MiniMax-M3 H100 AgentX EAGLE3 tuning on vLLM v0.27.1

benchmarks/single_node/agentic/minimaxm3_fp8_h100_mtp.sh

@@ -0,0 +1,225 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# MiniMax-M3 MXFP8 H100 AgentX (agentic-coding) recipe with EAGLE3 speculative
+# decoding — the spec-decoding=mtp variant of agentic/minimaxm3_fp8_h100.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, minimax_m3 parsers, vllm-router for
+# DP-attention), so the spec-decode delta is readable at equal concurrency.
+#
+# Speculative config: the current Inferact/MiniMax-M3-EAGLE3-GQA draft head
+# with three speculative tokens and the committed thinking-on golden AL.
+#
+# The drafter is pinned to FLASH_ATTN, as on every CUDA MiniMax-M3 MTP recipe:
+# the EAGLE3 head is MHA and FlashInfer only serves page size 128 through its
+# trtllm-gen kernel, which requires GQA/MQA. FLASH_ATTN accepts any
+# multiple-of-16 block size, so the mandatory 128 is fine for the draft. (The
+# ROCm recipes need no pin because their server runs TRITON_ATTN throughout.)
+#
+# 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 "$0")/../../benchmark_lib.sh"
+
+check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
+
+DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3-GQA"
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
+fi
+
+if [[ -n "${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="$MODEL"
+fi
+
+# The EAGLE3 draft is never pre-staged next to the target checkpoint; fetch it
+# into the shared HF cache. That cache is a network FS where concurrent
+# day-zero downloads hit huggingface_hub's WeakFileLock "[Errno 116] Stale file
+# handle" race, so retry (the download resumes) as the fixed-seq-len MTP
+# recipes do.
+for attempt in 1 2 3 4 5; do
+    hf download "$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 "$RESULT_DIR"
+
+OFFLOAD_ARGS=()
+MODEL_CPU_OFFLOAD_GB=26
+MODEL_CHECKPOINT_PAGE_CACHE_GIB=414
+MOONCAKE_LOCAL_BUFFER_GIB=4
+case "${KV_OFFLOAD_BACKEND:-}" in
+    "")
+        require_agentic_kv_offload_none
+        ;;
+    mooncake)
+        require_agentic_kv_offload_backend mooncake
+        TOTAL_CPU_DRAM_GIB=$((TOTAL_CPU_DRAM_GB * 1000000000 / 1073741824))
+        PER_RANK_GIB=$(((TOTAL_CPU_DRAM_GIB - MODEL_CHECKPOINT_PAGE_CACHE_GIB) / TP - MODEL_CPU_OFFLOAD_GB - MOONCAKE_LOCAL_BUFFER_GIB))
+        if (( PER_RANK_GIB <= 0 )); then
+            echo "Error: CPU DRAM budget is too small for checkpoint cache, model, 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",
+  "master_server_address": "127.0.0.1:$MOONCAKE_MASTER_PORT",
+  "global_segment_size": "${PER_RANK_GIB}GB",
+  "local_buffer_size": "${MOONCAKE_LOCAL_BUFFER_GIB}GB",
+  "protocol": "rdma",
+  "device_name": "",
+  "enable_offload": false
+}
+EOF
+        export MOONCAKE_CONFIG_PATH PYTHONHASHSEED=0 MC_SLICE_SIZE=1048576 MC_WORKERS_PER_CTX=4
+        export MC_ENABLE_DEST_DEVICE_AFFINITY=1
+        mooncake_master --port "$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 "$MOONCAKE_MASTER_PID"
+        OFFLOAD_ARGS=(
+            --kv-transfer-config
+            '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
+        )
+        ;;
+    *)
+        echo "Error: unsupported KV_OFFLOAD_BACKEND='$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'
+fi
+
+# The public endpoint is the router for DEP, so explicitly scrape the engine
+# endpoint. Pure TP/TEP deduplicates this URL against the automatic scrape.
+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. 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=2.78
+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\": $SYNTHETIC_ACCEPT_LEN}"
+fi
+
+# AgentX concurrency counts live session trees, not individual requests. DEP
+# splits those trees across eight data-parallel ranks; TP/TEP keeps them local.
+if [[ "$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_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 \
+    --cpu-offload-gb "$MODEL_CPU_OFFLOAD_GB" \
+    --kv-cache-dtype fp8 \
+    --attention-backend TRITON_ATTN \
+    --block-size 128 \
+    --language-model-only \
+    --enable-prefix-caching \
+    --enable-prompt-tokens-details \
+    --default-chat-template-kwargs '{"thinking_mode":"enabled"}' \
+    --max-num-seqs "$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 \
+    --safetensors-load-strategy lazy \
+    --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

configs/nvidia-master.yaml

@@ -7141,6 +7141,35 @@ minimaxm3-fp8-h100-vllm-agentic:
       - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [3, 4, 5, 6, 7, 8, 9, 10, 12, 14, 16] }
       - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [3, 4, 5, 6, 7, 8, 9, 10, 12, 14, 16] }
 
+# EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
+# minimaxm3-fp8-h100-vllm-agentic, pairing MiniMaxAI/MiniMax-M3-MXFP8 with the
+# Inferact/MiniMax-M3-EAGLE3-GQA draft head (3 speculative tokens, FLASH_ATTN
+# drafter) and pinning synthetic acceptance to the golden AL 2.78
+# (golden_al_distribution/minimaxm3_eagle3_gqa.yaml, thinking_on, K=3). Same TP8-only
+# layout and KV arms as the non-MTP entry so the spec-decode delta is readable at
+# equal concurrency, trimmed at the extreme-conc end: the draft head plus its KV
+# eat into the same HBM budget that already puts the GPU-resident cliff near
+# conc 6 on 80 GB H100s.
+minimaxm3-fp8-h100-vllm-agentic-mtp:
+  image: vllm/vllm-openai:v0.27.1
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: cluster:h100-dgxc
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    # The complete fast sweep places the GPU-resident cliff between c5 and c6.
+    # Retain the resident latency/knee curve through c5, then use Mooncake at
+    # c6/c8 to extend throughput without crossing that HBM cliff. TEP was
+    # dominated, DEP could not allocate its 1M-token KV cache on 80 GB H100s,
+    # and the vLLM-simple function screen did not justify a frontier row.
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 3, 4, 5] }
+      - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [6, 8] }
+
 minimaxm3-fp8-h200-vllm-agentic:
   image: vllm/vllm-openai:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
   model: MiniMaxAI/MiniMax-M3-MXFP8

perf-changelog.yaml

@@ -5818,3 +5818,13 @@
   description:
     - "Extend the SimpleCPUOffloadConnector grid to c8/c12/c16/c20/c24/c28/c32/c48/c64 to locate its crossover against the resident curve"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2475
+
+- config-keys:
+    - minimaxm3-fp8-h100-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Refresh H100 MiniMax-M3 MXFP8 AgentX with vLLM v0.27.1 and EAGLE3-GQA synthetic golden AL 2.78."
+    - "Use the complete fast sweep to retain TP8 c1-c5 and Mooncake TP8 c6/c8 for the strict Pareto sweep; TEP is dominated, DEP is infeasible at its required KV allocation, and vLLM-simple is not retained after its function screen."
+    - "Require vLLM server metrics and keep strict request and profile validation."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2564