feat(bench): add GLM-5.3-Flash FP8 and NVFP4 AgentX on GB300 / feat(bench): 新增 GB300 上的 GLM-5.3-Flash FP8 与 NVFP4 AgentX

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/glm5.3flash/sglang/gb300-fp4-mtp/agentic.yaml

@@ -0,0 +1,73 @@
+base:
+  schema: 2
+  name: glm5.3flash-fp4-gb300-sglang-agentic-mtp
+  identity:
+    model:
+      revision: 1d00738646cbb0c6d0a9eb403dac5f366f1473a7
+    container: {}
+    frameworks:
+      sglang: 0.5.22
+  resources:
+    gpu_type: gb300
+    gpus_per_node: 4
+  slurm:
+    time_limit: '4:00:00'
+  frontend:
+    type: sglang
+    enable_multiple_frontends: false
+  observability:
+    enabled: false
+    tachometer:
+      enabled: false
+  engine: sglang
+  roles:
+    agg:
+      nodes: 1
+      workers: 1
+      gpus: 4
+      args:
+        served-model-name: nvidia/GLM-5.3-Flash-NVFP4
+        tensor-parallel-size: 4
+        expert-parallel-size: 1
+        quantization: modelopt_fp4
+        context-length: 1048576
+        dsa-prefill-backend: trtllm
+        dsa-decode-backend: trtllm
+        kv-cache-dtype: fp8_e4m3
+        moe-runner-backend: flashinfer_trtllm
+        speculative-algorithm: EAGLE
+        speculative-num-steps: 5
+        speculative-eagle-topk: 1
+        speculative-num-draft-tokens: 6
+        reasoning-parser: auto
+        tool-call-parser: auto
+        watchdog-timeout: 1800
+        enable-metrics: true
+      env:
+        PYTHONNOUSERSITE: '1'
+        PYTHONUNBUFFERED: '1'
+        HF_HUB_OFFLINE: '1'
+        SGLANG_TIMEOUT_KEEP_ALIVE: '900'
+  health_check:
+    max_attempts: 1440
+    interval_seconds: 5
+  benchmark:
+    env:
+      AIPERF_HTTP_TCP_USER_TIMEOUT: '900000'
+      AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
+
+override_gpu_kv:
+  benchmark:
+    env:
+      KV_OFFLOADING: none
+
+override_hicache:
+  roles:
+    agg:
+      args:
+        enable-hierarchical-cache: true
+        # The 0.8 default sizes the host pool at about 1.0x the device KV pool on GB300.
+        hicache-host-memory-fraction: 0.90
+  benchmark:
+    env:
+      KV_OFFLOADING: dram

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/glm5.3flash/sglang/gb300-fp8-mtp/agentic.yaml

@@ -0,0 +1,72 @@
+base:
+  schema: 2
+  name: glm5.3flash-fp8-gb300-sglang-agentic-mtp
+  identity:
+    model:
+      revision: 46a050bec13002ad29ee795c16e61cf4fa3c4a9c
+    container: {}
+    frameworks:
+      sglang: 0.5.22
+  resources:
+    gpu_type: gb300
+    gpus_per_node: 4
+  slurm:
+    time_limit: '4:00:00'
+  frontend:
+    type: sglang
+    enable_multiple_frontends: false
+  observability:
+    enabled: false
+    tachometer:
+      enabled: false
+  engine: sglang
+  roles:
+    agg:
+      nodes: 1
+      workers: 1
+      gpus: 4
+      args:
+        served-model-name: zai-org/GLM-5.3-Flash
+        tensor-parallel-size: 4
+        expert-parallel-size: 1
+        context-length: 1048576
+        dsa-prefill-backend: trtllm
+        dsa-decode-backend: trtllm
+        kv-cache-dtype: fp8_e4m3
+        moe-runner-backend: flashinfer_trtllm
+        speculative-algorithm: EAGLE
+        speculative-num-steps: 5
+        speculative-eagle-topk: 1
+        speculative-num-draft-tokens: 6
+        reasoning-parser: auto
+        tool-call-parser: auto
+        watchdog-timeout: 1800
+        enable-metrics: true
+      env:
+        PYTHONNOUSERSITE: '1'
+        PYTHONUNBUFFERED: '1'
+        HF_HUB_OFFLINE: '1'
+        SGLANG_TIMEOUT_KEEP_ALIVE: '900'
+  health_check:
+    max_attempts: 1440
+    interval_seconds: 5
+  benchmark:
+    env:
+      AIPERF_HTTP_TCP_USER_TIMEOUT: '900000'
+      AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
+
+override_gpu_kv:
+  benchmark:
+    env:
+      KV_OFFLOADING: none
+
+override_hicache:
+  roles:
+    agg:
+      args:
+        enable-hierarchical-cache: true
+        # The 0.8 default sizes the host pool at about 1.0x the device KV pool on GB300.
+        hicache-host-memory-fraction: 0.90
+  benchmark:
+    env:
+      KV_OFFLOADING: dram

inferencex-e2e/configs/nvidia-master.yaml

@@ -5712,6 +5712,41 @@ glm5.2-fp4-gb300-dynamo-trt-agentic-disagg-mtp:
           tp: 8
           ep: 8
           dp-attn: true
+
+glm5.3flash-fp8-gb300-sglang-agentic-mtp:
+  image: lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae
+  model: zai-org/GLM-5.3-Flash
+  model-prefix: glm5.3flash
+  runner: cluster:gb300-nv
+  precision: fp8
+  framework: sglang
+  multinode: false
+  srt-recipe-dir: glm5.3flash/sglang/gb300-fp8-mtp
+  scenarios:
+    agentic-coding:
+    # HiCache auto-sizes at fraction 0.90; run 42026173061 measured 708.7-722.9 GB per node (4 ranks) of 964.7 GB.
+    - dram-utilization: 0.7495
+      search-space:
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 16, 24, 32], srt-recipe: "agentic.yaml:override_gpu_kv" }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 44, 48], srt-recipe: "agentic.yaml:override_hicache" }
+
+glm5.3flash-fp4-gb300-sglang-agentic-mtp:
+  image: lmsysorg/sglang:v0.5.22-cu130@sha256:e3a4f06001a29542fc4a8638536d9ffc5eb9c9942b08603c09b19a56eae4a1ae
+  model: nvidia/GLM-5.3-Flash-NVFP4
+  model-prefix: glm5.3flash
+  runner: cluster:gb300-nv
+  precision: fp4
+  framework: sglang
+  multinode: false
+  srt-recipe-dir: glm5.3flash/sglang/gb300-fp4-mtp
+  scenarios:
+    agentic-coding:
+    # HiCache auto-sizes at fraction 0.90; run 42026173061 measured 708.7-722.9 GB per node (4 ranks) of 964.7 GB.
+    - dram-utilization: 0.7495
+      search-space:
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 16, 24, 32], srt-recipe: "agentic.yaml:override_gpu_kv" }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 44, 48], srt-recipe: "agentic.yaml:override_hicache" }
+
 # H100 AgentX arm for DeepSeek-V4.1-Flash. H100 is not in the upstream hardware
 # table (h200, gb200, gb300, mi350x are). It needs its own script, not the
 # shared symlink: the sparse attention indexer's

inferencex-e2e/configs/runners.yaml

@@ -636,6 +636,8 @@ clusters:
         DeepSeek-R1-0528: {root: scratch, dir: DeepSeek-R1-0528}
         DeepSeek-R1-0528-NVFP4-v2: {root: scratch, dir: DeepSeek-R1-0528-NVFP4-v2}
         GLM-5.2-NVFP4: {root: scratch, dir: GLM-5.2-NVFP4}
+        GLM-5.3-Flash: {root: data-models, dir: GLM-5.3-Flash}
+        GLM-5.3-Flash-NVFP4: {root: data-models, dir: GLM-5.3-Flash-NVFP4}
         Kimi-K3.1-Pro: {root: scratch, dir: Kimi-K3.1-Pro}
         MiniMax-M3.1-Flash: {root: data-models, dir: MiniMax-M3.1-Flash}
         Muse-Spark-Open-816B-A42B-NVFP4: {root: data-models, dir: Muse-Spark-Open-816B-A42B-NVFP4}

inferencex-e2e/perf-changelog.yaml

@@ -12205,3 +12205,14 @@
     - "Retune the Muse Spark Open GB200 curve and set SGLANG_FLASHINFER_NVFP4_PER_TOKEN_ACTIVATION=0 at every concurrency; run TP4/EP4 at concurrency 1 without KV offload and at concurrency 4, 7, 10, 14, 16, 28, 36, 40 with HiCache, dropping concurrency 2, 6, 8, 20, 24 and 32."
     - "Use prefill-decode interval 64-256 and stream interval 20 at every HiCache point, shortest-prefill-first except at concurrency 7, request admission 2-6 at concurrency 4-16, 16K chunked prefill at concurrency 28-40, and at concurrency 40 mem-fraction-static 0.93, a 160 GB HiCache pool and CUDA graphs up to 48."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+
+- config-keys:
+    - glm5.3flash-fp8-gb300-sglang-agentic-mtp
+    - glm5.3flash-fp4-gb300-sglang-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add native SGLang GLM-5.3-Flash AgentX on one GB300 tray (TP4/EP1) with lmsysorg/sglang:v0.5.22-cu130, FP8 (zai-org/GLM-5.3-Flash) and NVFP4 (nvidia/GLM-5.3-Flash-NVFP4), embedded EAGLE MTP 5/1/6."
+    - "SGLang cookbook GB300 recipe: TRT-LLM DSA prefill and decode, FP8 E4M3 KV cache, flashinfer_trtllm MoE, auto reasoning and tool-call parsers; NVFP4 adds modelopt_fp4 quantization."
+    - "GPU-resident KV at concurrency 1, 4, 8, 16, 24 and 32; auto-sized HiCache host tier (--hicache-host-memory-fraction 0.90, recorded as 723 GB per node, the measured maximum) at concurrency 40, 44 and 48."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/4715