456c2dcb 100644
--- a/benchmarks/multi_node/amd_utils/models_vllm.yaml
+++ b/benchmarks/multi_node/amd_utils/models_vllm.yaml
@@ -42,7 +42,6 @@ MiniMax-M3-MXFP4:
   prefill_flags: "--tensor-parallel-size 8 --max-num-batched-tokens 32768 --max-num-seqs 512 --block-size 128 --language-model-only --attention-backend TRITON_ATTN --moe-backend aiter --no-enable-prefix-caching --gpu-memory-utilization 0.90 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice"
   decode_flags: "--tensor-parallel-size 8 --max-num-batched-tokens 32768 --max-num-seqs 512 --block-size 128 --language-model-only --attention-backend TRITON_ATTN --moe-backend aiter --no-enable-prefix-caching --gpu-memory-utilization 0.90 --tool-call-parser minimax_m3 --reasoning-parser minimax_m3 --enable-auto-tool-choice"
   env: "VLLM_USE_V1=1 VLLM_ROCM_USE_AITER=1 VLLM_ROCM_USE_AITER_MOE=1 VLLM_USE_BREAKABLE_CUDAGRAPH=0 VLLM_ENGINE_READY_TIMEOUT_S=3600"
-  prefill_env: "VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB=2048"
   hf_dir: "models--amd--MiniMax-M3-MXFP4"
 
 gpt-oss-120b:
diff --git a/benchmarks/multi_node/amd_utils/server_vllm.sh b/benchmarks/multi_node/amd_utils/server_vllm.sh
index 55154cd015..f19ce8560b 100755
--- a/benchmarks/multi_node/amd_utils/server_vllm.sh
+++ b/benchmarks/multi_node/amd_utils/server_vllm.sh
@@ -145,12 +145,10 @@ pf = bash_escape(m.get('prefill_flags', '--tensor-parallel-size 8'))
 df = bash_escape(m.get('decode_flags', '--tensor-parallel-size 8'))
 ev = bash_escape(m.get('env', ''))
 dev = bash_escape(m.get('decode_env', ''))
-pev = bash_escape(m.get('prefill_env', ''))
 print(f'PREFILL_SERVER_CONFIG=\"{pf}\"')
 print(f'DECODE_SERVER_CONFIG=\"{df}\"')
 print(f'MODEL_ENVS=\"{ev}\"')
 print(f'DECODE_MODEL_ENVS=\"{dev}\"')
-print(f'PREFILL_MODEL_ENVS=\"{pev}\"')
 ")"
 
 echo "Loaded model configuration for: $MODEL_NAME"
@@ -253,11 +251,6 @@ if [ "$NODE_RANK" -eq 0 ]; then
 
     setup_vllm_env
 
-    for env_pair in ${PREFILL_MODEL_ENVS}; do
-        export "$env_pair"
-        echo "[PREFILL_ENV] $env_pair"
-    done
-
     # Router is started as an external container by job.slurm (VLLM_ROUTER_IMAGE)
     echo "Using external vllm-router container (started by job.slurm on this node)"
 
@@ -427,11 +420,6 @@ elif [ "$NODE_RANK" -gt 0 ] && [ "$NODE_RANK" -lt "$xP" ]; then
 
     setup_vllm_env
 
-    for env_pair in ${PREFILL_MODEL_ENVS}; do
-        export "$env_pair"
-        echo "[PREFILL_ENV] $env_pair"
-    done
-
     SERVED_MODEL="${MODEL_NAME}"
     PREFILL_CMD="vllm serve ${MODEL_PATH} \
         --served-model-name ${SERVED_MODEL} \

From f1e314d52050e0f3c9154473b8db53f5f78d09ed Mon Sep 17 00:00:00 2001
From: Duyi-Wang <duyi.wang@amd.com>
Date: Thu, 25 Jun 2026 05:46:15 +0000
Subject: [PATCH 09/10] [AMD] minimaxm3-fp4 disagg: mori-0625 image, 1P1D conc
 1..512, re-add 2P1D

- image -> rocm/vllm-dev:vllm-0.23.1-rocm723-mi35x-mori-0625
- 1P1D TP4 conc sweep back to 1..512
- re-add 2P1D TP4 layout at conc 128/256/512
---
 .github/configs/amd-master.yaml | 23 ++++++++++++++++++++---
 1 file changed, 20 insertions(+), 3 deletions(-)

diff --git a/.github/configs/amd-master.yaml b/.github/configs/amd-master.yaml
index 698f2d02d3..6e06029ff1 100644
--- a/.github/configs/amd-master.yaml
+++ b/.github/configs/amd-master.yaml
@@ -2589,7 +2589,7 @@ minimaxm3-fp8-mi355x-vllm-mtp:
 
 # MiniMax-M3 MXFP4 MI355X vLLM disaggregated (prefill/decode) config.
 minimaxm3-fp4-mi355x-vllm-disagg:
-  image: rocm/sgl-dev:vllm-0.23.1-rocm723-mi35x-mori-0624-2
+  image: rocm/vllm-dev:vllm-0.23.1-rocm723-mi35x-mori-0625
   model: amd/MiniMax-M3-MXFP4
   model-prefix: minimaxm3
   runner: mi355x-disagg
@@ -2602,9 +2602,9 @@ minimaxm3-fp4-mi355x-vllm-disagg:
     - isl: 8192
       osl: 1024
       search-space:
-      # 1P TP4 + 1D TP4 (2 nodes total), conc sweep 1..128 (single job, looped)
+      # 1P TP4 + 1D TP4 (2 nodes total), conc sweep 1..512 (single job, looped)
       - spec-decoding: "none"
-        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128 ]
+        conc-list: [ 1, 2, 4, 8, 16, 32, 64, 128, 256, 512 ]
         prefill:
           num-worker: 1
           tp: 4
@@ -2619,6 +2619,23 @@ minimaxm3-fp4-mi355x-vllm-disagg:
           dp-attn: false
           additional-settings:
           - "DECODE_NODES=1"
+      # 2P TP4 + 1D TP4 (3 nodes total), conc 128/256/512 (single job, looped)
+      - spec-decoding: "none"
+        conc-list: [ 128, 256, 512 ]
+        prefill:
+          num-worker: 2
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "PREFILL_NODES=2"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "DECODE_NODES=1"
 
 # MiniMax-M3 MXFP4 MI355X atom recipe:
 # https://github.com/ROCm/ATOM/blob/5d42d49f9e4292e5b61475917e92e7ec1b1dacb7/recipes/MiniMax-M3.md

From eff31ec4e3deb96afbab282075cdc8302555f76a Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Thu, 25 Jun 2026 15:13:31 -0400
Subject: [PATCH 10/10] fix: keep perf-changelog.yaml append-only for reuse
 merge

PR 1914 had stripped trailing whitespace on 3 existing entries (1676,
1865, 1927), breaking the byte-for-byte append-only rule enforced by
validate_perf_changelog. Restore main's historical bytes verbatim and
append only the new minimaxm3-fp4-mi355x-vllm-disagg entry.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---
 perf-changelog.yaml | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 887abf5e53..f504fc4d92 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4146,7 +4146,7 @@
     - "Image: lmsysorg/sglang:nightly-dev-cu13-20260528-0abe6a85"
     - "8 topologies: low-latency 1p1d-tp8-tp8 + 1p6d-dep8-tp8; mid-curve 1p1d through 6p1d-dep8-dep16."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1676
-
+  
 - config-keys:
     - minimaxm3-fp8-b300-dynamo-vllm
   description:
@@ -4162,7 +4162,7 @@
     - "server_atom.sh: fix _MAX_CONC assignment before cudagraph size check; gate ATOM_MOE_GU_ITLV/AITER_BF16_FP8_MOE_BOUND on DeepSeek-V4-Pro only"
     - "Search space: ISL=8192 and ISL=1024, 1P1D TP4, conc 1-512"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1865
-
+  
 - config-keys:
     - minimaxm3-fp8-mi355x-vllm-disagg
   description:
@@ -4183,7 +4183,7 @@
     - "server_atom.sh: fix _MAX_CONC assignment before cudagraph size check; gate ATOM_MOE_GU_ITLV/AITER_BF16_FP8_MOE_BOUND on DeepSeek-V4-Pro only"
     - "Search space: ISL=8192 and ISL=1024, 1P1D TP4, conc 1-512"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1927
-
+  
 - config-keys:
     - dsv4-fp4-b200-dynamo-vllm
   description: