feat(config): add Qwen3.8 GB300 AgentX / feat(config): 新增 Qwen3.8 GB300 AgentX 配置

inferencex-e2e/benchmarks/single_node/srt-slurm-recipes/qwen3.8next/sglang/gb300-fp4-mtp/agentic.yaml

@@ -0,0 +1,156 @@
+# Stock ARM64 SGLang, one serving GPU per point. No source/package overlay.
+# Size the host cache against the public GB300 profile's per-worker DRAM allowance.
+# The matrix floors the allowance to 225 GB; H224 leaves pool-alignment headroom.
+base:
+  schema: 2
+  name: qwen3.8next-fp4-gb300-sglang-agentic-stock-v0521
+  identity:
+    model:
+      repo: nvidia/Qwen3.8-Flash-Next-NVFP4
+      revision: fc694b54fb0174e0913e6adf86691ef85a4ead47
+    container:
+      image: lmsysorg/sglang:v0.5.21-cu130@sha256:1f176b3a1dd1205d2a462ab80ea9fc17769a32e53cb6f3c690efa2dc2c384977
+  model:
+    path: hf:nvidia/Qwen3.8-Flash-Next-NVFP4
+    container: lmsysorg/sglang:v0.5.21-cu130@sha256:1f176b3a1dd1205d2a462ab80ea9fc17769a32e53cb6f3c690efa2dc2c384977
+    precision: fp4
+  resources:
+    gpu_type: gb300
+    gpus_per_node: 4
+  slurm:
+    time_limit: '6:00:00'
+  health_check:
+    max_attempts: 1440
+    interval_seconds: 5
+  frontend:
+    type: sglang
+    enable_multiple_frontends: false
+  engine: sglang
+  roles:
+    agg:
+      nodes: 1
+      workers: 1
+      gpus: 1
+      env:
+        PYTHONNOUSERSITE: '1'
+        PYTHONUNBUFFERED: '1'
+        NO_COLOR: '1'
+        SGLANG_TIMEOUT_KEEP_ALIVE: '1800'
+        # Explicitly disable the prohibited draft precision conversion path.
+        SGLANG_NVFP4_CKPT_FP8_NEXTN_MOE: '0'
+      args:
+        served-model-name: nvidia/Qwen3.8-Flash-Next-NVFP4
+        revision: fc694b54fb0174e0913e6adf86691ef85a4ead47
+        speculative-draft-model-revision: fc694b54fb0174e0913e6adf86691ef85a4ead47
+        trust-remote-code: true
+        quantization: modelopt_mixed
+        context-length: 262144
+        tensor-parallel-size: 1
+        pipeline-parallel-size: 1
+        data-parallel-size: 1
+        expert-parallel-size: 1
+        attention-backend: trtllm_mha
+        linear-attn-prefill-backend: flashinfer
+        linear-attn-decode-backend: flashinfer
+        linear-attn-verify-backend: triton
+        mamba-ssm-dtype: bfloat16
+        mamba-radix-cache-strategy: extra_buffer_lazy
+        mamba-track-interval: 256
+        max-mamba-cache-size: 480
+        mem-fraction-static: 0.88
+        max-running-requests: 96
+        page-size: 64
+        chunked-prefill-size: 16384
+        max-prefill-tokens: 16384
+        moe-runner-backend: flashinfer_trtllm
+        moe-a2a-backend: none
+        disable-shared-experts-fusion: true
+        speculative-algorithm: NEXTN
+        speculative-num-steps: 3
+        speculative-eagle-topk: 1
+        speculative-num-draft-tokens: 4
+        speculative-moe-runner-backend: flashinfer_trtllm
+        # The native connector injects model-specific golden AL for throughput;
+        # quality evaluations use real verification. No manual AL override.
+        enable-linear-replayssm-spec: true
+        cuda-graph-backend-prefill: disabled
+        cuda-graph-backend-decode: full
+        cuda-graph-max-bs-decode: 96
+        cuda-graph-bs-decode: [1, 2, 4, 8, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96]
+        kv-cache-dtype: bfloat16
+        speculative-draft-kv-cache-dtype: bfloat16
+        enable-hierarchical-cache: true
+        # One combined packed target/draft KV + Mamba host-pool budget.
+        # Verify actual allocations, not whole-node Slurm memory reservation.
+        hicache-size: 224
+        hicache-io-backend: kernel
+        hicache-mem-layout: page_first
+        hicache-write-policy: write_back
+        ple-offload-embedding: true
+        ple-offload-backend: pinned
+        enable-metrics: true
+        enable-cache-report: true
+        reasoning-parser: auto
+        tokenizer-worker-num: 1
+        scheduler-recv-interval: 1
+        decode-log-interval: 10
+        stream-interval: 20
+        watchdog-timeout: 1800
+        random-seed: 35
+  sbatch_directives:
+    mem: '0'
+    exclusive: ''
+  srun_options:
+    mem: '0'
+    container-remap-root: ''
+  benchmark:
+    type: custom
+    command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+    env:
+      MODEL: nvidia/Qwen3.8-Flash-Next-NVFP4
+      AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'sglang:'
+      WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126_256k
+      # Native single-node binding verifies these against the generated point.
+      TOTAL_CPU_DRAM_GB: '225'
+      KV_OFFLOADING: dram
+
+# Only client concurrency changes; serving/admission/graph/cache settings stay fixed.
+override_c1:
+  benchmark:
+    env:
+      CONC: '1'
+
+override_c2:
+  benchmark:
+    env:
+      CONC: '2'
+
+override_c4:
+  benchmark:
+    env:
+      CONC: '4'
+
+override_c8:
+  benchmark:
+    env:
+      CONC: '8'
+
+override_c16:
+  benchmark:
+    env:
+      CONC: '16'
+
+override_c32:
+  benchmark:
+    env:
+      CONC: '32'
+
+override_c48:
+  benchmark:
+    env:
+      CONC: '48'
+
+override_c64:
+  benchmark:
+    env:
+      CONC: '64'

inferencex-e2e/configs/nvidia-master.yaml

@@ -9374,3 +9374,27 @@ dsv41flash-fp4-gb300-dynamo-sglang-agentic-disagg:
           tp: 2
           ep: 2
           dp-attn: false
+
+# One serving GPU, stock ARM64 release, native AgentX client and proportional host cache.
+qwen3.8next-fp4-gb300-sglang-agentic-mtp:
+  image: lmsysorg/sglang:v0.5.21-cu130@sha256:1f176b3a1dd1205d2a462ab80ea9fc17769a32e53cb6f3c690efa2dc2c384977
+  model: nvidia/Qwen3.8-Flash-Next-NVFP4
+  model-prefix: qwen3.8next
+  runner: cluster:gb300-nv
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    # Use the GB300 profile's proportional per-worker host DRAM allowance.
+    # Generated allowance: 225 GB. Recipe H224 reserves alignment headroom.
+    - dram-utilization: 1.0
+      search-space:
+      - tp: 1
+        ep: 1
+        dp-attn: false
+        spec-decoding: mtp
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        conc-list: [1, 2, 4, 8, 16, 32, 48, 64]
+        srt-recipe: benchmarks/single_node/srt-slurm-recipes/qwen3.8next/sglang/gb300-fp4-mtp/agentic.yaml

inferencex-e2e/perf-changelog.yaml

@@ -9488,3 +9488,12 @@
     - "No data-type or precision change to the EAGLE3 draft model (Inferact/MiniMax-M3-EAGLE3-GQA): online_quant_config (ptpc_fp8), kv_cache_dtype fp8 and index-cache-dtype fp8 are unchanged; mono decode changes the target model's decode kernels, not draft precision."
     - "更新镜像至 rocm/atom-dev:nightly_202610070333（ATOM 0.1.7.dev60+g7fc0926c6，含 ROCm/ATOM#2483 的 mono step fence 修复）；去掉 TP4 C15/C20/C24/C28 的 ATOM_MONO_ENABLE=0，所有 TP4 点位启用 MiniMax-M3 mono decode（每步最多 16 个 token，即 EAGLE3 K3 下 4 个请求，覆盖 C1/C2/C4）；ATOM 只为 TP4 构建 mono，TP2 走非融合路径。新增 TP2 C4，LMCache 版本记录为 0.5.6.dev139+gf22dec28。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3750
+
+- config-keys:
+    - qwen3.8next-fp4-gb300-sglang-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add Qwen3.8-Flash-Next NVFP4 GB300 TP1/EP1 AgentX with native NEXTN MTP and SGLang HiCache."
+    - "Pin the stock SGLang v0.5.21-cu130 image and checkpoint revision; enable engine metrics and cache reporting."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3816