Add GLM-5.2 full-context H200 AgentX with SGLang PCP and HiSparse / 新增 GLM-5.2 H200 全上下文 AgentX SGLang PCP 与 HiSparse 配置

benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.2/agentic/disagg-h200-1p1d-pcp8-tp8-dp8-hisparse.yaml

@@ -0,0 +1,130 @@
+name: "disagg-h200-1p1d-pcp8-tp8-dp8-hisparse"
+
+# Experimental HiSparse decode. This is not HiCache: full KV lives in pinned
+# host memory while a fixed sparse working set remains in HBM.
+
+model:
+  path: "hf:zai-org/GLM-5.2-FP8"
+  container: "lmsysorg/sglang:nightly-dev-cu13-20260711-7de33ce8"
+  precision: "fp8"
+
+dynamo:
+  install: true
+  wheel: "1.3.0.dev1"
+
+slurm:
+  time_limit: "8:00:00"
+
+resources:
+  gpu_type: h200
+  gpus_per_node: 8
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: true
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Correlation-ID
+  env:
+    DYN_ROUTER_TEMPERATURE: "10000000"
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+  args:
+    router-mode: "kv"
+    router-reset-states: true
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+backend:
+  type: sglang
+  prefill_environment: &worker_environment
+    HF_HOME: "/hf_hub_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    PYTHONUNBUFFERED: "1"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+    SGLANG_OPT_USE_TOPK_V2: "1"
+    SGLANG_TIMEOUT_KEEP_ALIVE: "900"
+  decode_environment: *worker_environment
+
+  sglang_config:
+    prefill:
+      host: 0.0.0.0
+      model-path: /model/
+      served-model-name: zai-org/GLM-5.2-FP8
+      trust-remote-code: true
+      tool-call-parser: glm47
+      reasoning-parser: glm45
+      tp-size: 8
+      ep-size: 1
+      attn-cp-size: 8
+      enable-prefill-cp: true
+      cp-strategy: interleave
+      enable-dsa-cache-layer-split: true
+      disaggregation-transfer-backend: mooncake
+      disaggregation-mode: prefill
+      kv-cache-dtype: bfloat16
+      context-length: 1048576
+      max-total-tokens: 1048576
+      chunked-prefill-size: 32768
+      mem-fraction-static: 0.85
+      max-running-requests: 32
+      watchdog-timeout: 86400
+      stream-interval: 60
+      enable-metrics: true
+
+    decode:
+      host: 0.0.0.0
+      model-path: /model/
+      served-model-name: zai-org/GLM-5.2-FP8
+      trust-remote-code: true
+      tool-call-parser: glm47
+      reasoning-parser: glm45
+      tp-size: 8
+      dp-size: 8
+      ep-size: 1
+      enable-dp-attention: true
+      disaggregation-transfer-backend: mooncake
+      disaggregation-mode: decode
+      kv-cache-dtype: bfloat16
+      dsa-decode-backend: flashmla_sparse
+      context-length: 1048576
+      max-total-tokens: 1048576
+      mem-fraction-static: 0.85
+      max-running-requests: 200
+      page-size: 64
+      disable-radix-cache: true
+      enable-hisparse: true
+      hisparse-config: '{"top_k":2048,"device_buffer_size":4096,"host_to_device_ratio":5}'
+      watchdog-timeout: 86400
+      stream-interval: 60
+      enable-metrics: true
+
+sbatch_directives:
+  mem: "0"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "1"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"

configs/nvidia-master.yaml

@@ -7828,6 +7828,41 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           tp: 16
           ep: 16
           dp-attn: true
+
+# GLM-5.2 FP8 full-context AgentX on H200. PCP8 prefill processes
+# the 1M context while TP8/DPA decode uses HiSparse-backed host DRAM.
+glm5.2-fp8-h200-dynamo-sglang-agentic-hisparse-pareto:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260711-7de33ce8
+  model: zai-org/GLM-5.2-FP8
+  model-prefix: glm5.2
+  runner: cluster:h200-dgxc
+  precision: fp8
+  framework: dynamo-sglang
+  router: { name: dynamo-router, version: "1.3.0.dev1" }
+  kv-p2p-transfer: mooncake
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - spec-decoding: none
+        conc-list: [1, 2, 3, 4, 5, 6, 7, 8]
+        kv-offloading: dram
+        kv-offload-backend: { name: hisparse }
+        prefill:
+          num-worker: 1
+          tp: 1
+          pcp-size: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/glm5.2/agentic/disagg-h200-1p1d-pcp8-tp8-dp8-hisparse.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: true
 qwen3.5-fp8-gb300-dynamo-sglang:
   image: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
   model: Qwen/Qwen3.5-397B-A17B-FP8

perf-changelog.yaml

@@ -5011,3 +5011,13 @@
     - "dsv4-fp4-b300-vllm-agentic-mtp: MTP twins (--speculative-config {\"method\":\"mtp\",\"num_speculative_tokens\":3}) of the aggregate arms plus a new TP8 GPU-resident arm -- TP8 GPU-resident [1,2,4,6,8], TP4 GPU-resident [1,2,4,6,8,12,16,20], TP4 SimpleCPU [20,24,28,32,36,40], DEP4 [32,40,48,56], DEP8 [64,96,112,128,144,160,176,192,224]; routed via spec-decoding=mtp to dsv4_fp4_b300_vllm_mtp.sh (FULL_DECODE_ONLY cudagraph capture sizes in TOKENS = num_seqs*4)."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2258
 
+- config-keys:
+    - glm5.2-fp8-h200-dynamo-sglang-agentic-hisparse-pareto
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add GLM-5.2 FP8 full-context H200 AgentX with one PCP8 prefill worker and one TP8/DPA decode worker across two H200 nodes"
+    - "Use SGLang HiSparse to retain the full KV cache in pinned host DRAM while a sparse decode working set remains in HBM; this replaces the TP16 HiCache design that could not restore literal 1M-token prefixes reliably"
+    - "Run the complete concurrency curve [1, 2, 3, 4, 5, 6, 7, 8]; workflow run 29820102138 completed all eight points with zero benchmark errors, peaking at 312.63 total tok/s/GPU at concurrency 4"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2275
+

runners/launch_h200-dgxc-slurm.sh

@@ -9,13 +9,30 @@ set -x
 
 if [[ "$IS_MULTINODE" == "true" ]]; then
 
+    # Resolve InferenceX-owned recipes before entering the cloned srt-slurm
+    # checkout. AgentX also needs persistent host caches on every worker.
+    CONFIG_PATH="${CONFIG_FILE%%:*}"
+    LOCAL_CONFIG_FILE=""
+    LOCAL_CONFIG_PATH="$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/${CONFIG_PATH#recipes/}"
+    if [[ -n "${CONFIG_FILE:-}" && -f "$LOCAL_CONFIG_PATH" ]]; then
+        LOCAL_CONFIG_FILE="$LOCAL_CONFIG_PATH"
+    fi
+    AIPERF_MMAP_CACHE_HOST_PATH="/home/sa-shared/gharunners/ai-perf-cache"
+    HF_HUB_CACHE_HOST_PATH="/models/gharunners/hf-hub-cache"
+
     # MODEL_PATH: Override with pre-downloaded paths on H200 runner
     # The yaml files specify HuggingFace model IDs for portability, but we use
     # local paths to avoid repeated downloading on the shared H200 cluster.
     if [[ $FRAMEWORK == "dynamo-sglang" ]]; then
         if [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp8" ]]; then
             export MODEL_PATH="/models/DeepSeek-R1-0528"
             export SRT_SLURM_MODEL_PREFIX="dsr1-fp8"
+        elif [[ $MODEL_PREFIX == "glm5.2" && $PRECISION == "fp8" ]]; then
+            export MODEL_PATH="${GLM52_FP8_MODEL_PATH:-/models/GLM-5.2-FP8}"
+            if [[ ! -d "$MODEL_PATH" ]]; then
+                export MODEL_PATH="hf:zai-org/GLM-5.2-FP8"
+            fi
+            export SRT_SLURM_MODEL_PREFIX="glm5.2-fp8"
         else
             echo "Unsupported model prefix/precision for dynamo-sglang: $MODEL_PREFIX/$PRECISION"
             exit 1
@@ -42,7 +59,13 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
     fi
 
     # TODO(CJQ): make first class upon srt-slurm upstream refactor
-    if [[ "$IS_AGENTIC" == "1" ]]; then
+    if [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "glm5.2" ]]; then
+        # v1.0.10 contains the custom benchmark hook, session-affinity
+        # frontend, and large-request nginx fix used by SGLang AgentX recipes.
+        git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+        cd "$SRT_REPO_DIR"
+        git checkout v1.0.10
+    elif [[ "$IS_AGENTIC" == "1" ]]; then
         git clone --branch cam/sa-submission-q2-2026 --single-branch https://github.com/cquil11/srt-slurm-nv.git "$SRT_REPO_DIR"
         cd "$SRT_REPO_DIR"
     else
@@ -71,14 +94,36 @@ if [[ "$IS_MULTINODE" == "true" ]]; then
 
     if [[ $FRAMEWORK == "dynamo-sglang" ]]; then
         # SGLang container mapping
-        SQUASH_FILE="/data/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/+/g').sqsh"
+        if [[ $MODEL_PREFIX == "glm5.2" ]]; then
+            SQUASH_FILE="/data/gharunners/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
+        else
+            SQUASH_FILE="/data/containers/$(echo "$IMAGE" | sed 's/[\/:@#]/+/g').sqsh"
+        fi
         CONTAINER_KEY="$IMAGE"
     elif [[ $FRAMEWORK == "dynamo-trt" ]]; then
         # TRT-LLM container mapping - convert IMAGE to srt-slurm format (nvcr.io/ -> nvcr.io#)
         CONTAINER_KEY=$(echo "$IMAGE" | sed 's|nvcr.io/|nvcr.io#|')
         SQUASH_FILE="/data/containers/$(echo "$IMAGE" | sed 's|nvcr.io/||' | sed 's/[\/:@#]/+/g').sqsh"
     fi
 
+    if ! unsquashfs -l "$SQUASH_FILE" >/dev/null 2>&1; then
+        DOCKER_IMAGE=$(echo "$IMAGE" | sed 's/#/\//g')
+        LOCK_FILE="${SQUASH_FILE}.lock"
+        srun --partition="$SLURM_PARTITION" --account="$SLURM_ACCOUNT" \
+            --nodes=1 --ntasks=1 --time=30 --job-name="$RUNNER_NAME" \
+            bash -c "
+                exec 9>\"$LOCK_FILE\"
+                flock -w 1800 9 || { echo 'Failed to acquire $LOCK_FILE'; exit 1; }
+                if unsquashfs -l \"$SQUASH_FILE\" >/dev/null 2>&1; then
+                    exit 0
+                fi
+                rm -f \"$SQUASH_FILE\"
+                export ENROOT_CACHE_PATH=\${HOME}/.cache/enroot
+                mkdir -p \"\$ENROOT_CACHE_PATH\"
+                enroot import -o \"$SQUASH_FILE\" docker://$DOCKER_IMAGE
+            "
+    fi
+
     export ISL="$ISL"
     export OSL="$OSL"
     export EVAL_ONLY="${EVAL_ONLY:-false}"
@@ -98,6 +143,10 @@ gpus_per_node: 8
 network_interface: ""
 # Path to srtctl repo root (where the configs live)
 srtctl_root: "${SRTCTL_ROOT}"
+# Persistent AgentX dataset caches mounted into every worker container.
+default_mounts:
+  "${AIPERF_MMAP_CACHE_HOST_PATH}": "/aiperf_mmap_cache"
+  "${HF_HUB_CACHE_HOST_PATH}": "/hf_hub_cache"
 # Model path aliases
 model_paths:
   "${SRT_SLURM_MODEL_PREFIX}": "${MODEL_PATH}"
@@ -120,6 +169,13 @@ EOF
     echo "Running make setup..."
     make setup ARCH=x86_64
 
+    # make setup refreshes the srt-slurm recipe tree. Overlay the checked-in
+    # InferenceX recipe afterwards so it remains available to srtctl apply.
+    if [[ -n "$LOCAL_CONFIG_FILE" ]]; then
+        mkdir -p "$(dirname "$CONFIG_PATH")"
+        cp "$LOCAL_CONFIG_FILE" "$CONFIG_PATH"
+    fi
+
     # Export eval-related env vars for srt-slurm post-benchmark eval
     export INFMAX_WORKSPACE="$GITHUB_WORKSPACE"
 
@@ -132,10 +188,14 @@ EOF
     fi
 
     # Override the job name in the config file with the runner name
-    sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_FILE"
-    sed -i '/^health_check:/,/^[^ ]/{ /^health_check:/d; /^  /d; }' "${CONFIG_FILE%%:*}"
-    printf '\nhealth_check:\n  max_attempts: 720\n  interval_seconds: 10\n' >> "${CONFIG_FILE%%:*}"
-    SRTCTL_OUTPUT=$(srtctl apply -f "$CONFIG_FILE" --tags "h200,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)" 2>&1)
+    sed -i "s/^name:.*/name: \"${RUNNER_NAME}\"/" "$CONFIG_PATH"
+    sed -i '/^health_check:/,/^[^ ]/{ /^health_check:/d; /^  /d; }' "$CONFIG_PATH"
+    printf '\nhealth_check:\n  max_attempts: 720\n  interval_seconds: 10\n' >> "$CONFIG_PATH"
+    WORKLOAD_TAG="${ISL}x${OSL}"
+    if [[ "$IS_AGENTIC" == "1" ]]; then
+        WORKLOAD_TAG="agentic"
+    fi
+    SRTCTL_OUTPUT=$(srtctl apply -f "$CONFIG_FILE" --tags "h200,${MODEL_PREFIX},${PRECISION},${WORKLOAD_TAG},infmax-$(date +%Y%m%d)" 2>&1)
     echo "$SRTCTL_OUTPUT"
 
     # Extract JOB_ID from srtctl output