From 5733d3267717050d697e55b16547869bd99f6dc3 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Thu, 30 Jul 2026 01:09:18 -0400
Subject: [PATCH 1/7] feat(minimaxm3-h200-agentic-mtp): add EAGLE3 AgentX
 recipe with golden-AL synthetic acceptance
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add minimaxm3-fp8-h200-vllm-agentic-mtp, the spec-decoding=mtp variant of the
MiniMax-M3 MXFP8 H200 agentic-coding recipe: Inferact/MiniMax-M3-EAGLE3 draft
head, 3 speculative tokens, FLASH_ATTN drafter, and synthetic acceptance pinned
to the committed golden AL 2.83 (minimaxm3_eagle3.yaml, thinking_on, K=3).
EVAL_ONLY keeps real target verification. Serve shape, Mooncake offload and
search-space arms mirror the non-MTP sibling, trimmed at the top concurrencies.

中文：新增 minimaxm3-fp8-h200-vllm-agentic-mtp，即 MiniMax-M3 MXFP8 H200 智能体
编码配方的投机解码（spec-decoding=mtp）变体：Inferact/MiniMax-M3-EAGLE3 草稿头、
3 个投机 token、drafter 固定为 FLASH_ATTN，合成接受长度固定为黄金 AL 2.83
（minimaxm3_eagle3.yaml，thinking_on，K=3）。EVAL_ONLY 保留真实目标验证。服务参数、
Mooncake KV 卸载与搜索空间沿用非 MTP 版本，仅在高并发端裁剪。

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 .../agentic/minimaxm3_fp8_h200_mtp.sh         | 243 ++++++++++++++++++
 1 file changed, 243 insertions(+)
 create mode 100755 benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
new file mode 100755
index 0000000000..ad3b6a5115
--- /dev/null
+++ b/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
@@ -0,0 +1,243 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# MiniMax-M3 MXFP8 H200 AgentX (agentic-coding) recipe with EAGLE3 speculative
+# decoding — the spec-decoding=mtp variant of agentic/minimaxm3_fp8_h200.sh.
+# Everything outside the speculative block mirrors the non-MTP agentic sibling
+# (Mooncake host-DRAM KV offload, --block-size 128, --language-model-only,
+# --kv-cache-dtype fp8, TRITON_ATTN, gmu 0.92, minimax_m3 parsers, vllm-router
+# for DP-attention), so the spec-decode delta is readable at equal concurrency.
+#
+# Speculative config: Inferact/MiniMax-M3-EAGLE3 draft head, 3 speculative
+# tokens — the same draft/level as every merged MiniMax-M3 MTP recipe. The
+# drafter is pinned to FLASH_ATTN as on the other CUDA M3 MTP recipes: the
+# EAGLE3 head is MHA and FlashInfer only serves page size 128 through its
+# trtllm-gen kernel, which requires GQA/MQA.
+#
+# Throughput runs pin synthetic acceptance to the committed golden AL; the
+# EVAL_ONLY accuracy run keeps real target verification. See SYNTHETIC_ACCEPT_LEN.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
+
+DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3"
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
+fi
+
+resolve_complete_model_snapshot() {
+    python3 - "$1" <<'PY'
+import json
+import sys
+from pathlib import Path
+
+model_cache_dir = Path(sys.argv[1])
+try:
+    revision = model_cache_dir.joinpath("refs/main").read_text().strip()
+except OSError:
+    raise SystemExit
+
+if not revision or Path(revision).name != revision:
+    raise SystemExit
+
+snapshot = model_cache_dir / "snapshots" / revision
+index_path = snapshot / "model.safetensors.index.json"
+required_files = (
+    snapshot / "config.json",
+    snapshot / "tokenizer_config.json",
+    index_path,
+)
+if not all(path.is_file() for path in required_files):
+    raise SystemExit
+try:
+    weight_map = json.loads(index_path.read_text())["weight_map"]
+except (KeyError, json.JSONDecodeError, OSError):
+    raise SystemExit
+shards = {snapshot / filename for filename in weight_map.values()}
+if shards and all(path.is_file() for path in shards):
+    print(snapshot)
+PY
+}
+
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    MODEL_CACHE_ROOT="${HF_HUB_CACHE:-${HF_HOME:-$HOME/.cache/huggingface/hub}}"
+    MODEL_CACHE_DIR="$MODEL_CACHE_ROOT/models--${MODEL//\//--}"
+    mkdir -p "$MODEL_CACHE_ROOT"
+    MODEL_PATH=$(resolve_complete_model_snapshot "$MODEL_CACHE_DIR")
+    if [[ -z "$MODEL_PATH" ]]; then
+        exec 9>"$MODEL_CACHE_ROOT/.minimaxm3-download.lock"
+        flock -w 3600 9
+        MODEL_PATH=$(resolve_complete_model_snapshot "$MODEL_CACHE_DIR")
+        if [[ -z "$MODEL_PATH" ]]; then
+            DOWNLOADED_MODEL_PATH=$(hf download "$MODEL")
+            MODEL_PATH=$(resolve_complete_model_snapshot "$MODEL_CACHE_DIR")
+            if [[ -z "$MODEL_PATH" ]]; then
+                echo "Downloaded model snapshot is incomplete: $DOWNLOADED_MODEL_PATH" >&2
+                exit 1
+            fi
+        fi
+        flock -u 9
+    fi
+    echo "Using complete cached model snapshot: $MODEL_PATH"
+    export MODEL_PATH
+fi
+
+# The EAGLE3 draft is never pre-staged next to the target checkpoint; fetch it
+# into the shared HF cache. That cache is a network FS where concurrent
+# day-zero downloads hit huggingface_hub's WeakFileLock "[Errno 116] Stale file
+# handle" race, so retry (the download resumes) as the fixed-seq-len MTP
+# recipes do.
+for attempt in 1 2 3 4 5; do
+    hf download "$DRAFT_MODEL" && break
+    if [ "$attempt" = 5 ]; then echo "hf download of $DRAFT_MODEL failed after $attempt attempts" >&2; exit 1; fi
+    echo "hf download attempt $attempt failed; retrying in 60s" >&2
+    sleep 60
+done
+nvidia-smi
+
+export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_062126
+resolve_trace_source
+install_agentic_deps
+
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+export PYTHONNOUSERSITE=1
+
+SERVER_LOG="$RESULT_DIR/server.log"
+ROUTER_LOG="$RESULT_DIR/router.log"
+MOONCAKE_MASTER_LOG="$RESULT_DIR/mooncake_master.log"
+mkdir -p "$RESULT_DIR"
+
+OFFLOAD_ARGS=()
+if require_agentic_kv_offload_backend mooncake; then
+        PER_RANK_GB=$((TOTAL_CPU_DRAM_GB / TP))
+        MOONCAKE_VERSION=0.3.11.post1
+        agentic_pip_install --quiet --no-cache-dir --no-deps \
+            --force-reinstall "mooncake-transfer-engine-cuda13==$MOONCAKE_VERSION"
+        python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null
+        MOONCAKE_MASTER_PORT=$((PORT + 12000))
+        MOONCAKE_CONFIG_PATH="$RESULT_DIR/mooncake_config.json"
+        cat > "$MOONCAKE_CONFIG_PATH" <<EOF
+{
+  "mode": "embedded",
+  "metadata_server": "P2PHANDSHAKE",
+  "master_server_address": "127.0.0.1:$MOONCAKE_MASTER_PORT",
+  "global_segment_size": "${PER_RANK_GB}GB",
+  "local_buffer_size": "4GB",
+  "protocol": "rdma",
+  "device_name": "",
+  "enable_offload": false
+}
+EOF
+        export MOONCAKE_CONFIG_PATH PYTHONHASHSEED=0 MC_SLICE_SIZE=1048576 MC_WORKERS_PER_CTX=4
+        export MC_ENABLE_DEST_DEVICE_AFFINITY=1
+        mooncake_master --port "$MOONCAKE_MASTER_PORT" \
+            --eviction_high_watermark_ratio=0.80 \
+            --eviction_ratio=0.10 > "$MOONCAKE_MASTER_LOG" 2>&1 &
+        MOONCAKE_MASTER_PID=$!
+        sleep 2
+        kill -0 "$MOONCAKE_MASTER_PID"
+        OFFLOAD_ARGS=(
+            --kv-transfer-config
+            '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
+        )
+fi
+
+PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
+if [[ "$DP_ATTENTION" == "true" ]]; then
+    PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
+fi
+
+EP_ARGS=()
+if (( EP_SIZE > 1 )); then
+    EP_ARGS=(--enable-expert-parallel)
+fi
+
+VLLM_BACKEND_PORT="$PORT"
+if [[ "$DP_ATTENTION" == "true" ]]; then
+    VLLM_BACKEND_PORT=$((PORT + 1))
+    export AIPERF_HTTP_X_SESSION_ID_FROM_CORRELATION_ID=1
+    agentic_pip_install --quiet 'vllm-router==0.1.14'
+fi
+
+# use 3 speculative tokens for all configs, matching the MiniMax-M3 MTP recipes
+NUM_SPEC_TOKENS=3
+TOKENS_PER_SEQ=$((1 + NUM_SPEC_TOKENS))
+
+# AgentX pins acceptance to the committed golden AL so submissions are compared
+# on system performance at a fixed acceptance target rather than on draft-head
+# quality (golden_al_distribution/README.md). 2.83 is the MiniMax-M3 EAGLE3
+# curve at num_speculative_tokens=3, thinking_on
+# (golden_al_distribution/minimaxm3_eagle3.yaml). The separate
+# minimaxm3_eagle3_gqa.yaml curve belongs to the Inferact/MiniMax-M3-EAGLE3-GQA
+# draft and is not mixed in here.
+#
+# EVAL_ONLY switches back to real verification: synthetic acceptance commits
+# drafted tokens regardless of the target logits, so generated text is wrong and
+# the eval would score ~0 (same split as dsv4_fp4_b*_vllm_mtp.sh).
+SYNTHETIC_ACCEPT_LEN=2.83
+if [ "${EVAL_ONLY:-false}" = "true" ]; then
+    SPEC_CONFIG="{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\"}"
+else
+    SPEC_CONFIG="{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\", \"rejection_sample_method\": \"synthetic\", \"synthetic_acceptance_length\": $SYNTHETIC_ACCEPT_LEN}"
+fi
+
+# AgentX concurrency counts live session trees, not individual requests, so keep
+# the non-MTP recipe's 2x scheduler headroom for subagent fan-out.
+MAX_NUM_SEQS=$((2 * CONC))
+# Cudagraph capture sizes are in TOKENS: a decode batch of S sequences verifies
+# S*(1+NUM_SPEC_TOKENS) tokens, so cap capture at MAX_NUM_SEQS*(1+N) or the
+# FULL_DECODE_ONLY ladder tops out at MAX_NUM_SEQS/(1+N) sequences and the
+# largest decode batches fall back to eager.
+MAX_CUDAGRAPH_CAPTURE_SIZE=$((MAX_NUM_SEQS * TOKENS_PER_SEQ))
+
+vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
+    --host 0.0.0.0 \
+    --port "$VLLM_BACKEND_PORT" \
+    "${PARALLEL_ARGS[@]}" \
+    "${EP_ARGS[@]}" \
+    --gpu-memory-utilization 0.92 \
+    --kv-cache-dtype fp8 \
+    --attention-backend TRITON_ATTN \
+    --block-size 128 \
+    --language-model-only \
+    --enable-prefix-caching \
+    --max-num-seqs "$MAX_NUM_SEQS" \
+    --max-cudagraph-capture-size "$MAX_CUDAGRAPH_CAPTURE_SIZE" \
+    --speculative-config "$SPEC_CONFIG" \
+    --tool-call-parser minimax_m3 \
+    --reasoning-parser minimax_m3 \
+    --enable-auto-tool-choice \
+    --trust-remote-code \
+    "${OFFLOAD_ARGS[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+
+wait_for_server_ready --port "$VLLM_BACKEND_PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [[ "$DP_ATTENTION" == "true" ]]; then
+    vllm-router \
+        --worker-urls "http://localhost:$VLLM_BACKEND_PORT" \
+        --policy consistent_hash \
+        --intra-node-data-parallel-size "$TP" \
+        --host 0.0.0.0 \
+        --port "$PORT" \
+        --prometheus-host 127.0.0.1 \
+        --prometheus-port "$((PORT + 10000))" \
+        --request-timeout-secs 14400 \
+        --disable-retries > "$ROUTER_LOG" 2>&1 &
+    ROUTER_PID=$!
+    wait_for_server_ready --port "$PORT" --server-log "$ROUTER_LOG" --server-pid "$ROUTER_PID"
+fi
+
+if [ "${EVAL_ONLY}" = "true" ]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi

From 316facdaedbffcece15b0d4ba0b188fa78085d26 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 11 Aug 2026 16:53:20 -0500
Subject: [PATCH 2/7] perf(h200): add MiniMax-M3 AgentX fast search

---
 .../agentic/minimaxm3_fp8_h200_mtp.sh         | 43 +++++++++++++------
 configs/nvidia-master.yaml                    | 21 +++++++++
 2 files changed, 50 insertions(+), 14 deletions(-)

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
index ad3b6a5115..8713b738c7 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
@@ -9,8 +9,8 @@ set -x
 # --kv-cache-dtype fp8, TRITON_ATTN, gmu 0.92, minimax_m3 parsers, vllm-router
 # for DP-attention), so the spec-decode delta is readable at equal concurrency.
 #
-# Speculative config: Inferact/MiniMax-M3-EAGLE3 draft head, 3 speculative
-# tokens — the same draft/level as every merged MiniMax-M3 MTP recipe. The
+# Speculative config: Inferact/MiniMax-M3-EAGLE3-GQA, 3 speculative tokens,
+# and the committed thinking-on golden AL. The
 # drafter is pinned to FLASH_ATTN as on the other CUDA M3 MTP recipes: the
 # EAGLE3 head is MHA and FlashInfer only serves page size 128 through its
 # trtllm-gen kernel, which requires GQA/MQA.
@@ -22,7 +22,7 @@ source "$(dirname "$0")/../../benchmark_lib.sh"
 
 check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
 
-DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3"
+DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3-GQA"
 
 if [[ -n "${SLURM_JOB_ID:-}" ]]; then
     echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
@@ -115,8 +115,15 @@ MOONCAKE_MASTER_LOG="$RESULT_DIR/mooncake_master.log"
 mkdir -p "$RESULT_DIR"
 
 OFFLOAD_ARGS=()
+MODEL_CHECKPOINT_PAGE_CACHE_GIB=414
+MOONCAKE_LOCAL_BUFFER_GIB=4
 if require_agentic_kv_offload_backend mooncake; then
-        PER_RANK_GB=$((TOTAL_CPU_DRAM_GB / TP))
+        TOTAL_CPU_DRAM_GIB=$((TOTAL_CPU_DRAM_GB * 1000000000 / 1073741824))
+        PER_RANK_GB=$(((TOTAL_CPU_DRAM_GIB - MODEL_CHECKPOINT_PAGE_CACHE_GIB) / TP - MOONCAKE_LOCAL_BUFFER_GIB))
+        if (( PER_RANK_GB <= 0 )); then
+            echo "Error: CPU DRAM budget is too small for checkpoint cache and KV offload" >&2
+            exit 1
+        fi
         MOONCAKE_VERSION=0.3.11.post1
         agentic_pip_install --quiet --no-cache-dir --no-deps \
             --force-reinstall "mooncake-transfer-engine-cuda13==$MOONCAKE_VERSION"
@@ -166,31 +173,37 @@ if [[ "$DP_ATTENTION" == "true" ]]; then
     agentic_pip_install --quiet 'vllm-router==0.1.14'
 fi
 
+export AIPERF_SERVER_METRICS_URLS="http://localhost:${VLLM_BACKEND_PORT}/metrics"
+export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="vllm:"
+
 # use 3 speculative tokens for all configs, matching the MiniMax-M3 MTP recipes
 NUM_SPEC_TOKENS=3
 TOKENS_PER_SEQ=$((1 + NUM_SPEC_TOKENS))
 
 # AgentX pins acceptance to the committed golden AL so submissions are compared
 # on system performance at a fixed acceptance target rather than on draft-head
-# quality (golden_al_distribution/README.md). 2.83 is the MiniMax-M3 EAGLE3
-# curve at num_speculative_tokens=3, thinking_on
-# (golden_al_distribution/minimaxm3_eagle3.yaml). The separate
-# minimaxm3_eagle3_gqa.yaml curve belongs to the Inferact/MiniMax-M3-EAGLE3-GQA
-# draft and is not mixed in here.
+# quality. 2.78 is minimaxm3_eagle3_gqa.yaml thinking_on[3].
 #
 # EVAL_ONLY switches back to real verification: synthetic acceptance commits
 # drafted tokens regardless of the target logits, so generated text is wrong and
 # the eval would score ~0 (same split as dsv4_fp4_b*_vllm_mtp.sh).
-SYNTHETIC_ACCEPT_LEN=2.83
+SYNTHETIC_ACCEPT_LEN=2.78
 if [ "${EVAL_ONLY:-false}" = "true" ]; then
     SPEC_CONFIG="{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\"}"
 else
     SPEC_CONFIG="{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\", \"rejection_sample_method\": \"synthetic\", \"synthetic_acceptance_length\": $SYNTHETIC_ACCEPT_LEN}"
 fi
 
-# AgentX concurrency counts live session trees, not individual requests, so keep
-# the non-MTP recipe's 2x scheduler headroom for subagent fan-out.
-MAX_NUM_SEQS=$((2 * CONC))
+# DEP distributes the live AgentX session trees across its data-parallel ranks.
+if [[ "$DP_ATTENTION" == "true" ]]; then
+    if (( 2 * CONC % TP != 0 )); then
+        echo "DEP requires 2*CONC divisible by TP (CONC=$CONC TP=$TP)" >&2
+        exit 1
+    fi
+    MAX_NUM_SEQS=$((2 * CONC / TP))
+else
+    MAX_NUM_SEQS=$((2 * CONC))
+fi
 # Cudagraph capture sizes are in TOKENS: a decode batch of S sequences verifies
 # S*(1+NUM_SPEC_TOKENS) tokens, so cap capture at MAX_NUM_SEQS*(1+N) or the
 # FULL_DECODE_ONLY ladder tops out at MAX_NUM_SEQS/(1+N) sequences and the
@@ -202,12 +215,14 @@ vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
     --port "$VLLM_BACKEND_PORT" \
     "${PARALLEL_ARGS[@]}" \
     "${EP_ARGS[@]}" \
-    --gpu-memory-utilization 0.92 \
+    --gpu-memory-utilization 0.90 \
     --kv-cache-dtype fp8 \
     --attention-backend TRITON_ATTN \
     --block-size 128 \
     --language-model-only \
     --enable-prefix-caching \
+    --enable-prompt-tokens-details \
+    --default-chat-template-kwargs '{"thinking_mode":"enabled"}' \
     --max-num-seqs "$MAX_NUM_SEQS" \
     --max-cudagraph-capture-size "$MAX_CUDAGRAPH_CAPTURE_SIZE" \
     --speculative-config "$SPEC_CONFIG" \
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index a376a9a16b..a7afb17024 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7126,6 +7126,27 @@ minimaxm3-fp8-h200-vllm-agentic:
       - { tp: 8, ep: 8, kv-offloading: none, conc-list: [2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
       - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
 
+minimaxm3-fp8-h200-vllm-agentic-mtp:
+  image: vllm/vllm-openai:v0.27.1
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: cluster:h200-dgxc
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    # Broad fast discovery spans TP8, TEP8, and session-routed DEP8. Mooncake
+    # overlaps the expected HBM knee and continues into the host-tier tail.
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 6, 8, 10, 12, 14, 16, 20] }
+      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 6, 8, 10, 12, 14, 16, 20] }
+      - { tp: 8, ep: 8, dp-attn: true, spec-decoding: mtp, kv-offloading: none, conc-list: [8, 12, 16, 24, 32, 48] }
+      - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [8, 10, 12, 14, 16, 20, 24, 32] }
+      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [8, 10, 12, 14, 16, 20, 24, 32] }
+      - { tp: 8, ep: 8, dp-attn: true, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [24, 32, 48, 64] }
+
 qwen3.5-fp4-b200-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
   model: nvidia/Qwen3.5-397B-A17B-NVFP4

From 6f7795bc99f27d82fa66e62aa837e3ee1588a9cf Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 11 Aug 2026 16:55:06 -0500
Subject: [PATCH 3/7] docs(perf): document H200 MiniMax-M3 tuning

---
 perf-changelog.yaml | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 7ae8479441..b6297aac91 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5794,3 +5794,13 @@
     - "Keep DSpark K=2 probabilistic throughput on synthetic golden AL 2.51 while the generated EVAL_ONLY row uses real block verification"
     - "Cap the GPU-resident search at concurrency 8 after concurrency 16 failed to complete deterministic warmup; retain concurrency 16 for the DRAM-offload capacity tier"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2476
+
+- config-keys:
+    - minimaxm3-fp8-h200-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add H200 MiniMax-M3 MXFP8 AgentX on vLLM v0.27.1 with EAGLE3-GQA synthetic golden AL 2.78."
+    - "Screen TP8, TEP8, DEP8, and Mooncake DRAM-offload curves in AgentX fast mode before the final full sweep."
+    - "Require vLLM server metrics and session-aware DEP routing without relaxed request thresholds."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565

From cce3900df05a385abfe6708a897532fea4892ee1 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 11 Aug 2026 21:54:22 -0500
Subject: [PATCH 4/7] perf(h200): test MiniMax-M3 vLLM-simple offload
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

在 H200 MiniMax-M3 的拐点附近测试 vLLM SimpleCPUOffloadConnector。
---
 .../single_node/agentic/minimaxm3_fp8_h200.sh | 23 +++++++++++++++++--
 configs/nvidia-master.yaml                    | 12 ++++------
 perf-changelog.yaml                           |  8 +++++++
 3 files changed, 33 insertions(+), 10 deletions(-)

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp8_h200.sh b/benchmarks/single_node/agentic/minimaxm3_fp8_h200.sh
index 6b984cee03..5b9f049248 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp8_h200.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp8_h200.sh
@@ -85,7 +85,21 @@ MOONCAKE_MASTER_LOG="$RESULT_DIR/mooncake_master.log"
 mkdir -p "$RESULT_DIR"
 
 OFFLOAD_ARGS=()
-if require_agentic_kv_offload_backend mooncake; then
+case "${KV_OFFLOAD_BACKEND:-}" in
+    "")
+        require_agentic_kv_offload_none
+        ;;
+    vllm-simple)
+        require_agentic_kv_offload_backend vllm-simple
+        CPU_BYTES_PER_RANK=$((TOTAL_CPU_DRAM_GB * 1000 * 1000 * 1000 / TP))
+        export PYTHONHASHSEED=42
+        OFFLOAD_ARGS=(
+            --kv-transfer-config
+            "{\"kv_connector\":\"SimpleCPUOffloadConnector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"cpu_bytes_to_use_per_rank\":${CPU_BYTES_PER_RANK},\"lazy_offload\":false}}"
+        )
+        ;;
+    mooncake)
+        require_agentic_kv_offload_backend mooncake
         PER_RANK_GB=$((TOTAL_CPU_DRAM_GB / TP))
         MOONCAKE_VERSION=0.3.11.post1
         agentic_pip_install --quiet --no-cache-dir --no-deps \
@@ -117,7 +131,12 @@ EOF
             --kv-transfer-config
             '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
         )
-fi
+        ;;
+    *)
+        echo "Error: unsupported KV_OFFLOAD_BACKEND='$KV_OFFLOAD_BACKEND'" >&2
+        exit 1
+        ;;
+esac
 
 PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
 if [[ "$DP_ATTENTION" == "true" ]]; then
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index a7afb17024..7081344e0f 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7136,16 +7136,12 @@ minimaxm3-fp8-h200-vllm-agentic-mtp:
   multinode: false
   scenarios:
     agentic-coding:
-    # Broad fast discovery spans TP8, TEP8, and session-routed DEP8. Mooncake
-    # overlaps the expected HBM knee and continues into the host-tier tail.
+    # Broad fast run 31540120459 located the TP8 knee at c10-c14. Compare the
+    # built-in vLLM host tier at equal concurrency before selecting the final
+    # full-duration Pareto sweep.
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 6, 8, 10, 12, 14, 16, 20] }
-      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 6, 8, 10, 12, 14, 16, 20] }
-      - { tp: 8, ep: 8, dp-attn: true, spec-decoding: mtp, kv-offloading: none, conc-list: [8, 12, 16, 24, 32, 48] }
-      - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [8, 10, 12, 14, 16, 20, 24, 32] }
-      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [8, 10, 12, 14, 16, 20, 24, 32] }
-      - { tp: 8, ep: 8, dp-attn: true, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [24, 32, 48, 64] }
+      - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [10, 12, 14, 16] }
 
 qwen3.5-fp4-b200-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index b6297aac91..2fc5902043 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5804,3 +5804,11 @@
     - "Screen TP8, TEP8, DEP8, and Mooncake DRAM-offload curves in AgentX fast mode before the final full sweep."
     - "Require vLLM server metrics and session-aware DEP routing without relaxed request thresholds."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565
+
+- config-keys:
+    - minimaxm3-fp8-h200-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Broad fast run 31540120459 identifies the TP8 knee at c10-c14; compare vLLM-simple DRAM offload at c10/c12/c14/c16 before the final full sweep."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565

From f5c81c56483ba2b8fade8ecf52050f399b7e59ee Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 11 Aug 2026 22:05:34 -0500
Subject: [PATCH 5/7] fix(h200): route MiniMax-M3 offload through MTP
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将 vLLM SimpleCPUOffloadConnector 配置接入实际的 MiniMax-M3 MTP 运行时，并保留检查点缓存的主机内存预算。
---
 .../single_node/agentic/minimaxm3_fp8_h200.sh | 23 ++-------------
 .../agentic/minimaxm3_fp8_h200_mtp.sh         | 29 +++++++++++++++++--
 perf-changelog.yaml                           |  8 +++++
 3 files changed, 37 insertions(+), 23 deletions(-)

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp8_h200.sh b/benchmarks/single_node/agentic/minimaxm3_fp8_h200.sh
index 5b9f049248..6b984cee03 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp8_h200.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp8_h200.sh
@@ -85,21 +85,7 @@ MOONCAKE_MASTER_LOG="$RESULT_DIR/mooncake_master.log"
 mkdir -p "$RESULT_DIR"
 
 OFFLOAD_ARGS=()
-case "${KV_OFFLOAD_BACKEND:-}" in
-    "")
-        require_agentic_kv_offload_none
-        ;;
-    vllm-simple)
-        require_agentic_kv_offload_backend vllm-simple
-        CPU_BYTES_PER_RANK=$((TOTAL_CPU_DRAM_GB * 1000 * 1000 * 1000 / TP))
-        export PYTHONHASHSEED=42
-        OFFLOAD_ARGS=(
-            --kv-transfer-config
-            "{\"kv_connector\":\"SimpleCPUOffloadConnector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"cpu_bytes_to_use_per_rank\":${CPU_BYTES_PER_RANK},\"lazy_offload\":false}}"
-        )
-        ;;
-    mooncake)
-        require_agentic_kv_offload_backend mooncake
+if require_agentic_kv_offload_backend mooncake; then
         PER_RANK_GB=$((TOTAL_CPU_DRAM_GB / TP))
         MOONCAKE_VERSION=0.3.11.post1
         agentic_pip_install --quiet --no-cache-dir --no-deps \
@@ -131,12 +117,7 @@ EOF
             --kv-transfer-config
             '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
         )
-        ;;
-    *)
-        echo "Error: unsupported KV_OFFLOAD_BACKEND='$KV_OFFLOAD_BACKEND'" >&2
-        exit 1
-        ;;
-esac
+fi
 
 PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
 if [[ "$DP_ATTENTION" == "true" ]]; then
diff --git a/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
index 8713b738c7..6917307c76 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
@@ -117,7 +117,27 @@ mkdir -p "$RESULT_DIR"
 OFFLOAD_ARGS=()
 MODEL_CHECKPOINT_PAGE_CACHE_GIB=414
 MOONCAKE_LOCAL_BUFFER_GIB=4
-if require_agentic_kv_offload_backend mooncake; then
+case "${KV_OFFLOAD_BACKEND:-}" in
+    "")
+        require_agentic_kv_offload_none
+        ;;
+    vllm-simple)
+        require_agentic_kv_offload_backend vllm-simple
+        TOTAL_CPU_DRAM_GIB=$((TOTAL_CPU_DRAM_GB * 1000000000 / 1073741824))
+        CPU_OFFLOAD_GIB_PER_RANK=$(((TOTAL_CPU_DRAM_GIB - MODEL_CHECKPOINT_PAGE_CACHE_GIB) / TP))
+        if (( CPU_OFFLOAD_GIB_PER_RANK <= 0 )); then
+            echo "Error: CPU DRAM budget is too small for checkpoint cache and KV offload" >&2
+            exit 1
+        fi
+        CPU_BYTES_PER_RANK=$((CPU_OFFLOAD_GIB_PER_RANK * 1024 * 1024 * 1024))
+        export PYTHONHASHSEED=42
+        OFFLOAD_ARGS=(
+            --kv-transfer-config
+            "{\"kv_connector\":\"SimpleCPUOffloadConnector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"cpu_bytes_to_use_per_rank\":${CPU_BYTES_PER_RANK},\"lazy_offload\":false}}"
+        )
+        ;;
+    mooncake)
+        require_agentic_kv_offload_backend mooncake
         TOTAL_CPU_DRAM_GIB=$((TOTAL_CPU_DRAM_GB * 1000000000 / 1073741824))
         PER_RANK_GB=$(((TOTAL_CPU_DRAM_GIB - MODEL_CHECKPOINT_PAGE_CACHE_GIB) / TP - MOONCAKE_LOCAL_BUFFER_GIB))
         if (( PER_RANK_GB <= 0 )); then
@@ -154,7 +174,12 @@ EOF
             --kv-transfer-config
             '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
         )
-fi
+        ;;
+    *)
+        echo "Error: unsupported KV_OFFLOAD_BACKEND='$KV_OFFLOAD_BACKEND'" >&2
+        exit 1
+        ;;
+esac
 
 PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
 if [[ "$DP_ATTENTION" == "true" ]]; then
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index e879a58e3d..d9bf18de36 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5836,3 +5836,11 @@
   description:
     - "Broad fast run 31540120459 identifies the TP8 knee at c10-c14; compare vLLM-simple DRAM offload at c10/c12/c14/c16 before the final full sweep."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565
+
+- config-keys:
+    - minimaxm3-fp8-h200-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Route the vLLM-simple refinement through the MiniMax-M3 MTP runtime and preserve its checkpoint-cache DRAM reservation."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565

From d621268bed643afa3e33be318481c7e37871806a Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 11 Aug 2026 22:54:00 -0500
Subject: [PATCH 6/7] perf(h200): finalize MiniMax-M3 AgentX curve
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

保留 TP8 常驻 c1/c2/c4/c6/c8/c10 和 Mooncake c12/c14，依据完整快速扫描形成严格全量扫点。
---
 configs/nvidia-master.yaml | 10 ++++++----
 perf-changelog.yaml        |  8 ++++++++
 2 files changed, 14 insertions(+), 4 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 57df6b0c91..9a826653a2 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7171,12 +7171,14 @@ minimaxm3-fp8-h200-vllm-agentic-mtp:
   multinode: false
   scenarios:
     agentic-coding:
-    # Broad fast run 31540120459 located the TP8 knee at c10-c14. Compare the
-    # built-in vLLM host tier at equal concurrency before selecting the final
-    # full-duration Pareto sweep.
+    # Fast runs 31540120459 and 31558981228 locate the resident knee at c10
+    # and show Mooncake retaining throughput at c12-c14. TEP is dominated,
+    # DEP cannot allocate the 1M-token KV cache, and vLLM-simple falls after
+    # c10, so the strict sweep keeps only the measured Pareto candidates.
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [10, 12, 14, 16] }
+      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 6, 8, 10] }
+      - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [12, 14] }
 
 qwen3.5-fp4-b200-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d9bf18de36..50d329c160 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5844,3 +5844,11 @@
   description:
     - "Route the vLLM-simple refinement through the MiniMax-M3 MTP runtime and preserve its checkpoint-cache DRAM reservation."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565
+
+- config-keys:
+    - minimaxm3-fp8-h200-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Finalize the strict TP8 MTP sweep at resident c1/c2/c4/c6/c8/c10 and Mooncake c12/c14 from completed fast-run measurements."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565

From 6cd6832ff1281711c474e7d18cce08f549ffbe27 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 12 Aug 2026 09:02:55 -0500
Subject: [PATCH 7/7] chore(h200): simplify MiniMax-M3 benchmark

---
 benchmarks/benchmark_lib.sh                   |   8 +
 .../agentic/minimaxm3_fp8_h200_mtp.sh         | 240 ++++++------------
 perf-changelog.yaml                           |  28 +-
 3 files changed, 92 insertions(+), 184 deletions(-)

diff --git a/benchmarks/benchmark_lib.sh b/benchmarks/benchmark_lib.sh
index 4e2e7b7b17..d51ebece63 100644
--- a/benchmarks/benchmark_lib.sh
+++ b/benchmarks/benchmark_lib.sh
@@ -343,6 +343,14 @@ check_env_vars() {
     fi
 }
 
+# Record a shell-array command in a reproducible, shell-escaped form.
+write_command() {
+    local output_file="$1"
+    shift
+    printf '%q ' "$@" | tee "$output_file"
+    printf '\n' | tee -a "$output_file"
+}
+
 # Wait for server to be ready by polling the health endpoint
 # All parameters are required
 # Parameters:
diff --git a/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
index 6917307c76..28e7adc4a0 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp8_h200_mtp.sh
@@ -1,33 +1,15 @@
 #!/usr/bin/env bash
-set -euo pipefail
+set -eo pipefail
 set -x
 
-# MiniMax-M3 MXFP8 H200 AgentX (agentic-coding) recipe with EAGLE3 speculative
-# decoding — the spec-decoding=mtp variant of agentic/minimaxm3_fp8_h200.sh.
-# Everything outside the speculative block mirrors the non-MTP agentic sibling
-# (Mooncake host-DRAM KV offload, --block-size 128, --language-model-only,
-# --kv-cache-dtype fp8, TRITON_ATTN, gmu 0.92, minimax_m3 parsers, vllm-router
-# for DP-attention), so the spec-decode delta is readable at equal concurrency.
-#
-# Speculative config: Inferact/MiniMax-M3-EAGLE3-GQA, 3 speculative tokens,
-# and the committed thinking-on golden AL. The
-# drafter is pinned to FLASH_ATTN as on the other CUDA M3 MTP recipes: the
-# EAGLE3 head is MHA and FlashInfer only serves page size 128 through its
-# trtllm-gen kernel, which requires GQA/MQA.
-#
-# Throughput runs pin synthetic acceptance to the committed golden AL; the
-# EVAL_ONLY accuracy run keeps real target verification. See SYNTHETIC_ACCEPT_LEN.
+# H200 MiniMax-M3 MXFP8 AgentX with EAGLE3 and optional Mooncake DRAM KV offload.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
-check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
+check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION PORT EVAL_ONLY
 
 DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3-GQA"
 
-if [[ -n "${SLURM_JOB_ID:-}" ]]; then
-    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
-fi
-
 resolve_complete_model_snapshot() {
     python3 - "$1" <<'PY'
 import json
@@ -89,11 +71,7 @@ else
     export MODEL_PATH
 fi
 
-# The EAGLE3 draft is never pre-staged next to the target checkpoint; fetch it
-# into the shared HF cache. That cache is a network FS where concurrent
-# day-zero downloads hit huggingface_hub's WeakFileLock "[Errno 116] Stale file
-# handle" race, so retry (the download resumes) as the fixed-seq-len MTP
-# recipes do.
+# Concurrent downloads on the shared HF cache can hit transient stale handles.
 for attempt in 1 2 3 4 5; do
     hf download "$DRAFT_MODEL" && break
     if [ "$attempt" = 5 ]; then echo "hf download of $DRAFT_MODEL failed after $attempt attempts" >&2; exit 1; fi
@@ -110,47 +88,43 @@ export VLLM_ENGINE_READY_TIMEOUT_S=3600
 export PYTHONNOUSERSITE=1
 
 SERVER_LOG="$RESULT_DIR/server.log"
-ROUTER_LOG="$RESULT_DIR/router.log"
 MOONCAKE_MASTER_LOG="$RESULT_DIR/mooncake_master.log"
 mkdir -p "$RESULT_DIR"
 
+SERVER_PID=""
+MOONCAKE_MASTER_PID=""
+cleanup_services() {
+    local exit_code=$?
+    trap - EXIT INT TERM
+    set +e
+    stop_background_process_tree "$SERVER_PID" "vLLM server" 60
+    stop_background_process_tree "$MOONCAKE_MASTER_PID" "Mooncake master" 30
+    exit "$exit_code"
+}
+trap cleanup_services EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
 OFFLOAD_ARGS=()
 MODEL_CHECKPOINT_PAGE_CACHE_GIB=414
 MOONCAKE_LOCAL_BUFFER_GIB=4
-case "${KV_OFFLOAD_BACKEND:-}" in
-    "")
-        require_agentic_kv_offload_none
-        ;;
-    vllm-simple)
-        require_agentic_kv_offload_backend vllm-simple
-        TOTAL_CPU_DRAM_GIB=$((TOTAL_CPU_DRAM_GB * 1000000000 / 1073741824))
-        CPU_OFFLOAD_GIB_PER_RANK=$(((TOTAL_CPU_DRAM_GIB - MODEL_CHECKPOINT_PAGE_CACHE_GIB) / TP))
-        if (( CPU_OFFLOAD_GIB_PER_RANK <= 0 )); then
-            echo "Error: CPU DRAM budget is too small for checkpoint cache and KV offload" >&2
-            exit 1
-        fi
-        CPU_BYTES_PER_RANK=$((CPU_OFFLOAD_GIB_PER_RANK * 1024 * 1024 * 1024))
-        export PYTHONHASHSEED=42
-        OFFLOAD_ARGS=(
-            --kv-transfer-config
-            "{\"kv_connector\":\"SimpleCPUOffloadConnector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"cpu_bytes_to_use_per_rank\":${CPU_BYTES_PER_RANK},\"lazy_offload\":false}}"
-        )
-        ;;
-    mooncake)
-        require_agentic_kv_offload_backend mooncake
-        TOTAL_CPU_DRAM_GIB=$((TOTAL_CPU_DRAM_GB * 1000000000 / 1073741824))
-        PER_RANK_GB=$(((TOTAL_CPU_DRAM_GIB - MODEL_CHECKPOINT_PAGE_CACHE_GIB) / TP - MOONCAKE_LOCAL_BUFFER_GIB))
-        if (( PER_RANK_GB <= 0 )); then
-            echo "Error: CPU DRAM budget is too small for checkpoint cache and KV offload" >&2
-            exit 1
-        fi
-        MOONCAKE_VERSION=0.3.11.post1
-        agentic_pip_install --quiet --no-cache-dir --no-deps \
-            --force-reinstall "mooncake-transfer-engine-cuda13==$MOONCAKE_VERSION"
-        python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null
-        MOONCAKE_MASTER_PORT=$((PORT + 12000))
-        MOONCAKE_CONFIG_PATH="$RESULT_DIR/mooncake_config.json"
-        cat > "$MOONCAKE_CONFIG_PATH" <<EOF
+if [ "$KV_OFFLOADING" = "none" ]; then
+    require_agentic_kv_offload_none
+elif [ "$KV_OFFLOADING" = "dram" ]; then
+    require_agentic_kv_offload_backend mooncake
+    TOTAL_CPU_DRAM_GIB=$((TOTAL_CPU_DRAM_GB * 1000000000 / 1073741824))
+    PER_RANK_GB=$(((TOTAL_CPU_DRAM_GIB - MODEL_CHECKPOINT_PAGE_CACHE_GIB) / TP - MOONCAKE_LOCAL_BUFFER_GIB))
+    if (( PER_RANK_GB <= 0 )); then
+        echo "Error: CPU DRAM budget is too small for checkpoint cache and KV offload" >&2
+        exit 1
+    fi
+    MOONCAKE_VERSION=0.3.11.post1
+    agentic_pip_install --quiet --no-cache-dir --no-deps \
+        --force-reinstall "mooncake-transfer-engine-cuda13==$MOONCAKE_VERSION"
+    python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null
+    MOONCAKE_MASTER_PORT=$((PORT + 12000))
+    MOONCAKE_CONFIG_PATH="$RESULT_DIR/mooncake_config.json"
+    cat > "$MOONCAKE_CONFIG_PATH" <<EOF
 {
   "mode": "embedded",
   "metadata_server": "P2PHANDSHAKE",
@@ -162,120 +136,72 @@ case "${KV_OFFLOAD_BACKEND:-}" in
   "enable_offload": false
 }
 EOF
-        export MOONCAKE_CONFIG_PATH PYTHONHASHSEED=0 MC_SLICE_SIZE=1048576 MC_WORKERS_PER_CTX=4
-        export MC_ENABLE_DEST_DEVICE_AFFINITY=1
-        mooncake_master --port "$MOONCAKE_MASTER_PORT" \
-            --eviction_high_watermark_ratio=0.80 \
-            --eviction_ratio=0.10 > "$MOONCAKE_MASTER_LOG" 2>&1 &
-        MOONCAKE_MASTER_PID=$!
-        sleep 2
-        kill -0 "$MOONCAKE_MASTER_PID"
-        OFFLOAD_ARGS=(
-            --kv-transfer-config
-            '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
-        )
-        ;;
-    *)
-        echo "Error: unsupported KV_OFFLOAD_BACKEND='$KV_OFFLOAD_BACKEND'" >&2
-        exit 1
-        ;;
-esac
-
-PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
-if [[ "$DP_ATTENTION" == "true" ]]; then
-    PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
-fi
-
-EP_ARGS=()
-if (( EP_SIZE > 1 )); then
-    EP_ARGS=(--enable-expert-parallel)
-fi
-
-VLLM_BACKEND_PORT="$PORT"
-if [[ "$DP_ATTENTION" == "true" ]]; then
-    VLLM_BACKEND_PORT=$((PORT + 1))
-    export AIPERF_HTTP_X_SESSION_ID_FROM_CORRELATION_ID=1
-    agentic_pip_install --quiet 'vllm-router==0.1.14'
+    export MOONCAKE_CONFIG_PATH PYTHONHASHSEED=0 MC_SLICE_SIZE=1048576 MC_WORKERS_PER_CTX=4
+    export MC_ENABLE_DEST_DEVICE_AFFINITY=1
+    mooncake_master --port "$MOONCAKE_MASTER_PORT" \
+        --eviction_high_watermark_ratio=0.80 \
+        --eviction_ratio=0.10 > "$MOONCAKE_MASTER_LOG" 2>&1 &
+    MOONCAKE_MASTER_PID=$!
+    sleep 2
+    kill -0 "$MOONCAKE_MASTER_PID"
+    OFFLOAD_ARGS=(
+        --kv-transfer-config
+        '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
+    )
+else
+    echo "Error: unsupported KV_OFFLOADING='$KV_OFFLOADING'" >&2
+    exit 1
 fi
 
-export AIPERF_SERVER_METRICS_URLS="http://localhost:${VLLM_BACKEND_PORT}/metrics"
+export AIPERF_SERVER_METRICS_URLS="http://localhost:${PORT}/metrics"
 export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="vllm:"
 
-# use 3 speculative tokens for all configs, matching the MiniMax-M3 MTP recipes
 NUM_SPEC_TOKENS=3
 TOKENS_PER_SEQ=$((1 + NUM_SPEC_TOKENS))
 
-# AgentX pins acceptance to the committed golden AL so submissions are compared
-# on system performance at a fixed acceptance target rather than on draft-head
-# quality. 2.78 is minimaxm3_eagle3_gqa.yaml thinking_on[3].
-#
-# EVAL_ONLY switches back to real verification: synthetic acceptance commits
-# drafted tokens regardless of the target logits, so generated text is wrong and
-# the eval would score ~0 (same split as dsv4_fp4_b*_vllm_mtp.sh).
+# Golden AL is minimaxm3_eagle3_gqa.yaml thinking_on[3]; eval uses real verification.
 SYNTHETIC_ACCEPT_LEN=2.78
-if [ "${EVAL_ONLY:-false}" = "true" ]; then
+if [ "$EVAL_ONLY" = "true" ]; then
     SPEC_CONFIG="{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\"}"
 else
     SPEC_CONFIG="{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\", \"rejection_sample_method\": \"synthetic\", \"synthetic_acceptance_length\": $SYNTHETIC_ACCEPT_LEN}"
 fi
 
-# DEP distributes the live AgentX session trees across its data-parallel ranks.
-if [[ "$DP_ATTENTION" == "true" ]]; then
-    if (( 2 * CONC % TP != 0 )); then
-        echo "DEP requires 2*CONC divisible by TP (CONC=$CONC TP=$TP)" >&2
-        exit 1
-    fi
-    MAX_NUM_SEQS=$((2 * CONC / TP))
-else
-    MAX_NUM_SEQS=$((2 * CONC))
-fi
-# Cudagraph capture sizes are in TOKENS: a decode batch of S sequences verifies
-# S*(1+NUM_SPEC_TOKENS) tokens, so cap capture at MAX_NUM_SEQS*(1+N) or the
-# FULL_DECODE_ONLY ladder tops out at MAX_NUM_SEQS/(1+N) sequences and the
-# largest decode batches fall back to eager.
+MAX_NUM_SEQS=$((2 * CONC))
+# MTP verifies four tokens per sequence, so CUDA graph capture is token-sized.
 MAX_CUDAGRAPH_CAPTURE_SIZE=$((MAX_NUM_SEQS * TOKENS_PER_SEQ))
 
-vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
-    --host 0.0.0.0 \
-    --port "$VLLM_BACKEND_PORT" \
-    "${PARALLEL_ARGS[@]}" \
-    "${EP_ARGS[@]}" \
-    --gpu-memory-utilization 0.90 \
-    --kv-cache-dtype fp8 \
-    --attention-backend TRITON_ATTN \
-    --block-size 128 \
-    --language-model-only \
-    --enable-prefix-caching \
-    --enable-prompt-tokens-details \
-    --default-chat-template-kwargs '{"thinking_mode":"enabled"}' \
-    --max-num-seqs "$MAX_NUM_SEQS" \
-    --max-cudagraph-capture-size "$MAX_CUDAGRAPH_CAPTURE_SIZE" \
-    --speculative-config "$SPEC_CONFIG" \
-    --tool-call-parser minimax_m3 \
-    --reasoning-parser minimax_m3 \
-    --enable-auto-tool-choice \
-    --trust-remote-code \
-    "${OFFLOAD_ARGS[@]}" > "$SERVER_LOG" 2>&1 &
+VLLM_CMD=(
+    vllm serve "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --tensor-parallel-size "$TP"
+    --data-parallel-size 1
+    --gpu-memory-utilization 0.90
+    --kv-cache-dtype fp8
+    --attention-backend TRITON_ATTN
+    --block-size 128
+    --language-model-only
+    --enable-prefix-caching
+    --enable-prompt-tokens-details
+    --default-chat-template-kwargs '{"thinking_mode":"enabled"}'
+    --max-num-seqs "$MAX_NUM_SEQS"
+    --max-cudagraph-capture-size "$MAX_CUDAGRAPH_CAPTURE_SIZE"
+    --speculative-config "$SPEC_CONFIG"
+    --tool-call-parser minimax_m3
+    --reasoning-parser minimax_m3
+    --enable-auto-tool-choice
+    --trust-remote-code
+    "${OFFLOAD_ARGS[@]}"
+)
+write_command "$RESULT_DIR/vllm_command.txt" "${VLLM_CMD[@]}"
+"${VLLM_CMD[@]}" > "$SERVER_LOG" 2>&1 &
 SERVER_PID=$!
 
-wait_for_server_ready --port "$VLLM_BACKEND_PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
-
-if [[ "$DP_ATTENTION" == "true" ]]; then
-    vllm-router \
-        --worker-urls "http://localhost:$VLLM_BACKEND_PORT" \
-        --policy consistent_hash \
-        --intra-node-data-parallel-size "$TP" \
-        --host 0.0.0.0 \
-        --port "$PORT" \
-        --prometheus-host 127.0.0.1 \
-        --prometheus-port "$((PORT + 10000))" \
-        --request-timeout-secs 14400 \
-        --disable-retries > "$ROUTER_LOG" 2>&1 &
-    ROUTER_PID=$!
-    wait_for_server_ready --port "$PORT" --server-log "$ROUTER_LOG" --server-pid "$ROUTER_PID"
-fi
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
 
-if [ "${EVAL_ONLY}" = "true" ]; then
+if [ "$EVAL_ONLY" = "true" ]; then
     run_eval --port "$PORT"
 else
     build_replay_cmd "$RESULT_DIR"
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 50d329c160..a80a48f248 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5824,31 +5824,5 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add H200 MiniMax-M3 MXFP8 AgentX on vLLM v0.27.1 with EAGLE3-GQA synthetic golden AL 2.78."
-    - "Screen TP8, TEP8, DEP8, and Mooncake DRAM-offload curves in AgentX fast mode before the final full sweep."
-    - "Require vLLM server metrics and session-aware DEP routing without relaxed request thresholds."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565
-
-- config-keys:
-    - minimaxm3-fp8-h200-vllm-agentic-mtp
-  scenario-type:
-    - agentic-coding
-  description:
-    - "Broad fast run 31540120459 identifies the TP8 knee at c10-c14; compare vLLM-simple DRAM offload at c10/c12/c14/c16 before the final full sweep."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565
-
-- config-keys:
-    - minimaxm3-fp8-h200-vllm-agentic-mtp
-  scenario-type:
-    - agentic-coding
-  description:
-    - "Route the vLLM-simple refinement through the MiniMax-M3 MTP runtime and preserve its checkpoint-cache DRAM reservation."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565
-
-- config-keys:
-    - minimaxm3-fp8-h200-vllm-agentic-mtp
-  scenario-type:
-    - agentic-coding
-  description:
-    - "Finalize the strict TP8 MTP sweep at resident c1/c2/c4/c6/c8/c10 and Mooncake c12/c14 from completed fast-run measurements."
+    - "Add MiniMax-M3 MXFP8 AgentX on H200 with vLLM v0.27.1, EAGLE3 golden AL 2.78, resident TP8 c1/c2/c4/c6/c8/c10, and Mooncake DRAM offload c12/c14."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2565