diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/install-mooncake-efa-cu13.sh b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/install-mooncake-efa-cu13.sh
new file mode 100755
index 0000000000..252168b822
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/configs/install-mooncake-efa-cu13.sh
@@ -0,0 +1,5 @@
+#!/usr/bin/env bash
+set -eo pipefail
+
+python3 -m pip uninstall --break-system-packages -y mooncake-transfer-engine-cuda13 mooncake-transfer-engine-efa-cuda13
+python3 -m pip install --break-system-packages --no-deps mooncake-transfer-engine-efa-cuda13==0.3.13.post1
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/agg-tp4-c4-mtp.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/agg-tp4-c4-mtp.yaml
new file mode 100644
index 0000000000..d4d6a4b6ec
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/agg-tp4-c4-mtp.yaml
@@ -0,0 +1,128 @@
+schema: 2
+name: "agg-b300-tp4-c4-mtp"
+
+# AgentX aggregate topology: one TP4 worker occupies half of one eight-GPU
+# B300 node and serves both prefill and decode with DSpark K=6.
+
+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"
+
+dynamo:
+  install: false
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b300"
+  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: 4
+
+    env:
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_DEFAULT_THINKING: "1"
+      SGLANG_DSV4_REASONING_EFFORT: high
+      PIP_BREAK_SYSTEM_PACKAGES: "1"
+      SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
+      SGLANG_JIT_DEEPGEMM_FAST_WARMUP: "1"
+      SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "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: 4
+      ep-size: 1
+      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"
+  exclude: "dsxe-sa-b300-prd0-gpu-00,dsxe-sa-b300-prd0-gpu-01,dsxe-sa-b300-prd0-gpu-04,dsxe-sa-b300-prd0-gpu-06,dsxe-sa-b300-prd0-gpu-11,dsxe-sa-b300-prd0-gpu-16"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "false"
+    TP: "4"
+    PP_SIZE: "1"
+    PCP_SIZE: "1"
+    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/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/agg-tp8-c1-mtp.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/agg-tp8-c1-mtp.yaml
new file mode 100644
index 0000000000..a5f30a1bff
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/agg-tp8-c1-mtp.yaml
@@ -0,0 +1,128 @@
+schema: 2
+name: "agg-b300-tp8-c1-mtp"
+
+# AgentX aggregate topology: one TP8 worker occupies one eight-GPU B300 node
+# and serves both prefill and decode with DSpark K=6.
+
+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"
+
+dynamo:
+  install: false
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "b300"
+  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_JIT_DEEPGEMM_FAST_WARMUP: "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
+
+sbatch_directives:
+  mem: "0"
+  exclude: "dsxe-sa-b300-prd0-gpu-00,dsxe-sa-b300-prd0-gpu-01,dsxe-sa-b300-prd0-gpu-04,dsxe-sa-b300-prd0-gpu-06,dsxe-sa-b300-prd0-gpu-11,dsxe-sa-b300-prd0-gpu-16"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "false"
+    TP: "8"
+    PP_SIZE: "1"
+    PCP_SIZE: "1"
+    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/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p1d-dep8-dep8-c240-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p1d-dep8-dep8-c240-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..4a0db2f37b
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p1d-dep8-dep8-c240-mtp-kvoffload.yaml
@@ -0,0 +1,226 @@
+schema: 2
+name: "disagg-b300-1p1d-dep8-dep8-c240-mtp-kvoffload"
+
+# Agentic-coding SGLang disaggregated recipe for DeepSeek-V4-Pro-0813 on B300
+# (1P x DEP8 / 1D x DEP8, DSpark K=6 + hierarchical-cache KV offload),
+# tuned for concurrency 240. Concurrency is exported by the master config.
+
+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"
+
+dynamo:
+  install: false
+
+setup_script: install-mooncake-efa-cu13.sh
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: b300
+  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"
+    DYN_TCP_REQUEST_TIMEOUT: "60"
+  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_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      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_CUMEM_ENABLE: '1'
+      MOONCAKE_PROTOCOL: efa
+      MC_MS_FILTERS: rdmap191s0
+      WITH_NVIDIA_PEERMEM: '0'
+      NCCL_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
+      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-prefill: 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: 3
+      hicache-io-backend: direct
+
+  decode:
+    nodes: 1
+    workers: 1
+    gpus: 8
+
+    env:
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      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_CUMEM_ENABLE: '1'
+      MOONCAKE_PROTOCOL: efa
+      MC_MS_FILTERS: rdmap191s0
+      WITH_NVIDIA_PEERMEM: '0'
+      NCCL_TIMEOUT: '100000'
+      SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+      SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: '8'
+      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-mode: decode
+      load-balance-method: total_tokens
+      mem-fraction-static: 0.9
+      page-size: 256
+      swa-full-tokens-ratio: 0.02
+      max-running-requests: 480
+      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"
+  exclude: "dsxe-sa-b300-prd0-gpu-00,dsxe-sa-b300-prd0-gpu-01,dsxe-sa-b300-prd0-gpu-04,dsxe-sa-b300-prd0-gpu-06,dsxe-sa-b300-prd0-gpu-11,dsxe-sa-b300-prd0-gpu-16"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/srt_agentic.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/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p1d-dep8-dep8-c64-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p1d-dep8-dep8-c64-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..23f215cf11
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p1d-dep8-dep8-c64-mtp-kvoffload.yaml
@@ -0,0 +1,226 @@
+schema: 2
+name: "disagg-b300-1p1d-dep8-dep8-c64-mtp-kvoffload"
+
+# Agentic-coding SGLang disaggregated recipe for DeepSeek-V4-Pro-0813 on B300
+# (1P x DEP8 / 1D x DEP8, DSpark K=6 + hierarchical-cache KV offload),
+# tuned for concurrency 64. Concurrency is exported by the master config.
+
+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"
+
+dynamo:
+  install: false
+
+setup_script: install-mooncake-efa-cu13.sh
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: b300
+  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"
+    DYN_TCP_REQUEST_TIMEOUT: "60"
+  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_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      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_CUMEM_ENABLE: '1'
+      MOONCAKE_PROTOCOL: efa
+      MC_MS_FILTERS: rdmap191s0
+      WITH_NVIDIA_PEERMEM: '0'
+      NCCL_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
+      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-prefill: 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: 3
+      hicache-io-backend: direct
+
+  decode:
+    nodes: 1
+    workers: 1
+    gpus: 8
+
+    env:
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      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_CUMEM_ENABLE: '1'
+      MOONCAKE_PROTOCOL: efa
+      MC_MS_FILTERS: rdmap191s0
+      WITH_NVIDIA_PEERMEM: '0'
+      NCCL_TIMEOUT: '100000'
+      SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+      SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: '8'
+      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-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"
+  exclude: "dsxe-sa-b300-prd0-gpu-00,dsxe-sa-b300-prd0-gpu-01,dsxe-sa-b300-prd0-gpu-04,dsxe-sa-b300-prd0-gpu-06,dsxe-sa-b300-prd0-gpu-11,dsxe-sa-b300-prd0-gpu-16"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/srt_agentic.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/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-2p1d-dep8-dep8-c480-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-2p1d-dep8-dep8-c480-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..fc9ba21e43
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/b300-fp4/agentx/disagg-2p1d-dep8-dep8-c480-mtp-kvoffload.yaml
@@ -0,0 +1,228 @@
+schema: 2
+name: "disagg-b300-2p1d-dep8-dep8-c480-mtp-kvoffload"
+
+# Agentic-coding SGLang disaggregated recipe for DeepSeek-V4-Pro-0813 on B300
+# (2P x DEP8 / 1D x DEP8, DSpark K=6 + hierarchical-cache KV offload),
+# tuned for concurrency 480. Concurrency is exported by the master config.
+
+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"
+
+dynamo:
+  install: false
+
+setup_script: install-mooncake-efa-cu13.sh
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: b300
+  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"
+    DYN_TCP_REQUEST_TIMEOUT: "60"
+  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_ENABLE_PREFILL_WAR_READ_DONE: "1"
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PYTHONUNBUFFERED: '1'
+      OMP_NUM_THREADS: '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_CUMEM_ENABLE: '1'
+      MOONCAKE_PROTOCOL: efa
+      MC_MS_FILTERS: rdmap191s0
+      WITH_NVIDIA_PEERMEM: '0'
+      NCCL_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
+      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-prefill: 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: 3
+      hicache-io-backend: direct
+
+  decode:
+    nodes: 1
+    workers: 1
+    gpus: 8
+    env:
+      SGLANG_RAGGED_VERIFY_MODE: "static"
+      SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+      PIP_BREAK_SYSTEM_PACKAGES: '1'
+      PYTHONUNBUFFERED: '1'
+      OMP_NUM_THREADS: '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_CUMEM_ENABLE: '1'
+      MOONCAKE_PROTOCOL: efa
+      MC_MS_FILTERS: rdmap191s0
+      WITH_NVIDIA_PEERMEM: '0'
+      NCCL_TIMEOUT: '100000'
+      SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
+      SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: '8'
+      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-mode: decode
+      load-balance-method: total_tokens
+      mem-fraction-static: 0.9
+      page-size: 256
+      swa-full-tokens-ratio: 0.02
+      max-running-requests: 960
+      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"
+  exclude: "dsxe-sa-b300-prd0-gpu-00,dsxe-sa-b300-prd0-gpu-01,dsxe-sa-b300-prd0-gpu-04,dsxe-sa-b300-prd0-gpu-06,dsxe-sa-b300-prd0-gpu-11,dsxe-sa-b300-prd0-gpu-16"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/srt_agentic.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/inferencex-e2e/configs/nvidia-master.yaml b/inferencex-e2e/configs/nvidia-master.yaml
index 07c47e6e40..f9b9e315fd 100644
--- a/inferencex-e2e/configs/nvidia-master.yaml
+++ b/inferencex-e2e/configs/nvidia-master.yaml
@@ -1161,6 +1161,104 @@ dsv4-fp4-b300-sglang-agentic-hicache-eagle:
       - { tp: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv4/sglang/b300-fp4-eagle/agentic.yaml }
       - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: hicache }, spec-decoding: mtp, conc-list: [64, 128, 256, 384, 512], router: { name: sglang-router, version: "0.3.2" }, srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv4/sglang/b300-fp4-eagle/agentic.yaml }
 
+dsv4-fp4-b300-dynamo-sglang-agentic-agg:
+  image: lmsysorg/sglang:nightly-dev-20260916-c9a8fba9
+  model: deepseek-ai/DeepSeek-V4-Pro-0813
+  model-prefix: dsv4
+  runner: cluster:b300-dsxe
+  precision: fp4
+  framework: dynamo-sglang
+  router: { name: dynamo-router, version: "1.5.0.dev20260914" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      - spec-decoding: draft_model
+        conc-list: [1]
+        num-nodes: 1
+        worker:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4/sglang/b300-fp4/agentx/agg-tp8-c1-mtp.yaml"
+      - spec-decoding: draft_model
+        conc-list: [4]
+        num-nodes: 1
+        worker:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4/sglang/b300-fp4/agentx/agg-tp4-c4-mtp.yaml"
+
+dsv4-fp4-b300-dynamo-sglang-agentic-disagg:
+  image: lmsysorg/sglang:nightly-dev-20260916-c9a8fba9
+  model: deepseek-ai/DeepSeek-V4-Pro-0813
+  model-prefix: dsv4
+  runner: cluster:b300-dsxe
+  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: draft_model
+        conc-list: [64]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p1d-dep8-dep8-c64-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      - spec-decoding: draft_model
+        conc-list: [240]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4/sglang/b300-fp4/agentx/disagg-1p1d-dep8-dep8-c240-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      - spec-decoding: draft_model
+        conc-list: [480]
+        kv-offloading: dram
+        kv-offload-backend: { name: hicache }
+        prefill:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv4/sglang/b300-fp4/agentx/disagg-2p1d-dep8-dep8-c480-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+
 qwen3.5-fp8-b200-sglang:
   image: lmsysorg/sglang:nightly-dev-cu13-20260918-20518d85
   model: Qwen/Qwen3.5-397B-A17B-FP8
diff --git a/inferencex-e2e/configs/runners.yaml b/inferencex-e2e/configs/runners.yaml
index cf67501a63..7f25a1221b 100644
--- a/inferencex-e2e/configs/runners.yaml
+++ b/inferencex-e2e/configs/runners.yaml
@@ -541,6 +541,7 @@ clusters:
       cpus-per-gpu: 24
       salloc-args: [--mem=0]
       volumes:
+        aiperf-cache: {path: /data/home/sa-gha-runner/aiperf-cache}
         hf-home: {path: ~/.cache/huggingface}
         hf-hub-cache: {path: ~/.cache/huggingface/hub}
         scratch: {path: /scratch/models, visibility: node-local}
diff --git a/inferencex-e2e/infx/launch/drivers/srt/models.py b/inferencex-e2e/infx/launch/drivers/srt/models.py
index a8f1dad3fc..267f4596b9 100644
--- a/inferencex-e2e/infx/launch/drivers/srt/models.py
+++ b/inferencex-e2e/infx/launch/drivers/srt/models.py
@@ -40,7 +40,7 @@ class Override:
 OVERRIDES: dict[str, tuple[Override, ...]] = {
     "b300-dsxe": (
         Override(
-            Match(frameworks=any_of("vllm"), model_glob="*/DeepSeek-V4-Pro-0813"),
+            Match(frameworks=any_of("vllm", "dynamo-sglang"), model_glob="*/DeepSeek-V4-Pro-0813"),
             entry="DeepSeek-V4-Pro-0813@scratch",
         ),
         Override(Match(model_glob="*/DeepSeek-V4-Pro-0813"), entry="DeepSeek-V4-Pro-0813"),
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index d7fc64f0d6..af3dcd6e62 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9149,3 +9149,19 @@
     - "Use AITER attention and allreduce fusion, FP8 KV (fp8_e4m3), and EAGLE MTP (3 steps, topk 1, 4 draft tokens); do not enable ROCm INT4 quick all-reduce. HiCache uses ratio 1.5, write_through_selective, kernel I/O and page_first layout. The matching cookbook recipe is sgl-project/sglang#41849."
     - "The embedded MTP head runs at its stored precision: block-FP8 expert weights and checkpoint dtype for mtp.fc and gates. No separate draft, draft dtype override or submission-side quantization is used."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3602
+
+- config-keys:
+    - dsv4-fp4-b300-dynamo-sglang-agentic-agg
+    - dsv4-fp4-b300-dynamo-sglang-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add DeepSeek-V4-Pro-0813 FP4 B300 Dynamo+SGLang AgentX aggregate TP8 c1 and TP4 c4 recipes, plus disaggregated 1P1D DEP8 c64/c240 and 2P1D DEP8 c480 recipes."
+    - "Use Mooncake for disaggregated KV transfer and install mooncake-transfer-engine-efa-cuda13 0.3.13.post1 during container setup; rely on the DSXE-managed fabric injection without explicit host EFA/OFI mounts."
+    - "Use the DSXE cluster's 24 CPUs per GPU, HiCache ratio 3, decode max-running-requests at twice concurrency, persistent AgentX/Hugging Face caches, and exclude the known degraded DSXE node."
+    - "Serve B300 DSXE Dynamo+SGLang DeepSeek-V4-Pro-0813 weights from the node-local NVMe copy (DeepSeek-V4-Pro-0813@scratch, as for vLLM) instead of the shared /data/models copy, which slowed startup by hours when several points loaded concurrently."
+    - "新增 DeepSeek-V4-Pro-0813 FP4 B300 Dynamo+SGLang AgentX 配方：聚合式 TP8 c1、TP4 c4，以及分离式 1P1D DEP8 c64/c240 和 2P1D DEP8 c480。"
+    - "分离式 KV 传输使用 Mooncake，并在容器初始化时安装 mooncake-transfer-engine-efa-cuda13 0.3.13.post1；依赖 DSXE 管理的网络栈注入，不显式挂载主机 EFA/OFI 目录。"
+    - "使用 DSXE 集群的每 GPU 24 个 CPU、HiCache 比例 3，decode 的 max-running-requests 为并发数两倍，挂载持久化 AgentX/Hugging Face 缓存，并排除已知异常 DSXE 节点。"
+    - "B300 DSXE Dynamo+SGLang 的 DeepSeek-V4-Pro-0813 权重改为从节点本地 NVMe 副本（DeepSeek-V4-Pro-0813@scratch，与 vLLM 相同）加载，而非共享的 /data/models 副本；后者在多个测试点并发加载时会使启动慢数小时。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3190