From ba7e5579a52bb070b046f2450f4cf639ddf653a3 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Thu, 30 Jul 2026 01:10:33 -0400
Subject: [PATCH 1/5] feat(minimaxm3-mi300x-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-mi300x-vllm-agentic-mtp, the spec-decoding=mtp variant of the
MiniMax-M3 MXFP8 MI300X agentic-coding recipe: Inferact/MiniMax-M3-EAGLE3 draft
head, 3 speculative tokens, no drafter backend pin (server-wide TRITON_ATTN),
and synthetic acceptance pinned to the committed golden AL 2.83
(minimaxm3_eagle3.yaml, thinking_on, K=3). EVAL_ONLY keeps real verification.
The pinned ROCm nightly postdates vllm-project/vllm#45546, so no in-place
SupportsEagle3 patch is needed.

中文：新增 minimaxm3-fp8-mi300x-vllm-agentic-mtp，即 MiniMax-M3 MXFP8 MI300X 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；由于服务端整体使用 TRITON_ATTN，drafter 无需固定注意力后端；合成
接受长度固定为黄金 AL 2.83（minimaxm3_eagle3.yaml，thinking_on，K=3）。EVAL_ONLY
保留真实验证。所选 ROCm nightly 镜像晚于 vllm-project/vllm#45546，因此无需运行时
SupportsEagle3 补丁。

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 .../agentic/minimaxm3_fp8_mi300x_mtp.sh       | 269 ++++++++++++++++++
 configs/amd-master.yaml                       |  28 ++
 perf-changelog.yaml                           |  11 +
 3 files changed, 308 insertions(+)
 create mode 100755 benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh
new file mode 100755
index 0000000000..ff253568fd
--- /dev/null
+++ b/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh
@@ -0,0 +1,269 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# MiniMax-M3 MXFP8 MI300X (gfx942) AgentX (agentic-coding) recipe with EAGLE3
+# speculative decoding — the spec-decoding=mtp variant of
+# agentic/minimaxm3_fp8_mi300x.sh. Everything outside the speculative block
+# mirrors the non-MTP agentic sibling (Mooncake host-DRAM KV offload over TCP,
+# --block-size 128, --language-model-only, --kv-cache-dtype fp8, TRITON_ATTN,
+# minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0), so the spec-decode delta
+# is readable at equal concurrency.
+#
+# Speculative config: Inferact/MiniMax-M3-EAGLE3 draft head, 3 speculative
+# tokens. Unlike the CUDA recipes the drafter needs no attention_backend
+# override — the FlashInfer "page size 128 requires GQA/MQA" limitation that
+# forces FLASH_ATTN for the MHA EAGLE3 head is FlashInfer/CUDA-specific, and
+# here the whole server runs TRITON_ATTN, which serves the MHA draft fine
+# (same reasoning as fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh).
+#
+# No in-place SupportsEagle3 patch: the pinned ROCm nightly is built from
+# upstream 2026-06-23, after vllm-project/vllm#45546 (EAGLE3 support on the AMD
+# MiniMax-M3 model) merged on 2026-06-14. The older vllm-openai-rocm:minimax-m3
+# bring-up image still needs that runtime patch.
+#
+# 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 ${SLURMD_NODENAME:-unknown}"
+fi
+
+if [[ -n "${ROCR_VISIBLE_DEVICES:-}" ]]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+
+# The EAGLE3 draft is never pre-staged next to the target checkpoint; fetch it
+# into the shared HF cache, retrying because that cache is a network FS where
+# concurrent downloads hit huggingface_hub's WeakFileLock stale-handle race.
+for attempt in 1 2 3 4 5; do
+    hf download "$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 $attempt failed; retrying in 60s" >&2
+    sleep 60
+done
+rocm-smi || true
+amd-smi || true
+
+export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_062126
+resolve_trace_source
+install_agentic_deps
+
+export VLLM_ENGINE_READY_TIMEOUT_S=3600
+export VLLM_USE_BREAKABLE_CUDAGRAPH=0
+export PYTHONNOUSERSITE=1
+
+SERVER_LOG="$RESULT_DIR/server.log"
+ROUTER_LOG="$RESULT_DIR/router.log"
+MOONCAKE_MASTER_LOG="$RESULT_DIR/mooncake_master.log"
+mkdir -p "$RESULT_DIR"
+
+install_mooncake_rocm() {
+    local mooncake_tag="v0.3.11.post1"
+    local mooncake_src="/tmp/Mooncake-$mooncake_tag"
+    local mooncake_stage="/tmp/mooncake-stage-$mooncake_tag"
+    local build_jobs
+    local cache_root
+    local cache_key
+    local cache_archive
+    local cache_tmp
+    local engine_path
+    local os_version
+    local python_abi
+    local rocm_version
+
+    build_jobs=$(nproc)
+    if ((build_jobs > 32)); then
+        build_jobs=32
+    fi
+
+    os_version=$(. /etc/os-release && printf '%s-%s' "$ID" "$VERSION_ID")
+    python_abi=$(python3 -c 'import sys; print(f"cp{sys.version_info.major}{sys.version_info.minor}")')
+    rocm_version=$(sed -n '1p' /opt/rocm/.info/version 2>/dev/null || true)
+    if [[ -z "$rocm_version" ]]; then
+        rocm_version=$(hipconfig --version)
+    fi
+    rocm_version=${rocm_version//[^[:alnum:]._-]/_}
+    cache_root="${HF_HUB_CACHE:?HF_HUB_CACHE must be set}/inferencex/mooncake"
+    cache_key="${mooncake_tag}-${os_version}-${python_abi}-${rocm_version}-$(uname -m)-hip"
+    cache_archive="$cache_root/$cache_key.tar.gz"
+    mkdir -p "$cache_root"
+
+    apt-get update
+    DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
+        build-essential cmake git libasio-dev libboost-dev libcurl4-openssl-dev \
+        libgflags-dev libgoogle-glog-dev libibverbs-dev libjsoncpp-dev \
+        libnuma-dev libpython3-dev libssl-dev libunwind-dev liburing-dev \
+        libxxhash-dev libyaml-cpp-dev libzstd-dev ninja-build pybind11-dev
+
+    exec 9>"$cache_archive.lock"
+    flock -w 1800 9
+    if [[ -f "$cache_archive" ]] && ! tar -tzf "$cache_archive" >/dev/null 2>&1; then
+        rm -f "$cache_archive"
+    fi
+    if [[ ! -f "$cache_archive" ]]; then
+        echo "Building HIP Mooncake cache artifact: $cache_archive"
+        rm -rf "$mooncake_src" "$mooncake_stage"
+        git clone --depth 1 --branch "$mooncake_tag" --recurse-submodules \
+            --shallow-submodules https://github.com/kvcache-ai/Mooncake.git "$mooncake_src"
+        cmake -S "$mooncake_src/extern/yalantinglibs" \
+            -B "$mooncake_src/extern/yalantinglibs/build" \
+            -DBUILD_EXAMPLES=OFF -DBUILD_BENCHMARK=OFF -DBUILD_UNIT_TESTS=OFF
+        cmake --build "$mooncake_src/extern/yalantinglibs/build" -j "$build_jobs"
+        cmake --install "$mooncake_src/extern/yalantinglibs/build"
+        cmake -S "$mooncake_src" -B "$mooncake_src/build" -G Ninja \
+            -DCMAKE_BUILD_TYPE=Release -DUSE_CUDA=OFF -DUSE_HIP=ON \
+            -DWITH_EP=OFF -DWITH_STORE=ON -DWITH_STORE_RUST=OFF \
+            -DWITH_RUST_EXAMPLE=OFF -DBUILD_EXAMPLES=OFF -DBUILD_UNIT_TESTS=OFF
+        cmake --build "$mooncake_src/build" -j "$build_jobs"
+        mkdir -p "$mooncake_stage"
+        DESTDIR="$mooncake_stage" cmake --install "$mooncake_src/build"
+        cache_tmp=$(mktemp "$cache_root/$cache_key.tmp.XXXXXX")
+        tar -C "$mooncake_stage" -czf "$cache_tmp" .
+        mv -f "$cache_tmp" "$cache_archive"
+    else
+        echo "Using HIP Mooncake cache artifact: $cache_archive"
+    fi
+    tar -C / -xzf "$cache_archive"
+    engine_path=$(python3 -c 'import mooncake.engine; print(mooncake.engine.__file__)')
+    ldd "$engine_path" | grep -q 'libamdhip64.so'
+    exec 9>&-
+}
+
+OFFLOAD_ARGS=()
+if require_agentic_kv_offload_backend mooncake; then
+        PER_RANK_GB=$((TOTAL_CPU_DRAM_GB / TP))
+        if ! python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null 2>&1; then
+            install_mooncake_rocm
+        fi
+        python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null
+        MOONCAKE_MASTER_PORT=$((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": "tcp",
+  "device_name": "",
+  "enable_offload": false
+}
+EOF
+        export MOONCAKE_CONFIG_PATH PYTHONHASHSEED=0 MC_SLICE_SIZE=1048576 MC_WORKERS_PER_CTX=4
+        export MC_TCP_ENABLE_CONNECTION_POOL=1
+        mooncake_master --port "$MOONCAKE_MASTER_PORT" \
+            --default_kv_lease_ttl=120000 \
+            --eviction_high_watermark_ratio=0.80 \
+            --eviction_ratio=0.10 > "$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
+
+# 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\": $NUM_SPEC_TOKENS}"
+else
+    SPEC_CONFIG="{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"rejection_sample_method\": \"synthetic\", \"synthetic_acceptance_length\": $SYNTHETIC_ACCEPT_LEN}"
+fi
+
+# AgentX concurrency counts live session trees, not individual requests, so keep
+# the non-MTP recipe's 2x scheduler headroom for subagent fan-out. Cudagraph
+# capture is left at the sibling's default: the ROCm MiniMax-M3 MTP recipes run
+# with VLLM_USE_BREAKABLE_CUDAGRAPH=0 and no explicit capture ceiling.
+MAX_NUM_SEQS=$((2 * CONC))
+vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
+    --host 0.0.0.0 \
+    --port "$VLLM_BACKEND_PORT" \
+    "${PARALLEL_ARGS[@]}" \
+    "${EP_ARGS[@]}" \
+    --gpu-memory-utilization 0.95 \
+    --block-size 128 \
+    --language-model-only \
+    --attention-backend TRITON_ATTN \
+    --kv-cache-dtype fp8 \
+    --enable-prefix-caching \
+    --max-num-seqs "$MAX_NUM_SEQS" \
+    --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
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index b15ae5bccd..dd4904d8b3 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2204,6 +2204,34 @@ minimaxm3-fp8-mi300x-vllm-agentic:
       - { tp: 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] }
       - { 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] }
 
+# EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
+# minimaxm3-fp8-mi300x-vllm-agentic, pairing MiniMaxAI/MiniMax-M3-MXFP8 with the
+# Inferact/MiniMax-M3-EAGLE3 draft head (3 speculative tokens) and pinning
+# synthetic acceptance to the golden AL 2.83
+# (golden_al_distribution/minimaxm3_eagle3.yaml, thinking_on, K=3). Unlike the
+# CUDA recipes the drafter needs no attention_backend override: the whole server
+# runs TRITON_ATTN, which serves the MHA EAGLE3 head. The pinned ROCm nightly is
+# built from upstream 2026-06-23, after vllm-project/vllm#45546 landed
+# SupportsEagle3 on the AMD MiniMax-M3 model, so no in-place patch is needed.
+# Same TP8-only arms as the non-MTP entry (gfx942 192 GB is memory-tight),
+# trimmed at the extreme-conc end for the draft head's extra KV footprint.
+minimaxm3-fp8-mi300x-vllm-agentic-mtp:
+  image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: cluster:mi300x-amds
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 16] }
+      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 16] }
+      - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 14, 16] }
+      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 14, 16] }
+
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
   model: MiniMaxAI/MiniMax-M3-MXFP8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index c1736ae420..cef3786c4c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5311,3 +5311,14 @@
     - "Search space mirrors the non-MTP entry's KV arms -- TP8 GPU-resident and TP8 host-DRAM offload -- so the spec-decoding delta is readable at equal concurrency, but stops at conc 16 rather than 24. The non-MTP bring-up sweep (run 30326393603) showed the GPU-resident arm already thrashing at conc >= 16 (prefix cache hit rate 2.7%, TTFT p50 86-191s) because GPU KV holds only ~3.1 max-length requests, so conc 24 would spend a full job per arm re-measuring that regime; conc 16 still exercises the DRAM tier meaningfully (62% external prefix cache hit rate). TP8-only for the same memory reason: a ~1.5 TB MXFP4 checkpoint needs ~188 GB/GPU across 8 B300s and does not fit below 8 GPUs."
     - "Sets VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION=1, which the upstream recipe requires on both its blackwell and nvidia paths and which this repo had never set. It defaults to 0 and is threaded into LatentMoERunner as runner_args={\"enable_k3_latent_moe_tail_fusion\": ...} at vllm/models/kimi_k3/nvidia/model.py:549 -- the same runner whose shared-experts output buffer asserted (fused_moe/runner/shared_experts.py:165, all 8 TP ranks at once) when the wider flag alignment was attempted, so every K3 run here so far has been exercising a MoE tail path upstream does not use. Enabled on its own, ahead of re-attempting gpu-memory-utilization 0.95, max-num-seqs 32, --no-enable-flashinfer-autotune or VLLM_USE_V2_MODEL_RUNNER=1, to isolate its effect."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2418
+
+- config-keys:
+    - minimaxm3-fp8-mi300x-vllm-agentic-mtp
+  description:
+    - "Add the EAGLE3 speculative-decoding (spec-decoding: mtp) variant of minimaxm3-fp8-mi300x-vllm-agentic: MiniMax-M3 MXFP8 on MI300X (gfx942) with vLLM, agentic-coding scenario only, routed via spec-decoding=mtp to benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh."
+    - "Speculative config: method eagle3 on the Inferact/MiniMax-M3-EAGLE3 draft head, num_speculative_tokens 3. No attention_backend override for the drafter, unlike the CUDA recipes: the FlashInfer page-size-128-requires-GQA/MQA limitation that forces FLASH_ATTN for the MHA EAGLE3 head is FlashInfer-specific, and this server runs TRITON_ATTN throughout (same reasoning as fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh)."
+    - "Throughput runs pin rejection_sample_method synthetic with synthetic_acceptance_length 2.83 -- the committed golden AL for MiniMax-M3 EAGLE3 at K=3, thinking_on (golden_al_distribution/minimaxm3_eagle3.yaml), per the AgentX policy in golden_al_distribution/README.md. The minimaxm3_eagle3_gqa.yaml curve belongs to the Inferact/MiniMax-M3-EAGLE3-GQA draft (run 29784780049) and is not mixed in. EVAL_ONLY runs drop synthetic acceptance and keep real target verification."
+    - "No in-place SupportsEagle3 patch is applied. vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914 is built from upstream 2026-06-23, after vllm-project/vllm#45546 (EAGLE3 support on the AMD MiniMax-M3 model) merged on 2026-06-14; only the older vllm-openai-rocm:minimax-m3 bring-up image still needs the runtime patch that fixed_seq_len/minimaxm3_fp8_mi325x_mtp.sh carries."
+    - "Serve shape is otherwise identical to the non-MTP agentic sibling (Mooncake host-DRAM KV offload over TCP with the HIP-built transfer engine, --block-size 128, --language-model-only, --kv-cache-dtype fp8, --attention-backend TRITON_ATTN, gpu-memory-utilization 0.95, minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0, max-num-seqs 2*CONC). Cudagraph capture is left at the sibling default rather than raised to MAX_NUM_SEQS*(1+3) tokens as on the CUDA agentic MTP recipes, matching the ROCm MiniMax-M3 MTP recipes already merged."
+    - "Search space mirrors the non-MTP entry's four arms (TP8 and TP8/EP8, GPU-resident and Mooncake) trimmed at the extreme-conc end: GPU-resident 1-16 (vs 1-20) and Mooncake 5-16 (vs 5-20). MI300X has ~2.76M active FP8 KV tokens against a 269k-token service-time-weighted request, so the cliff is near conc 10 before the draft head and its KV take their share. TP8-only for the same memory reason as the non-MTP entry."
+  pr-link: TBD

From 18c53556b49721be6403f5cabcf20c12b417f9ba Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Thu, 30 Jul 2026 01:10:52 -0400
Subject: [PATCH 2/5] chore: fill perf-changelog pr-link for #2425
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：补充 perf-changelog 中 #2425 的 pr-link。

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index cef3786c4c..3ca823ff75 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5321,4 +5321,4 @@
     - "No in-place SupportsEagle3 patch is applied. vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914 is built from upstream 2026-06-23, after vllm-project/vllm#45546 (EAGLE3 support on the AMD MiniMax-M3 model) merged on 2026-06-14; only the older vllm-openai-rocm:minimax-m3 bring-up image still needs the runtime patch that fixed_seq_len/minimaxm3_fp8_mi325x_mtp.sh carries."
     - "Serve shape is otherwise identical to the non-MTP agentic sibling (Mooncake host-DRAM KV offload over TCP with the HIP-built transfer engine, --block-size 128, --language-model-only, --kv-cache-dtype fp8, --attention-backend TRITON_ATTN, gpu-memory-utilization 0.95, minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0, max-num-seqs 2*CONC). Cudagraph capture is left at the sibling default rather than raised to MAX_NUM_SEQS*(1+3) tokens as on the CUDA agentic MTP recipes, matching the ROCm MiniMax-M3 MTP recipes already merged."
     - "Search space mirrors the non-MTP entry's four arms (TP8 and TP8/EP8, GPU-resident and Mooncake) trimmed at the extreme-conc end: GPU-resident 1-16 (vs 1-20) and Mooncake 5-16 (vs 5-20). MI300X has ~2.76M active FP8 KV tokens against a 269k-token service-time-weighted request, so the cliff is near conc 10 before the draft head and its KV take their share. TP8-only for the same memory reason as the non-MTP entry."
-  pr-link: TBD
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2425

From 5c9c40e4d2dff7707a3b3b2a8f62ecd8562c948e Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Thu, 30 Jul 2026 01:40:44 -0400
Subject: [PATCH 3/5] fix(minimaxm3-mi300x-agentic-mtp): drop
 gpu-memory-utilization to 0.90 to fit the EAGLE3 draft
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The H100 twin of this recipe died mid-warmup with torch.OutOfMemoryError on all
8 ranks (run 30515793863) because it inherited the non-MTP agentic sibling's
gpu-memory-utilization, leaving no headroom for the draft head and its KV.
fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh passes no gmu flag at all, i.e. vLLM's
0.90 default; match that here instead of the sibling's 0.95.

中文：本配方的 H100 版本在预热阶段因 8 个 rank 全部 OOM 而失败（run 30515793863），
原因是沿用了非 MTP 版本的 gpu-memory-utilization，未为草稿头及其 KV 预留显存。
fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh 未设置该参数（即使用 vLLM 默认的 0.90），
此处与之对齐，不再使用 0.95。

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 .../single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh       | 8 +++++++-
 perf-changelog.yaml                                       | 2 +-
 2 files changed, 8 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh
index ff253568fd..e12ca3e5c3 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh
@@ -10,6 +10,12 @@ set -x
 # minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0), so the spec-decode delta
 # is readable at equal concurrency.
 #
+# One deliberate exception: --gpu-memory-utilization is 0.90, not the non-MTP
+# sibling's 0.95. fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh passes no gmu flag at
+# all, i.e. it runs vLLM's 0.90 default, and that is the proven MTP setting here.
+# The H100 twin of this recipe died mid-warmup at 0.95 with torch.OutOfMemoryError
+# on every rank once the EAGLE3 head and its KV were resident (run 30515793863).
+#
 # Speculative config: Inferact/MiniMax-M3-EAGLE3 draft head, 3 speculative
 # tokens. Unlike the CUDA recipes the drafter needs no attention_backend
 # override — the FlashInfer "page size 128 requires GQA/MQA" limitation that
@@ -229,7 +235,7 @@ vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
     --port "$VLLM_BACKEND_PORT" \
     "${PARALLEL_ARGS[@]}" \
     "${EP_ARGS[@]}" \
-    --gpu-memory-utilization 0.95 \
+    --gpu-memory-utilization 0.90 \
     --block-size 128 \
     --language-model-only \
     --attention-backend TRITON_ATTN \
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 3ca823ff75..3516a9533e 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5319,6 +5319,6 @@
     - "Speculative config: method eagle3 on the Inferact/MiniMax-M3-EAGLE3 draft head, num_speculative_tokens 3. No attention_backend override for the drafter, unlike the CUDA recipes: the FlashInfer page-size-128-requires-GQA/MQA limitation that forces FLASH_ATTN for the MHA EAGLE3 head is FlashInfer-specific, and this server runs TRITON_ATTN throughout (same reasoning as fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh)."
     - "Throughput runs pin rejection_sample_method synthetic with synthetic_acceptance_length 2.83 -- the committed golden AL for MiniMax-M3 EAGLE3 at K=3, thinking_on (golden_al_distribution/minimaxm3_eagle3.yaml), per the AgentX policy in golden_al_distribution/README.md. The minimaxm3_eagle3_gqa.yaml curve belongs to the Inferact/MiniMax-M3-EAGLE3-GQA draft (run 29784780049) and is not mixed in. EVAL_ONLY runs drop synthetic acceptance and keep real target verification."
     - "No in-place SupportsEagle3 patch is applied. vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914 is built from upstream 2026-06-23, after vllm-project/vllm#45546 (EAGLE3 support on the AMD MiniMax-M3 model) merged on 2026-06-14; only the older vllm-openai-rocm:minimax-m3 bring-up image still needs the runtime patch that fixed_seq_len/minimaxm3_fp8_mi325x_mtp.sh carries."
-    - "Serve shape is otherwise identical to the non-MTP agentic sibling (Mooncake host-DRAM KV offload over TCP with the HIP-built transfer engine, --block-size 128, --language-model-only, --kv-cache-dtype fp8, --attention-backend TRITON_ATTN, gpu-memory-utilization 0.95, minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0, max-num-seqs 2*CONC). Cudagraph capture is left at the sibling default rather than raised to MAX_NUM_SEQS*(1+3) tokens as on the CUDA agentic MTP recipes, matching the ROCm MiniMax-M3 MTP recipes already merged."
+    - "Serve shape is otherwise identical to the non-MTP agentic sibling (Mooncake host-DRAM KV offload over TCP with the HIP-built transfer engine, --block-size 128, --language-model-only, --kv-cache-dtype fp8, --attention-backend TRITON_ATTN, minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0, max-num-seqs 2*CONC), with one deliberate exception: gpu-memory-utilization is 0.90 rather than the sibling's 0.95. fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh passes no gmu flag at all (vLLM's 0.90 default), which is the proven MTP setting here; the H100 twin of this recipe died mid-warmup at 0.95 with torch.OutOfMemoryError on all 8 ranks once the EAGLE3 head and its KV were resident (run 30515793863). Cudagraph capture is left at the sibling default rather than raised to MAX_NUM_SEQS*(1+3) tokens as on the CUDA agentic MTP recipes, matching the ROCm MiniMax-M3 MTP recipes already merged."
     - "Search space mirrors the non-MTP entry's four arms (TP8 and TP8/EP8, GPU-resident and Mooncake) trimmed at the extreme-conc end: GPU-resident 1-16 (vs 1-20) and Mooncake 5-16 (vs 5-20). MI300X has ~2.76M active FP8 KV tokens against a 269k-token service-time-weighted request, so the cliff is near conc 10 before the draft head and its KV take their share. TP8-only for the same memory reason as the non-MTP entry."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2425

From 514008562afd4a3684fd09dd1f6e703e3a12aacf Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Thu, 30 Jul 2026 01:42:02 -0400
Subject: [PATCH 4/5] =?UTF-8?q?fix:=20AgentX=20MTP=20conc=20grid=20?=
 =?UTF-8?q?=E2=80=94=20steps=20>=3D2,=20hard=20stop=20at=20conc=2016?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Replace the dense integer sampling inherited from the non-MTP agentic siblings
with the AgentX MTP grid: GPU-resident [1,4,8,12,16], offload [4,8,12,16].
Single-step sampling cannot separate configs by more than noise on the agentic
corpus, and past conc 16 these SKUs are in the post-cliff thrashing regime.

中文：将沿用自非 MTP 智能体配方的逐 1 递增并发采样，替换为 AgentX MTP 并发网格：
GPU 常驻 [1,4,8,12,16]，KV 卸载 [4,8,12,16]。相邻并发间隔至少为 2，且一律止于 16 ——
逐 1 采样在智能体语料上无法区分出超过噪声的差异，并发超过 16 后这些 SKU 已进入
显存瓶颈后的抖动区间。

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 configs/amd-master.yaml | 8 ++++----
 perf-changelog.yaml     | 2 +-
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index dd4904d8b3..f8671cf3d6 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2227,10 +2227,10 @@ minimaxm3-fp8-mi300x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 16] }
-      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 14, 16] }
-      - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 14, 16] }
-      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 14, 16] }
+      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16] }
+      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16] }
+      - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [4, 8, 12, 16] }
+      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [4, 8, 12, 16] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 3516a9533e..7b36023666 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5320,5 +5320,5 @@
     - "Throughput runs pin rejection_sample_method synthetic with synthetic_acceptance_length 2.83 -- the committed golden AL for MiniMax-M3 EAGLE3 at K=3, thinking_on (golden_al_distribution/minimaxm3_eagle3.yaml), per the AgentX policy in golden_al_distribution/README.md. The minimaxm3_eagle3_gqa.yaml curve belongs to the Inferact/MiniMax-M3-EAGLE3-GQA draft (run 29784780049) and is not mixed in. EVAL_ONLY runs drop synthetic acceptance and keep real target verification."
     - "No in-place SupportsEagle3 patch is applied. vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914 is built from upstream 2026-06-23, after vllm-project/vllm#45546 (EAGLE3 support on the AMD MiniMax-M3 model) merged on 2026-06-14; only the older vllm-openai-rocm:minimax-m3 bring-up image still needs the runtime patch that fixed_seq_len/minimaxm3_fp8_mi325x_mtp.sh carries."
     - "Serve shape is otherwise identical to the non-MTP agentic sibling (Mooncake host-DRAM KV offload over TCP with the HIP-built transfer engine, --block-size 128, --language-model-only, --kv-cache-dtype fp8, --attention-backend TRITON_ATTN, minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0, max-num-seqs 2*CONC), with one deliberate exception: gpu-memory-utilization is 0.90 rather than the sibling's 0.95. fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh passes no gmu flag at all (vLLM's 0.90 default), which is the proven MTP setting here; the H100 twin of this recipe died mid-warmup at 0.95 with torch.OutOfMemoryError on all 8 ranks once the EAGLE3 head and its KV were resident (run 30515793863). Cudagraph capture is left at the sibling default rather than raised to MAX_NUM_SEQS*(1+3) tokens as on the CUDA agentic MTP recipes, matching the ROCm MiniMax-M3 MTP recipes already merged."
-    - "Search space mirrors the non-MTP entry's four arms (TP8 and TP8/EP8, GPU-resident and Mooncake) trimmed at the extreme-conc end: GPU-resident 1-16 (vs 1-20) and Mooncake 5-16 (vs 5-20). MI300X has ~2.76M active FP8 KV tokens against a 269k-token service-time-weighted request, so the cliff is near conc 10 before the draft head and its KV take their share. TP8-only for the same memory reason as the non-MTP entry."
+    - "Search space now runs on the AgentX MTP concurrency grid: GPU-resident rows [1, 4, 8, 12, 16] and Mooncake rows [4, 8, 12, 16]. Every step is at least 2 concurrency apart and every arm stops hard at conc 16 -- single-step sampling cannot separate configurations by more than run-to-run noise on the agentic corpus, and past conc 16 these SKUs are into the post-HBM-cliff thrashing regime that the non-MTP sweeps already characterized, which is not worth one GPU job per point."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2425

From 018fb6c399f2d1660161e7beceff2eb061ee3413 Mon Sep 17 00:00:00 2001
From: functionstackx <47992694+functionstackx@users.noreply.github.com>
Date: Thu, 30 Jul 2026 02:02:41 -0400
Subject: [PATCH 5/5] =?UTF-8?q?fix(minimaxm3-mi300x-agentic-mtp):=20drop?=
 =?UTF-8?q?=20Mooncake=20=E2=80=94=20GPU-resident=20only=20on=20AMD?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The dram-offload job (TP8 mtp mooncake c8, run 30517360090) failed: the
HIP-built Mooncake transfer engine runs over TCP on this cluster and the store
collapses mid-replay, spewing "ServerSession::readBody failed ... Connection
reset by peer" until AIPerf aborts the run.

Mooncake is not used on AMD. Remove both host-DRAM arms and the in-script
Mooncake build/launch plumbing; the recipe is now GPU-resident only (TP8 and
TP8/EP8, 10 jobs) and hard-fails if a dram row is ever pointed at it rather
than silently serving GPU-resident under a dram label.

中文：dram 卸载作业（TP8 mtp mooncake c8，run 30517360090）失败：基于 HIP 构建的
Mooncake 传输引擎在该集群上走 TCP，运行中 store 会崩溃并持续输出
"ServerSession::readBody failed ... Connection reset by peer"，直至 AIPerf 中止。
AMD 上不再使用 Mooncake：删除两条主机 DRAM 分支及脚本内的 Mooncake 构建/启动逻辑，
本配方改为仅 GPU 常驻（TP8 与 TP8/EP8，共 10 个作业）；若仍传入 dram 配置将直接报错
退出，而不会在 dram 标签下静默执行 GPU 常驻测试。

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---
 .../agentic/minimaxm3_fp8_mi300x_mtp.sh       | 120 ++----------------
 configs/amd-master.yaml                       |  11 +-
 perf-changelog.yaml                           |   4 +-
 3 files changed, 18 insertions(+), 117 deletions(-)

diff --git a/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh
index e12ca3e5c3..8244427509 100755
--- a/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh
+++ b/benchmarks/single_node/agentic/minimaxm3_fp8_mi300x_mtp.sh
@@ -5,8 +5,7 @@ set -x
 # MiniMax-M3 MXFP8 MI300X (gfx942) AgentX (agentic-coding) recipe with EAGLE3
 # speculative decoding — the spec-decoding=mtp variant of
 # agentic/minimaxm3_fp8_mi300x.sh. Everything outside the speculative block
-# mirrors the non-MTP agentic sibling (Mooncake host-DRAM KV offload over TCP,
-# --block-size 128, --language-model-only, --kv-cache-dtype fp8, TRITON_ATTN,
+# mirrors the non-MTP agentic sibling (--block-size 128, --language-model-only, --kv-cache-dtype fp8, TRITON_ATTN,
 # minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0), so the spec-decode delta
 # is readable at equal concurrency.
 #
@@ -76,116 +75,19 @@ 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"
 
-install_mooncake_rocm() {
-    local mooncake_tag="v0.3.11.post1"
-    local mooncake_src="/tmp/Mooncake-$mooncake_tag"
-    local mooncake_stage="/tmp/mooncake-stage-$mooncake_tag"
-    local build_jobs
-    local cache_root
-    local cache_key
-    local cache_archive
-    local cache_tmp
-    local engine_path
-    local os_version
-    local python_abi
-    local rocm_version
-
-    build_jobs=$(nproc)
-    if ((build_jobs > 32)); then
-        build_jobs=32
-    fi
-
-    os_version=$(. /etc/os-release && printf '%s-%s' "$ID" "$VERSION_ID")
-    python_abi=$(python3 -c 'import sys; print(f"cp{sys.version_info.major}{sys.version_info.minor}")')
-    rocm_version=$(sed -n '1p' /opt/rocm/.info/version 2>/dev/null || true)
-    if [[ -z "$rocm_version" ]]; then
-        rocm_version=$(hipconfig --version)
-    fi
-    rocm_version=${rocm_version//[^[:alnum:]._-]/_}
-    cache_root="${HF_HUB_CACHE:?HF_HUB_CACHE must be set}/inferencex/mooncake"
-    cache_key="${mooncake_tag}-${os_version}-${python_abi}-${rocm_version}-$(uname -m)-hip"
-    cache_archive="$cache_root/$cache_key.tar.gz"
-    mkdir -p "$cache_root"
-
-    apt-get update
-    DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
-        build-essential cmake git libasio-dev libboost-dev libcurl4-openssl-dev \
-        libgflags-dev libgoogle-glog-dev libibverbs-dev libjsoncpp-dev \
-        libnuma-dev libpython3-dev libssl-dev libunwind-dev liburing-dev \
-        libxxhash-dev libyaml-cpp-dev libzstd-dev ninja-build pybind11-dev
-
-    exec 9>"$cache_archive.lock"
-    flock -w 1800 9
-    if [[ -f "$cache_archive" ]] && ! tar -tzf "$cache_archive" >/dev/null 2>&1; then
-        rm -f "$cache_archive"
-    fi
-    if [[ ! -f "$cache_archive" ]]; then
-        echo "Building HIP Mooncake cache artifact: $cache_archive"
-        rm -rf "$mooncake_src" "$mooncake_stage"
-        git clone --depth 1 --branch "$mooncake_tag" --recurse-submodules \
-            --shallow-submodules https://github.com/kvcache-ai/Mooncake.git "$mooncake_src"
-        cmake -S "$mooncake_src/extern/yalantinglibs" \
-            -B "$mooncake_src/extern/yalantinglibs/build" \
-            -DBUILD_EXAMPLES=OFF -DBUILD_BENCHMARK=OFF -DBUILD_UNIT_TESTS=OFF
-        cmake --build "$mooncake_src/extern/yalantinglibs/build" -j "$build_jobs"
-        cmake --install "$mooncake_src/extern/yalantinglibs/build"
-        cmake -S "$mooncake_src" -B "$mooncake_src/build" -G Ninja \
-            -DCMAKE_BUILD_TYPE=Release -DUSE_CUDA=OFF -DUSE_HIP=ON \
-            -DWITH_EP=OFF -DWITH_STORE=ON -DWITH_STORE_RUST=OFF \
-            -DWITH_RUST_EXAMPLE=OFF -DBUILD_EXAMPLES=OFF -DBUILD_UNIT_TESTS=OFF
-        cmake --build "$mooncake_src/build" -j "$build_jobs"
-        mkdir -p "$mooncake_stage"
-        DESTDIR="$mooncake_stage" cmake --install "$mooncake_src/build"
-        cache_tmp=$(mktemp "$cache_root/$cache_key.tmp.XXXXXX")
-        tar -C "$mooncake_stage" -czf "$cache_tmp" .
-        mv -f "$cache_tmp" "$cache_archive"
-    else
-        echo "Using HIP Mooncake cache artifact: $cache_archive"
-    fi
-    tar -C / -xzf "$cache_archive"
-    engine_path=$(python3 -c 'import mooncake.engine; print(mooncake.engine.__file__)')
-    ldd "$engine_path" | grep -q 'libamdhip64.so'
-    exec 9>&-
-}
-
-OFFLOAD_ARGS=()
-if require_agentic_kv_offload_backend mooncake; then
-        PER_RANK_GB=$((TOTAL_CPU_DRAM_GB / TP))
-        if ! python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null 2>&1; then
-            install_mooncake_rocm
-        fi
-        python3 -c "from mooncake.store import MooncakeDistributedStore" >/dev/null
-        MOONCAKE_MASTER_PORT=$((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": "tcp",
-  "device_name": "",
-  "enable_offload": false
-}
-EOF
-        export MOONCAKE_CONFIG_PATH PYTHONHASHSEED=0 MC_SLICE_SIZE=1048576 MC_WORKERS_PER_CTX=4
-        export MC_TCP_ENABLE_CONNECTION_POOL=1
-        mooncake_master --port "$MOONCAKE_MASTER_PORT" \
-            --default_kv_lease_ttl=120000 \
-            --eviction_high_watermark_ratio=0.80 \
-            --eviction_ratio=0.10 > "$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}}'
-        )
+# GPU-resident only: no Mooncake on AMD. The HIP-built Mooncake transfer engine's
+# TCP transport collapses mid-run on gfx942 -- the store spews
+# "ServerSession::readBody failed ... Connection reset by peer" until the replay
+# aborts (run 30517360090, MI300X TP8 mtp dram c8). This recipe therefore carries
+# no host-DRAM KV offload arm, and fails loudly rather than silently serving a
+# GPU-resident run under a dram label.
+if [[ "${KV_OFFLOADING}" != "none" ]]; then
+    echo "Error: KV_OFFLOADING=$KV_OFFLOADING requested, but this recipe is GPU-resident only (Mooncake is not used on AMD)" >&2
+    exit 1
 fi
+OFFLOAD_ARGS=()
 
 PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
 if [[ "$DP_ATTENTION" == "true" ]]; then
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index f8671cf3d6..4c3b45223b 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -2213,8 +2213,10 @@ minimaxm3-fp8-mi300x-vllm-agentic:
 # runs TRITON_ATTN, which serves the MHA EAGLE3 head. The pinned ROCm nightly is
 # built from upstream 2026-06-23, after vllm-project/vllm#45546 landed
 # SupportsEagle3 on the AMD MiniMax-M3 model, so no in-place patch is needed.
-# Same TP8-only arms as the non-MTP entry (gfx942 192 GB is memory-tight),
-# trimmed at the extreme-conc end for the draft head's extra KV footprint.
+# GPU-resident only: no Mooncake on AMD. The HIP-built Mooncake transfer engine's
+# TCP transport collapses mid-run on gfx942 (run 30517360090), so the host-DRAM
+# offload arms the non-MTP entry carries are dropped here rather than shipped
+# broken. TP8-only, as on the non-MTP entry (gfx942 192 GB is memory-tight).
 minimaxm3-fp8-mi300x-vllm-agentic-mtp:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
   model: MiniMaxAI/MiniMax-M3-MXFP8
@@ -2225,12 +2227,9 @@ minimaxm3-fp8-mi300x-vllm-agentic-mtp:
   multinode: false
   scenarios:
     agentic-coding:
-    - dram-utilization: 0.80
-      search-space:
+    - search-space:
       - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16] }
       - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16] }
-      - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [4, 8, 12, 16] }
-      - { tp: 8, ep: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [4, 8, 12, 16] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 7b36023666..58b8ff502c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5319,6 +5319,6 @@
     - "Speculative config: method eagle3 on the Inferact/MiniMax-M3-EAGLE3 draft head, num_speculative_tokens 3. No attention_backend override for the drafter, unlike the CUDA recipes: the FlashInfer page-size-128-requires-GQA/MQA limitation that forces FLASH_ATTN for the MHA EAGLE3 head is FlashInfer-specific, and this server runs TRITON_ATTN throughout (same reasoning as fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh)."
     - "Throughput runs pin rejection_sample_method synthetic with synthetic_acceptance_length 2.83 -- the committed golden AL for MiniMax-M3 EAGLE3 at K=3, thinking_on (golden_al_distribution/minimaxm3_eagle3.yaml), per the AgentX policy in golden_al_distribution/README.md. The minimaxm3_eagle3_gqa.yaml curve belongs to the Inferact/MiniMax-M3-EAGLE3-GQA draft (run 29784780049) and is not mixed in. EVAL_ONLY runs drop synthetic acceptance and keep real target verification."
     - "No in-place SupportsEagle3 patch is applied. vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914 is built from upstream 2026-06-23, after vllm-project/vllm#45546 (EAGLE3 support on the AMD MiniMax-M3 model) merged on 2026-06-14; only the older vllm-openai-rocm:minimax-m3 bring-up image still needs the runtime patch that fixed_seq_len/minimaxm3_fp8_mi325x_mtp.sh carries."
-    - "Serve shape is otherwise identical to the non-MTP agentic sibling (Mooncake host-DRAM KV offload over TCP with the HIP-built transfer engine, --block-size 128, --language-model-only, --kv-cache-dtype fp8, --attention-backend TRITON_ATTN, minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0, max-num-seqs 2*CONC), with one deliberate exception: gpu-memory-utilization is 0.90 rather than the sibling's 0.95. fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh passes no gmu flag at all (vLLM's 0.90 default), which is the proven MTP setting here; the H100 twin of this recipe died mid-warmup at 0.95 with torch.OutOfMemoryError on all 8 ranks once the EAGLE3 head and its KV were resident (run 30515793863). Cudagraph capture is left at the sibling default rather than raised to MAX_NUM_SEQS*(1+3) tokens as on the CUDA agentic MTP recipes, matching the ROCm MiniMax-M3 MTP recipes already merged."
-    - "Search space now runs on the AgentX MTP concurrency grid: GPU-resident rows [1, 4, 8, 12, 16] and Mooncake rows [4, 8, 12, 16]. Every step is at least 2 concurrency apart and every arm stops hard at conc 16 -- single-step sampling cannot separate configurations by more than run-to-run noise on the agentic corpus, and past conc 16 these SKUs are into the post-HBM-cliff thrashing regime that the non-MTP sweeps already characterized, which is not worth one GPU job per point."
+    - "Serve shape is otherwise identical to the non-MTP agentic sibling (--block-size 128, --language-model-only, --kv-cache-dtype fp8, --attention-backend TRITON_ATTN, minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0, max-num-seqs 2*CONC), with one deliberate exception: gpu-memory-utilization is 0.90 rather than the sibling's 0.95. fixed_seq_len/minimaxm3_fp8_mi300x_mtp.sh passes no gmu flag at all (vLLM's 0.90 default), which is the proven MTP setting here; the H100 twin of this recipe died mid-warmup at 0.95 with torch.OutOfMemoryError on all 8 ranks once the EAGLE3 head and its KV were resident (run 30515793863). Cudagraph capture is left at the sibling default rather than raised to MAX_NUM_SEQS*(1+3) tokens as on the CUDA agentic MTP recipes, matching the ROCm MiniMax-M3 MTP recipes already merged."
+    - "Search space is GPU-resident only -- TP8 and TP8/EP8 on the AgentX MTP concurrency grid [1, 4, 8, 12, 16], steps of at least 2 with a hard stop at conc 16. The non-MTP entry's two Mooncake host-DRAM arms are dropped: Mooncake is not used on AMD. Its HIP-built transfer engine runs over TCP on this cluster and the store collapses mid-replay (ServerSession::readBody failed ... Connection reset by peer, repeating until AIPerf aborts; run 30517360090, TP8 mtp dram c8), so a dram arm here would measure the transport, not the engine."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2425