diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-15p1d-dep1-dep16-c606-b32-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-15p1d-dep1-dep16-c606-b32-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..1ef28e9a55
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-15p1d-dep1-dep16-c606-b32-mtp-kvoffload.yaml
@@ -0,0 +1,217 @@
+name: disagg-gb300-15p1d-dep1-dep16-c606-b32-mtp-kvoffload
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  precision: fp4
+
+dynamo:
+  install: true
+  wheel: 1.4.0.dev20260807
+
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  frameworks:
+    tensorrt_llm: 1.3.0rc24
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 4
+  prefill_workers: 15
+  gpus_per_prefill: 1
+  decode_nodes: 4
+  decode_workers: 1
+  gpus_per_decode: 16
+backend:
+  type: trtllm
+  sequential_node_start: 2
+  prefill_environment:
+    CUDA_SCALE_LAUNCH_QUEUES: 4x
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  decode_environment:
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  trtllm_config:
+    prefill:
+      attention_dp_config:
+        kv_cache_routing_conversation_affinity: true
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: false
+      enable_attention_dp: true
+      enable_chunked_prefill: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      kv_cache_config:
+        block_reuse_config:
+          max_num_turns: 3
+          policy: per_conversation
+        dtype: fp8
+        enable_block_reuse: true
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.8
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        mamba_state_config:
+          additional_snapshot_offsets_from_end:
+          - 2
+          periodic_snapshot_interval: 0
+        pool_ratio:
+        - 0.8
+        - 0.2
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 8192
+      max_seq_len: 262144
+      moe_config:
+        backend: CUTEDSL
+      moe_expert_parallel_size: 1
+      num_serve_frontends: 8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+      tensor_parallel_size: 1
+      torch_compile_config:
+        capture_num_tokens:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 32
+        - 64
+        - 128
+        - 256
+        - 512
+        - 1024
+        - 2048
+        - 4096
+        - 8192
+        enable_fullgraph: true
+        enable_piecewise_cuda_graph: true
+      trust_remote_code: true
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 24
+        - 32
+        enable_padding: true
+      enable_attention_dp: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        avg_seq_len: 102150
+        dtype: fp8
+        enable_block_reuse: false
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.85
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 32
+      max_num_tokens: 224
+      max_seq_len: 262148
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 16
+      num_postprocess_workers: 4
+      num_serve_frontends: 4
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: GUARANTEED_NO_EVICT
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 6
+      stream_interval: 20
+      tensor_parallel_size: 16
+      trust_remote_code: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: "120"
+    DYN_TOKENIZER_CACHE: "1"
+    DYN_TOKENIZER_CACHE_BYTES: "8000000000"
+    DYN_TOKENIZER: "fastokens"
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '14400'
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    SERVED_MODEL_NAME: Qwen3.5-397B-A17B-NVFP4-V2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-17p3d-dep2-dep8-c1452-b64-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-17p3d-dep2-dep8-c1452-b64-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..5377338e5b
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-17p3d-dep2-dep8-c1452-b64-mtp-kvoffload.yaml
@@ -0,0 +1,219 @@
+name: disagg-gb300-17p3d-dep2-dep8-c1452-b64-mtp-kvoffload
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  precision: fp4
+
+dynamo:
+  install: true
+  wheel: 1.4.0.dev20260807
+
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  frameworks:
+    tensorrt_llm: 1.3.0rc24
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 9
+  prefill_workers: 17
+  gpus_per_prefill: 2
+  decode_nodes: 6
+  decode_workers: 3
+  gpus_per_decode: 8
+backend:
+  type: trtllm
+  sequential_node_start: 2
+  prefill_environment:
+    CUDA_SCALE_LAUNCH_QUEUES: 4x
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  decode_environment:
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  trtllm_config:
+    prefill:
+      attention_dp_config:
+        kv_cache_routing_conversation_affinity: true
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: false
+      enable_attention_dp: true
+      enable_chunked_prefill: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      kv_cache_config:
+        block_reuse_config:
+          max_num_turns: 3
+          policy: per_conversation
+        dtype: fp8
+        enable_block_reuse: true
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.8
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        mamba_state_config:
+          additional_snapshot_offsets_from_end:
+          - 2
+          periodic_snapshot_interval: 0
+        pool_ratio:
+        - 0.8
+        - 0.2
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 8192
+      max_seq_len: 262144
+      moe_config:
+        backend: CUTEDSL
+      moe_expert_parallel_size: 2
+      num_serve_frontends: 8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+      tensor_parallel_size: 2
+      torch_compile_config:
+        capture_num_tokens:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 32
+        - 64
+        - 128
+        - 256
+        - 512
+        - 1024
+        - 2048
+        - 4096
+        - 8192
+        enable_fullgraph: true
+        enable_piecewise_cuda_graph: true
+      trust_remote_code: true
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 24
+        - 32
+        - 40
+        - 48
+        - 56
+        - 64
+        enable_padding: true
+      enable_attention_dp: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        avg_seq_len: 102150
+        dtype: fp8
+        enable_block_reuse: false
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.85
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 448
+      max_seq_len: 262148
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 8
+      num_postprocess_workers: 4
+      num_serve_frontends: 4
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: GUARANTEED_NO_EVICT
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 6
+      stream_interval: 20
+      tensor_parallel_size: 8
+      trust_remote_code: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: "120"
+    DYN_TOKENIZER_CACHE: "1"
+    DYN_TOKENIZER_CACHE_BYTES: "8000000000"
+    DYN_TOKENIZER: "fastokens"
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '14400'
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    SERVED_MODEL_NAME: Qwen3.5-397B-A17B-NVFP4-V2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-19p1d-dep1-dep16-c225-b16-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-19p1d-dep1-dep16-c225-b16-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..46b7341885
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-19p1d-dep1-dep16-c225-b16-mtp-kvoffload.yaml
@@ -0,0 +1,215 @@
+name: disagg-gb300-19p1d-dep1-dep16-c225-b16-mtp-kvoffload
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  precision: fp4
+
+dynamo:
+  install: true
+  wheel: 1.4.0.dev20260807
+
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  frameworks:
+    tensorrt_llm: 1.3.0rc24
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 5
+  prefill_workers: 19
+  gpus_per_prefill: 1
+  decode_nodes: 4
+  decode_workers: 1
+  gpus_per_decode: 16
+backend:
+  type: trtllm
+  sequential_node_start: 2
+  prefill_environment:
+    CUDA_SCALE_LAUNCH_QUEUES: 4x
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.41'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  decode_environment:
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.41'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  trtllm_config:
+    prefill:
+      attention_dp_config:
+        kv_cache_routing_conversation_affinity: true
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: false
+      enable_attention_dp: true
+      enable_chunked_prefill: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      kv_cache_config:
+        block_reuse_config:
+          max_num_turns: 3
+          policy: per_conversation
+        dtype: fp8
+        enable_block_reuse: true
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.8
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        mamba_state_config:
+          additional_snapshot_offsets_from_end:
+          - 2
+          periodic_snapshot_interval: 0
+        pool_ratio:
+        - 0.8
+        - 0.2
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 8192
+      max_seq_len: 262144
+      moe_config:
+        backend: CUTEDSL
+      moe_expert_parallel_size: 1
+      num_serve_frontends: 8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+      tensor_parallel_size: 1
+      torch_compile_config:
+        capture_num_tokens:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 32
+        - 64
+        - 128
+        - 256
+        - 512
+        - 1024
+        - 2048
+        - 4096
+        - 8192
+        enable_fullgraph: true
+        enable_piecewise_cuda_graph: true
+      trust_remote_code: true
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        enable_padding: true
+      enable_attention_dp: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        avg_seq_len: 102150
+        dtype: fp8
+        enable_block_reuse: false
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.85
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 16
+      max_num_tokens: 96
+      max_seq_len: 262148
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 16
+      num_postprocess_workers: 4
+      num_serve_frontends: 4
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: GUARANTEED_NO_EVICT
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 5
+      stream_interval: 20
+      tensor_parallel_size: 16
+      trust_remote_code: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: "120"
+    DYN_TOKENIZER_CACHE: "1"
+    DYN_TOKENIZER_CACHE_BYTES: "8000000000"
+    DYN_TOKENIZER: "fastokens"
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '14400'
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    SERVED_MODEL_NAME: Qwen3.5-397B-A17B-NVFP4-V2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-19p1d-dep1-dep16-c449-b16-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-19p1d-dep1-dep16-c449-b16-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..c462435df8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-19p1d-dep1-dep16-c449-b16-mtp-kvoffload.yaml
@@ -0,0 +1,215 @@
+name: disagg-gb300-19p1d-dep1-dep16-c449-b16-mtp-kvoffload
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  precision: fp4
+
+dynamo:
+  install: true
+  wheel: 1.4.0.dev20260807
+
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  frameworks:
+    tensorrt_llm: 1.3.0rc24
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 5
+  prefill_workers: 19
+  gpus_per_prefill: 1
+  decode_nodes: 4
+  decode_workers: 1
+  gpus_per_decode: 16
+backend:
+  type: trtllm
+  sequential_node_start: 2
+  prefill_environment:
+    CUDA_SCALE_LAUNCH_QUEUES: 4x
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.41'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  decode_environment:
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.41'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    TRT_LLM_DISABLE_LOAD_WEIGHTS_IN_PARALLEL: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  trtllm_config:
+    prefill:
+      attention_dp_config:
+        kv_cache_routing_conversation_affinity: true
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: false
+      enable_attention_dp: true
+      enable_chunked_prefill: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      kv_cache_config:
+        block_reuse_config:
+          max_num_turns: 3
+          policy: per_conversation
+        dtype: fp8
+        enable_block_reuse: true
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.8
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        mamba_state_config:
+          additional_snapshot_offsets_from_end:
+          - 2
+          periodic_snapshot_interval: 0
+        pool_ratio:
+        - 0.8
+        - 0.2
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 8192
+      max_seq_len: 262144
+      moe_config:
+        backend: CUTEDSL
+      moe_expert_parallel_size: 1
+      num_serve_frontends: 8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+      tensor_parallel_size: 1
+      torch_compile_config:
+        capture_num_tokens:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 32
+        - 64
+        - 128
+        - 256
+        - 512
+        - 1024
+        - 2048
+        - 4096
+        - 8192
+        enable_fullgraph: true
+        enable_piecewise_cuda_graph: true
+      trust_remote_code: true
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        enable_padding: true
+      enable_attention_dp: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        avg_seq_len: 102150
+        dtype: fp8
+        enable_block_reuse: false
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.85
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 16
+      max_num_tokens: 96
+      max_seq_len: 262148
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 16
+      num_postprocess_workers: 4
+      num_serve_frontends: 4
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: GUARANTEED_NO_EVICT
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 5
+      stream_interval: 20
+      tensor_parallel_size: 16
+      trust_remote_code: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: "120"
+    DYN_TOKENIZER_CACHE: "1"
+    DYN_TOKENIZER_CACHE_BYTES: "8000000000"
+    DYN_TOKENIZER: "fastokens"
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '14400'
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    SERVED_MODEL_NAME: Qwen3.5-397B-A17B-NVFP4-V2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-1p1d-tep2-tep4-c8-b128-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-1p1d-tep2-tep4-c8-b128-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..fddf6697eb
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-1p1d-tep2-tep4-c8-b128-mtp-kvoffload.yaml
@@ -0,0 +1,229 @@
+name: disagg-gb300-1p1d-tep2-tep4-c8-b128-mtp-kvoffload
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  precision: fp4
+
+dynamo:
+  install: true
+  wheel: 1.4.0.dev20260807
+
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  frameworks:
+    tensorrt_llm: 1.3.0rc24
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 2
+  decode_nodes: 1
+  decode_workers: 1
+  gpus_per_decode: 4
+backend:
+  type: trtllm
+  prefill_environment:
+    CUDA_SCALE_LAUNCH_QUEUES: 4x
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_KV_TRANSFER_NUM_THREADS: '8'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  decode_environment:
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_KV_TRANSFER_NUM_THREADS: '8'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  trtllm_config:
+    prefill:
+      attention_dp_config:
+        kv_cache_routing_conversation_affinity: true
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: false
+      enable_attention_dp: false
+      enable_chunked_prefill: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      kv_cache_config:
+        block_reuse_config:
+          max_num_turns: 3
+          policy: per_conversation
+        dtype: fp8
+        enable_block_reuse: true
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.8
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        mamba_state_config:
+          additional_snapshot_offsets_from_end:
+          - 2
+          periodic_snapshot_interval: 0
+        pool_ratio:
+        - 0.8
+        - 0.2
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 16384
+      max_seq_len: 262144
+      moe_config:
+        backend: CUTEDSL
+      moe_expert_parallel_size: 2
+      num_serve_frontends: 8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+      tensor_parallel_size: 2
+      torch_compile_config:
+        capture_num_tokens:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 32
+        - 64
+        - 128
+        - 256
+        - 512
+        - 1024
+        - 2048
+        - 4096
+        - 8192
+        - 16384
+        enable_fullgraph: true
+        enable_piecewise_cuda_graph: true
+      trust_remote_code: true
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 24
+        - 32
+        - 40
+        - 48
+        - 56
+        - 64
+        - 72
+        - 80
+        - 88
+        - 96
+        - 104
+        - 112
+        - 120
+        - 128
+        enable_padding: true
+      enable_attention_dp: false
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        avg_seq_len: 102150
+        dtype: fp8
+        enable_block_reuse: false
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.85
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 128
+      max_num_tokens: 896
+      max_seq_len: 262148
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 4
+      num_postprocess_workers: 4
+      num_serve_frontends: 4
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: GUARANTEED_NO_EVICT
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 6
+      stream_interval: 20
+      tensor_parallel_size: 4
+      trust_remote_code: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: "120"
+    DYN_TOKENIZER_CACHE: "1"
+    DYN_TOKENIZER_CACHE_BYTES: "8000000000"
+    DYN_TOKENIZER: "fastokens"
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '14400'
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    SERVED_MODEL_NAME: Qwen3.5-397B-A17B-NVFP4-V2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep4-c24-b128-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep4-c24-b128-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..a3cec9ad8a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep4-c24-b128-mtp-kvoffload.yaml
@@ -0,0 +1,230 @@
+name: disagg-gb300-2p3d-tep2-tep4-c24-b128-mtp-kvoffload
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  precision: fp4
+
+dynamo:
+  install: true
+  wheel: 1.4.0.dev20260807
+
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  frameworks:
+    tensorrt_llm: 1.3.0rc24
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 2
+  gpus_per_prefill: 2
+  decode_nodes: 3
+  decode_workers: 3
+  gpus_per_decode: 4
+backend:
+  type: trtllm
+  sequential_node_start: 1
+  prefill_environment:
+    CUDA_SCALE_LAUNCH_QUEUES: 4x
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_KV_TRANSFER_NUM_THREADS: '8'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  decode_environment:
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_KV_TRANSFER_NUM_THREADS: '8'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  trtllm_config:
+    prefill:
+      attention_dp_config:
+        kv_cache_routing_conversation_affinity: true
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: false
+      enable_attention_dp: false
+      enable_chunked_prefill: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      kv_cache_config:
+        block_reuse_config:
+          max_num_turns: 3
+          policy: per_conversation
+        dtype: fp8
+        enable_block_reuse: true
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.8
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        mamba_state_config:
+          additional_snapshot_offsets_from_end:
+          - 2
+          periodic_snapshot_interval: 0
+        pool_ratio:
+        - 0.8
+        - 0.2
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 16384
+      max_seq_len: 262144
+      moe_config:
+        backend: CUTEDSL
+      moe_expert_parallel_size: 2
+      num_serve_frontends: 8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+      tensor_parallel_size: 2
+      torch_compile_config:
+        capture_num_tokens:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 32
+        - 64
+        - 128
+        - 256
+        - 512
+        - 1024
+        - 2048
+        - 4096
+        - 8192
+        - 16384
+        enable_fullgraph: true
+        enable_piecewise_cuda_graph: true
+      trust_remote_code: true
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 24
+        - 32
+        - 40
+        - 48
+        - 56
+        - 64
+        - 72
+        - 80
+        - 88
+        - 96
+        - 104
+        - 112
+        - 120
+        - 128
+        enable_padding: true
+      enable_attention_dp: false
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        avg_seq_len: 102150
+        dtype: fp8
+        enable_block_reuse: false
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.85
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 128
+      max_num_tokens: 896
+      max_seq_len: 262148
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 4
+      num_postprocess_workers: 4
+      num_serve_frontends: 4
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: GUARANTEED_NO_EVICT
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 6
+      stream_interval: 20
+      tensor_parallel_size: 4
+      trust_remote_code: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: "120"
+    DYN_TOKENIZER_CACHE: "1"
+    DYN_TOKENIZER_CACHE_BYTES: "8000000000"
+    DYN_TOKENIZER: "fastokens"
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '14400'
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    SERVED_MODEL_NAME: Qwen3.5-397B-A17B-NVFP4-V2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep4-c6-b128-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep4-c6-b128-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..68bcb278d1
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep4-c6-b128-mtp-kvoffload.yaml
@@ -0,0 +1,230 @@
+name: disagg-gb300-2p3d-tep2-tep4-c6-b128-mtp-kvoffload
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  precision: fp4
+
+dynamo:
+  install: true
+  wheel: 1.4.0.dev20260807
+
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  frameworks:
+    tensorrt_llm: 1.3.0rc24
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 2
+  gpus_per_prefill: 2
+  decode_nodes: 3
+  decode_workers: 3
+  gpus_per_decode: 4
+backend:
+  type: trtllm
+  sequential_node_start: 1
+  prefill_environment:
+    CUDA_SCALE_LAUNCH_QUEUES: 4x
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_KV_TRANSFER_NUM_THREADS: '8'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  decode_environment:
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_KV_TRANSFER_NUM_THREADS: '8'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  trtllm_config:
+    prefill:
+      attention_dp_config:
+        kv_cache_routing_conversation_affinity: true
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: false
+      enable_attention_dp: false
+      enable_chunked_prefill: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      kv_cache_config:
+        block_reuse_config:
+          max_num_turns: 3
+          policy: per_conversation
+        dtype: fp8
+        enable_block_reuse: true
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.8
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        mamba_state_config:
+          additional_snapshot_offsets_from_end:
+          - 2
+          periodic_snapshot_interval: 0
+        pool_ratio:
+        - 0.8
+        - 0.2
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 16384
+      max_seq_len: 262144
+      moe_config:
+        backend: CUTEDSL
+      moe_expert_parallel_size: 2
+      num_serve_frontends: 8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+      tensor_parallel_size: 2
+      torch_compile_config:
+        capture_num_tokens:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 32
+        - 64
+        - 128
+        - 256
+        - 512
+        - 1024
+        - 2048
+        - 4096
+        - 8192
+        - 16384
+        enable_fullgraph: true
+        enable_piecewise_cuda_graph: true
+      trust_remote_code: true
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 24
+        - 32
+        - 40
+        - 48
+        - 56
+        - 64
+        - 72
+        - 80
+        - 88
+        - 96
+        - 104
+        - 112
+        - 120
+        - 128
+        enable_padding: true
+      enable_attention_dp: false
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        avg_seq_len: 102150
+        dtype: fp8
+        enable_block_reuse: false
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.85
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 128
+      max_num_tokens: 896
+      max_seq_len: 262148
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 4
+      num_postprocess_workers: 4
+      num_serve_frontends: 4
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: GUARANTEED_NO_EVICT
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 6
+      stream_interval: 20
+      tensor_parallel_size: 4
+      trust_remote_code: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: "120"
+    DYN_TOKENIZER_CACHE: "1"
+    DYN_TOKENIZER_CACHE_BYTES: "8000000000"
+    DYN_TOKENIZER: "fastokens"
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '14400'
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    SERVED_MODEL_NAME: Qwen3.5-397B-A17B-NVFP4-V2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep8-c6-b128-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep8-c6-b128-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..6be4e07c9a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep8-c6-b128-mtp-kvoffload.yaml
@@ -0,0 +1,230 @@
+name: disagg-gb300-2p3d-tep2-tep8-c6-b128-mtp-kvoffload
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  precision: fp4
+
+dynamo:
+  install: true
+  wheel: 1.4.0.dev20260807
+
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  frameworks:
+    tensorrt_llm: 1.3.0rc24
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 2
+  gpus_per_prefill: 2
+  decode_nodes: 6
+  decode_workers: 3
+  gpus_per_decode: 8
+backend:
+  type: trtllm
+  sequential_node_start: 1
+  prefill_environment:
+    CUDA_SCALE_LAUNCH_QUEUES: 4x
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_KV_TRANSFER_NUM_THREADS: '8'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  decode_environment:
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_KV_TRANSFER_NUM_THREADS: '8'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  trtllm_config:
+    prefill:
+      attention_dp_config:
+        kv_cache_routing_conversation_affinity: true
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: false
+      enable_attention_dp: false
+      enable_chunked_prefill: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      kv_cache_config:
+        block_reuse_config:
+          max_num_turns: 3
+          policy: per_conversation
+        dtype: fp8
+        enable_block_reuse: true
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.8
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        mamba_state_config:
+          additional_snapshot_offsets_from_end:
+          - 2
+          periodic_snapshot_interval: 0
+        pool_ratio:
+        - 0.8
+        - 0.2
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 16384
+      max_seq_len: 262144
+      moe_config:
+        backend: CUTEDSL
+      moe_expert_parallel_size: 2
+      num_serve_frontends: 8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+      tensor_parallel_size: 2
+      torch_compile_config:
+        capture_num_tokens:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 32
+        - 64
+        - 128
+        - 256
+        - 512
+        - 1024
+        - 2048
+        - 4096
+        - 8192
+        - 16384
+        enable_fullgraph: true
+        enable_piecewise_cuda_graph: true
+      trust_remote_code: true
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 24
+        - 32
+        - 40
+        - 48
+        - 56
+        - 64
+        - 72
+        - 80
+        - 88
+        - 96
+        - 104
+        - 112
+        - 120
+        - 128
+        enable_padding: true
+      enable_attention_dp: false
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      enable_lm_head_tp_in_adp: false
+      kv_cache_config:
+        avg_seq_len: 102150
+        dtype: fp8
+        enable_block_reuse: false
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.85
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 128
+      max_num_tokens: 896
+      max_seq_len: 262148
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 8
+      num_postprocess_workers: 4
+      num_serve_frontends: 4
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: GUARANTEED_NO_EVICT
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 6
+      stream_interval: 20
+      tensor_parallel_size: 8
+      trust_remote_code: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: "120"
+    DYN_TOKENIZER_CACHE: "1"
+    DYN_TOKENIZER_CACHE_BYTES: "8000000000"
+    DYN_TOKENIZER: "fastokens"
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '14400'
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    SERVED_MODEL_NAME: Qwen3.5-397B-A17B-NVFP4-V2
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-4p1d-dep4-dep8-c896-b64-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-4p1d-dep4-dep8-c896-b64-mtp-kvoffload.yaml
new file mode 100644
index 0000000000..19d999cf63
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-4p1d-dep4-dep8-c896-b64-mtp-kvoffload.yaml
@@ -0,0 +1,218 @@
+name: disagg-gb300-4p1d-dep4-dep8-c896-b64-mtp-kvoffload
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  precision: fp4
+
+dynamo:
+  install: true
+  wheel: 1.4.0.dev20260807
+
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  frameworks:
+    tensorrt_llm: 1.3.0rc24
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 4
+  prefill_workers: 4
+  gpus_per_prefill: 4
+  decode_nodes: 2
+  decode_workers: 1
+  gpus_per_decode: 8
+backend:
+  type: trtllm
+  prefill_environment:
+    CUDA_SCALE_LAUNCH_QUEUES: 4x
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  decode_environment:
+    MIMALLOC_ARENA_RESERVE: '0'
+    MIMALLOC_PURGE_DELAY: ''
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    PYTORCH_CUDA_ALLOC_CONF: expandable_segments:True
+    TLLM_LOG_LEVEL: INFO
+    TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS: '3.80'
+    TRTLLM_ENABLE_PDL: '1'
+    TRTLLM_PINNED_WEIGHT_STAGING: '1'
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    UCX_TLS: cuda_ipc,cuda_copy,sm,self,tcp
+    DYN_ENGINE_CONV_AFFINITY: "1"
+  trtllm_config:
+    prefill:
+      attention_dp_config:
+        kv_cache_routing_conversation_affinity: true
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config: null
+      disable_overlap_scheduler: false
+      enable_attention_dp: true
+      enable_chunked_prefill: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      kv_cache_config:
+        block_reuse_config:
+          max_num_turns: 3
+          policy: per_conversation
+        dtype: fp8
+        enable_block_reuse: true
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.8
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        mamba_state_config:
+          additional_snapshot_offsets_from_end:
+          - 2
+          periodic_snapshot_interval: 0
+        pool_ratio:
+        - 0.8
+        - 0.2
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 8192
+      max_seq_len: 262144
+      moe_config:
+        backend: CUTEDSL
+      moe_expert_parallel_size: 4
+      num_serve_frontends: 8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+      tensor_parallel_size: 4
+      torch_compile_config:
+        capture_num_tokens:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 32
+        - 64
+        - 128
+        - 256
+        - 512
+        - 1024
+        - 2048
+        - 4096
+        - 8192
+        enable_fullgraph: true
+        enable_piecewise_cuda_graph: true
+      trust_remote_code: true
+    decode:
+      cache_transceiver_config:
+        backend: NIXL
+        kv_transfer_sender_future_timeout_ms: 20
+        kv_transfer_timeout_ms: 600000
+        transceiver_runtime: PYTHON
+      cuda_graph_config:
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 24
+        - 32
+        - 40
+        - 48
+        - 56
+        - 64
+        enable_padding: true
+      enable_attention_dp: true
+      enable_iter_perf_stats: false
+      enable_iter_req_stats: false
+      enable_lm_head_tp_in_adp: true
+      kv_cache_config:
+        avg_seq_len: 102150
+        dtype: fp8
+        enable_block_reuse: false
+        event_buffer_max_size: 0
+        free_gpu_memory_fraction: 0.85
+        host_cache_size: 137438953472
+        iteration_stats_interval: 1
+        tokens_per_block: 64
+        use_kv_cache_manager_v2: true
+      max_batch_size: 64
+      max_num_tokens: 448
+      max_seq_len: 262148
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      moe_expert_parallel_size: 8
+      num_postprocess_workers: 4
+      num_serve_frontends: 4
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      return_perf_metrics: false
+      scheduler_config:
+        capacity_scheduler_policy: GUARANTEED_NO_EVICT
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 6
+      stream_interval: 20
+      tensor_parallel_size: 8
+      trust_remote_code: true
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: "120"
+    DYN_TOKENIZER_CACHE: "1"
+    DYN_TOKENIZER_CACHE_BYTES: "8000000000"
+    DYN_TOKENIZER: "fastokens"
+  args:
+    router-mode: kv
+    router-session-affinity-ttl-secs: '14400'
+    active-decode-blocks-threshold: None
+    active-prefill-tokens-threshold: None
+    active-prefill-tokens-threshold-frac: None
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+    RESULT_DIR: /logs/agentic
+    PORT: '8000'
+    IS_MULTINODE: 'true'
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: 'true'
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: '0'
+    AIPERF_DATASET_MMAP_CACHE_DIR: /aiperf_mmap_cache
+    HF_HUB_CACHE: /hf_hub_cache
+    WEKA_LOADER_OVERRIDE: semianalysis_cc_traces_weka_062126
+    SERVED_MODEL_NAME: Qwen3.5-397B-A17B-NVFP4-V2
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 5c049d9f0c..0f0b34c696 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -6724,6 +6724,156 @@ qwen3.5-fp4-gb300-dynamo-trt-mtp:
           ep: 16
           dp-attn: true
 
+qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg:
+  image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24
+  model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  model-prefix: qwen3.5
+  runner: cluster:gb300-nv
+  precision: fp4
+  framework: dynamo-trt
+  router: { name: dynamo-router, version: "1.4.0.dev20260807" }
+  kv-p2p-transfer: nixl
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - search-space:
+      # 1P1D: 1 prefill (TP2/EP2), 1 decode (TP4/EP4), conc=8
+      - spec-decoding: "mtp"
+        conc-list: [8]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-1p1d-tep2-tep4-c8-b128-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 4
+          dp-attn: false
+      # 2P3D: 2 prefill (TP2/EP2), 3 decode (TP4/EP4), conc=6
+      - spec-decoding: "mtp"
+        conc-list: [6]
+        prefill:
+          num-worker: 2
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep4-c6-b128-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 3
+          tp: 4
+          ep: 4
+          dp-attn: false
+      # 2P3D: 2 prefill (TP2/EP2), 3 decode (TP8/EP8), conc=6
+      - spec-decoding: "mtp"
+        conc-list: [6]
+        prefill:
+          num-worker: 2
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep8-c6-b128-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 3
+          tp: 8
+          ep: 8
+          dp-attn: false
+      # 2P3D: 2 prefill (TP2/EP2), 3 decode (TP4/EP4), conc=24
+      - spec-decoding: "mtp"
+        conc-list: [24]
+        prefill:
+          num-worker: 2
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-2p3d-tep2-tep4-c24-b128-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 3
+          tp: 4
+          ep: 4
+          dp-attn: false
+      # 4P1D: 4 prefill (TP4/EP4/dp-attn), 1 decode (TP8/EP8/dp-attn), conc=896
+      - spec-decoding: "mtp"
+        conc-list: [896]
+        prefill:
+          num-worker: 4
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-4p1d-dep4-dep8-c896-b64-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 8
+          ep: 8
+          dp-attn: true
+      # 15P1D: 15 prefill (TP1/EP1/dp-attn), 1 decode (TP16/EP16/dp-attn), conc=606
+      - spec-decoding: "mtp"
+        conc-list: [606]
+        prefill:
+          num-worker: 15
+          tp: 1
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-15p1d-dep1-dep16-c606-b32-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+      # 17P3D: 17 prefill (TP2/EP2/dp-attn), 3 decode (TP8/EP8/dp-attn), conc=1452
+      - spec-decoding: "mtp"
+        conc-list: [1452]
+        prefill:
+          num-worker: 17
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-17p3d-dep2-dep8-c1452-b64-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 3
+          tp: 8
+          ep: 8
+          dp-attn: true
+      # 19P1D: 19 prefill (TP1/EP1/dp-attn), 1 decode (TP16/EP16/dp-attn), conc=225
+      - spec-decoding: "mtp"
+        conc-list: [225]
+        prefill:
+          num-worker: 19
+          tp: 1
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-19p1d-dep1-dep16-c225-b16-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+      # 19P1D: 19 prefill (TP1/EP1/dp-attn), 1 decode (TP16/EP16/dp-attn), conc=449
+      - spec-decoding: "mtp"
+        conc-list: [449]
+        prefill:
+          num-worker: 19
+          tp: 1
+          ep: 1
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx/disagg-gb300-19p1d-dep1-dep16-c449-b16-mtp-kvoffload.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+
 qwen3.5-fp4-gb300-dynamo-sglang-agentic-agg:
   image: lmsysorg/sglang:nightly-dev-cu13-20260724-433429b1
   model: nvidia/Qwen3.5-397B-A17B-NVFP4
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index f118b99aa7..a2d8470ec6 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5909,3 +5909,13 @@
     - "Increase MTP speculative steps from 3 to 5 (num-draft-tokens 4→6) and raise SGLANG_SIMULATE_ACC_LEN from 2.99 to 3.61 to reflect higher measured acceptance rate, targeting ~15% throughput improvement"
     - "Replace the TP8/EP8 low-concurrency arm (conc [1,2,4]) hicache offload with kv-offloading: none to reduce per-request latency at low load; conc [1,2,4] are now tested on both tp=4+hicache and tp=8+no-offload so SemiAnalysis can select the Pareto-optimal point per concurrency"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2570
+
+- config-keys:
+    - qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add Qwen3.5-397B-A17B-NVFP4-V2 GB300 TRT-LLM disaggregated AgentX with Dynamo router 1.4.0.dev20260807 and NIXL KV transfer."
+    - "Cover 9 disaggregated topologies (1P1D through 19P1D/17P3D) with MTP speculative decoding across conc 6-1452 using TRT-LLM release:1.3.0rc24."
+    - "Attention-DP enabled (dep shapes) and disabled (tep shapes) variants, with recipe CONFIG_FILEs from recipes/trtllm/qwen3.5-fp4/agentx/submission/."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2584
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index 27154c404e..a5038c572a 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -184,7 +184,14 @@ SRT_REPO_DIR="${GITHUB_WORKSPACE}/srt-slurm-${GITHUB_RUN_ID:-manual}-${GITHUB_RU
 SRTCTL_SETUP_SCRIPT=""
 rm -rf "$SRT_REPO_DIR"
 
-if [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "qwen3.5" ]]; then
+if [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-trt" && $MODEL_PREFIX == "qwen3.5" ]]; then
+    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+    cd "$SRT_REPO_DIR"
+    git checkout v1.0.48
+    mkdir -p recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx" \
+        recipes/trtllm/qwen3.5/gb300-fp4/disagg/agentx
+elif [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "qwen3.5" ]]; then
     # Qwen3.5 agentic uses NVIDIA/srt-slurm v1.0.38: the two features the
     # cquil11 fork was pinned for are merged upstream (present in v1.0.36) —
     #   - `srtctl apply --no-preflight` (skip the in-process model FS check):