diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb200-dep8-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb200-dep8-mtp-agentic.yaml
new file mode 100644
index 0000000000..9548372f23
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb200-dep8-mtp-agentic.yaml
@@ -0,0 +1,98 @@
+name: dsv4-gb200-vllm-agentic-mtp-agg-dep8
+
+model:
+  path: deepseek-v4-pro
+  container: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106
+  precision: fp4
+
+identity:
+  model: { repo: deepseek-ai/DeepSeek-V4-Pro }
+  container: { image: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106 }
+  frameworks: { dynamo: "1.3.1" }
+
+dynamo:
+  version: "1.3.1"
+  install: true
+
+setup_script: vllm-container-deps.sh
+environment: { ETCD_LEASE_TTL: "7200" }
+slurm: { time_limit: "8:00:00" }
+health_check: { max_attempts: 2160, interval_seconds: 10 }
+
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    router-mode: random
+    router-session-affinity-ttl-secs: 900
+
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  aggregated_environment:
+    HF_HUB_CACHE: /hf_hub_cache
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "32768"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "rc,cuda_copy"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+  vllm_config:
+    aggregated:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      kv-cache-dtype: fp8
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      moe-backend: deep_gemm_mega_moe
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
+      max-model-len: 1048576
+      max-num-seqs: 256
+      max-num-batched-tokens: 8192
+      trust-remote-code: true
+      block-size: 256
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}'
+      speculative-config: '{"method":"mtp","num_speculative_tokens":2}'
+      gpu-memory-utilization: 0.90
+      stream-interval: 10
+      no-disable-hybrid-kv-cache-manager: true
+      tokenizer-mode: deepseek_v4
+
+sbatch_directives: { cpus-per-task: "144", mem: "0" }
+srun_options: { container-remap-root: "" }
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb200-tp8-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb200-tp8-mtp-agentic.yaml
new file mode 100644
index 0000000000..bcbcd22059
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/agg-gb200-tp8-mtp-agentic.yaml
@@ -0,0 +1,108 @@
+name: dsv4-gb200-vllm-agentic-mtp-agg-tp8
+
+model:
+  path: deepseek-v4-pro
+  container: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106
+  precision: fp4
+
+identity:
+  model:
+    repo: deepseek-ai/DeepSeek-V4-Pro
+  container:
+    image: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106
+  frameworks:
+    dynamo: "1.3.1"
+
+dynamo:
+  version: "1.3.1"
+  install: true
+
+setup_script: vllm-container-deps.sh
+
+environment:
+  ETCD_LEASE_TTL: "7200"
+
+slurm:
+  time_limit: "8:00:00"
+
+health_check:
+  max_attempts: 2160
+  interval_seconds: 10
+
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    router-mode: random
+    router-session-affinity-ttl-secs: 900
+
+backend:
+  type: vllm
+  connector: null
+  aggregated_environment:
+    HF_HUB_CACHE: /hf_hub_cache
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "32768"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "rc,cuda_copy"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+  vllm_config:
+    aggregated:
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      kv-cache-dtype: fp8
+      tensor-parallel-size: 8
+      pipeline-parallel-size: 1
+      disable-custom-all-reduce: true
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
+      max-model-len: 1048576
+      max-num-seqs: 128
+      max-num-batched-tokens: 8192
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      block-size: 256
+      speculative-config: '{"method":"mtp","num_speculative_tokens":2}'
+      gpu-memory-utilization: 0.90
+      stream-interval: 10
+      no-disable-hybrid-kv-cache-manager: true
+      tokenizer-mode: deepseek_v4
+
+sbatch_directives:
+  cpus-per-task: "144"
+  mem: "0"
+
+srun_options:
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep8-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep8-mtp-agentic.yaml
new file mode 100644
index 0000000000..5226a8dd02
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep8-mtp-agentic.yaml
@@ -0,0 +1,118 @@
+name: dsv4-gb200-vllm-agentic-mtp-disagg-1p1d-dep8-dep8
+
+model:
+  path: deepseek-v4-pro
+  container: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106
+  precision: fp4
+
+identity:
+  model: { repo: deepseek-ai/DeepSeek-V4-Pro }
+  container: { image: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106 }
+  frameworks: { dynamo: "1.3.1" }
+
+dynamo:
+  version: "1.3.1"
+  install: true
+
+setup_script: vllm-container-deps.sh
+environment: { ETCD_LEASE_TTL: "7200" }
+slurm: { time_limit: "8:00:00" }
+health_check: { max_attempts: 2160, interval_seconds: 10 }
+
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 2
+  prefill_workers: 1
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    # Long AgentX prefills can exceed Dynamo's request-plane default while
+    # the healthy DEP8 worker is still computing the first response.
+    DYN_TCP_REQUEST_TIMEOUT: "60"
+  args:
+    router-mode: random
+    router-session-affinity-ttl-secs: 900
+
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  prefill_environment: &worker_environment
+    HF_HUB_CACHE: /hf_hub_cache
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "32768"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "rc,cuda_copy"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+  decode_environment: *worker_environment
+  vllm_config:
+    prefill: &dep8_config
+      kv-transfer-config: '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      kv-cache-dtype: fp8
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      moe-backend: deep_gemm_mega_moe
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
+      max-model-len: 1048576
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      trust-remote-code: true
+      enforce-eager: true
+      block-size: 256
+      gpu-memory-utilization: 0.95
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+      speculative-config: '{"method":"mtp","num_speculative_tokens":2}'
+    decode:
+      <<: *dep8_config
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_fp4_indexer_cache":true}'
+      enforce-eager: false
+      gpu-memory-utilization: 0.90
+      max-num-seqs: 1024
+      max-num-batched-tokens: 1024
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      max-cudagraph-capture-size: 1024
+      stream-interval: 10
+
+sbatch_directives: { cpus-per-task: "144", mem: "0" }
+srun_options: { container-remap-root: "" }
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-mtp-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-mtp-agentic.yaml
new file mode 100644
index 0000000000..c07d458126
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-mtp-agentic.yaml
@@ -0,0 +1,113 @@
+name: dsv4-gb200-vllm-agentic-mtp-disagg-2p1d-dep8-dep8
+
+model:
+  path: deepseek-v4-pro
+  container: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106
+  precision: fp4
+
+identity:
+  model: { repo: deepseek-ai/DeepSeek-V4-Pro }
+  container: { image: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106 }
+  frameworks: { dynamo: "1.3.1" }
+
+dynamo: { version: "1.3.1", install: true }
+setup_script: vllm-container-deps.sh
+environment: { ETCD_LEASE_TTL: "7200" }
+slurm: { time_limit: "8:00:00" }
+health_check: { max_attempts: 2160, interval_seconds: 10 }
+
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  prefill_nodes: 4
+  decode_nodes: 2
+  prefill_workers: 2
+  decode_workers: 1
+  gpus_per_prefill: 8
+  gpus_per_decode: 8
+
+infra: { etcd_nats_dedicated_node: false, nats_max_payload_mb: 32 }
+
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    # Long AgentX prefills can exceed Dynamo's request-plane default while
+    # the healthy DEP8 workers are still computing their first responses.
+    DYN_TCP_REQUEST_TIMEOUT: "60"
+  args:
+    router-mode: random
+    router-session-affinity-ttl-secs: 900
+
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  prefill_environment: &worker_environment
+    HF_HUB_CACHE: /hf_hub_cache
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_LOG_STATS_INTERVAL: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "32768"
+    VLLM_USE_NCCL_SYMM_MEM: "1"
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_NET_DEVICES: "mlx5_0:1,mlx5_1:1,mlx5_2:1,mlx5_3:1"
+    UCX_TLS: "rc,cuda_copy"
+    NCCL_IB_HCA: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+  decode_environment: *worker_environment
+  vllm_config:
+    prefill: &dep8_config
+      kv-transfer-config: '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      kv-cache-dtype: fp8
+      tensor-parallel-size: 1
+      pipeline-parallel-size: 1
+      data-parallel-size: 8
+      enable-expert-parallel: true
+      enable-ep-weight-filter: true
+      moe-backend: deep_gemm_mega_moe
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true,"use_fp4_indexer_cache":true}'
+      max-model-len: 1048576
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      trust-remote-code: true
+      enforce-eager: true
+      block-size: 256
+      gpu-memory-utilization: 0.95
+      no-disable-hybrid-kv-cache-manager: true
+      enable-sleep-mode: true
+      tokenizer-mode: deepseek_v4
+      speculative-config: '{"method":"mtp","num_speculative_tokens":2}'
+    decode:
+      <<: *dep8_config
+      attention-config: '{"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_fp4_indexer_cache":true}'
+      enforce-eager: false
+      gpu-memory-utilization: 0.90
+      max-num-seqs: 1024
+      max-num-batched-tokens: 1024
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      max-cudagraph-capture-size: 1024
+      stream-interval: 10
+
+sbatch_directives: { cpus-per-task: "144", mem: "0" }
+srun_options: { container-remap-root: "" }
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 5c049d9f0c..10c79ea22a 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7545,6 +7545,103 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-2p1d-dep8-dep8:
           ep: 8
           dp-attn: true
 
+dsv4-fp4-gb200-dynamo-vllm-agentic-mtp-agg:
+  image: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:gb200-nv
+  precision: fp4
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.3.1" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - search-space:
+      # Match the B200 TP8 latency points for direct normalized-interactivity
+      # comparison, then use the public vLLM GB200 DEP8 topology for throughput.
+      - spec-decoding: mtp
+        conc-list: [1, 2, 4, 6, 8]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/agg-gb200-tp8-mtp-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 1
+          dp-attn: false
+      - spec-decoding: mtp
+        conc-list: [8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 60, 64, 68, 72, 76, 80]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/agg-gb200-dep8-mtp-agentic.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 8
+          dp-attn: true
+
+dsv4-fp4-gb200-dynamo-vllm-agentic-mtp-disagg:
+  image: vllm/vllm-openai:nightly-3ee2df30337a301164c46ae444b76ee67e71c106
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:gb200-nv
+  precision: fp4
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "1.3.1" }
+  kv-p2p-transfer: nixl
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - search-space:
+      # One two-tray DEP8 prefill and decode worker covers the balanced region.
+      - spec-decoding: mtp
+        conc-list: [32, 64, 96, 128, 192, 256]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep8-mtp-agentic.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      # A second DEP8 prefill worker extends the high-concurrency region.
+      - spec-decoding: mtp
+        conc-list: [192, 256, 384, 512]
+        prefill:
+          num-worker: 2
+          tp: 8
+          ep: 8
+          dp-attn: true
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=2.49"
+          - "CONFIG_FILE=recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep8-mtp-agentic.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+
 dsv4-fp4-gb300-dynamo-vllm-agentic-mtp-agg:
   image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-426e59f
   model: deepseek-ai/DeepSeek-V4-Pro
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 676fc73e7f..d082ced48f 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5902,3 +5902,14 @@
     - "Use NIXL for KV transfer with EAGLE speculative decoding and chat-formatted benchmark inputs."
     - "Image: lmsysorg/sglang:nightly-dev-cu13-20260710-cfc66e05"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2554
+
+- config-keys:
+    - dsv4-fp4-gb200-dynamo-vllm-agentic-mtp-agg
+    - dsv4-fp4-gb200-dynamo-vllm-agentic-mtp-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add GB200 DeepSeek-V4-Pro FP4 Dynamo-vLLM AgentX with K=2 MTP and golden synthetic acceptance length 2.49."
+    - "Compare the B200 TP8 and DEP8 concurrency range and extend throughput with 1P1D and 2P1D DEP8 disaggregation."
+    - "Retain prefix-cache blocks for 32,768 scheduling steps and export every logical vLLM server-metrics endpoint."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2567
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index 660ae078d5..ed4cbed78b 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -391,18 +391,24 @@ if [ -d "$SRT_REPO_DIR" ]; then
     rm -rf "$SRT_REPO_DIR"
 fi
 
-# Qwen3.5 FP4 AgentX uses the latest released srt-slurm. v1.0.45 injects
-# the aggregate logical-worker Prometheus endpoint into custom benchmarks.
-if [[ "$IS_AGENTIC" == "1" && "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "dynamo-sglang" ]]; then
+# These AgentX submissions use released srt-slurm custom-benchmark metrics
+# discovery so AIPerf receives every logical worker endpoint.
+if [[ "$IS_AGENTIC" == "1" && (( "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "dynamo-sglang" ) || ( "$MODEL_PREFIX" == "dsv4" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "dynamo-vllm" )) ]]; then
     git clone --branch v1.0.45 --single-branch https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
     test "$(git rev-parse HEAD)" = "9d8d92b20c350a5d42f0709f5a0b64e30eb37d33" || {
         echo "Error: NVIDIA/srt-slurm v1.0.45 resolved to an unexpected commit" >&2
         exit 1
     }
-    mkdir -p recipes/sglang/qwen3.5/gb200-fp4/agentic
-    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4/agentic" \
-        recipes/sglang/qwen3.5/gb200-fp4/agentic
+    if [[ "$MODEL_PREFIX" == "qwen3.5" ]]; then
+        mkdir -p recipes/sglang/qwen3.5/gb200-fp4/agentic
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4/agentic" \
+            recipes/sglang/qwen3.5/gb200-fp4/agentic
+    else
+        mkdir -p recipes/vllm/deepseek-v4/agentic
+        cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
+            recipes/vllm/deepseek-v4/agentic
+    fi
 # TODO(CJQ): migrate the remaining Agentic model paths to released srt-slurm.
 elif [[ "$IS_AGENTIC" == "1" ]]; then
     # Agentic multi-node pins cquil11/srt-slurm-nv revisions that provide: