diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-dspark4-maxseq2-mooncake-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-dspark4-maxseq2-mooncake-agentic.yaml
new file mode 100644
index 000000000..195125d25
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-dspark4-maxseq2-mooncake-agentic.yaml
@@ -0,0 +1,142 @@
+# Measured p90-ITL Pareto family: OCI jobs 471335-471337 and 471341-471342,
+# using one aggregate TP8 x DCP8 worker, DSpark K=4, and max-num-seqs 2.
+name: "kimi-k3-vllm-agg-gb300-dcp8-dspark4-maxseq2-mooncake-agentic"
+
+model:
+  path: "moonshotai/Kimi-K3"
+  container: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  frameworks:
+    dynamo: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
+
+dynamo:
+  hash: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
+  install: true
+
+slurm:
+  time_limit: "04:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  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
+  env:
+    DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600"
+    DYN_TOKENIZER_CACHE_BYTES: "8589934592"
+
+backend:
+  type: vllm
+  connector: null
+  mooncake_kv_store:
+    master_extra_args:
+      - "--eviction_high_watermark_ratio=0.95"
+      - "--eviction_ratio=0.10"
+    store_config:
+      metadata_server: "P2PHANDSHAKE"
+      global_segment_size: "150GB"
+      local_buffer_size: "4GB"
+      protocol: "rdma"
+      # Preserve the proven GB300 runner fabric rather than copying PR #211's
+      # single mlx5_0 rail, which is specific to the rail-isolated B300 fleet.
+      device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      mode: "embedded"
+      enable_offload: false
+  aggregated_environment:
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    # Preserve the measured collective fallback when moving to the built image.
+    VLLM_USE_DIRECT_DCP_A2A: "0"
+    VLLM_USE_DIRECT_DCP_Q_GATHER: "0"
+    VLLM_USE_DIRECT_DCP_KV_GATHER: "0"
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
+    VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: "0"
+    PYTHONNOUSERSITE: "1"
+    TORCH_CUDA_ARCH_LIST: "10.0"
+    PYTHONHASHSEED: "42"
+    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: "900"
+    MC_GID_INDEX: "3"
+    MC_STORE_MEMCPY: "1"
+    MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
+    MC_SLICE_SIZE: "1048576"
+    MC_WORKERS_PER_CTX: "4"
+    WITH_NVIDIA_PEERMEM: "0"
+    VLLM_MOONCAKE_LOAD_RECV_THREADS: "4"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 8
+      decode-context-parallel-size: 8
+      dcp-comm-backend: "a2a"
+      max-num-seqs: 2
+      gpu-memory-utilization: 0.90
+      max-num-batched-tokens: 8192
+      trust-remote-code: true
+      language-model-only: true
+      load-format: "fastsafetensors"
+      moe-backend: "auto"
+      no-enable-flashinfer-autotune: true
+      enable-cumem-allocator: true
+      enable-prefix-caching: true
+      prefix-match-unit: 128
+      kv-cache-dtype: "fp8"
+      stream-interval: 10
+      attention-backend: "TOKENSPEED_MLA"
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
+      # Throughput jobs inject the committed K=4 golden AL (3.36); EVAL_ONLY
+      # preserves this real target-verification configuration.
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":4,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+      max-cudagraph-capture-size: 1024
+      kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_offload":false}}'
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+  comment: >-
+    '{"OccupiedIdleGPUsJobReaper":{"exemptIdleTimeMins":"60","reason":"model_loading","description":"Very large model will take extra loading time."}}'
+
+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"
+    AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
+    AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_HTTP_TCP_USER_TIMEOUT: "900000"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-nospec-mooncake-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-nospec-mooncake-agentic.yaml
new file mode 100644
index 000000000..2b1c1d8cd
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-nospec-mooncake-agentic.yaml
@@ -0,0 +1,135 @@
+# Measured p90-ITL Pareto family: OCI jobs 471604 (c32), 471605 (c48), and
+# 471606 (c64), using one aggregate TP8 x DCP8 worker without a draft model.
+name: "kimi-k3-vllm-agg-gb300-dcp8-nospec-mooncake-agentic"
+
+model:
+  path: "moonshotai/Kimi-K3"
+  container: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  frameworks:
+    dynamo: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
+
+dynamo:
+  hash: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
+  install: true
+
+slurm:
+  time_limit: "04:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  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
+  env:
+    DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600"
+    DYN_TOKENIZER_CACHE_BYTES: "8589934592"
+
+backend:
+  type: vllm
+  connector: null
+  mooncake_kv_store:
+    master_extra_args:
+      - "--eviction_high_watermark_ratio=0.95"
+      - "--eviction_ratio=0.10"
+    store_config:
+      metadata_server: "P2PHANDSHAKE"
+      global_segment_size: "150GB"
+      local_buffer_size: "4GB"
+      protocol: "rdma"
+      device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      mode: "embedded"
+      enable_offload: false
+  aggregated_environment:
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    # Preserve the measured collective fallback when moving to the built image.
+    VLLM_USE_DIRECT_DCP_A2A: "0"
+    VLLM_USE_DIRECT_DCP_Q_GATHER: "0"
+    VLLM_USE_DIRECT_DCP_KV_GATHER: "0"
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_RPC_TIMEOUT: "600000"
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
+    VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: "0"
+    PYTHONNOUSERSITE: "1"
+    TORCH_CUDA_ARCH_LIST: "10.0"
+    PYTHONHASHSEED: "42"
+    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: "900"
+    MC_GID_INDEX: "3"
+    MC_STORE_MEMCPY: "1"
+    MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
+    MC_SLICE_SIZE: "1048576"
+    MC_WORKERS_PER_CTX: "4"
+    WITH_NVIDIA_PEERMEM: "0"
+    VLLM_MOONCAKE_LOAD_RECV_THREADS: "4"
+  vllm_config:
+    aggregated:
+      served-model-name: "moonshotai/Kimi-K3"
+      tensor-parallel-size: 8
+      decode-context-parallel-size: 8
+      dcp-comm-backend: "a2a"
+      gpu-memory-utilization: 0.90
+      max-num-batched-tokens: 16384
+      trust-remote-code: true
+      language-model-only: true
+      load-format: "fastsafetensors"
+      moe-backend: "auto"
+      no-enable-flashinfer-autotune: true
+      enable-prefix-caching: true
+      prefix-match-unit: 128
+      kv-cache-dtype: "fp8"
+      stream-interval: 10
+      attention-backend: "FLASHINFER_MLA"
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
+      max-cudagraph-capture-size: 1024
+      kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_offload":false}}'
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+  comment: >-
+    '{"OccupiedIdleGPUsJobReaper":{"exemptIdleTimeMins":"60","reason":"model_loading","description":"Very large model will take extra loading time."}}'
+
+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"
+    AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
+    AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
+    AIPERF_HTTP_TCP_USER_TIMEOUT: "900000"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p1d-dcp8-dcp8-dspark4-mooncake-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p1d-dcp8-dcp8-dspark4-mooncake-agentic.yaml
new file mode 100644
index 000000000..c79fbdd17
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p1d-dcp8-dcp8-dspark4-mooncake-agentic.yaml
@@ -0,0 +1,204 @@
+# Built vLLM image containing the Kimi-K3 DCP, DSpark, and PMU fixes.
+# Measured Pareto family: OCI jobs 469790 (c32), 469796 (c48), and 469792
+# (c64), using DSpark K=4 with synthetic acceptance length 3.36.
+name: "kimi-k3-vllm-disagg-gb300-1p1d-dcp8-dcp8-dspark4-mooncake-agentic"
+
+model:
+  path: "moonshotai/Kimi-K3"
+  container: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  frameworks:
+    dynamo: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
+
+dynamo:
+  hash: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
+  install: true
+
+slurm:
+  time_limit: "04:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  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
+  args:
+    router-mode: "random"
+    router-session-affinity-ttl-secs: 900
+  env:
+    DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600"
+    DYN_TOKENIZER_CACHE_BYTES: "8589934592"
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  kv_events_config:
+    prefill: true
+  mooncake_kv_store:
+    store_config:
+      metadata_server: "P2PHANDSHAKE"
+      # Keep the colleague's measured 150 GB Mooncake segment setting.
+      global_segment_size: "150GB"
+      local_buffer_size: "4GB"
+      protocol: "rdma"
+      device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      mode: "embedded"
+      enable_offload: false
+  prefill_environment: &kimi_env
+    # Preserve the measured collective fallback when moving to the built image.
+    VLLM_USE_DIRECT_DCP_A2A: "0"
+    VLLM_USE_DIRECT_DCP_Q_GATHER: "0"
+    VLLM_USE_DIRECT_DCP_KV_GATHER: "0"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    # No-op in this recipe: shared-expert sequence-parallel sharding requires
+    # expert parallelism, which is disabled on both roles.
+    VLLM_KIMI_K3_SHARD_SP_SHARED_EXPERT: "0"
+    DYN_REQUEST_PLANE: "tcp"
+    ETCD_LEASE_TTL: "600"
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_RPC_TIMEOUT: "600000"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "0"
+    # TORCH_SYMMMEM deliberately NOT set to NVSHMEM here.
+    # MegaMoE (decode role) allocates its symm buffer with device='cuda' (no
+    # index, deep_gemm/mega/__init__.py:48) and calls symm_mem.rendezvous. On
+    # the NVSHMEM backend that reaches a process-global TeamManager singleton
+    # pinned to the first caller's device (nvshmem_team_manager.hpp:36-44), so a
+    # TP>1 engine whose workers span devices 0-3 dies with
+    #   "Detected use of TeamManager on multiple devices."
+    # Measured: fine at TP=1 (DEP16 decode, job 2649641), fails at TP>1 (job
+    # 2649921 prefill). Leaving it unset selects torch's default CUDA symm-mem
+    # backend, which has no such singleton.
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_MOONCAKE_LOAD_RECV_THREADS: "4"
+    VLLM_CONNECTOR_PREFETCH_DEPTH: "8"
+    VLLM_CONNECTOR_PREFETCH_KV_CAP: "0.65"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
+    NCCL_P2P_LEVEL: NVL
+    MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
+    WITH_NVIDIA_PEERMEM: "0"
+    # NIXL (venv wheel) bundles its own UCX; NCCL's rdma_sharp_plugin loads
+    # /opt/hpcx UCX. Both in one process share UCM memory hooks and corrupt the
+    # heap -- NixlConnector then dies with
+    #   nixl_agent.cpp:470 registerMem: registration failed ... NIXL_ERR_BACKEND
+    # (and a "corrupted size vs. prev_size" abort at teardown). Disabling the
+    # external NCCL net plugin leaves exactly one UCX in the process; NCCL falls
+    # back to its built-in IB/NVLink transports. Verified on theia0263: 4 ranks,
+    # hpcx libs mapped = 0, 93-region registerMem OK on every rank.
+    NCCL_NET_PLUGIN: "none"
+    # UCX_MEMTYPE_CACHE=n stops UCX misclassifying a CUDA pointer it did not
+    # allocate as host memory; REG_WHOLE=n keeps it from registering the whole
+    # enclosing region instead of the requested range.
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_RCACHE_MAX_UNRELEASED: "1024"
+    UCX_TCP_AF_PRIO: "inet"
+    # K3's linear-attention layers hold Mamba conv state. NixlConnector's
+    # 3-read conv transfer asserts the DS (dim, state_len) layout at worker
+    # init; without this both roles die with "requires DS conv state layout".
+    # Must match on prefill and decode, hence the shared anchor.
+    VLLM_SSM_CONV_STATE_LAYOUT: "DS"
+  decode_environment:
+    <<: *kimi_env
+
+  vllm_config:
+    prefill:
+      # MultiConnector: Nixl (P->D handshake transfer) + Mooncake (external DRAM KV).
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      served-model-name: "moonshotai/Kimi-K3"
+      enable-prefix-caching: true
+      prefix-match-unit: 128
+      load-format: fastsafetensors
+      kv-cache-dtype: fp8
+      tensor-parallel-size: 8
+      # The point of this recipe: DCP reuses the 8 TP ranks and shards the MLA
+      # KV cache across them instead of replicating it, giving the prefill role
+      # ~7.5x the KV pool it has under plain TP8.
+      decode-context-parallel-size: 8
+      enable-cumem-allocator: true
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      max-cudagraph-capture-size: 512
+      stream-interval: 10
+      max_num_batched_tokens: 16384
+      language-model-only: true
+      attention-backend: "TOKENSPEED_MLA"
+      attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED", "use_prefill_query_quantization": true}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":4,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+    decode:
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      served-model-name: "moonshotai/Kimi-K3"
+      enable-prefix-caching: true
+      prefix-match-unit: 128
+      load-format: fastsafetensors
+      kv-cache-dtype: fp8
+      # Must equal the prefill TP for the NIXL MLA rank-to-rank mapping (see
+      # the header); DCP must likewise match at 8.
+      tensor-parallel-size: 8
+      decode-context-parallel-size: 8
+      enable-cumem-allocator: true
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      max-cudagraph-capture-size: 1024
+      stream-interval: 10
+      max-num-seqs: 256
+      language-model-only: true
+      attention-backend: "TOKENSPEED_MLA"
+      attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED", "use_prefill_query_quantization": true}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":4,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+  comment: >-
+    '{"OccupiedIdleGPUsJobReaper":{"exemptIdleTimeMins":"60","reason":"model_loading","description":"Very large model will take extra loading time."}}'
+
+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_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
+    # MODEL, topology, concurrency, duration, offload metadata, and result name
+    # are inherited from the generated InferenceMAX matrix.
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p1d-dcp8-dcp8-dspark7-mooncake-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p1d-dcp8-dcp8-dspark7-mooncake-agentic.yaml
new file mode 100644
index 000000000..6bd7a7fb6
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p1d-dcp8-dcp8-dspark7-mooncake-agentic.yaml
@@ -0,0 +1,204 @@
+# Built vLLM image containing the Kimi-K3 DCP, DSpark, and PMU fixes.
+# Measured Pareto family: OCI jobs 471040 (c1) and 471041 (c2), using DSpark
+# K=7 with synthetic acceptance length 3.84.
+name: "kimi-k3-vllm-disagg-gb300-1p1d-dcp8-dcp8-dspark7-mooncake-agentic"
+
+model:
+  path: "moonshotai/Kimi-K3"
+  container: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  frameworks:
+    dynamo: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
+
+dynamo:
+  hash: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
+  install: true
+
+slurm:
+  time_limit: "04:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  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
+  args:
+    router-mode: "random"
+    router-session-affinity-ttl-secs: 900
+  env:
+    DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600"
+    DYN_TOKENIZER_CACHE_BYTES: "8589934592"
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  kv_events_config:
+    prefill: true
+  mooncake_kv_store:
+    store_config:
+      metadata_server: "P2PHANDSHAKE"
+      # Keep the colleague's measured 150 GB Mooncake segment setting.
+      global_segment_size: "150GB"
+      local_buffer_size: "4GB"
+      protocol: "rdma"
+      device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      mode: "embedded"
+      enable_offload: false
+  prefill_environment: &kimi_env
+    # Preserve the measured collective fallback when moving to the built image.
+    VLLM_USE_DIRECT_DCP_A2A: "0"
+    VLLM_USE_DIRECT_DCP_Q_GATHER: "0"
+    VLLM_USE_DIRECT_DCP_KV_GATHER: "0"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    # No-op in this recipe: shared-expert sequence-parallel sharding requires
+    # expert parallelism, which is disabled on both roles.
+    VLLM_KIMI_K3_SHARD_SP_SHARED_EXPERT: "0"
+    DYN_REQUEST_PLANE: "tcp"
+    ETCD_LEASE_TTL: "600"
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_RPC_TIMEOUT: "600000"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "0"
+    # TORCH_SYMMMEM deliberately NOT set to NVSHMEM here.
+    # MegaMoE (decode role) allocates its symm buffer with device='cuda' (no
+    # index, deep_gemm/mega/__init__.py:48) and calls symm_mem.rendezvous. On
+    # the NVSHMEM backend that reaches a process-global TeamManager singleton
+    # pinned to the first caller's device (nvshmem_team_manager.hpp:36-44), so a
+    # TP>1 engine whose workers span devices 0-3 dies with
+    #   "Detected use of TeamManager on multiple devices."
+    # Measured: fine at TP=1 (DEP16 decode, job 2649641), fails at TP>1 (job
+    # 2649921 prefill). Leaving it unset selects torch's default CUDA symm-mem
+    # backend, which has no such singleton.
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_MOONCAKE_LOAD_RECV_THREADS: "4"
+    VLLM_CONNECTOR_PREFETCH_DEPTH: "8"
+    VLLM_CONNECTOR_PREFETCH_KV_CAP: "0.65"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
+    NCCL_P2P_LEVEL: NVL
+    MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
+    WITH_NVIDIA_PEERMEM: "0"
+    # NIXL (venv wheel) bundles its own UCX; NCCL's rdma_sharp_plugin loads
+    # /opt/hpcx UCX. Both in one process share UCM memory hooks and corrupt the
+    # heap -- NixlConnector then dies with
+    #   nixl_agent.cpp:470 registerMem: registration failed ... NIXL_ERR_BACKEND
+    # (and a "corrupted size vs. prev_size" abort at teardown). Disabling the
+    # external NCCL net plugin leaves exactly one UCX in the process; NCCL falls
+    # back to its built-in IB/NVLink transports. Verified on theia0263: 4 ranks,
+    # hpcx libs mapped = 0, 93-region registerMem OK on every rank.
+    NCCL_NET_PLUGIN: "none"
+    # UCX_MEMTYPE_CACHE=n stops UCX misclassifying a CUDA pointer it did not
+    # allocate as host memory; REG_WHOLE=n keeps it from registering the whole
+    # enclosing region instead of the requested range.
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_RCACHE_MAX_UNRELEASED: "1024"
+    UCX_TCP_AF_PRIO: "inet"
+    # K3's linear-attention layers hold Mamba conv state. NixlConnector's
+    # 3-read conv transfer asserts the DS (dim, state_len) layout at worker
+    # init; without this both roles die with "requires DS conv state layout".
+    # Must match on prefill and decode, hence the shared anchor.
+    VLLM_SSM_CONV_STATE_LAYOUT: "DS"
+  decode_environment:
+    <<: *kimi_env
+
+  vllm_config:
+    prefill:
+      # MultiConnector: Nixl (P->D handshake transfer) + Mooncake (external DRAM KV).
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      served-model-name: "moonshotai/Kimi-K3"
+      enable-prefix-caching: true
+      prefix-match-unit: 128
+      load-format: fastsafetensors
+      kv-cache-dtype: fp8
+      tensor-parallel-size: 8
+      # The point of this recipe: DCP reuses the 8 TP ranks and shards the MLA
+      # KV cache across them instead of replicating it, giving the prefill role
+      # ~7.5x the KV pool it has under plain TP8.
+      decode-context-parallel-size: 8
+      enable-cumem-allocator: true
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      max-cudagraph-capture-size: 512
+      stream-interval: 10
+      max_num_batched_tokens: 16384
+      language-model-only: true
+      attention-backend: "TOKENSPEED_MLA"
+      attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED", "use_prefill_query_quantization": true}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":7,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+    decode:
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      served-model-name: "moonshotai/Kimi-K3"
+      enable-prefix-caching: true
+      prefix-match-unit: 128
+      load-format: fastsafetensors
+      kv-cache-dtype: fp8
+      # Must equal the prefill TP for the NIXL MLA rank-to-rank mapping (see
+      # the header); DCP must likewise match at 8.
+      tensor-parallel-size: 8
+      decode-context-parallel-size: 8
+      enable-cumem-allocator: true
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      max-cudagraph-capture-size: 1024
+      stream-interval: 10
+      max-num-seqs: 256
+      language-model-only: true
+      attention-backend: "TOKENSPEED_MLA"
+      attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED", "use_prefill_query_quantization": true}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":7,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+  comment: >-
+    '{"OccupiedIdleGPUsJobReaper":{"exemptIdleTimeMins":"60","reason":"model_loading","description":"Very large model will take extra loading time."}}'
+
+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_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
+    # MODEL, topology, concurrency, duration, offload metadata, and result name
+    # are inherited from the generated InferenceMAX matrix.
diff --git a/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p2d-dcp8-dcp8-dspark4-mooncake-agentic.yaml b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p2d-dcp8-dcp8-dspark4-mooncake-agentic.yaml
new file mode 100644
index 000000000..c015f228d
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p2d-dcp8-dcp8-dspark4-mooncake-agentic.yaml
@@ -0,0 +1,204 @@
+# Built vLLM image containing the Kimi-K3 DCP, DSpark, and PMU fixes.
+# Measured p90-ITL Pareto point: OCI job 489701 (c48), using one DCP8
+# prefill worker, two DCP8 decode workers, DSpark K=4, and synthetic AL 3.36.
+name: "kimi-k3-vllm-disagg-gb300-1p2d-dcp8-dcp8-dspark4-mooncake-agentic"
+
+model:
+  path: "moonshotai/Kimi-K3"
+  container: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "moonshotai/Kimi-K3"
+  container:
+    image: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  frameworks:
+    dynamo: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
+
+dynamo:
+  hash: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
+  install: true
+
+slurm:
+  time_limit: "04:00:00"
+
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  prefill_nodes: 2
+  decode_nodes: 4
+  prefill_workers: 1
+  decode_workers: 2
+  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
+  args:
+    router-mode: "random"
+    router-session-affinity-ttl-secs: 900
+  env:
+    DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600"
+    DYN_TOKENIZER_CACHE_BYTES: "8589934592"
+backend:
+  type: vllm
+  connector: null
+  dp_launch_mode: per_node
+  kv_events_config:
+    prefill: true
+  mooncake_kv_store:
+    store_config:
+      metadata_server: "P2PHANDSHAKE"
+      # Keep the colleague's measured 150 GB Mooncake segment setting.
+      global_segment_size: "150GB"
+      local_buffer_size: "4GB"
+      protocol: "rdma"
+      device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      mode: "embedded"
+      enable_offload: false
+  prefill_environment: &kimi_env
+    # Preserve the measured collective fallback when moving to the built image.
+    VLLM_USE_DIRECT_DCP_A2A: "0"
+    VLLM_USE_DIRECT_DCP_Q_GATHER: "0"
+    VLLM_USE_DIRECT_DCP_KV_GATHER: "0"
+    VLLM_ALLREDUCE_USE_FLASHINFER: "1"
+    # No-op in this recipe: shared-expert sequence-parallel sharding requires
+    # expert parallelism, which is disabled on both roles.
+    VLLM_KIMI_K3_SHARD_SP_SHARED_EXPERT: "0"
+    DYN_REQUEST_PLANE: "tcp"
+    ETCD_LEASE_TTL: "600"
+    VLLM_ENGINE_READY_TIMEOUT_S: "3600"
+    VLLM_RPC_TIMEOUT: "600000"
+    TILELANG_CLEANUP_TEMP_FILES: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "0"
+    # TORCH_SYMMMEM deliberately NOT set to NVSHMEM here.
+    # MegaMoE (decode role) allocates its symm buffer with device='cuda' (no
+    # index, deep_gemm/mega/__init__.py:48) and calls symm_mem.rendezvous. On
+    # the NVSHMEM backend that reaches a process-global TeamManager singleton
+    # pinned to the first caller's device (nvshmem_team_manager.hpp:36-44), so a
+    # TP>1 engine whose workers span devices 0-3 dies with
+    #   "Detected use of TeamManager on multiple devices."
+    # Measured: fine at TP=1 (DEP16 decode, job 2649641), fails at TP>1 (job
+    # 2649921 prefill). Leaving it unset selects torch's default CUDA symm-mem
+    # backend, which has no such singleton.
+    NCCL_CUMEM_ENABLE: "1"
+    NCCL_MNNVL_ENABLE: "1"
+    NCCL_NVLS_ENABLE: "1"
+    VLLM_SERVER_DEV_MODE: "1"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_MOONCAKE_LOAD_RECV_THREADS: "4"
+    VLLM_CONNECTOR_PREFETCH_DEPTH: "8"
+    VLLM_CONNECTOR_PREFETCH_KV_CAP: "0.65"
+    VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
+    VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
+    NCCL_P2P_LEVEL: NVL
+    MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
+    WITH_NVIDIA_PEERMEM: "0"
+    # NIXL (venv wheel) bundles its own UCX; NCCL's rdma_sharp_plugin loads
+    # /opt/hpcx UCX. Both in one process share UCM memory hooks and corrupt the
+    # heap -- NixlConnector then dies with
+    #   nixl_agent.cpp:470 registerMem: registration failed ... NIXL_ERR_BACKEND
+    # (and a "corrupted size vs. prev_size" abort at teardown). Disabling the
+    # external NCCL net plugin leaves exactly one UCX in the process; NCCL falls
+    # back to its built-in IB/NVLink transports. Verified on theia0263: 4 ranks,
+    # hpcx libs mapped = 0, 93-region registerMem OK on every rank.
+    NCCL_NET_PLUGIN: "none"
+    # UCX_MEMTYPE_CACHE=n stops UCX misclassifying a CUDA pointer it did not
+    # allocate as host memory; REG_WHOLE=n keeps it from registering the whole
+    # enclosing region instead of the requested range.
+    UCX_MEMTYPE_CACHE: "n"
+    UCX_MEMTYPE_REG_WHOLE: "n"
+    UCX_RCACHE_MAX_UNRELEASED: "1024"
+    UCX_TCP_AF_PRIO: "inet"
+    # K3's linear-attention layers hold Mamba conv state. NixlConnector's
+    # 3-read conv transfer asserts the DS (dim, state_len) layout at worker
+    # init; without this both roles die with "requires DS conv state layout".
+    # Must match on prefill and decode, hence the shared anchor.
+    VLLM_SSM_CONV_STATE_LAYOUT: "DS"
+  decode_environment:
+    <<: *kimi_env
+
+  vllm_config:
+    prefill:
+      # MultiConnector: Nixl (P->D handshake transfer) + Mooncake (external DRAM KV).
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      served-model-name: "moonshotai/Kimi-K3"
+      enable-prefix-caching: true
+      prefix-match-unit: 128
+      load-format: fastsafetensors
+      kv-cache-dtype: fp8
+      tensor-parallel-size: 8
+      # The point of this recipe: DCP reuses the 8 TP ranks and shards the MLA
+      # KV cache across them instead of replicating it, giving the prefill role
+      # ~7.5x the KV pool it has under plain TP8.
+      decode-context-parallel-size: 8
+      enable-cumem-allocator: true
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      max-cudagraph-capture-size: 512
+      stream-interval: 10
+      max_num_batched_tokens: 16384
+      language-model-only: true
+      attention-backend: "TOKENSPEED_MLA"
+      attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED", "use_prefill_query_quantization": true}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":4,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+    decode:
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      served-model-name: "moonshotai/Kimi-K3"
+      enable-prefix-caching: true
+      prefix-match-unit: 128
+      load-format: fastsafetensors
+      kv-cache-dtype: fp8
+      # Must equal the prefill TP for the NIXL MLA rank-to-rank mapping (see
+      # the header); DCP must likewise match at 8.
+      tensor-parallel-size: 8
+      decode-context-parallel-size: 8
+      enable-cumem-allocator: true
+      trust-remote-code: true
+      no-enable-flashinfer-autotune: true
+      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
+      max-cudagraph-capture-size: 1024
+      stream-interval: 10
+      max-num-seqs: 256
+      language-model-only: true
+      attention-backend: "TOKENSPEED_MLA"
+      attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED", "use_prefill_query_quantization": true}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":4,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "72"
+  comment: >-
+    '{"OccupiedIdleGPUsJobReaper":{"exemptIdleTimeMins":"60","reason":"model_loading","description":"Very large model will take extra loading time."}}'
+
+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_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_DYNAMO_SESSION_TIMEOUT_SECONDS: "14400"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
+    WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
+    # MODEL, topology, concurrency, duration, offload metadata, and result name
+    # are inherited from the generated InferenceMAX matrix.
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index c8d55ad5d..77265b9cc 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8575,6 +8575,116 @@ qwen3.5-fp8-gb200-dynamo-sglang-mtp:
           ep: 16
           dp-attn: true
 
+# Kimi-K3 GB300 p90-ITL Pareto points. Each DCP8 worker spans two 4-GPU
+# trays: 1P1D uses 16 GPUs and 1P2D uses 24 GPUs.
+kimik3-fp4-gb300-dynamo-vllm-agentic-dspark-mooncake-dcp8-disagg:
+  image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:gb300-nv
+  precision: fp4
+  framework: dynamo-vllm
+  router: { name: dynamo-router, version: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1" }
+  kv-p2p-transfer: nixl
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    # 0.1664 resolves to 150 GB of host DRAM per 4-GPU server node on
+    # cluster:gb300-nv, matching the recipe's Mooncake segment size.
+    - dram-utilization: 0.1664
+      search-space:
+      - spec-decoding: mtp
+        kv-offloading: dram
+        kv-offload-backend: { name: mooncake, version: "0.3.12.post1" }
+        conc-list: [64]
+        prefill:
+          num-worker: 1
+          tp: 8
+          dcp-size: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.36"
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/disagg-gb300-1p1d-dcp8-dcp8-dspark4-mooncake-agentic.yaml"
+        decode: { num-worker: 1, tp: 8, dcp-size: 8, ep: 1, dp-attn: false }
+      - spec-decoding: mtp
+        kv-offloading: dram
+        kv-offload-backend: { name: mooncake, version: "0.3.12.post1" }
+        conc-list: [48]
+        prefill:
+          num-worker: 1
+          tp: 8
+          dcp-size: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.36"
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/disagg-gb300-1p2d-dcp8-dcp8-dspark4-mooncake-agentic.yaml"
+        decode: { num-worker: 2, tp: 8, dcp-size: 8, ep: 1, dp-attn: false }
+
+# Aggregate p90-ITL Pareto points. One TP8 x DCP8 worker spans two 4-GPU nodes
+# and serves both phases; decode num-worker is zero so GPUs are counted once.
+kimik3-fp4-gb300-dynamo-vllm-agentic-dspark-mooncake-dcp8-agg:
+  image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:gb300-nv
+  precision: fp4
+  framework: dynamo-vllm
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.1664
+      search-space:
+      - spec-decoding: mtp
+        kv-offloading: dram
+        kv-offload-backend: { name: mooncake, version: "0.3.12.post1" }
+        conc-list: [1, 4]
+        prefill:
+          num-worker: 1
+          tp: 8
+          dcp-size: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.36"
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-dspark4-maxseq2-mooncake-agentic.yaml"
+        decode: { num-worker: 0, tp: 8, dcp-size: 8, ep: 1, dp-attn: false }
+
+# High-throughput aggregate p90-ITL Pareto points without speculative decoding
+# or a max-num-seqs override.
+kimik3-fp4-gb300-dynamo-vllm-agentic-mooncake-dcp8-agg:
+  image: vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:gb300-nv
+  precision: fp4
+  framework: dynamo-vllm
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.1664
+      search-space:
+      - kv-offloading: dram
+        kv-offload-backend: { name: mooncake, version: "0.3.12.post1" }
+        conc-list: [48, 64, 70]
+        prefill:
+          num-worker: 1
+          tp: 8
+          dcp-size: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-nospec-mooncake-agentic.yaml"
+        decode: { num-worker: 0, tp: 8, dcp-size: 8, ep: 1, dp-attn: false }
+
+
 # Kimi-K3 MXFP4 B200 aggregated vLLM via Dynamo (TP8 x PP2, 2 nodes / 16
 # GPUs), agentic bring-up. The native MXFP4 checkpoint (2.8T total params,
 # ~1.4TB weights) does not fit one 8xB200 node, so TP8 shards attention/dense
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 73648d9b9..e834238a4 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6066,6 +6066,16 @@
   description:
     - "Add GB200 DeepSeek-V4-Pro FP4 Dynamo-vLLM AgentX mirroring the GB300 PR #2571 MTP tuning, with every GB300 4-GPU worker sized to 8 GPUs on GB200."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2619
+
+- config-keys:
+    - kimik3-fp4-gb300-dynamo-vllm-agentic-dspark-mooncake-dcp8-disagg
+    - kimik3-fp4-gb300-dynamo-vllm-agentic-dspark-mooncake-dcp8-agg
+    - kimik3-fp4-gb300-dynamo-vllm-agentic-mooncake-dcp8-agg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add Kimi-K3 GB300 AgentX configurations."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2637
   
 - config-keys:
     - dsv4-fp4-gb200-dynamo-vllm-agentic-mtp2-agg
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index 88b964a8a..9ac6980d6 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -86,6 +86,9 @@ elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" ]]; then
 elif [[ $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
     export MODEL_PATH=/scratch/models/Kimi-K2.5-NVFP4
     export SRT_SLURM_MODEL_PREFIX="nvidia/Kimi-K2.5-NVFP4"
+elif [[ $MODEL_PREFIX == "kimik3" && $PRECISION == "fp4" ]]; then
+    export MODEL_PATH=/scratch/models/Kimi-K3
+    export SRT_SLURM_MODEL_PREFIX="moonshotai/Kimi-K3"
 elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp4" ]]; then
     # SRT_SLURM_MODEL_PREFIX must match the model.path alias used in our
     # Qwen3.5 sglang recipes (qwen3.5-fp4).
@@ -97,7 +100,7 @@ elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp8" ]]; then
     export MODEL_PATH=/scratch/models/Qwen3.5-397B-A17B-FP8
     export SRT_SLURM_MODEL_PREFIX="qwen3.5-fp8"
 else
-    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4, glm5-fp4, glm5-fp8, glm5.2-fp4, minimaxm2.5-fp4, minimaxm2.5-fp8, kimik2.5-fp4, qwen3.5-fp4, qwen3.5-fp8"
+    echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4, glm5-fp4, glm5-fp8, glm5.2-fp4, minimaxm2.5-fp4, minimaxm2.5-fp8, kimik2.5-fp4, kimik3-fp4, qwen3.5-fp4, qwen3.5-fp8"
     exit 1
 fi
 
@@ -257,6 +260,9 @@ elif [[ "$IS_AGENTIC" == "1" ]]; then
     mkdir -p recipes/vllm/deepseek-v4/agentic || exit 1
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic" \
         recipes/vllm/deepseek-v4/agentic || exit 1
+    mkdir -p recipes/vllm/kimi-k3/agentic || exit 1
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic" \
+        recipes/vllm/kimi-k3/agentic || exit 1
 elif [[ $FRAMEWORK == "dynamo-vllm" && $MODEL_PREFIX == "dsv4" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"