diff --git a/benchmarks/single_node/agentic/apply_atom_pr2106_patch.sh b/benchmarks/single_node/agentic/apply_atom_pr2106_patch.sh
new file mode 100755
index 0000000000..b044b0e15b
--- /dev/null
+++ b/benchmarks/single_node/agentic/apply_atom_pr2106_patch.sh
@@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Apply ROCm/ATOM PR #2106 to the pinned image before serving.
+# The patch fixes the MiniMax-M3 EAGLE3 draft KV layout and block-table
+# freshness issues that can cause prefix-cache warmup OOMs or stale-KV reads.
+
+ATOM_ROOT="${ATOM_ROOT:-/app/ATOM}"
+# Pin the PR's current base/head range so a later PR update cannot silently
+# change an already reviewed benchmark run.
+PATCH_URL="https://github.com/ROCm/ATOM/compare/86476f716e9887a7cb2e423deb88712737f702f5...c63ab67e58d39f1c887d2b6af13ac0b9034a1a1a.diff"
+PATCH_FILE="$(mktemp /tmp/atom-pr2106.XXXXXX.patch)"
+trap 'rm -f "$PATCH_FILE"' EXIT
+PATCH_EXCLUDES=(
+    --exclude=atom/model_ops/attentions/backends.py
+    --exclude=atom/spec_decode/eagle_proposer.py
+)
+
+if [[ ! -d "$ATOM_ROOT" ]]; then
+    echo "ERROR: ATOM source tree not found at $ATOM_ROOT" >&2
+    exit 1
+fi
+
+curl -fsSL "$PATCH_URL" -o "$PATCH_FILE"
+
+if git -C "$ATOM_ROOT" rev-parse --is-inside-work-tree >/dev/null 2>&1; then
+    if git -C "$ATOM_ROOT" apply --reverse --check "$PATCH_FILE" >/dev/null 2>&1; then
+        echo "ATOM PR #2106 already applied"
+        exit 0
+    fi
+
+    # These files were structurally rewritten after the pinned image's ATOM
+    # commit. Their freshness guard is orthogonal to the draft-KV OOM fix and
+    # is intentionally omitted until it can be ported against that revision.
+    git -C "$ATOM_ROOT" apply --check "${PATCH_EXCLUDES[@]}" "$PATCH_FILE"
+    git -C "$ATOM_ROOT" apply "${PATCH_EXCLUDES[@]}" "$PATCH_FILE"
+else
+    if [[ -f "$ATOM_ROOT/atom/spec_decode/draft_kv_layout.py" ]]; then
+        echo "ATOM PR #2106 already applied"
+        exit 0
+    fi
+    if patch --dry-run -p1 -d "$ATOM_ROOT" < "$PATCH_FILE" >/dev/null 2>&1; then
+        patch -p1 -d "$ATOM_ROOT" < "$PATCH_FILE"
+    else
+        echo "ERROR: ATOM PR #2106 does not apply cleanly to $ATOM_ROOT" >&2
+        exit 1
+    fi
+fi
+echo "Applied ROCm/ATOM PR #2106 to $ATOM_ROOT"
diff --git a/benchmarks/single_node/agentic/apply_atom_pr2147_patch.sh b/benchmarks/single_node/agentic/apply_atom_pr2147_patch.sh
new file mode 100644
index 0000000000..cedf2e3df8
--- /dev/null
+++ b/benchmarks/single_node/agentic/apply_atom_pr2147_patch.sh
@@ -0,0 +1,75 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# Apply ROCm/ATOM PR #2147 on top of the pinned MiniMax image checkout.
+#
+# PR #2147 supersedes PR #2106 for the MiniMax-M3 EAGLE3 draft-KV layout work:
+# both fix the draft/target KV pool disagreement that can OOM during prefix-cache
+# warmup. The pinned image predates #2147's merge base, so we fast-forward the
+# in-image tree to the current PR head rather than applying the PR diff alone.
+#
+# A small MiniMax EAGLE3 follow-up fix is applied after the forward diff because
+# the upstream check compared kernel block sizes (16 vs 128) instead of the
+# scheduler block size both sides actually index (128).
+
+ATOM_ROOT="${ATOM_ROOT:-/app/ATOM}"
+PINNED_ATOM_SHA="00760297ef69af7ab5d345af9c8fc6da00f5314d"
+PR2147_HEAD_SHA="71ad8c69c5377738657907e24c5aff8672c0f004"
+PATCH_URL="https://github.com/ROCm/ATOM/compare/${PINNED_ATOM_SHA}...${PR2147_HEAD_SHA}.diff"
+PATCH_FILE="$(mktemp /tmp/atom-pr2147.XXXXXX.patch)"
+MARKER_FILE="$ATOM_ROOT/.inferencex-pr2147-applied"
+DRAFT_KV_FILE="$ATOM_ROOT/atom/spec_decode/draft_kv.py"
+
+if [[ ! -d "$ATOM_ROOT" ]]; then
+    echo "ERROR: ATOM source tree not found at $ATOM_ROOT" >&2
+    exit 1
+fi
+
+trap 'rm -f "$PATCH_FILE"' EXIT
+
+if [[ -f "$MARKER_FILE" ]]; then
+    echo "ATOM PR #2147 already applied ($(cat "$MARKER_FILE"))"
+    exit 0
+fi
+
+curl -fsSL "$PATCH_URL" -o "$PATCH_FILE"
+git -C "$ATOM_ROOT" apply --check "$PATCH_FILE"
+git -C "$ATOM_ROOT" apply "$PATCH_FILE"
+
+python3 - "$DRAFT_KV_FILE" <<'PY'
+from pathlib import Path
+import sys
+
+path = Path(sys.argv[1])
+text = path.read_text()
+old = """        if pool.block_size != target.block_size:
+            raise ValueError(
+                f"the draft's blocks are {pool.block_size} tokens and the "
+                f"target's {target.block_size}; a draft is indexed by the "
+                "target's block tables, so it cannot yet block at anything else"
+            )"""
+new = """        pool_scheduler_block = pool.block_size * (
+            runner.block_size // pool.block_size
+        )
+        target_scheduler_block = target.block_size * target.block_ratio
+        if pool_scheduler_block != target_scheduler_block:
+            raise ValueError(
+                f"the draft indexes {pool_scheduler_block}-token scheduler "
+                f"blocks but the target indexes {target_scheduler_block}-token "
+                "scheduler blocks; a draft is indexed by the target's block "
+                "tables, so the two must agree at scheduler granularity"
+            )"""
+if old not in text:
+    if "pool_scheduler_block" in text:
+        print("MiniMax EAGLE3 scheduler-block fix already present")
+    else:
+        raise SystemExit(
+            "ERROR: expected draft_kv.py block-size guard not found after PR #2147 apply"
+        )
+else:
+    path.write_text(text.replace(old, new, 1))
+    print("Applied MiniMax EAGLE3 scheduler-block alignment fix to draft_kv.py")
+PY
+
+printf '%s\n' "$PR2147_HEAD_SHA" > "$MARKER_FILE"
+echo "Applied ROCm/ATOM PR #2147 (${PINNED_ATOM_SHA} -> ${PR2147_HEAD_SHA}) to $ATOM_ROOT"
diff --git a/benchmarks/single_node/agentic/minimaxm3_fp4_mi355x_atom_mtp.sh b/benchmarks/single_node/agentic/minimaxm3_fp4_mi355x_atom_mtp.sh
new file mode 100644
index 0000000000..fe194e73b1
--- /dev/null
+++ b/benchmarks/single_node/agentic/minimaxm3_fp4_mi355x_atom_mtp.sh
@@ -0,0 +1,345 @@
+#!/usr/bin/env bash
+set -eo pipefail
+set -x
+
+# Agentic trace replay benchmark for MiniMax-M3 MXFP4 on MI355X / MI350X (gfx950)
+# using the official ATOM MiniMax-M3 launch settings.
+#
+# Companion to minimaxm3_fp4_mi355x_mtp.sh, which runs the same checkpoint under
+# vLLM, so the two arms are directly comparable.
+#
+# TP2 and TP4 follow the official ATOM MiniMax-M3 MXFP4 recipe. TP8 is also
+# accepted for larger-memory variants and manual smoke tests.
+#
+# The ATOM image is launched with the official MiniMax attention/index-cache
+# settings; Kimi-specific vLLM patches are not sourced here.
+#
+# Required env vars:
+#   MODEL, MODEL_PATH, TP, DCP_SIZE, CONC, KV_OFFLOADING, KV_OFFLOAD_BACKEND,
+#   TOTAL_CPU_DRAM_GB, RESULT_DIR, DURATION, EP_SIZE, DP_ATTENTION
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
+
+export RESULT_FILENAME="${RESULT_FILENAME:-minimaxm3_agentic}"
+
+echo "MODEL=$MODEL TP=$TP DCP_SIZE=${DCP_SIZE:-1} CONC=$CONC KV_OFFLOADING=$KV_OFFLOADING TOTAL_CPU_DRAM_GB=$TOTAL_CPU_DRAM_GB RESULT_DIR=$RESULT_DIR DURATION=$DURATION EP_SIZE=$EP_SIZE DP_ATTENTION=$DP_ATTENTION"
+
+if [[ -v SLURM_JOB_ID ]]; then
+    echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
+fi
+
+if [ "$TP" -ne 2 ] && [ "$TP" -ne 4 ] && [ "$TP" -ne 8 ]; then
+    echo "Error: MiniMax-M3 MXFP4 supports TP2, TP4, or TP8 on 288 GB gfx950 parts." >&2
+    exit 1
+fi
+
+if [[ -v 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="$MODEL"
+fi
+
+DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3-GQA"
+hf download "$DRAFT_MODEL"
+
+wait_for_amd_gpu_clean
+
+rocm-smi || true
+amd-smi || true
+
+resolve_trace_source
+install_agentic_deps
+# ATOM's server runs from the image's system venv rather than the temporary
+# AIPerf venv created by install_agentic_deps. MiniMax's tokenizer fallback
+# requires these packages in that server environment.
+ATOM_RUNTIME_DEPS=/tmp/inferencex-atom-runtime-deps
+/opt/venv/bin/python -m pip install --quiet --target "$ATOM_RUNTIME_DEPS" --no-deps sentencepiece tiktoken
+
+# The pinned ATOM image predates upstream KV-pool fixes. Apply #2147 when set,
+# otherwise keep the reviewed #2106 patch for the current recipe image.
+case "${ATOM_PATCH:-2147}" in
+    2147)
+        bash "$(dirname "$0")/apply_atom_pr2147_patch.sh"
+        ;;
+    2106)
+        bash "$(dirname "$0")/apply_atom_pr2106_patch.sh"
+        ;;
+    *)
+        echo "Unsupported ATOM_PATCH=${ATOM_PATCH} (expected 2106 or 2147)" >&2
+        exit 1
+        ;;
+esac
+
+# Require the ATOM Prometheus stream in every official result. AIPerf
+# deduplicates this endpoint against its automatic localhost discovery.
+export AIPERF_SERVER_METRICS_URLS="http://localhost:${PORT}/metrics"
+export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="atom:"
+
+# Long agentic turns against a 1M context: keep the client from timing out
+# mid-request while the server is prefill-bound. Matches the vLLM K3 arm.
+export AIPERF_HTTP_TCP_USER_TIMEOUT=900000
+
+# VRAM space check
+wait_for_amd_gpu_clean
+
+# ---- Server config ----------------------------------------------------------
+SERVER_LOG="$RESULT_DIR/server.log"
+mkdir -p "$RESULT_DIR"
+
+SERVER_PID=""
+cleanup_agentic_services() {
+    local exit_code=$?
+    trap - EXIT INT TERM
+    set +e
+    stop_background_process_tree "$SERVER_PID" "ATOM server" 60
+    exit "$exit_code"
+}
+trap cleanup_agentic_services EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
+# ---- Per-concurrency knobs --------------------------------------------------
+# Concurrency 1-4 is the latency floor: everything GPU-resident, no decode
+# context parallelism, the deepest draft the golden curve publishes, and an
+# 8192-token prefill step.
+# Higher concurrency retains ATOM state replay checkpoints; KV offloading is
+# selected independently by KV_OFFLOADING/KV_OFFLOAD_BACKEND below.
+#
+# STATE_OFFLOAD_CPU_GIB is the per-rank slice of the CPU budget reserved for
+# the MiniMax-M3 state; 0 leaves the whole budget to the paged KV.
+case "$CONC" in
+    # No KV offload; the working set fits in HBM.
+    1|2|4|5)
+        MAX_NUM_SEQS=32
+        MAX_NUM_BATCHED_TOKENS=8192
+        GPU_MEM_UTIL=0.88
+        ATOM_ENABLE_REPLAYSSM=0
+        STATE_CHECKPOINT_SLOTS=""
+        NUM_SPEC_TOKENS=3
+        SPEC_DECODE_AL=2.78
+        STATE_OFFLOAD_CPU_GIB=0
+        ;;
+    # Higher concurrency settings retain ATOM's state replay checkpoints. KV
+    # offloading itself is selected by KV_OFFLOADING/KV_OFFLOAD_BACKEND.
+    8|10|12|14|15|20|24|28)
+        MAX_NUM_SEQS=32
+        MAX_NUM_BATCHED_TOKENS=4096
+        GPU_MEM_UTIL=0.88
+        ATOM_ENABLE_REPLAYSSM=1
+        STATE_CHECKPOINT_SLOTS=96
+        NUM_SPEC_TOKENS=3
+        SPEC_DECODE_AL=2.78
+        STATE_OFFLOAD_CPU_GIB=0
+        ;;
+    # LMCache paged-KV tier plus ATOM's CPU state tier, 32 GB/rank carved out
+    # for the attention state.
+    16)
+        MAX_NUM_SEQS=32
+        MAX_NUM_BATCHED_TOKENS=8192
+        GPU_MEM_UTIL=0.86
+        ATOM_ENABLE_REPLAYSSM=0
+        STATE_CHECKPOINT_SLOTS=""
+        NUM_SPEC_TOKENS=3
+        SPEC_DECODE_AL=2.78
+        STATE_OFFLOAD_CPU_GIB=32
+        ;;
+    32)
+        MAX_NUM_SEQS=64
+        MAX_NUM_BATCHED_TOKENS=8192
+        GPU_MEM_UTIL=0.86
+        ATOM_ENABLE_REPLAYSSM=0
+        STATE_CHECKPOINT_SLOTS=""
+        NUM_SPEC_TOKENS=3
+        SPEC_DECODE_AL=2.78
+        STATE_OFFLOAD_CPU_GIB=32
+        ;;
+    40)
+        MAX_NUM_SEQS=80
+        MAX_NUM_BATCHED_TOKENS=8192
+        GPU_MEM_UTIL=0.86
+        ATOM_ENABLE_REPLAYSSM=0
+        STATE_CHECKPOINT_SLOTS=""
+        NUM_SPEC_TOKENS=0
+        SPEC_DECODE_AL=0
+        STATE_OFFLOAD_CPU_GIB=32
+        ;;
+    56)
+        MAX_NUM_SEQS=72
+        MAX_NUM_BATCHED_TOKENS=4096
+        GPU_MEM_UTIL=0.88
+        ATOM_ENABLE_REPLAYSSM=0
+        STATE_CHECKPOINT_SLOTS=""
+        NUM_SPEC_TOKENS=0
+        SPEC_DECODE_AL=0
+        STATE_OFFLOAD_CPU_GIB=32
+        ;;
+    *)
+        echo "Unsupported CONC=$CONC" >&2
+        exit 2
+        ;;
+esac
+# Official MiniMax-M3 ATOM launch settings. Concurrency remains a benchmark
+# input, while the server capacity knobs follow the validated reference.
+MAX_NUM_SEQS=$((2 * CONC))
+MAX_NUM_BATCHED_TOKENS=32768
+GPU_MEM_UTIL=0.9
+export ATOM_ENABLE_REPLAYSSM
+
+# Extra in-GPU state checkpoint slots beyond the in-flight floor. Checkpoints
+# and live requests share one pool, so without this the room to retain a
+# checkpoint is whatever max-num-seqs happens to leave.
+STATE_CKPT_ARGS=()
+if [ -n "$STATE_CHECKPOINT_SLOTS" ]; then
+    STATE_CKPT_ARGS=(--state-checkpoint-slots "$STATE_CHECKPOINT_SLOTS")
+fi
+
+# ---- KV offload -------------------------------------------------------------
+# K3 is a hybrid: MiniMax-M3 attention carries a per-request recurrent state
+# alongside the paged KV. The paged KV rides this LMCache tier from
+# concurrency 8 up; from concurrency 16 up the CPU state tier is switched on
+# alongside it, because the state tier is what makes a resumed agentic turn
+# cheap and the paged KV tier alone cannot restore one.
+OFFLOAD_ARGS=()
+
+case "$KV_OFFLOAD_BACKEND" in
+    "")
+        require_agentic_kv_offload_none
+        ;;
+    lmcache)
+        require_agentic_kv_offload_backend lmcache
+
+        # TOTAL_CPU_DRAM_GB is the AGGREGATE budget from the matrix generator.
+        # LMCACHE_MAX_LOCAL_CPU_SIZE and OFFLOAD_STATE_CPU_SIZE are per rank and
+        # every rank allocates its own, so the aggregate is divided by TP as the
+        # agentic README requires. Handing a rank the whole aggregate does not
+        # just overcommit -- it never finishes pinning and hangs the launch
+        # partway through.
+        PER_RANK_CPU_GB="$((TOTAL_CPU_DRAM_GB / TP))"
+        LMCACHE_CPU_GB="$((PER_RANK_CPU_GB - STATE_OFFLOAD_CPU_GIB))"
+
+        export PYTHONHASHSEED=0
+        export LMCACHE_LOCAL_CPU=True
+        export LMCACHE_MAX_LOCAL_CPU_SIZE="$LMCACHE_CPU_GB"
+        # DCP-locked: the offload hash block is block-size(128) x dcp(8) = 1024,
+        # so the KV grid and the state-checkpoint grid coincide and the joint
+        # load aims both legs at one boundary. 512 or 2048 misaligns it.
+        export LMCACHE_CHUNK_SIZE=1024
+        export OFFLOAD_KV_FOR_HYBRID=1
+        # Statistics only -- per-step offload counters in the connector. Kept on
+        # because the submitted numbers were measured with it on.
+        export OFFLOAD_PROFILE=1
+
+        if [ "$STATE_OFFLOAD_CPU_GIB" -gt 0 ]; then
+            # CPU state-offload tier for the attention state.
+            export OFFLOAD_STATE=1
+            export OFFLOAD_STATE_CPU_SIZE="$STATE_OFFLOAD_CPU_GIB"
+            export OFFLOAD_STATE_STAGING_GROUPS=8
+            export OFFLOAD_STATE_MIN_LOAD_TOKENS=0
+            # Must be set: the staging buffer defaults to 2 chunks (8 MiB), one
+            # K3 state entry is 54.78 MiB, and a buffer too small to hold one
+            # entry makes the tier decline to build -- one log line, then
+            # nothing offloads, which reads exactly like a tier that is on and
+            # idle.
+            export OFFLOAD_GPU_STAGING_CHUNKS=32
+        fi
+
+        OFFLOAD_ARGS=(
+            --kv-transfer-config
+            "{\"kv_connector\":\"lmcache_offload\",\"kv_role\":\"offload\"}"
+        )
+        ;;
+    *)
+        echo "Unsupported KV_OFFLOAD_BACKEND: $KV_OFFLOAD_BACKEND (expected empty or lmcache)" >&2
+        exit 1
+        ;;
+esac
+
+# ---- ATOM env ---------------------------------------------------------------
+echo "Starting atom server..."
+export PYTHONNOUSERSITE=1
+
+# Required by ATOM: without it the aiter kernel logs flood the server log for
+# the whole 3600 s replay.
+export AITER_LOG_LEVEL="${AITER_LOG_LEVEL:-WARNING}"
+export AITER_SITUV2_A4W4=1
+export AITER_QUICK_REDUCE_QUANTIZATION=INT4
+export AITER_FLYDSL_STAGE2_FP8=1
+export ATOM_FORCE_ATTN_TRITON=1
+# Anchor-only state checkpointing: the demand rung is 47% of checkpoint writes
+# but reads back 2.8% of the time, against 85.2% for a prompt-end anchor, so it
+# costs more in evictions than its reuse is worth on these traces.
+
+# ---- Speculative ------------------------------------------------------------
+# golden_al_distribution/minimaxm3_eagle3_gqa.yaml:
+# minimax-m3.thinking_on[3] -> AL 2.78
+# Match the MI355X vLLM arm: EAGLE3-GQA draft, synthetic acceptance length on
+# throughput runs, and real target verification on eval-only runs.
+SPEC_ARGS=()
+if [ "$NUM_SPEC_TOKENS" -gt 0 ]; then
+    SPEC_ARGS=(
+        --method eagle3
+        --draft-model "$DRAFT_MODEL"
+        --num-speculative-tokens "$NUM_SPEC_TOKENS"
+    )
+    if [ "${EVAL_ONLY}" != "true" ]; then
+        SPEC_ARGS+=(--spec-decode-acceptance-length "$SPEC_DECODE_AL")
+    fi
+fi
+echo "SPEC_DECODE_AL=$SPEC_DECODE_AL NUM_SPEC_TOKENS=$NUM_SPEC_TOKENS"
+
+# ---- LLM server -------------------------------------------------------------
+ATOM_CMD=(
+    python -m atom.entrypoints.openai_server
+    --model "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --server-port "$PORT"
+    --trust-remote-code
+    --tensor-parallel-size "$TP"
+    --kv_cache_dtype fp8
+    --block-size 128
+    --max-num-seqs "$MAX_NUM_SEQS"
+    --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS"
+    --gpu-memory-utilization "$GPU_MEM_UTIL"
+    --index-cache-dtype fp8
+    --hf-overrides '{"use_index_cache": true, "index_topk_freq": 4}'
+    --online_quant_config '{"global_quant_config":"ptpc_fp8","exclude_layer":["lm_head","model.embed_tokens","vision_tower","multi_modal_projector","patch_merge_mlp","*block_sparse_moe"]}'
+    --default-chat-template-kwargs '{"thinking_mode":"enabled"}'
+    "${SPEC_ARGS[@]}"
+    "${OFFLOAD_ARGS[@]}"
+)
+if [[ "${ENABLE_PREFIX_CACHING:-true}" != "true" ]]; then
+    ATOM_CMD+=(--no-enable_prefix_caching)
+fi
+if declare -F write_command >/dev/null 2>&1; then
+    write_command "$RESULT_DIR/server_command.txt" "${ATOM_CMD[@]}"
+else
+    printf '%q ' "${ATOM_CMD[@]}" > "$RESULT_DIR/server_command.txt"
+    printf '\n' >> "$RESULT_DIR/server_command.txt"
+fi
+PYTHONPATH="$ATOM_RUNTIME_DEPS${PYTHONPATH:+:$PYTHONPATH}" \
+    "${ATOM_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+echo "Server PID: $SERVER_PID"
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+# ---- Run benchmark ----------------------------------------------------------
+if [ "${EVAL_ONLY}" = "true" ]; then
+    export EVAL_FRAMEWORK="${EVAL_FRAMEWORK:-lm-eval}"
+    export EVAL_TASKS_DIR="${EVAL_TASKS_DIR:-utils/evals/gsm8k.yaml}"
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --apply-chat-template"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 62daacca6f..452f929f79 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -685,6 +685,21 @@ kimik3-fp4-mi355x-atom-agentic-mtp:
       search-space:
       - { tp: 8, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.4.5" }, conc-list: [32, 40, 56], spec-decoding: mtp }
 
+minimaxm3-fp4-mi355x-atom-agentic-mtp:
+  image: rocm/atom-dev@sha256:1cd6abf29958cf0c332041b945249f7303ac86a36352f976b26d844ad72b4f5c
+  model: amd/MiniMax-M3-MXFP4
+  model-prefix: minimaxm3
+  runner: cluster:mi355x-amds
+  precision: fp4
+  framework: atom
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.20
+      search-space:
+      - { tp: 4, kv-offloading: none, conc-list: [1, 2, 4, 5, 8, 10, 12, 15, 20, 24, 28, 32], spec-decoding: mtp }
+      - { tp: 2, kv-offloading: none, conc-list: [1, 2, 5], spec-decoding: mtp }
+
 dsr1-fp4-mi355x-sglang-disagg:
   image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519
   model: amd/DeepSeek-R1-0528-MXFP4-v2
diff --git a/docs/waiver/2106.md b/docs/waiver/2106.md
new file mode 100644
index 0000000000..0107a81df7
--- /dev/null
+++ b/docs/waiver/2106.md
@@ -0,0 +1,62 @@
+# Inference-engine patch waiver — PR #2106
+
+Filed per [`docs/PR_REVIEW_CHECKLIST.md`](../PR_REVIEW_CHECKLIST.md) (image-provenance /
+patch item): the MiniMax-M3 ATOM benchmark applies an upstream serving-stack patch to
+the pinned image before serving.
+
+## Config covered
+
+- **Master config entry:** `minimaxm3-fp4-mi355x-atom-agentic-mtp` in
+  [`configs/amd-master.yaml`](../../configs/amd-master.yaml)
+- **Pinned image:** `rocm/atom-dev@sha256:1cd6abf29958cf0c332041b945249f7303ac86a36352f976b26d844ad72b4f5c`
+- **Patch entrypoint:** `ATOM_PATCH` selects the patch script from
+  [`benchmarks/single_node/agentic/minimaxm3_fp4_mi355x_atom_mtp.sh`](../../benchmarks/single_node/agentic/minimaxm3_fp4_mi355x_atom_mtp.sh)
+  (`2147` default, `2106` retained for comparison)
+- **Patch scripts:**
+  [`apply_atom_pr2147_patch.sh`](../../benchmarks/single_node/agentic/apply_atom_pr2147_patch.sh)
+  (default) and
+  [`apply_atom_pr2106_patch.sh`](../../benchmarks/single_node/agentic/apply_atom_pr2106_patch.sh)
+
+## What is patched
+
+### Default: PR #2147
+
+The default patch script fast-forwards the pinned image checkout to
+[ROCm/ATOM #2147](https://github.com/ROCm/ATOM/pull/2147)
+(`00760297ef69af7ab5d345af9c8fc6da00f5314d` →
+`71ad8c69c5377738657907e24c5aff8672c0f004`) and applies a MiniMax EAGLE3
+scheduler-block alignment follow-up in `draft_kv.py`. GSM8K and infra smoke
+validated this path on MI355X.
+
+### Legacy: PR #2106
+
+The legacy patch script downloads the pinned base-to-head patch for
+[ROCm/ATOM #2106](https://github.com/ROCm/ATOM/pull/2106)
+(`86476f716e9887a7cb2e423deb88712737f702f5` →
+`c63ab67e58d39f1c887d2b6af13ac0b9034a1a1a`), verifies it applies cleanly, and
+applies it idempotently to the ATOM source tree in `/app/ATOM`.
+The upstream change allocates the MiniMax-M3 EAGLE3 MHA draft KV pool in flash
+layout and validates block-table freshness, avoiding the full-pool V-layout
+conversion that can OOM during prefix-cache warmup.
+
+The pinned image contains a later, structurally divergent ATOM checkout. The
+flash-layout portion applies cleanly; the two freshness-guard files
+(`attentions/backends.py` and `spec_decode/eagle_proposer.py`) are excluded
+until those guards are ported to the image revision. The excluded guard is
+correctness hardening, not part of the warmup OOM fix.
+
+## Why the unmodified upstream image cannot run this benchmark
+
+The pinned image predates PR #2106. With MiniMax-M3 MXFP4, EAGLE3, prefix
+caching, FP8 KV/index cache, and long AgentX traces, the unmodified image can
+attempt a multi-gigabyte draft KV conversion during warmup. The observed
+`conc=24` run stopped with requests in flight and no progress, consistent with
+the OOM condition documented by the upstream PR. The patch is required for this
+benchmark path to run reliably.
+
+## Removal plan
+
+Remove `apply_atom_pr2106_patch.sh`, its invocation, and this waiver once a
+published ATOM image contains PR #2106. At that point update the pinned image
+digest in `configs/amd-master.yaml` and verify the MiniMax AgentX smoke,
+full-sweep, and GSM8K evaluation paths without the patch.
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 2f58d1b442..e120870aea 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6730,3 +6730,131 @@
     - "Configure model-specific tool and reasoning parsers for the Kimi-K3 and MiniMax-M3 serving paths so opt-in verifier requests produce structured OpenAI tool calls."
     - "Limit shared AgentX readiness and synthetic-acceptance rewrites to eval-only runs so default throughput recipe rendering remains unchanged."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2634
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add MiniMax-M3 MXFP4 on MI355X with ATOM agentic speculative decoding and LMCache DRAM offload."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Align MiniMax-M3 ATOM agentic MTP concurrency and LMCache DRAM tiers with the validated Kimi-K3 MI355X recipe."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use the official ATOM MiniMax-M3 MXFP4 FP8 index-cache and online quantization settings for the TP8 agentic path."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Remove the 32K serving-only context cap from the MiniMax-M3 ATOM AgentX path so long agentic traces use the model context limit."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Switch the MiniMax-M3 ATOM AgentX recipe to TP4 and enable prefix caching for the MI355X run."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Expand the TP4 no-offload MiniMax-M3 ATOM AgentX sweep to the requested concurrency points from 1 through 32."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Pin the MiniMax-M3 ATOM AgentX recipe to the latest image validated by the TP4 prefix-cache smoke."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Replace the floating ATOM latest tag with the exact image digest validated on MI355X."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Enable prefix caching by default in the MiniMax-M3 ATOM AgentX entrypoint for generated Actions runs."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Scale ATOM max-num-seqs with concurrency using two scheduler slots per AgentX lane."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Apply ROCm/ATOM PR #2106 before serving to fix MiniMax-M3 EAGLE3 draft KV warmup OOM and stale block-table reads."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add the TP2 no-offload MiniMax-M3 ATOM AgentX MTP search space at concurrencies 1, 2, and 5."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Allow the MiniMax-M3 ATOM AgentX launcher to accept TP2 search-space configurations."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Raise the MiniMax-M3 ATOM AgentX GPU memory utilization target to 0.9."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Align MiniMax-M3 ATOM AgentX with the MI355X vLLM arm: EAGLE3-GQA draft, synthetic acceptance length 2.78, and thinking_mode enabled."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2816
+
+- config-keys:
+    - minimaxm3-fp4-mi355x-atom-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Default to ROCm/ATOM PR #2147 KV-pool forward diff plus MiniMax EAGLE3 scheduler-block alignment; keep PR #2106 selectable via ATOM_PATCH and default EVAL_ONLY to lm-eval gsm8k."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2816
diff --git a/utils/matrix_logic/test_generate_sweep_configs.py b/utils/matrix_logic/test_generate_sweep_configs.py
index 520e4105e0..df47dc376e 100644
--- a/utils/matrix_logic/test_generate_sweep_configs.py
+++ b/utils/matrix_logic/test_generate_sweep_configs.py
@@ -2359,9 +2359,9 @@ def test_kimi_minimax_trimmed_eval_matrix_covers_current_configs(
             json.dumps(manifest, separators=(',', ':')).encode()
         ).hexdigest()
 
-        assert len(manifest) == 63
+        assert len(manifest) == 65
         assert manifest_digest == (
-            '1630cdd6fbf77302ee3286b118710576aa090e1502b3e5ec482b7f537a3f1132'
+            '65c954899fe920db1ca592bc19bd464018d99d85d48d85cccdf712e9f9696913'
         ), manifest_digest
         for row in rows:
             if isinstance(row['conc'], list):