diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen1tep16_mtp3_c48.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen1tep16_mtp3_c48.yaml
new file mode 100644
index 000000000..9438243e2
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen1tep16_mtp3_c48.yaml
@@ -0,0 +1,167 @@
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/1ctxDEP2c128_1genTEP16c32_E2E_C32/conc32
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx1dep2_gen1tep16_mtp3_c48
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 2
+  decode_nodes: 4
+  decode_workers: 1
+  gpus_per_decode: 16
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 16
+      moe_expert_parallel_size: 16
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      pipeline_parallel_size: 1
+      max_batch_size: 40
+      max_num_tokens: 4096
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 20
+        - 24
+        - 32
+        - 64
+        - 128
+        - 256
+        - 384
+        - 512
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "48"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen4tp16_mtp3_c4.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen4tp16_mtp3_c4.yaml
new file mode 100644
index 000000000..c1fda5499
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen4tp16_mtp3_c4.yaml
@@ -0,0 +1,156 @@
+# Derived from config #01 for one-request-per-decode-worker minimum-TPOT measurement.
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/1ctxDEP2c128_4genTP16c4_E2E_C16/conc16
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx1dep2_gen4tp16_mtp3_c4
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 2
+  decode_nodes: 16
+  decode_workers: 4
+  gpus_per_decode: 16
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 16
+      moe_expert_parallel_size: 1
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      pipeline_parallel_size: 1
+      max_batch_size: 1
+      max_num_tokens: 4
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "4"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen7tep8_mtp3_c42.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen7tep8_mtp3_c42.yaml
new file mode 100644
index 000000000..f32dbc95a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen7tep8_mtp3_c42.yaml
@@ -0,0 +1,168 @@
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/1ctxDEP2c128_7genTEP8c4_E2E_C28/conc28
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx1dep2_gen7tep8_mtp3_c42
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 1
+  gpus_per_prefill: 2
+  decode_nodes: 14
+  decode_workers: 7
+  gpus_per_decode: 8
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 8
+      moe_expert_parallel_size: 8
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      pipeline_parallel_size: 1
+      max_batch_size: 5
+      max_num_tokens: 4096
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 20
+        - 24
+        - 32
+        - 64
+        - 128
+        - 256
+        - 384
+        - 512
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      allreduce_strategy: MNNVL
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "42"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx2dep2_gen15tp4_mtp3_c90.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx2dep2_gen15tp4_mtp3_c90.yaml
new file mode 100644
index 000000000..7cb1d46b1
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx2dep2_gen15tp4_mtp3_c90.yaml
@@ -0,0 +1,167 @@
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/2ctxDEP2c128_15genTP4c4_E2E_C60/conc60
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx2dep2_gen15tp4_mtp3_c90
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 1
+  prefill_workers: 2
+  gpus_per_prefill: 2
+  decode_nodes: 15
+  decode_workers: 15
+  gpus_per_decode: 4
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 4
+      moe_expert_parallel_size: 1
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      pipeline_parallel_size: 1
+      max_batch_size: 5
+      max_num_tokens: 4096
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 20
+        - 24
+        - 32
+        - 64
+        - 128
+        - 256
+        - 384
+        - 512
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "90"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx3dep2_gen4tep8_mtp3_c192.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx3dep2_gen4tep8_mtp3_c192.yaml
new file mode 100644
index 000000000..2a088dc3b
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx3dep2_gen4tep8_mtp3_c192.yaml
@@ -0,0 +1,168 @@
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/3ctxDEP2c128_4genTEP8c32_E2E_C128/conc128
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx3dep2_gen4tep8_mtp3_c192
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 2
+  prefill_workers: 3
+  gpus_per_prefill: 2
+  decode_nodes: 8
+  decode_workers: 4
+  gpus_per_decode: 8
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 8
+      moe_expert_parallel_size: 8
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      pipeline_parallel_size: 1
+      max_batch_size: 40
+      max_num_tokens: 4096
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 20
+        - 24
+        - 32
+        - 64
+        - 128
+        - 256
+        - 384
+        - 512
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      allreduce_strategy: MNNVL
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "192"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx3dep2_gen7tep8_mtp3_c168.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx3dep2_gen7tep8_mtp3_c168.yaml
new file mode 100644
index 000000000..12b6138e6
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx3dep2_gen7tep8_mtp3_c168.yaml
@@ -0,0 +1,168 @@
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/3ctxDEP2c128_7genTEP8c16_E2E_C112/conc112
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx3dep2_gen7tep8_mtp3_c168
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 2
+  prefill_workers: 3
+  gpus_per_prefill: 2
+  decode_nodes: 14
+  decode_workers: 7
+  gpus_per_decode: 8
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 8
+      moe_expert_parallel_size: 8
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      pipeline_parallel_size: 1
+      max_batch_size: 20
+      max_num_tokens: 4096
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 20
+        - 24
+        - 32
+        - 64
+        - 128
+        - 256
+        - 384
+        - 512
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      allreduce_strategy: MNNVL
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "168"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx4dep2_gen1dep16_mtp3_c384.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx4dep2_gen1dep16_mtp3_c384.yaml
new file mode 100644
index 000000000..6a3802ef4
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx4dep2_gen1dep16_mtp3_c384.yaml
@@ -0,0 +1,167 @@
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/4ctxDEP2c128_1genDEP16c256_E2E_C256/conc256
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx4dep2_gen1dep16_mtp3_c384
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 2
+  prefill_workers: 4
+  gpus_per_prefill: 2
+  decode_nodes: 4
+  decode_workers: 1
+  gpus_per_decode: 16
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 16
+      moe_expert_parallel_size: 16
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      pipeline_parallel_size: 1
+      max_batch_size: 20
+      max_num_tokens: 4096
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 20
+        - 24
+        - 32
+        - 64
+        - 128
+        - 256
+        - 384
+        - 512
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "384"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx6dep2_gen5tep8_mtp3_c480.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx6dep2_gen5tep8_mtp3_c480.yaml
new file mode 100644
index 000000000..14ff198a0
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx6dep2_gen5tep8_mtp3_c480.yaml
@@ -0,0 +1,168 @@
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/6ctxDEP2c128_5genTEP8c64_E2E_C320/conc320
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx6dep2_gen5tep8_mtp3_c480
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 3
+  prefill_workers: 6
+  gpus_per_prefill: 2
+  decode_nodes: 10
+  decode_workers: 5
+  gpus_per_decode: 8
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 64
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 8
+      moe_expert_parallel_size: 8
+      enable_attention_dp: false
+      enable_lm_head_tp_in_adp: false
+      pipeline_parallel_size: 1
+      max_batch_size: 80
+      max_num_tokens: 4096
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 20
+        - 24
+        - 32
+        - 64
+        - 128
+        - 256
+        - 384
+        - 512
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: TRTLLM
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      allreduce_strategy: MNNVL
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "480"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx9dep2_gen1dep16_mtp3_c1536.yaml b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx9dep2_gen1dep16_mtp3_c1536.yaml
new file mode 100644
index 000000000..6e0195f2d
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx9dep2_gen1dep16_mtp3_c1536.yaml
@@ -0,0 +1,167 @@
+# Reconstructed from successful source result: /lustre/fsw/coreai_comparch_infbench/minma/sa_qwen_disagg/exp3_8k1k_mtp3_0609_tot/results/e2e-clean/9ctxDEP2c2048_1genDEP16c1024_E2E_C1024/conc1024
+# Variant runtime: srt-slurm f6eb42aee4664207dcf2ec601e3bd57bd527efd6 with official TensorRT-LLM 1.3.0rc18 container.
+# Unlisted model, topology, resource, frontend, and engine fields remain identical to original-configs.
+name: ctx9dep2_gen1dep16_mtp3_c1536
+identity:
+  model:
+    repo: nvidia/Qwen3.5-397B-A17B-NVFP4
+  container:
+    image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  frameworks:
+    tensorrt_llm: 1.3.0rc18
+model:
+  path: qwen3.5-fp4
+  container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  precision: fp4
+resources:
+  gpu_type: gb300
+  gpus_per_node: 4
+  prefill_nodes: 5
+  prefill_workers: 9
+  gpus_per_prefill: 2
+  decode_nodes: 4
+  decode_workers: 1
+  gpus_per_decode: 16
+  het_jobs: false
+backend:
+  type: trtllm
+  prefill_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '1'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  decode_environment:
+    TRTLLM_SERVER_DISABLE_GC: '1'
+    TRTLLM_WORKER_DISABLE_GC: '1'
+    NCCL_GRAPH_MIXING_SUPPORT: '0'
+    TRTLLM_ENABLE_PDL: '1'
+    TLLM_LOG_LEVEL: INFO
+    TLLM_PROFILE_LOG_RANKS: all
+    TLLM_USE_FLASHINFER_GDN_PREFILL: '0'
+    HF_HUB_OFFLINE: '1'
+    PYTHONNOUSERSITE: '1'
+  trtllm_config:
+    prefill:
+      trust_remote_code: true
+      max_batch_size: 128
+      max_num_tokens: 32768
+      max_seq_len: 8395
+      tensor_parallel_size: 2
+      moe_expert_parallel_size: 2
+      enable_attention_dp: true
+      pipeline_parallel_size: 1
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      cuda_graph_config: null
+      disable_overlap_scheduler: true
+      enable_chunked_prefill: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.85
+        dtype: fp8
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+    decode:
+      trust_remote_code: true
+      tensor_parallel_size: 16
+      moe_expert_parallel_size: 16
+      enable_attention_dp: true
+      enable_lm_head_tp_in_adp: true
+      pipeline_parallel_size: 1
+      max_batch_size: 80
+      max_num_tokens: 4096
+      max_seq_len: 9419
+      cuda_graph_config:
+        enable_padding: true
+        batch_sizes:
+        - 1
+        - 2
+        - 4
+        - 8
+        - 16
+        - 20
+        - 24
+        - 32
+        - 64
+        - 128
+        - 256
+        - 384
+        - 512
+      print_iter_log: true
+      enable_iter_perf_stats: true
+      kv_cache_config:
+        enable_block_reuse: false
+        free_gpu_memory_fraction: 0.8
+        dtype: fp8
+      nvfp4_gemm_config:
+        allowed_backends:
+        - cutlass
+        - cublaslt
+        - cutedsl
+        - cuda_core
+      moe_config:
+        backend: CUTEDSL
+        use_low_precision_moe_combine: true
+      cache_transceiver_config:
+        max_tokens_in_buffer: 8192
+        backend: UCX
+        # Match NVIDIA's 8k/1k high-concurrency disagg deadline and avoid rc18's unsafe 60s cancellation path.
+        kv_transfer_timeout_ms: 600000
+      stream_interval: 100
+      num_postprocess_workers: 4
+      scheduler_config:
+        capacity_scheduler_policy: MAX_UTILIZATION
+        context_chunking_policy: FIRST_COME_FIRST_SERVED
+      speculative_config:
+        decoding_type: MTP
+        max_draft_len: 3
+        use_mtp_vanilla: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  env:
+    ETCD_LEASE_TTL: '120'
+    DYN_TOKENIZER: "fastokens"
+benchmark:
+  type: "sa-bench"
+  isl: 8192
+  osl: 1024
+  concurrencies: "1536"
+  req_rate: "inf"
+  num_prompts_mult: 16
+dynamo:
+  install: true
+  version: 1.3.0.dev20260708
+  request_plane: tcp
+infra:
+  etcd_nats_dedicated_node: false
+profiling:
+  type: none
+health_check:
+  max_attempts: 720
+  interval_seconds: 10
+enable_config_dump: true
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 73327351e..1b5cbe0c7 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -11486,6 +11486,157 @@ qwen3.5-fp4-gb300-dynamo-sglang:
           ep: 16
           dp-attn: true
 
+qwen3.5-fp4-gb300-dynamo-trt-mtp:
+  image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18
+  model: nvidia/Qwen3.5-397B-A17B-NVFP4
+  model-prefix: qwen3.5
+  runner: gb300
+  precision: fp4
+  framework: dynamo-trt
+  multinode: true
+  disagg: true
+  kv-p2p-transfer: nixl
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      # 1P1D: 1 prefill (TP2/EP2/dp-attn), 1 decode (TP16/EP16/no-dp-attn), conc=48
+      - spec-decoding: "mtp"
+        conc-list: [48]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen1tep16_mtp3_c48.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: false
+      # 1P4D: 1 prefill (TP2/EP2/dp-attn), 4 decode (TP16/EP1/no-dp-attn), conc=4
+      - spec-decoding: "mtp"
+        conc-list: [4]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen4tp16_mtp3_c4.yaml"
+        decode:
+          num-worker: 4
+          tp: 16
+          ep: 1
+          dp-attn: false
+      # 1P7D: 1 prefill (TP2/EP2/dp-attn), 7 decode (TP8/EP8/no-dp-attn), conc=42
+      - spec-decoding: "mtp"
+        conc-list: [42]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx1dep2_gen7tep8_mtp3_c42.yaml"
+        decode:
+          num-worker: 7
+          tp: 8
+          ep: 8
+          dp-attn: false
+      # 2P15D: 2 prefill (TP2/EP2/dp-attn), 15 decode (TP4/EP1/no-dp-attn), conc=90
+      - spec-decoding: "mtp"
+        conc-list: [90]
+        prefill:
+          num-worker: 2
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx2dep2_gen15tp4_mtp3_c90.yaml"
+        decode:
+          num-worker: 15
+          tp: 4
+          ep: 1
+          dp-attn: false
+      # 3P4D: 3 prefill (TP2/EP2/dp-attn), 4 decode (TP8/EP8/no-dp-attn), conc=192
+      - spec-decoding: "mtp"
+        conc-list: [192]
+        prefill:
+          num-worker: 3
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx3dep2_gen4tep8_mtp3_c192.yaml"
+        decode:
+          num-worker: 4
+          tp: 8
+          ep: 8
+          dp-attn: false
+      # 3P7D: 3 prefill (TP2/EP2/dp-attn), 7 decode (TP8/EP8/no-dp-attn), conc=168
+      - spec-decoding: "mtp"
+        conc-list: [168]
+        prefill:
+          num-worker: 3
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx3dep2_gen7tep8_mtp3_c168.yaml"
+        decode:
+          num-worker: 7
+          tp: 8
+          ep: 8
+          dp-attn: false
+      # 4P1D: 4 prefill (TP2/EP2/dp-attn), 1 decode (TP16/EP16/dp-attn), conc=384
+      - spec-decoding: "mtp"
+        conc-list: [384]
+        prefill:
+          num-worker: 4
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx4dep2_gen1dep16_mtp3_c384.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+      # 6P5D: 6 prefill (TP2/EP2/dp-attn), 5 decode (TP8/EP8/no-dp-attn), conc=480
+      - spec-decoding: "mtp"
+        conc-list: [480]
+        prefill:
+          num-worker: 6
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx6dep2_gen5tep8_mtp3_c480.yaml"
+        decode:
+          num-worker: 5
+          tp: 8
+          ep: 8
+          dp-attn: false
+      # 9P1D: 9 prefill (TP2/EP2/dp-attn), 1 decode (TP16/EP16/dp-attn), conc=1536
+      - spec-decoding: "mtp"
+        conc-list: [1536]
+        prefill:
+          num-worker: 9
+          tp: 2
+          ep: 2
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/trtllm/qwen3.5/gb300-fp4/disagg/8k1k/mtp/ctx9dep2_gen1dep16_mtp3_c1536.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+
 glm5-fp4-gb300-dynamo-sglang:
   image: lmsysorg/sglang:v0.5.11-cu130
   model: nvidia/GLM-5.1-NVFP4
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d77c35f1f..d09ef9ec4 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4781,3 +4781,12 @@
     - "Bump image to lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260708"
     - "Clean the export envs"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2198
+
+- config-keys:
+    - qwen3.5-fp4-gb300-dynamo-trt-mtp
+  description:
+    - "Add Qwen3.5-397B-A17B-NVFP4 GB300 disaggregated TensorRT-LLM+Dynamo MTP (3-token) benchmarks for 8k1k."
+    - "Image: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc18 with Dynamo installed at runtime (version 1.3.0.dev20260708, TCP request plane)."
+    - "9 MTP configs: 1P1D DEP2+1D TEP16 (conc 48), 1P4D DEP2+4D TP16 (conc 4), 1P7D DEP2+7D TEP8 (conc 42), 2P15D DEP2+15D TP4 (conc 90), 3P4D DEP2+4D TEP8 (conc 192), 3P7D DEP2+7D TEP8 (conc 168), 4P1D DEP2+1D DEP16 (conc 384), 6P5D DEP2+5D TEP8 (conc 480), 9P1D DEP2+1D DEP16 (conc 1536)."
+    - "Runner: launch_gb300-nv.sh updated to clone NVIDIA/srt-slurm@sa-submission-q2-2026 and overlay local recipes from benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg/ for the dynamo-trt+qwen3.5+fp4 path; --no-preflight added for model on compute-node /scratch."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2218
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index 1381bd5db..494b6500e 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -229,6 +229,13 @@ elif [[ $FRAMEWORK == "dynamo-trt" && $MODEL_PREFIX == "dsv4" ]]; then
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
     git checkout sa-submission-q2-2026
+elif [[ $FRAMEWORK == "dynamo-trt" && $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp4" ]]; then
+    git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
+    cd "$SRT_REPO_DIR"
+    git checkout v1.0.29
+    mkdir -p recipes/trtllm/qwen3.5/gb300-fp4/disagg
+    cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/trtllm/qwen3.5/gb300-fp4/disagg" \
+        recipes/trtllm/qwen3.5/gb300-fp4/disagg
 else
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
@@ -337,7 +344,7 @@ SRTCTL_APPLY_ARGS=(
     -f "$CONFIG_FILE"
     --tags "gb300,${MODEL_PREFIX},${PRECISION},${ISL}x${OSL},infmax-$(date +%Y%m%d)"
 )
-if [[ "$IS_AGENTIC" == "1" || "$MODEL_PREFIX" == "glm5.1" || ( "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp8" ) ]]; then
+if [[ "$IS_AGENTIC" == "1" || "$MODEL_PREFIX" == "glm5.1" || ( "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp8" ) || ( "$MODEL_PREFIX" == "qwen3.5" && "$PRECISION" == "fp4" && "$FRAMEWORK" == "dynamo-trt" ) ]]; then
     SRTCTL_APPLY_ARGS+=(--no-preflight)
 fi
 if [[ -n "$SRTCTL_SETUP_SCRIPT" ]]; then