diff --git a/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c1-mtp-hicache.yaml b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c1-mtp-hicache.yaml
new file mode 100644
index 0000000000..f2368fa8a8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c1-mtp-hicache.yaml
@@ -0,0 +1,136 @@
+schema: 2
+name: "agg-b200-tp8-c1-mtp-hicache"
+
+# B200 AgentX aggregate topology: one TP8 worker uses one eight-GPU node
+# and serves both prefill and decode with bundled DSpark and HiCache.
+
+model:
+  path: "deepseek-v4-pro-0813"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
+  frameworks:
+    dynamo: "1.5.0.dev20260914"
+
+dynamo:
+  install: true
+  source:
+    wheel: "1.5.0.dev20260914"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: infra
+  - name: nats
+    type: nats
+    placement:
+      node: infra
+    options:
+      max_payload_mb: 32
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: false
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    DYN_NATS_REQUEST_TIMEOUT_SECS: "1800"
+  args:
+    router-mode: "kv"
+    router-session-affinity-ttl-secs: "3600"
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+engine: sglang
+roles:
+  agg:
+    nodes: 1
+    workers: 1
+    gpus: 8
+
+    env:
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      SGLANG_DEFAULT_THINKING: "1"
+      SGLANG_DSV4_REASONING_EFFORT: high
+      PIP_BREAK_SYSTEM_PACKAGES: "1"
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+      SGLANG_OPT_USE_ONLINE_COMPRESS: "0"
+      SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
+      SGLANG_OPT_USE_JIT_NORM: "1"
+      SGLANG_OPT_USE_TOPK_V2: "True"
+
+    args:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813"
+      enable-metrics: true
+      enable-cache-report: true
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      stream-interval: 10
+      watchdog-timeout: 1000000
+      mem-fraction-static: 0.90
+      page-size: 256
+      chunked-prefill-size: 8192
+      max-prefill-tokens: 8192
+      moe-runner-backend: "flashinfer_mxfp4"
+      enable-deepseek-v4-fp4-indexer: true
+      disable-flashinfer-autotune: true
+      swa-full-tokens-ratio: 0.1
+      max-running-requests: 2
+      cuda-graph-max-bs-decode: 2
+      scheduler-recv-interval: 30
+      dp-size: 1
+      tp-size: 8
+      ep-size: 1
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+      enable-hierarchical-cache: true
+      hicache-ratio: 2.75
+      hicache-write-policy: write_through
+      hicache-io-backend: direct
+      hicache-mem-layout: page_first_direct
+
+sbatch_directives:
+  mem: "0"
+  # NScale B200 nodes expose 192 logical CPUs. Request the full node CPU set;
+  # mem=0 reserves all 1.7 TiB of allocatable host DRAM for HiCache.
+  cpus-per-task: "192"
+
+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: "false"
+    TP: "8"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c4-mtp-hicache.yaml b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c4-mtp-hicache.yaml
new file mode 100644
index 0000000000..5cc7a371c4
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c4-mtp-hicache.yaml
@@ -0,0 +1,136 @@
+schema: 2
+name: "agg-b200-tp8-c4-mtp-hicache"
+
+# B200 AgentX aggregate topology: one TP8 worker uses one eight-GPU node
+# and serves both prefill and decode with bundled DSpark and HiCache.
+
+model:
+  path: "deepseek-v4-pro-0813"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
+  frameworks:
+    dynamo: "1.5.0.dev20260914"
+
+dynamo:
+  install: true
+  source:
+    wheel: "1.5.0.dev20260914"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: infra
+  - name: nats
+    type: nats
+    placement:
+      node: infra
+    options:
+      max_payload_mb: 32
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: false
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    DYN_NATS_REQUEST_TIMEOUT_SECS: "1800"
+  args:
+    router-mode: "kv"
+    router-session-affinity-ttl-secs: "3600"
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+engine: sglang
+roles:
+  agg:
+    nodes: 1
+    workers: 1
+    gpus: 8
+
+    env:
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      SGLANG_DEFAULT_THINKING: "1"
+      SGLANG_DSV4_REASONING_EFFORT: high
+      PIP_BREAK_SYSTEM_PACKAGES: "1"
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+      SGLANG_OPT_USE_ONLINE_COMPRESS: "0"
+      SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
+      SGLANG_OPT_USE_JIT_NORM: "1"
+      SGLANG_OPT_USE_TOPK_V2: "True"
+
+    args:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813"
+      enable-metrics: true
+      enable-cache-report: true
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      stream-interval: 10
+      watchdog-timeout: 1000000
+      mem-fraction-static: 0.90
+      page-size: 256
+      chunked-prefill-size: 8192
+      max-prefill-tokens: 8192
+      moe-runner-backend: "flashinfer_mxfp4"
+      enable-deepseek-v4-fp4-indexer: true
+      disable-flashinfer-autotune: true
+      swa-full-tokens-ratio: 0.1
+      max-running-requests: 8
+      cuda-graph-max-bs-decode: 8
+      scheduler-recv-interval: 30
+      dp-size: 1
+      tp-size: 8
+      ep-size: 1
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+      enable-hierarchical-cache: true
+      hicache-ratio: 2.75
+      hicache-write-policy: write_through
+      hicache-io-backend: direct
+      hicache-mem-layout: page_first_direct
+
+sbatch_directives:
+  mem: "0"
+  # NScale B200 nodes expose 192 logical CPUs. Request the full node CPU set;
+  # mem=0 reserves all 1.7 TiB of allocatable host DRAM for HiCache.
+  cpus-per-task: "192"
+
+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: "false"
+    TP: "8"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c8-mtp-hicache.yaml b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c8-mtp-hicache.yaml
new file mode 100644
index 0000000000..e1f28fa87d
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c8-mtp-hicache.yaml
@@ -0,0 +1,136 @@
+schema: 2
+name: "agg-b200-tp8-c8-mtp-hicache"
+
+# B200 AgentX aggregate topology: one TP8 worker uses one eight-GPU node
+# and serves both prefill and decode with bundled DSpark and HiCache.
+
+model:
+  path: "deepseek-v4-pro-0813"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
+  frameworks:
+    dynamo: "1.5.0.dev20260914"
+
+dynamo:
+  install: true
+  source:
+    wheel: "1.5.0.dev20260914"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b200"
+  gpus_per_node: 8
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: infra
+  - name: nats
+    type: nats
+    placement:
+      node: infra
+    options:
+      max_payload_mb: 32
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: false
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    DYN_NATS_REQUEST_TIMEOUT_SECS: "1800"
+  args:
+    router-mode: "kv"
+    router-session-affinity-ttl-secs: "3600"
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+engine: sglang
+roles:
+  agg:
+    nodes: 1
+    workers: 1
+    gpus: 8
+
+    env:
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      SGLANG_DEFAULT_THINKING: "1"
+      SGLANG_DSV4_REASONING_EFFORT: high
+      PIP_BREAK_SYSTEM_PACKAGES: "1"
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+      SGLANG_OPT_USE_ONLINE_COMPRESS: "0"
+      SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
+      SGLANG_OPT_USE_JIT_NORM: "1"
+      SGLANG_OPT_USE_TOPK_V2: "True"
+
+    args:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813"
+      enable-metrics: true
+      enable-cache-report: true
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      stream-interval: 10
+      watchdog-timeout: 1000000
+      mem-fraction-static: 0.90
+      page-size: 256
+      chunked-prefill-size: 8192
+      max-prefill-tokens: 8192
+      moe-runner-backend: "flashinfer_mxfp4"
+      enable-deepseek-v4-fp4-indexer: true
+      disable-flashinfer-autotune: true
+      swa-full-tokens-ratio: 0.1
+      max-running-requests: 16
+      cuda-graph-max-bs-decode: 16
+      scheduler-recv-interval: 30
+      dp-size: 1
+      tp-size: 8
+      ep-size: 1
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+      enable-hierarchical-cache: true
+      hicache-ratio: 2.75
+      hicache-write-policy: write_through
+      hicache-io-backend: direct
+      hicache-mem-layout: page_first_direct
+
+sbatch_directives:
+  mem: "0"
+  # NScale B200 nodes expose 192 logical CPUs. Request the full node CPU set;
+  # mem=0 reserves all 1.7 TiB of allocatable host DRAM for HiCache.
+  cpus-per-task: "192"
+
+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: "false"
+    TP: "8"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-1p1d-dep8-dep8-c128-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-1p1d-dep8-dep8-c128-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..305bf5d7ef
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-1p1d-dep8-dep8-c128-mtp-kvoffload.yaml
@@ -0,0 +1,234 @@
+schema: 2
+name: "disagg-b200-1p1d-dep8-dep8-c128-mtp-kvoffload"
+
+# B200 AgentX SGLang disaggregated recipe for DeepSeek-V4-Pro-0813
+# (1P x DEP8 / 1D x DEP8, bundled DSpark + HiCache KV offload), tuned for concurrency 128.
+# Each DEP8 worker occupies one eight-GPU B200 node.
+
+model:
+  path: "deepseek-v4-pro-0813"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
+  frameworks:
+    dynamo: "1.5.0.dev20260914"
+
+dynamo:
+  install: true
+  source:
+    wheel: "1.5.0.dev20260914"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: b200
+  gpus_per_node: 8
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: dedicated
+  - name: nats
+    type: nats
+    placement:
+      node: dedicated
+    options:
+      max_payload_mb: 32
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: true
+  num_additional_frontends: 4
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+  args:
+    router-mode: "kv"
+    router-session-affinity-ttl-secs: "3600"
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+engine: sglang
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 8
+    env:
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      SGLANG_DSV4_MHC_PREWARM: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PYTHONUNBUFFERED: '1'
+      SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK: '0'
+      SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: '1'
+      SGLANG_JIT_DEEPGEMM_FAST_WARMUP: '1'
+      SGLANG_DEFAULT_THINKING: '1'
+      SGLANG_DSV4_REASONING_EFFORT: high
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: '1'
+      SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: '1'
+      SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '9216'
+      SGLANG_OPT_USE_ONLINE_COMPRESS: '0'
+      NCCL_MNNVL_ENABLE: '1'
+      NCCL_CUMEM_ENABLE: '1'
+    # P/D workers are on separate HGX B200 nodes, so use Mooncake's standard
+    # RDMA-registered CUDA buffers rather than an NVL72 custom memory pool.
+      NCCL_TIMEOUT: '100000'
+      NCCL_MNNVL_UUID_QUERY_TIMEOUT: '100000'
+      SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+      DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+      DYN_LOG: info,dynamo_runtime::pipeline::network::ingress::push_handler=warn
+      SGLANG_LOG_FORWARD_ITERS: '1'
+      SGLANG_LOG_MS: '1'
+      SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+      SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
+
+    args:
+      host: 0.0.0.0
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
+      enable-metrics: true
+      enable-cache-report: true
+      model-path: /model/
+      trust-remote-code: true
+      watchdog-timeout: 86400
+      stream-interval: 60
+      tp-size: 8
+      dp-size: 8
+      ep-size: 8
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      enable-deepseek-v4-fp4-indexer: true
+      enable-w4a4-mxfp4-megamoe: true
+      disaggregation-transfer-backend: mooncake
+      # Use the eight external HCAs. mlx5_6..9 are on a separate low-LID
+      # fabric, while mlx5_bond_0 caused cross-fabric QP RTR timeouts.
+      disaggregation-ib-device: mlx5_0,mlx5_1,mlx5_2,mlx5_3,mlx5_4,mlx5_5,mlx5_10,mlx5_11
+      disaggregation-mode: prefill
+      load-balance-method: total_tokens
+      mem-fraction-static: 0.85
+      page-size: 256
+      # Favor the full-attention pool; measured SWA utilization remained low.
+      swa-full-tokens-ratio: 0.01
+      max-running-requests: 256
+      cuda-graph-max-bs-decode: 256
+      chunked-prefill-size: 65536
+      disable-flashinfer-autotune: true
+      model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
+      kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}'
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+      enable-hierarchical-cache: true
+      hicache-write-policy: write_back
+      hicache-ratio: 8
+      hicache-io-backend: direct
+      hicache-mem-layout: page_first_direct
+
+  decode:
+    nodes: 1
+    workers: 1
+    gpus: 8
+
+    env:
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PYTHONUNBUFFERED: '1'
+      SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK: '0'
+      SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: '1'
+      SGLANG_JIT_DEEPGEMM_FAST_WARMUP: '1'
+      SGLANG_DEFAULT_THINKING: '1'
+      SGLANG_DSV4_REASONING_EFFORT: high
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_DEEPGEMM_MEGA_MOE: '1'
+      SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '4096'
+      SGLANG_OPT_USE_ONLINE_COMPRESS: '0'
+      NCCL_MNNVL_ENABLE: '1'
+      NCCL_CUMEM_ENABLE: '1'
+      NCCL_TIMEOUT: '100000'
+      NCCL_MNNVL_UUID_QUERY_TIMEOUT: '100000'
+      SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+      DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+      DYN_LOG: info,dynamo_runtime::pipeline::network::ingress::push_handler=warn
+      SGLANG_LOG_FORWARD_ITERS: '1'
+      SGLANG_LOG_MS: '1'
+      SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+
+    args:
+      host: 0.0.0.0
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
+      enable-metrics: true
+      enable-cache-report: true
+      model-path: /model/
+      trust-remote-code: true
+      watchdog-timeout: 86400
+      stream-interval: 60
+      tp-size: 8
+      dp-size: 8
+      ep-size: 8
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      enable-deepseek-v4-fp4-indexer: true
+      enable-w4a4-mxfp4-megamoe: true
+      disaggregation-transfer-backend: mooncake
+      disaggregation-ib-device: mlx5_0,mlx5_1,mlx5_2,mlx5_3,mlx5_4,mlx5_5,mlx5_10,mlx5_11
+      disaggregation-mode: decode
+      load-balance-method: total_tokens
+      mem-fraction-static: 0.9
+      page-size: 256
+      swa-full-tokens-ratio: 0.02
+      max-running-requests: 256
+      cuda-graph-max-bs-decode: 256
+      disable-flashinfer-autotune: true
+      model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
+      kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5567"}'
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+
+
+sbatch_directives:
+  mem: "0"
+  # NScale B200 nodes expose 192 logical CPUs. Request the full node CPU set;
+  # mem=0 reserves all 1.7 TiB of allocatable host DRAM for HiCache.
+  cpus-per-task: "192"
+
+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_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-1p1d-dep8-dep8-c64-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-1p1d-dep8-dep8-c64-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..16f1ed6c8e
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-1p1d-dep8-dep8-c64-mtp-kvoffload.yaml
@@ -0,0 +1,234 @@
+schema: 2
+name: "disagg-b200-1p1d-dep8-dep8-c64-mtp-kvoffload"
+
+# B200 AgentX SGLang disaggregated recipe for DeepSeek-V4-Pro-0813
+# (1P x DEP8 / 1D x DEP8, bundled DSpark + HiCache KV offload), tuned for concurrency 64.
+# Each DEP8 worker occupies one eight-GPU B200 node.
+
+model:
+  path: "deepseek-v4-pro-0813"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
+  frameworks:
+    dynamo: "1.5.0.dev20260914"
+
+dynamo:
+  install: true
+  source:
+    wheel: "1.5.0.dev20260914"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: b200
+  gpus_per_node: 8
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: dedicated
+  - name: nats
+    type: nats
+    placement:
+      node: dedicated
+    options:
+      max_payload_mb: 32
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: true
+  num_additional_frontends: 4
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+  args:
+    router-mode: "kv"
+    router-session-affinity-ttl-secs: "3600"
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+engine: sglang
+roles:
+  prefill:
+    nodes: 1
+    workers: 1
+    gpus: 8
+    env:
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      SGLANG_DSV4_MHC_PREWARM: '1'
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PYTHONUNBUFFERED: '1'
+      SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK: '0'
+      SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: '1'
+      SGLANG_JIT_DEEPGEMM_FAST_WARMUP: '1'
+      SGLANG_DEFAULT_THINKING: '1'
+      SGLANG_DSV4_REASONING_EFFORT: high
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: '1'
+      SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: '1'
+      SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '9216'
+      SGLANG_OPT_USE_ONLINE_COMPRESS: '0'
+      NCCL_MNNVL_ENABLE: '1'
+      NCCL_CUMEM_ENABLE: '1'
+    # P/D workers are on separate HGX B200 nodes, so use Mooncake's standard
+    # RDMA-registered CUDA buffers rather than an NVL72 custom memory pool.
+      NCCL_TIMEOUT: '100000'
+      NCCL_MNNVL_UUID_QUERY_TIMEOUT: '100000'
+      SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+      DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+      DYN_LOG: info,dynamo_runtime::pipeline::network::ingress::push_handler=warn
+      SGLANG_LOG_FORWARD_ITERS: '1'
+      SGLANG_LOG_MS: '1'
+      SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+      SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
+
+    args:
+      host: 0.0.0.0
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
+      enable-metrics: true
+      enable-cache-report: true
+      model-path: /model/
+      trust-remote-code: true
+      watchdog-timeout: 86400
+      stream-interval: 60
+      tp-size: 8
+      dp-size: 8
+      ep-size: 8
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      enable-deepseek-v4-fp4-indexer: true
+      enable-w4a4-mxfp4-megamoe: true
+      disaggregation-transfer-backend: mooncake
+      # Use the eight external HCAs. mlx5_6..9 are on a separate low-LID
+      # fabric, while mlx5_bond_0 caused cross-fabric QP RTR timeouts.
+      disaggregation-ib-device: mlx5_0,mlx5_1,mlx5_2,mlx5_3,mlx5_4,mlx5_5,mlx5_10,mlx5_11
+      disaggregation-mode: prefill
+      load-balance-method: total_tokens
+      mem-fraction-static: 0.85
+      page-size: 256
+      # Favor the full-attention pool; measured SWA utilization remained low.
+      swa-full-tokens-ratio: 0.01
+      max-running-requests: 256
+      cuda-graph-max-bs-decode: 256
+      chunked-prefill-size: 65536
+      disable-flashinfer-autotune: true
+      model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
+      kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}'
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+      enable-hierarchical-cache: true
+      hicache-write-policy: write_back
+      hicache-ratio: 8
+      hicache-io-backend: direct
+      hicache-mem-layout: page_first_direct
+
+  decode:
+    nodes: 1
+    workers: 1
+    gpus: 8
+
+    env:
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PYTHONUNBUFFERED: '1'
+      SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK: '0'
+      SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: '1'
+      SGLANG_JIT_DEEPGEMM_FAST_WARMUP: '1'
+      SGLANG_DEFAULT_THINKING: '1'
+      SGLANG_DSV4_REASONING_EFFORT: high
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_DEEPGEMM_MEGA_MOE: '1'
+      SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '4096'
+      SGLANG_OPT_USE_ONLINE_COMPRESS: '0'
+      NCCL_MNNVL_ENABLE: '1'
+      NCCL_CUMEM_ENABLE: '1'
+      NCCL_TIMEOUT: '100000'
+      NCCL_MNNVL_UUID_QUERY_TIMEOUT: '100000'
+      SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+      DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+      DYN_LOG: info,dynamo_runtime::pipeline::network::ingress::push_handler=warn
+      SGLANG_LOG_FORWARD_ITERS: '1'
+      SGLANG_LOG_MS: '1'
+      SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+
+    args:
+      host: 0.0.0.0
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
+      enable-metrics: true
+      enable-cache-report: true
+      model-path: /model/
+      trust-remote-code: true
+      watchdog-timeout: 86400
+      stream-interval: 60
+      tp-size: 8
+      dp-size: 8
+      ep-size: 8
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      enable-deepseek-v4-fp4-indexer: true
+      enable-w4a4-mxfp4-megamoe: true
+      disaggregation-transfer-backend: mooncake
+      disaggregation-ib-device: mlx5_0,mlx5_1,mlx5_2,mlx5_3,mlx5_4,mlx5_5,mlx5_10,mlx5_11
+      disaggregation-mode: decode
+      load-balance-method: total_tokens
+      mem-fraction-static: 0.9
+      page-size: 256
+      swa-full-tokens-ratio: 0.02
+      max-running-requests: 128
+      cuda-graph-max-bs-decode: 256
+      disable-flashinfer-autotune: true
+      model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
+      kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5567"}'
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+
+
+sbatch_directives:
+  mem: "0"
+  # NScale B200 nodes expose 192 logical CPUs. Request the full node CPU set;
+  # mem=0 reserves all 1.7 TiB of allocatable host DRAM for HiCache.
+  cpus-per-task: "192"
+
+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_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-2p1d-dep8-dep8-c256-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-2p1d-dep8-dep8-c256-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..4ac6c19f63
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-2p1d-dep8-dep8-c256-mtp-kvoffload.yaml
@@ -0,0 +1,239 @@
+schema: 2
+name: "disagg-b200-2p1d-dep8-dep8-c256-mtp-kvoffload"
+
+# B200 AgentX SGLang disaggregated recipe for DeepSeek-V4-Pro-0813
+# (2P x DEP8 / 1D x DEP8, bundled DSpark + HiCache KV offload), tuned for concurrency 256.
+# Each DEP8 worker occupies one eight-GPU B200 node.
+
+model:
+  path: "deepseek-v4-pro-0813"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-20260916-c9a8fba9"
+  frameworks:
+    dynamo: "1.5.0.dev20260914"
+
+dynamo:
+  install: true
+  source:
+    wheel: "1.5.0.dev20260914"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: b200
+  gpus_per_node: 8
+services:
+  - name: etcd
+    type: etcd
+    placement:
+      node: dedicated
+  - name: nats
+    type: nats
+    placement:
+      node: dedicated
+    options:
+      max_payload_mb: 32
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: true
+  num_additional_frontends: 4
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+  args:
+    router-mode: "kv"
+    router-session-affinity-ttl-secs: "3600"
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+engine: sglang
+roles:
+  prefill:
+    nodes: 2
+    workers: 2
+    gpus: 8
+    env:
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PYTHONUNBUFFERED: '1'
+      SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK: '0'
+      SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: '1'
+      SGLANG_JIT_DEEPGEMM_FAST_WARMUP: '1'
+      SGLANG_DEFAULT_THINKING: '1'
+      SGLANG_DSV4_REASONING_EFFORT: high
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: '1'
+      SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: '1'
+      SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '9216'
+      SGLANG_OPT_USE_ONLINE_COMPRESS: '0'
+      NCCL_MNNVL_ENABLE: '1'
+      NCCL_CUMEM_ENABLE: '1'
+    # P/D workers are on separate HGX B200 nodes, so use Mooncake's standard
+    # RDMA-registered CUDA buffers rather than an NVL72 custom memory pool.
+      NCCL_TIMEOUT: '100000'
+      NCCL_MNNVL_UUID_QUERY_TIMEOUT: '100000'
+      SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+      DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+      DYN_LOG: info,dynamo_runtime::pipeline::network::ingress::push_handler=warn
+      SGLANG_LOG_FORWARD_ITERS: '1'
+      SGLANG_LOG_MS: '1'
+      SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+      SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
+      SGLANG_DSV4_MHC_PREWARM: '1'
+
+    args:
+      host: 0.0.0.0
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
+      enable-metrics: true
+      enable-cache-report: true
+      model-path: /model/
+      trust-remote-code: true
+      watchdog-timeout: 86400
+      stream-interval: 60
+      tp-size: 8
+      dp-size: 8
+      ep-size: 8
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      enable-deepseek-v4-fp4-indexer: true
+      enable-w4a4-mxfp4-megamoe: true
+      disaggregation-transfer-backend: mooncake
+      # Use the eight external HCAs. mlx5_6..9 are on a separate low-LID
+      # fabric, while mlx5_bond_0 caused cross-fabric QP RTR timeouts.
+      disaggregation-ib-device: mlx5_0,mlx5_1,mlx5_2,mlx5_3,mlx5_4,mlx5_5,mlx5_10,mlx5_11
+      disaggregation-mode: prefill
+      load-balance-method: total_tokens
+      mem-fraction-static: 0.85
+      page-size: 256
+      swa-full-tokens-ratio: 0.02
+      max-running-requests: 256
+      cuda-graph-max-bs-decode: 256
+      chunked-prefill-size: 65536
+      disable-flashinfer-autotune: true
+      model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
+      kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}'
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+      enable-hierarchical-cache: true
+      hicache-write-policy: write_back
+      hicache-ratio: 8
+      hicache-io-backend: direct
+      hicache-mem-layout: page_first_direct
+
+  decode:
+    nodes: 1
+    workers: 1
+    gpus: 8
+
+
+    env:
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PYTHONUNBUFFERED: '1'
+      SGLANG_ENABLE_TP_MEMORY_INBALANCE_CHECK: '0'
+      SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: '1'
+      SGLANG_JIT_DEEPGEMM_FAST_WARMUP: '1'
+      SGLANG_DEFAULT_THINKING: '1'
+      SGLANG_DSV4_REASONING_EFFORT: high
+      SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
+      SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+      SGLANG_OPT_DEEPGEMM_MEGA_MOE: '1'
+      SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '4096'
+      SGLANG_OPT_USE_ONLINE_COMPRESS: '0'
+      NCCL_MNNVL_ENABLE: '1'
+      NCCL_CUMEM_ENABLE: '1'
+      NCCL_TIMEOUT: '100000'
+      NCCL_MNNVL_UUID_QUERY_TIMEOUT: '100000'
+      SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+      DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
+      DYN_LOG: info,dynamo_runtime::pipeline::network::ingress::push_handler=warn
+      SGLANG_LOG_FORWARD_ITERS: '1'
+      SGLANG_LOG_MS: '1'
+      SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+      SGLANG_DSV4_MHC_PREWARM: '1'
+
+    args:
+      host: 0.0.0.0
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
+      enable-metrics: true
+      enable-cache-report: true
+      model-path: /model/
+      trust-remote-code: true
+      watchdog-timeout: 86400
+      stream-interval: 60
+      tp-size: 8
+      dp-size: 8
+      ep-size: 8
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      moe-dense-tp-size: 1
+      moe-a2a-backend: megamoe
+      enable-deepseek-v4-fp4-indexer: true
+      enable-w4a4-mxfp4-megamoe: true
+      disaggregation-transfer-backend: mooncake
+      disaggregation-ib-device: mlx5_0,mlx5_1,mlx5_2,mlx5_3,mlx5_4,mlx5_5,mlx5_10,mlx5_11
+      disaggregation-mode: decode
+      load-balance-method: total_tokens
+      # Leave enough decode activation headroom while expanding the KV pools
+      # for transient DP-rank imbalance at concurrency 256.
+      mem-fraction-static: 0.91
+      page-size: 256
+      # Favor the full-attention pool while retaining enough SWA capacity for
+      # the busiest decode rank.
+      swa-full-tokens-ratio: 0.005
+      max-running-requests: 512
+      cuda-graph-max-bs-decode: 256
+      disable-flashinfer-autotune: true
+      model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
+      kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5567"}'
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+
+
+sbatch_directives:
+  mem: "0"
+  # NScale B200 nodes expose 192 logical CPUs. Request the full node CPU set;
+  # mem=0 reserves all 1.7 TiB of allocatable host DRAM for HiCache.
+  cpus-per-task: "192"
+
+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_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index f28e74de34..02fa5c45b9 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8150,3 +8150,130 @@ dsv41flash-fp4-h200-vllm-agentic-dspark:
       # 8x141 GB holds the 511 GB checkpoint minus the UVA-offloaded Engram
       # tables, so the KV cache stays GPU-resident across the full range.
       - { tp: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128] }
+
+dsv4-fp4-b200-dynamo-sglang-agentic-agg:
+  image: lmsysorg/sglang:nightly-dev-20260916-c9a8fba9
+  model: deepseek-ai/DeepSeek-V4-Pro-0813
+  model-prefix: dsv4
+  runner: cluster:b200-nscale
+  precision: fp4
+  framework: dynamo-sglang
+  router: { name: dynamo-router, version: "1.5.0.dev20260914" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - spec-decoding: mtp
+        conc-list: [1]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        num-nodes: 1
+        worker:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c1-mtp-hicache.yaml"
+      - spec-decoding: mtp
+        conc-list: [4]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        num-nodes: 1
+        worker:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c4-mtp-hicache.yaml"
+      - spec-decoding: mtp
+        conc-list: [8]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        num-nodes: 1
+        worker:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/dsv4/sglang/b200-fp4/agentx/agg-b200-tp8-c8-mtp-hicache.yaml"
+
+dsv4-fp4-b200-dynamo-sglang-agentic-disagg:
+  image: lmsysorg/sglang:nightly-dev-20260916-c9a8fba9
+  model: deepseek-ai/DeepSeek-V4-Pro-0813
+  model-prefix: dsv4
+  runner: cluster:b200-nscale
+  precision: fp4
+  framework: dynamo-sglang
+  router: { name: dynamo-router, version: "1.5.0.dev20260914" }
+  kv-p2p-transfer: mooncake
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - spec-decoding: mtp
+        conc-list: [64]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-1p1d-dep8-dep8-c64-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      - spec-decoding: mtp
+        conc-list: [128]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-1p1d-dep8-dep8-c128-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      - spec-decoding: mtp
+        conc-list: [256]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        prefill:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/dsv4/sglang/b200-fp4/agentx/disagg-b200-2p1d-dep8-dep8-c256-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index a112ba56db..f42e0480f4 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -8082,3 +8082,23 @@
     - "Capture only full decode CUDA graphs (VLLM_USE_BREAKABLE_CUDAGRAPH=0 with cudagraph_mode FULL_DECODE_ONLY, as the MiniMax-M3 gfx942 arm does) and restore --moe-backend aiter: on gfx942 every worker segfaulted during piecewise graph capture with both the Triton W4A16 MoE kernel (run 35305045778) and the auto-selected unfused Triton kernel (run 35306398350), so the capture mode rather than the MoE kernel is the failing piece; prefill runs eagerly"
     - "仅捕获完整的 decode CUDA graph（VLLM_USE_BREAKABLE_CUDAGRAPH=0 并设置 cudagraph_mode FULL_DECODE_ONLY，与 MiniMax-M3 gfx942 配方一致）并恢复 --moe-backend aiter：在 gfx942 上，无论使用 Triton W4A16 MoE 内核（运行 35305045778）还是自动选择的未融合 Triton 内核（运行 35306398350），所有 worker 都在 piecewise graph 捕获期间段错误，说明问题在于捕获模式而非 MoE 内核；prefill 以 eager 方式运行"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3247
+
+- config-keys:
+    - dsv4-fp4-b200-dynamo-sglang-agentic-agg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add B200 Dynamo+SGLang AgentX TP8 aggregate HiCache recipes at concurrency 1, 4, and 8."
+    - "Use lmsysorg/sglang:nightly-dev-20260916-c9a8fba9 with bundled DSpark K=6, full-node CPU/DRAM allocation, and per-point request/CUDA-graph limits."
+    - "Map the exact DeepSeek-V4-Pro-0813 recipe alias to the pre-staged NScale checkpoint."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3257
+
+- config-keys:
+    - dsv4-fp4-b200-dynamo-sglang-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add B200 Dynamo+SGLang AgentX DEP8/DEP8 HiCache recipes for 1P1D at concurrency 64 and 128, plus 2P1D at concurrency 256."
+    - "Limit decode max-running-requests to 128, 256, and 512, respectively. Set decode swa-full-tokens-ratio to 0.01 for 1P1D and 0.005 for 2P1D c256, and use a 0.91 decode memory fraction for c256 so both KV pools retain headroom under transient DP-rank imbalance."
+    - "Map the exact DeepSeek-V4-Pro-0813 recipe alias to the pre-staged NScale checkpoint."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3257
diff --git a/runners/launch_b200-nscale-slurm.sh b/runners/launch_b200-nscale-slurm.sh
index a6a7992fa5..02c7482ba6 100755
--- a/runners/launch_b200-nscale-slurm.sh
+++ b/runners/launch_b200-nscale-slurm.sh
@@ -37,7 +37,10 @@ if [[ "$FRAMEWORK" == "tilert" && "${IS_AGENTIC}" != "1" ]]; then
     run_compat_launcher
 fi
 
-if [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
+if [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" && $MODEL == "deepseek-ai/DeepSeek-V4-Pro-0813" ]]; then
+    export MODEL_PATH="$NSCALE_MODEL_ROOT/DeepSeek-V4-Pro-0813"
+    export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro-0813"
+elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
     check_env_vars MODEL_PATH
     export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro"
 elif [[ $MODEL_PREFIX == "kimik2.6" && $PRECISION == "fp4" ]]; then