Add Qwen3.5 FP4 B200 AgentX MTP

benchmarks/single_node/agentic/qwen3.5_fp4_b200_sglang_mtp.sh

@@ -0,0 +1,173 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# AgentX trace replay for Qwen3.5-397B-A17B NVFP4 on B200 with SGLang
+# native NEXTN MTP. Throughput uses the committed golden synthetic AL; evals
+# retain real target-model verification.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL TP CONC EP_SIZE KV_OFFLOADING \
+    TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
+
+SCHEDULER_RECV_INTERVAL=${SCHEDULER_RECV_INTERVAL:-10}
+
+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="$MODEL"
+fi
+nvidia-smi
+
+export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_062126_256k
+resolve_trace_source
+install_agentic_deps
+
+SERVER_LOG="$RESULT_DIR/server.log"
+mkdir -p "$RESULT_DIR"
+
+CACHE_ARGS=()
+if require_agentic_kv_offload_backend hicache; then
+    REQUESTED_HICACHE_TOTAL_GB="${HICACHE_TOTAL_CPU_DRAM_GB:-$TOTAL_CPU_DRAM_GB}"
+    if [ "$REQUESTED_HICACHE_TOTAL_GB" -gt "$TOTAL_CPU_DRAM_GB" ]; then
+        echo "Error: requested HiCache pool ${REQUESTED_HICACHE_TOTAL_GB} GB exceeds configured capacity ${TOTAL_CPU_DRAM_GB} GB" >&2
+        exit 1
+    fi
+    TOTAL_CPU_DRAM_GB="$REQUESTED_HICACHE_TOTAL_GB"
+    # SGLang applies --hicache-size independently to Qwen's target KV and
+    # Mamba pools. Native NEXTN also creates a draft KV pool with the same
+    # slot count; its one attention layer adds 1/15 of the target KV bytes.
+    # Reserve 1 GB/rank for page alignment and enforce H * 31/15 per rank.
+    HICACHE_ALIGNMENT_RESERVE_GB=$TP
+    HICACHE_USABLE_TOTAL_GB=$((TOTAL_CPU_DRAM_GB - HICACHE_ALIGNMENT_RESERVE_GB))
+    if [ "$HICACHE_USABLE_TOTAL_GB" -lt 1 ]; then
+        echo "Error: insufficient DRAM after HiCache alignment reserve" >&2
+        exit 1
+    fi
+    MAX_HICACHE_SIZE_GB=$((HICACHE_USABLE_TOTAL_GB * 15 / TP / 31))
+    HICACHE_SIZE_GB="${HICACHE_SIZE_GB:-$MAX_HICACHE_SIZE_GB}"
+    if [ "$HICACHE_SIZE_GB" -lt 1 ] || [ "$HICACHE_SIZE_GB" -gt "$MAX_HICACHE_SIZE_GB" ]; then
+        echo "Error: HICACHE_SIZE_GB=$HICACHE_SIZE_GB outside 1..$MAX_HICACHE_SIZE_GB" >&2
+        exit 1
+    fi
+    PROJECTED_HICACHE_TOTAL_GB=$(((HICACHE_SIZE_GB * TP * 31 + 14) / 15 + HICACHE_ALIGNMENT_RESERVE_GB))
+    if [ "$PROJECTED_HICACHE_TOTAL_GB" -gt "$TOTAL_CPU_DRAM_GB" ]; then
+        echo "Error: projected HiCache use ${PROJECTED_HICACHE_TOTAL_GB} GB exceeds configured capacity ${TOTAL_CPU_DRAM_GB} GB" >&2
+        exit 1
+    fi
+    echo "HiCache CPU pools: ${HICACHE_SIZE_GB} GB target + Mamba + 1/15 draft per rank across TP=${TP}; projected node total ${PROJECTED_HICACHE_TOTAL_GB} GB <= ${TOTAL_CPU_DRAM_GB} GB"
+    CACHE_ARGS=(
+        --page-size 64
+        --enable-hierarchical-cache
+        --hicache-size "$HICACHE_SIZE_GB"
+        --hicache-io-backend kernel
+        --hicache-mem-layout page_first
+        --hicache-write-policy write_through_selective
+    )
+fi
+
+PARALLEL_ARGS=(
+    --tp "$TP"
+    --dp 1
+    --ep-size "$EP_SIZE"
+)
+
+# TP4 needs parallel tokenization to keep 256k AgentX warmups below the client
+# request timeout. Keep TP2 on SGLang's single-worker default: multi-tokenizer
+# startup races with the TP2 HiCache shared-memory initialization path.
+TOKENIZER_ARGS=()
+if [ "$TP" -ge 4 ]; then
+    TOKENIZER_ARGS=(--tokenizer-worker-num 6)
+fi
+
+# AgentX concurrency counts live session trees rather than individual HTTP
+# requests. Leave room for subagent fan-out and avoid spending HBM on graphs
+# above the batch sizes that remain useful for this long-context workload.
+MAX_RUNNING_REQUESTS=$((2 * CONC))
+CUDA_GRAPH_MAX_BS="$CONC"
+[ "$CUDA_GRAPH_MAX_BS" -gt 64 ] && CUDA_GRAPH_MAX_BS=64
+
+export TORCH_CUDA_ARCH_LIST="10.0"
+export PYTHONNOUSERSITE=1
+export NCCL_NVLS_ENABLE=1
+export SGL_ENABLE_JIT_DEEPGEMM=false
+export SGLANG_ENABLE_FLASHINFER_GEMM=true
+
+if [ "${EVAL_ONLY:-false}" != "true" ]; then
+    export SGLANG_SIMULATE_ACC_LEN=3.39
+    export SGLANG_SIMULATE_ACC_METHOD=match-expected
+    export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
+fi
+
+SGLANG_CMD=(
+    python3 -m sglang.launch_server
+    --model-path "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --trust-remote-code
+    "${PARALLEL_ARGS[@]}"
+    --enable-symm-mem
+    --quantization modelopt_fp4
+    --fp4-gemm-backend flashinfer_cutlass
+    --kv-cache-dtype fp8_e4m3
+    --mamba-ssm-dtype bfloat16
+    --attention-backend trtllm_mha
+    --moe-runner-backend flashinfer_trtllm
+    --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
+    --max-prefill-tokens 16384
+    --chunked-prefill-size 16384
+    --mem-fraction-static 0.80
+    --stream-interval 50
+    --scheduler-recv-interval "$SCHEDULER_RECV_INTERVAL"
+    "${TOKENIZER_ARGS[@]}"
+    --tokenizer-path "$MODEL"
+    --reasoning-parser qwen3
+    --tool-call-parser qwen3_coder
+    --speculative-algorithm NEXTN
+    --speculative-num-steps 3
+    --speculative-eagle-topk 1
+    --speculative-num-draft-tokens 4
+    --enable-metrics
+    --enable-cache-report
+    "${CACHE_ARGS[@]}"
+)
+
+printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
+"${SGLANG_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+
+capture_cache_metrics() {
+    {
+        echo "=== SGLang cache metrics snapshot $(date --iso-8601=seconds) ==="
+        curl -fsS "http://localhost:$PORT/metrics" 2>/dev/null \
+            | grep -E '^(sglang:(cache_hit_rate|cached_tokens_total|prompt_tokens_total|hicache_host_used_tokens|hicache_host_total_tokens|token_usage|num_requests_running|num_requests_waiting))' \
+            || true
+        echo "============================================================"
+    } >> "$SERVER_LOG"
+}
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+capture_cache_metrics
+trap capture_cache_metrics EXIT
+
+if [ "${EVAL_ONLY:-false}" = "true" ]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --trace-idle-gap-cap-seconds 300"
+    REPLAY_CMD+=" --server-metrics http://localhost:$PORT/metrics"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi

configs/nvidia-master.yaml

@@ -8478,6 +8478,37 @@ minimaxm3-fp8-h200-vllm-agentic:
       - { tp: 8, ep: 8, kv-offloading: none, conc-list: [2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
       - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
 
+# Qwen3.5 has 15 full-attention layers with two 256-wide KV heads, so its
+# target FP8 KV is 15,360 bytes/token before the native MTP layer. TP4
+# replicates the two KV heads and measured 30.5 KB/token node-wide; TP2
+# measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
+# p75 ISL, TP2 no-offload is already GPU-KV-bound by c18; measured c20
+# HiCache improves both its throughput and latency, so offload takes over
+# after c14 rather than repeating dominated no-offload points. TP2 c40 is
+# capped to 25 running requests by the available HBM and matches c32 throughput
+# with 15-second median TTFT, so the HiCache arm ends at c32.
+# TP4 retains 83.9% prefix hits at c64. No-offload c66 and c68 are already
+# throughput/latency dominated, while c70 collapses to 6% hits and 96% KV
+# usage, so HiCache takes over at c64 and densely covers the boundary through
+# c76. At c80 GPU KV reaches 100% with 19-second median TTFT, and c88 falls to
+# 10% prefix hits, so the dominated upper tail is excluded.
+qwen3.5-fp4-b200-sglang-agentic-mtp:
+  image: lmsysorg/sglang:v0.5.16-cu130
+  model: nvidia/Qwen3.5-397B-A17B-NVFP4
+  model-prefix: qwen3.5
+  runner: cluster:b200-dgxc
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 60, 62, 64] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 66, 68, 70, 72, 76] }
+      - { tp: 2, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 14] }
+      - { tp: 2, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 18, 20, 22, 24, 28, 32] }
+
 dsv4-fp4-b200-sglang-agentic-hicache:
   image: lmsysorg/sglang:nightly-dev-cu13-20260707-b4155233
   model: deepseek-ai/DeepSeek-V4-Pro

perf-changelog.yaml

@@ -5339,3 +5339,12 @@
     - "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:
+    - qwen3.5-fp4-b200-sglang-agentic-mtp
+  description:
+    - "Add Qwen3.5-397B-A17B NVFP4 AgentX benchmark on B200 with SGLang native NEXTN MTP"
+    - "Use the 256k trace dataset and golden synthetic acceptance length 3.39"
+    - "Pin AIPerf trace idle-gap support and cap per-trace idle gaps at 300 seconds"
+    - "Image: lmsysorg/sglang:v0.5.16-cu130"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2420

runners/launch_b200-dgxc.sh

@@ -11,8 +11,8 @@ set -x
 # portability, but we resolve to /lustre/fsw/models/* here to avoid repeated
 # downloading on every dgxc node. Runs for both single-node and multinode
 # launches.
-# NOTE: per-node /raid/models/* would be faster but is only populated on a
-# subset of dgxc nodes today, so we use Lustre for reliability.
+# Single-node Qwen3.5 FP4 runs probe the allocated node for a complete local
+# /raid checkpoint below and use this shared path only as the fallback.
 if [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp4" ]]; then
     export MODEL_PATH="/scratch/fsw/models/DeepSeek-R1-0528-NVFP4-v2"
     export SRT_SLURM_MODEL_PREFIX="dsr1"
@@ -437,10 +437,6 @@ EOF
 else
 
     SQUASH_FILE="/home/sa-shared/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
-    # Point the bench script at the local MODEL_PATH resolved above instead of
-    # pulling from the HF hub cache. Bench scripts skip `hf download` when
-    # MODEL is a local path.
-    export MODEL="$MODEL_PATH"
     FRAMEWORK_SUFFIX=$([[ "$FRAMEWORK" == "trt" ]] && printf '_trt' || printf '')
     SPEC_SUFFIX=$([[ "$SPEC_DECODING" == "mtp" ]] && printf '_mtp' || printf '')
     # Prefer a framework-tagged script (e.g. dsv4_fp4_b200_vllm.sh) so models
@@ -474,6 +470,41 @@ else
     salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
 
+    # Qwen3.5 FP4 may be staged on local NVMe on a DGXC node.
+    # Resolve this only after Slurm assigns the node; checking on the login
+    # host would select the wrong filesystem. Validate every shard named by
+    # the safetensor index so a partial staging directory is never selected.
+    if [[ "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp4" ]]; then
+        LOCAL_MODEL_PATH="/raid/models/Qwen3.5-397B-A17B-NVFP4"
+        if srun --jobid="$JOB_ID" --nodes=1 --ntasks=1 \
+            python3 - "$LOCAL_MODEL_PATH" <<'PY'; then
+import json
+from pathlib import Path
+import sys
+
+model_path = Path(sys.argv[1])
+index_path = model_path / "model.safetensors.index.json"
+if not (model_path / "config.json").is_file() or not index_path.is_file():
+    raise SystemExit(1)
+
+with index_path.open() as index_file:
+    shards = set(json.load(index_file).get("weight_map", {}).values())
+if not shards or any(not (model_path / shard).is_file() for shard in shards):
+    raise SystemExit(1)
+PY
+            MODEL_PATH="$LOCAL_MODEL_PATH"
+            export MODEL_PATH
+            echo "Using node-local NVMe checkpoint: $MODEL_PATH"
+        else
+            echo "Node-local Qwen3.5 FP4 checkpoint unavailable; using shared checkpoint: $MODEL_PATH"
+        fi
+    fi
+
+    # Point the bench script at the resolved local/shared MODEL_PATH instead of
+    # pulling from the HF hub cache. Bench scripts skip `hf download` when
+    # MODEL is a local path.
+    export MODEL="$MODEL_PATH"
+
     # Use flock to serialize concurrent imports to the same squash file
     # Override ENROOT_CACHE_PATH to avoid permission issues with system-wide cache on worker nodes
     srun --jobid=$JOB_ID bash -c "