[Klaud Cold] minimaxm3-fp8-mi325x-vllm-agentic-mtp: MiniMax-M3 MI325X AgentX EAGLE3 recipe with golden-AL synthetic acceptance / MiniMax-M3 MI325X AgentX EAGLE3 配方（黄金 AL 合成接受）

benchmarks/single_node/agentic/minimaxm3_fp8_mi325x_mtp.sh

@@ -0,0 +1,184 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# MiniMax-M3 MXFP8 MI325X (gfx942) AgentX (agentic-coding) recipe with EAGLE3
+# speculative decoding — the spec-decoding=mtp variant of
+# agentic/minimaxm3_fp8_mi325x.sh. Everything outside the speculative block
+# mirrors the non-MTP agentic sibling (--block-size 128, --language-model-only, TRITON_ATTN, FP8 KV on the nightly
+# images, minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0, TP4 gmu bump), 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_mi325x_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 fixed-seq-len MI325X MTP recipe
+# still carries that runtime patch because it runs the older
+# vllm-openai-rocm:minimax-m3 bring-up image.
+#
+# 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 "$0")/../../benchmark_lib.sh"
+
+check_env_vars MODEL IMAGE TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
+
+DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3"
+
+if [[ -n "${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="$RESULT_DIR/router.log"
+mkdir -p "$RESULT_DIR"
+
+# 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
+    PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
+fi
+
+EP_ARGS=()
+if (( EP_SIZE > 1 )); then
+    EP_ARGS=(--enable-expert-parallel)
+fi
+
+KV_CACHE_ARGS=()
+if [[ "$IMAGE" == vllm/vllm-openai-rocm:nightly-* ]]; then
+    KV_CACHE_ARGS=(--kv-cache-dtype fp8)
+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))
+# 0.90, not the non-MTP sibling's 0.95/0.98: fixed_seq_len/minimaxm3_fp8_mi325x_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 every rank once the EAGLE3 head and its KV were
+# resident (run 30515793863), so the TP4 0.98 bump is dropped as well.
+GPU_MEMORY_UTILIZATION=0.90
+
+vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
+    --host 0.0.0.0 \
+    --port "$VLLM_BACKEND_PORT" \
+    "${PARALLEL_ARGS[@]}" \
+    "${EP_ARGS[@]}" \
+    --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION" \
+    --block-size 128 \
+    --language-model-only \
+    --attention-backend TRITON_ATTN \
+    --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 \
+    "${KV_CACHE_ARGS[@]}" \
+    "${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/amd-master.yaml

@@ -2230,6 +2230,38 @@ minimaxm3-fp8-mi325x-vllm-agentic:
       - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 32] }
       - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [24, 32, 36, 40, 44, 48, 52, 56, 60, 64, 72, 80, 96], router: { name: vllm-router, version: "0.1.14" } }
 
+# EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
+# minimaxm3-fp8-mi325x-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). 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 this recipe needs no in-place patch (unlike the
+# fixed-seq-len MI325X MTP entry, which still runs the older minimax-m3 image).
+# 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 non-MTP
+# entry's three host-DRAM arms are dropped rather than shipped broken. The
+# remaining TP4/TP8/DEP arms run the AgentX MTP concurrency grid: every conc-list
+# steps by at least 2 and stops hard at 16.
+minimaxm3-fp8-mi325x-vllm-agentic-mtp:
+  image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
+  model: MiniMaxAI/MiniMax-M3-MXFP8
+  model-prefix: minimaxm3
+  runner: cluster:mi325x-amds
+  precision: fp8
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16] }
+      - { tp: 4, ep: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 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, ep: 8, dp-attn: true, spec-decoding: mtp, kv-offloading: none, conc-list: [8, 12, 16], router: { name: vllm-router, version: "0.1.14" } }
+
 minimaxm3-fp4-mi355x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-dcfebf93f4eccf30f71872283331eee757915daf
   model: amd/MiniMax-M3-MXFP4

perf-changelog.yaml

@@ -5339,3 +5339,14 @@
     - "Pass --use-chat-template to the benchmark, as required for EAGLE-style speculative decoding"
     - "Validated on the node through the real launcher, every request successful: TEP4 concurrency 64 640/640 at 6895 tok/s (11% above the non-MTP arm's 6198 tok/s, mean TTFT 27.2s to 9.6s), and TEP4 concurrency 1 10/10 at 1233 tok/s (83% above the non-MTP arm's 672 tok/s, mean TPOT 12.6ms to 6.5ms)"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2312
+
+- config-keys:
+    - minimaxm3-fp8-mi325x-vllm-agentic-mtp
+  description:
+    - "Add the EAGLE3 speculative-decoding (spec-decoding: mtp) variant of minimaxm3-fp8-mi325x-vllm-agentic: MiniMax-M3 MXFP8 on MI325X (gfx942) with vLLM, agentic-coding scenario only, routed via spec-decoding=mtp to benchmarks/single_node/agentic/minimaxm3_fp8_mi325x_mtp.sh."
+    - "Speculative config: method eagle3 on the Inferact/MiniMax-M3-EAGLE3 draft head, num_speculative_tokens 3, with no attention_backend override for the drafter -- the FlashInfer page-size-128-requires-GQA/MQA limitation that forces FLASH_ATTN on CUDA is FlashInfer-specific and this server runs TRITON_ATTN throughout (same reasoning as fixed_seq_len/minimaxm3_fp8_mi325x_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, unlike fixed_seq_len/minimaxm3_fp8_mi325x_mtp.sh: this recipe runs the agentic sibling's vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914 (upstream 2026-06-23), which postdates vllm-project/vllm#45546 (merged 2026-06-14); the fixed-seq-len entry still needs the runtime patch because it pins the older vllm-openai-rocm:minimax-m3 bring-up image."
+    - "Serve shape is otherwise identical to the non-MTP agentic sibling (--block-size 128, --language-model-only, --attention-backend TRITON_ATTN, FP8 KV on nightly-* images, gpu-memory-utilization 0.95 with the 0.98 TP4 bump, minimax_m3 parsers, VLLM_USE_BREAKABLE_CUDAGRAPH=0, max-num-seqs 2*CONC, vllm-router for DP-attention). Cudagraph capture stays at the sibling default rather than the CUDA recipes' MAX_NUM_SEQS*(1+3) token ceiling, matching the merged ROCm MiniMax-M3 MTP recipes."
+    - "Search space is GPU-resident only -- TP4, TP4/EP4, TP8, TP8/EP8 on the AgentX MTP concurrency grid [1, 4, 8, 12, 16] and TP8/DEP8 dp-attn on [8, 12, 16], steps of at least 2 with a hard stop at conc 16. The non-MTP entry's three Mooncake host-DRAM arms are dropped: Mooncake is not used on AMD. Its HIP-built transfer engine runs over TCP here and the store collapses mid-replay (ServerSession::readBody failed ... Connection reset by peer until AIPerf aborts; observed on the MI300X sibling, run 30517360090), so a dram arm would measure the transport, not the engine."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2426