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"
+
+# HiSparse decode keeps full KV 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: fp8_e4m3
+      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: fp8_e4m3
+      dsa-decode-backend: flashmla_kv
+      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"

benchmarks/multi_node/srt-slurm-recipes/sglang/glm5.2/agentic/disagg-h200-1p2d-pcp8-tp8-dp8-hisparse.yaml

@@ -0,0 +1,130 @@
+name: "disagg-h200-1p2d-pcp8-tp8-dp8-hisparse"
+
+# HiSparse decode keeps full KV 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: 2
+  prefill_workers: 1
+  decode_workers: 2
+  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: fp8_e4m3
+      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: fp8_e4m3
+      dsa-decode-backend: flashmla_kv
+      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

@@ -8010,6 +8010,58 @@ 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
+      - spec-decoding: none
+        conc-list: [4]
+        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-h200-1p2d-pcp8-tp8-dp8-hisparse.yaml"
+        decode:
+          num-worker: 2
+          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

@@ -5060,3 +5060,15 @@
     - "Re-pin VLLM_ROUTER_IMAGE to vllm/vllm-router:nightly-20260716-1fbcde7 (previous nightly-20260629-e667ebb was garbage-collected from Docker Hub)"
     - "Exclude known-bad nodes mia1-p01-g09,g14 from the disagg node pool"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2301
+
+- 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 PCP8 prefill and TP8/DPA decode, covering the complete 1P1D concurrency curve [1, 2, 3, 4, 5, 6, 7, 8] plus the non-dominated 1P2D concurrency-4 scale point"
+    - "Use FP8 KV cache, flashmla_kv, 32K chunked prefill, and SGLang HiSparse top_k=2048/device_buffer_size=4096 to retain full KV in pinned host DRAM while a sparse decode working set remains in HBM"
+    - "Literal 1.04M-token canaries and GSM8K gates passed on the selected HiSparse configuration; TP8/DCP4 and TP8/DCP8 decode failed in SGLang attn_mqa_for_dcp_decode, while prefill-only HiCache, BF16 KV, 64K chunking, and alternate HiSparse buffers were slower or failed"
+    - "Unofficial full curves 29881640438 and 29882624421 completed with zero benchmark errors; the selected 1P1D c4 point reached 546.33 total tok/s/GPU at 11.05s mean TTFT, while scale run 30104463995 reached 567.82 total tok/s/GPU at 3.89s mean TTFT on 1P2D c4"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2275
+