diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c1.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c1.yaml
new file mode 100644
index 000000000..0704f959f
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c1.yaml
@@ -0,0 +1,169 @@
+# Qwen3.5 NVFP4 V2 benchmark operating point.
+# Topology and serving knobs are pinned for reproducibility.
+name: qwen35-v2-gb200-c1-1p1d-tp4-mtp3-baseline
+dynamo:
+  install: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    enforce-disagg: true
+model:
+  path: qwen3.5-fp4
+  container: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
+  precision: fp4
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+backend:
+  type: sglang
+  prefill_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  decode_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    MC_TE_METRIC: 'true'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_DECODE_BOOTSTRAP_TIMEOUT: '1000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  sglang_config:
+    prefill:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 2048
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disaggregation-bootstrap-port: 31000
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 128000
+      max-running-requests: 128
+      cuda-graph-max-bs: 4
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 10
+      stream-interval: 30
+      load-balance-method: round_robin
+      page-size: 64
+      watchdog-timeout: 1000000
+      log-level: info
+    decode:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 128
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disaggregation-bootstrap-port: 31000
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 1500000
+      max-mamba-cache-size: 256
+      max-running-requests: 128
+      cuda-graph-max-bs: 256
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 10
+      stream-interval: 30
+      page-size: 64
+      watchdog-timeout: 1000000
+      decode-log-interval: 1
+health_check:
+  max_attempts: 240
+  interval_seconds: 10
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: '1'
+  req_rate: inf
+  random_range_ratio: 0.8
+  num_prompts_mult: 10
+  num_warmup_mult: 2
+  use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c4.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c4.yaml
new file mode 100644
index 000000000..1f9e110f9
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c4.yaml
@@ -0,0 +1,169 @@
+# Qwen3.5 NVFP4 V2 benchmark operating point.
+# Topology and serving knobs are pinned for reproducibility.
+name: qwen35-v2-gb200-c4-1p1d-tp4-mtp3-prefill-recv1
+dynamo:
+  install: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    enforce-disagg: true
+model:
+  path: qwen3.5-fp4
+  container: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
+  precision: fp4
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+backend:
+  type: sglang
+  prefill_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  decode_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    MC_TE_METRIC: 'true'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_DECODE_BOOTSTRAP_TIMEOUT: '1000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  sglang_config:
+    prefill:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 2048
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disaggregation-bootstrap-port: 31000
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 128000
+      max-running-requests: 128
+      cuda-graph-max-bs: 4
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 1
+      stream-interval: 30
+      load-balance-method: round_robin
+      page-size: 64
+      watchdog-timeout: 1000000
+      log-level: info
+    decode:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 128
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disaggregation-bootstrap-port: 31000
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 1500000
+      max-mamba-cache-size: 256
+      max-running-requests: 128
+      cuda-graph-max-bs: 256
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 10
+      stream-interval: 30
+      page-size: 64
+      watchdog-timeout: 1000000
+      decode-log-interval: 1
+health_check:
+  max_attempts: 240
+  interval_seconds: 10
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: '4'
+  req_rate: inf
+  random_range_ratio: 0.8
+  num_prompts_mult: 10
+  num_warmup_mult: 2
+  use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c8.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c8.yaml
new file mode 100644
index 000000000..03cdcb9d8
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c8.yaml
@@ -0,0 +1,169 @@
+# Qwen3.5 NVFP4 V2 benchmark operating point.
+# Topology and serving knobs are pinned for reproducibility.
+name: qwen35-v2-gb200-c8-1p1d-tp4-mtp3-decode-stream50
+dynamo:
+  install: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    enforce-disagg: true
+model:
+  path: qwen3.5-fp4
+  container: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
+  precision: fp4
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  prefill_nodes: 1
+  decode_nodes: 1
+  prefill_workers: 1
+  decode_workers: 1
+backend:
+  type: sglang
+  prefill_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  decode_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    MC_TE_METRIC: 'true'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_DECODE_BOOTSTRAP_TIMEOUT: '1000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  sglang_config:
+    prefill:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 2048
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disaggregation-bootstrap-port: 31000
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 128000
+      max-running-requests: 128
+      cuda-graph-max-bs: 4
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 10
+      stream-interval: 30
+      load-balance-method: round_robin
+      page-size: 64
+      watchdog-timeout: 1000000
+      log-level: info
+    decode:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 128
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disaggregation-bootstrap-port: 31000
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 1500000
+      max-mamba-cache-size: 256
+      max-running-requests: 128
+      cuda-graph-max-bs: 256
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 10
+      stream-interval: 50
+      page-size: 64
+      watchdog-timeout: 1000000
+      decode-log-interval: 1
+health_check:
+  max_attempts: 240
+  interval_seconds: 10
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: '8'
+  req_rate: inf
+  random_range_ratio: 0.8
+  num_prompts_mult: 10
+  num_warmup_mult: 2
+  use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/2p2d-tp2-tp2-c512.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/2p2d-tp2-tp2-c512.yaml
new file mode 100644
index 000000000..d1f13b24a
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/2p2d-tp2-tp2-c512.yaml
@@ -0,0 +1,172 @@
+# Qwen3.5 NVFP4 V2 benchmark operating point.
+# Topology and serving knobs are pinned for reproducibility.
+name: qwen35-v2-gb200-c512-2p2d-tp2-mtp3-cap84
+dynamo:
+  install: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  env:
+    ETCD_LEASE_TTL: '120'
+  args:
+    enforce-disagg: true
+model:
+  path: qwen3.5-fp4
+  container: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
+  precision: fp4
+resources:
+  gpu_type: gb200
+  gpus_per_node: 2
+  prefill_nodes: 2
+  decode_nodes: 2
+  prefill_workers: 2
+  decode_workers: 2
+  spread_workers: true
+backend:
+  type: sglang
+  prefill_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    ETCD_LEASE_TTL: '120'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  decode_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    ETCD_LEASE_TTL: '120'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    MC_TE_METRIC: 'true'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_DECODE_BOOTSTRAP_TIMEOUT: '1000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  sglang_config:
+    prefill:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 2
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 2048
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 64000
+      max-running-requests: 64
+      cuda-graph-max-bs: 4
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      scheduler-recv-interval: 10
+      stream-interval: 30
+      load-balance-method: round_robin
+      page-size: 64
+      watchdog-timeout: 1000000
+      log-level: info
+    decode:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 2
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 128
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 750000
+      max-mamba-cache-size: 128
+      max-running-requests: 84
+      cuda-graph-max-bs: 88
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 10
+      stream-interval: 30
+      page-size: 64
+      watchdog-timeout: 1000000
+      decode-log-interval: 1
+health_check:
+  max_attempts: 240
+  interval_seconds: 10
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: '512'
+  req_rate: inf
+  random_range_ratio: 0.8
+  num_prompts_mult: 10
+  num_warmup_mult: 2
+  use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/2p3d-tp2-tp2-c768.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/2p3d-tp2-tp2-c768.yaml
new file mode 100644
index 000000000..384a7c166
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/2p3d-tp2-tp2-c768.yaml
@@ -0,0 +1,168 @@
+# Qwen3.5 NVFP4 V2 benchmark operating point.
+# Topology, collective settings, and serving knobs are pinned for reproducibility.
+name: qwen35-v2-gb200-c768-2p3d-tp2-mtp3-cap86
+dynamo:
+  install: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    enforce-disagg: true
+model:
+  path: qwen3.5-fp4
+  container: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
+  precision: fp4
+resources:
+  gpu_type: gb200
+  gpus_per_node: 2
+  prefill_nodes: 2
+  decode_nodes: 3
+  prefill_workers: 2
+  decode_workers: 3
+  spread_workers: true
+backend:
+  type: sglang
+  prefill_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  decode_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    MC_TE_METRIC: 'true'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_DECODE_BOOTSTRAP_TIMEOUT: '1000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  sglang_config:
+    prefill:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 2
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 2048
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 64000
+      max-running-requests: 64
+      cuda-graph-max-bs: 4
+      chunked-prefill-size: 16384
+      max-prefill-tokens: 16384
+      scheduler-recv-interval: 10
+      stream-interval: 30
+      load-balance-method: round_robin
+      page-size: 64
+      watchdog-timeout: 1000000
+      log-level: info
+    decode:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 2
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 128
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 750000
+      max-mamba-cache-size: 128
+      max-running-requests: 86
+      cuda-graph-max-bs: 88
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 10
+      stream-interval: 30
+      page-size: 64
+      watchdog-timeout: 1000000
+      decode-log-interval: 1
+health_check:
+  max_attempts: 240
+  interval_seconds: 10
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: '768'
+  req_rate: inf
+  random_range_ratio: 0.8
+  num_prompts_mult: 10
+  num_warmup_mult: 2
+  use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/3p5d-tp4-tp4-c128.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/3p5d-tp4-tp4-c128.yaml
new file mode 100644
index 000000000..450e091ad
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/3p5d-tp4-tp4-c128.yaml
@@ -0,0 +1,169 @@
+# Qwen3.5 NVFP4 V2 benchmark operating point.
+# Topology and serving knobs are pinned for reproducibility.
+name: qwen35-v2-gb200-c128-3p5d-tp4-mtp3
+dynamo:
+  install: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    enforce-disagg: true
+model:
+  path: qwen3.5-fp4
+  container: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
+  precision: fp4
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  prefill_nodes: 3
+  decode_nodes: 5
+  prefill_workers: 3
+  decode_workers: 5
+backend:
+  type: sglang
+  prefill_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  decode_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    MC_TE_METRIC: 'true'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_DECODE_BOOTSTRAP_TIMEOUT: '1000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  sglang_config:
+    prefill:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 2048
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disaggregation-bootstrap-port: 31000
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 128000
+      max-running-requests: 128
+      cuda-graph-max-bs: 4
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 10
+      stream-interval: 30
+      load-balance-method: round_robin
+      page-size: 64
+      watchdog-timeout: 1000000
+      log-level: info
+    decode:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 1
+      expert-parallel-size: 1
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 128
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 3
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 4
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disaggregation-bootstrap-port: 31000
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 1500000
+      max-mamba-cache-size: 256
+      max-running-requests: 128
+      cuda-graph-max-bs: 256
+      chunked-prefill-size: 32768
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 10
+      stream-interval: 30
+      page-size: 64
+      watchdog-timeout: 1000000
+      decode-log-interval: 1
+health_check:
+  max_attempts: 240
+  interval_seconds: 10
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: '128'
+  req_rate: inf
+  random_range_ratio: 0.8
+  num_prompts_mult: 10
+  num_warmup_mult: 2
+  use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p1d-dep4-dep16-c5120.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p1d-dep4-dep16-c5120.yaml
new file mode 100644
index 000000000..c1c23b5be
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p1d-dep4-dep16-c5120.yaml
@@ -0,0 +1,179 @@
+# V2 mixed quantization uses the CuteDSL MoE path at this saturated wide-EP point.
+# Topology and serving knobs are pinned for reproducibility.
+name: qwen35-v2-gb200-c5120-7p1d-dep4-dep16-mtp2-cap2048
+dynamo:
+  install: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 2
+  nginx_container: nginx
+  args:
+    enforce-disagg: true
+model:
+  path: qwen3.5-fp4
+  container: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
+  precision: fp4
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  prefill_nodes: 7
+  decode_nodes: 4
+  prefill_workers: 7
+  decode_workers: 1
+backend:
+  type: sglang
+  prefill_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  decode_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    MC_TE_METRIC: 'true'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_MOE_NVFP4_DISPATCH: '1'
+    SGLANG_FLASHINFER_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '1024'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_DECODE_BOOTSTRAP_TIMEOUT: '1000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  sglang_config:
+    prefill:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 2048
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 128000
+      max-running-requests: 128
+      cuda-graph-max-bs: 4
+      chunked-prefill-size: 65536
+      max-prefill-tokens: 65536
+      scheduler-recv-interval: 1
+      stream-interval: 50
+      load-balance-method: round_robin
+      page-size: 64
+      watchdog-timeout: 1000000
+      log-level: info
+    decode:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 16
+      data-parallel-size: 16
+      expert-parallel-size: 16
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 128
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-a2a-backend: flashinfer
+      moe-runner-backend: flashinfer_cutedsl
+      speculative-moe-a2a-backend: none
+      speculative-moe-runner-backend: flashinfer_trtllm
+      disable-shared-experts-fusion: true
+      linear-attn-decode-backend: flashinfer
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 2
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 3
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 2200000
+      max-mamba-cache-size: 2048
+      max-running-requests: 2048
+      chunked-prefill-size: 4096
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 1
+      stream-interval: 50
+      page-size: 64
+      watchdog-timeout: 1000000
+      decode-log-interval: 50
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: '5120'
+  req_rate: inf
+  random_range_ratio: 0.8
+  num_prompts_mult: 10
+  num_warmup_mult: 2
+  use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p1d-dep4-dep16-c8192.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p1d-dep4-dep16-c8192.yaml
new file mode 100644
index 000000000..b5ba39378
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p1d-dep4-dep16-c8192.yaml
@@ -0,0 +1,179 @@
+# V2 mixed quantization uses the CuteDSL MoE path at this saturated wide-EP point.
+# Topology and serving knobs are pinned for reproducibility.
+name: qwen35-v2-gb200-c8192-7p1d-dep4-dep16-mtp2-cap3413
+dynamo:
+  install: false
+frontend:
+  type: dynamo
+  enable_multiple_frontends: true
+  num_additional_frontends: 2
+  nginx_container: nginx
+  args:
+    enforce-disagg: true
+model:
+  path: qwen3.5-fp4
+  container: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
+  precision: fp4
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  prefill_nodes: 7
+  decode_nodes: 4
+  prefill_workers: 7
+  decode_workers: 1
+backend:
+  type: sglang
+  prefill_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  decode_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    MC_TE_METRIC: 'true'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_MOE_NVFP4_DISPATCH: '1'
+    SGLANG_FLASHINFER_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '1024'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_DECODE_BOOTSTRAP_TIMEOUT: '1000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  sglang_config:
+    prefill:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 2048
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 128000
+      max-running-requests: 128
+      cuda-graph-max-bs: 4
+      chunked-prefill-size: 65536
+      max-prefill-tokens: 65536
+      scheduler-recv-interval: 1
+      stream-interval: 50
+      load-balance-method: round_robin
+      page-size: 64
+      watchdog-timeout: 1000000
+      log-level: info
+    decode:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 16
+      data-parallel-size: 16
+      expert-parallel-size: 16
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 128
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-a2a-backend: flashinfer
+      moe-runner-backend: flashinfer_cutedsl
+      speculative-moe-a2a-backend: none
+      speculative-moe-runner-backend: flashinfer_trtllm
+      disable-shared-experts-fusion: true
+      linear-attn-decode-backend: flashinfer
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 2
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 3
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      cuda-graph-max-bs: 256
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 2200000
+      max-mamba-cache-size: 3413
+      max-running-requests: 3413
+      chunked-prefill-size: 4096
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 1
+      stream-interval: 50
+      page-size: 64
+      watchdog-timeout: 1000000
+      decode-log-interval: 50
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: '8192'
+  req_rate: inf
+  random_range_ratio: 0.8
+  num_prompts_mult: 10
+  num_warmup_mult: 2
+  use_chat_template: true
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p4d-dep4-dep4-c12288.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p4d-dep4-dep4-c12288.yaml
new file mode 100644
index 000000000..bfa84d6cb
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p4d-dep4-dep4-c12288.yaml
@@ -0,0 +1,191 @@
+# Qwen3.5 NVFP4 V2 benchmark operating point.
+# Seven DEP4 prefill workers feed four DEP4 MTP1 decode workers.
+name: qwen35-v2-gb200-c12288-7p4d-dep4-mtp1-cap1056
+dynamo:
+  install: true
+  wheel: "1.2.1"
+  request_plane: "nats"
+frontend:
+  type: dynamo
+  enable_multiple_frontends: false
+  args:
+    enforce-disagg: true
+model:
+  path: qwen3.5-fp4
+  container: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
+  precision: fp4
+resources:
+  gpu_type: gb200
+  gpus_per_node: 4
+  prefill_nodes: 7
+  decode_nodes: 4
+  prefill_workers: 7
+  decode_workers: 4
+backend:
+  type: sglang
+  prefill_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+  decode_environment:
+    NO_COLOR: '1'
+    PYTHONUNBUFFERED: '1'
+    HF_HUB_OFFLINE: '1'
+    TRANSFORMERS_OFFLINE: '1'
+    TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: '3600'
+    TORCH_NCCL_WATCHDOG_TIMEOUT_SEC: '3600'
+    TORCH_NCCL_HEARTBEAT_TIMEOUT_SEC: '3600'
+    NCCL_DEBUG: INFO
+    NCCL_DEBUG_SUBSYS: COLL
+    NCCL_MNNVL_ENABLE: '1'
+    NCCL_NVLS_ENABLE: '0'
+    NCCL_CUMEM_ENABLE: '1'
+    MC_FORCE_MNNVL: '1'
+    MC_TE_METRIC: 'true'
+    SGLANG_ENABLE_SPEC_V2: '1'
+    SGLANG_DG_CACHE_DIR: /configs/deepgemm-cache
+    FLASHINFER_WORKSPACE_BASE: /configs/flashinfer-cache
+    SGLANG_ENABLE_JIT_DEEPGEMM: 'true'
+    SGLANG_ENABLE_FLASHINFER_GEMM: 'true'
+    FLASHINFER_DISABLE_VERSION_CHECK: '1'
+    SGLANG_MOE_NVFP4_DISPATCH: '1'
+    SGLANG_FLASHINFER_NUM_MAX_DISPATCH_TOKENS_PER_RANK: '1024'
+    SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
+    SGLANG_DISAGGREGATION_HEARTBEAT_MAX_FAILURE: '100000'
+    SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT: '100000'
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: '100000'
+    SGLANG_DECODE_BOOTSTRAP_TIMEOUT: '1000'
+    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
+    SGLANG_USE_MESSAGE_QUEUE_BROADCASTER: '0'
+    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
+    SGLANG_HEALTH_CHECK_TIMEOUT: '3600'
+    SGLANG_HEALTH_STARTING_OK: '1'
+    SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION: '0'
+    CUDA_ENABLE_COREDUMP_ON_EXCEPTION: '1'
+    CUDA_ENABLE_CPU_COREDUMP_ON_EXCEPTION: '0'
+    CUDA_COREDUMP_FILE: /logs/cuda_coredump_%h_%p
+    CUDA_COREDUMP_GENERATION_FLAGS: skip_nonrelocated_elf_images,skip_global_memory,skip_shared_memory,skip_local_memory,skip_constbank_memory
+  sglang_config:
+    prefill:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+      enable-dp-attention: true
+      enable-dp-attention-local-control-broadcast: true
+      enable-dp-lm-head: true
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 2048
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-runner-backend: flashinfer_trtllm
+      linear-attn-decode-backend: flashinfer
+      disaggregation-mode: prefill
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      mem-fraction-static: 0.8
+      context-length: 9236
+      max-total-tokens: 256000
+      max-running-requests: 128
+      cuda-graph-max-bs: 4
+      chunked-prefill-size: 262144
+      max-prefill-tokens: 131072
+      scheduler-recv-interval: 1
+      stream-interval: 50
+      load-balance-method: round_robin
+      page-size: 64
+      watchdog-timeout: 1000000
+      log-level: info
+    decode:
+      served-model-name: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+      model-path: /model
+      trust-remote-code: true
+      reasoning-parser: qwen3
+      tool-call-parser: qwen3_coder
+      quantization: modelopt_mixed
+      fp4-gemm-backend: flashinfer_cutlass
+      kv-cache-dtype: fp8_e4m3
+      tensor-parallel-size: 4
+      data-parallel-size: 4
+      expert-parallel-size: 4
+      enable-dp-attention: true
+      enable-dp-lm-head: true
+      mamba-scheduler-strategy: no_buffer
+      mamba-ssm-dtype: bfloat16
+      mamba-track-interval: 128
+      attention-backend: trtllm_mha
+      mm-attention-backend: triton_attn
+      moe-a2a-backend: none
+      moe-runner-backend: flashinfer_trtllm
+      speculative-moe-a2a-backend: none
+      speculative-moe-runner-backend: flashinfer_trtllm
+      disable-shared-experts-fusion: true
+      linear-attn-decode-backend: flashinfer
+      speculative-algorithm: NEXTN
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 2
+      disable-flashinfer-autotune: true
+      disaggregation-mode: decode
+      disaggregation-transfer-backend: mooncake
+      disable-radix-cache: true
+      cuda-graph-max-bs: 512
+      mem-fraction-static: 0.94
+      context-length: 9236
+      max-total-tokens: 2052736
+      max-mamba-cache-size: 1056
+      max-running-requests: 1056
+      chunked-prefill-size: 4096
+      max-prefill-tokens: 32768
+      scheduler-recv-interval: 1
+      disaggregation-decode-polling-interval: 8
+      stream-interval: 50
+      page-size: 64
+      soft-watchdog-timeout: 240
+      watchdog-timeout: 300
+      decode-log-interval: 50
+health_check:
+  max_attempts: 360
+  interval_seconds: 10
+benchmark:
+  type: sa-bench
+  isl: 8192
+  osl: 1024
+  concurrencies: '12288'
+  req_rate: inf
+  random_range_ratio: 0.8
+  num_prompts_mult: 10
+  num_warmup_mult: 2
+  use_chat_template: true
+  reuse_http_connections: true
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index caa0a7e08..6c3dea03c 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -8084,6 +8084,148 @@ glm5.2-fp4-b200-sglang-agentic-mtp:
       search-space:
       - { tp: 8, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [1, 4, 8, 12, 16] }
 
+qwen3.5-fp4-gb200-dynamo-sglang-mtp:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928
+  model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
+  model-prefix: qwen3.5
+  runner: gb200
+  precision: fp4
+  framework: dynamo-sglang
+  kv-p2p-transfer: mooncake
+  multinode: true
+  disagg: true
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - spec-decoding: "mtp"
+        conc-list: [1]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c1.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - spec-decoding: "mtp"
+        conc-list: [4]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c4.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - spec-decoding: "mtp"
+        conc-list: [8]
+        prefill:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/1p1d-tp4-tp4-c8.yaml"
+        decode:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - spec-decoding: "mtp"
+        conc-list: [512]
+        prefill:
+          num-worker: 2
+          tp: 2
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/2p2d-tp2-tp2-c512.yaml"
+        decode:
+          num-worker: 2
+          tp: 2
+          ep: 1
+          dp-attn: false
+      - spec-decoding: "mtp"
+        conc-list: [128]
+        prefill:
+          num-worker: 3
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/3p5d-tp4-tp4-c128.yaml"
+        decode:
+          num-worker: 5
+          tp: 4
+          ep: 1
+          dp-attn: false
+      - spec-decoding: "mtp"
+        conc-list: [768]
+        prefill:
+          num-worker: 2
+          tp: 2
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/2p3d-tp2-tp2-c768.yaml"
+        decode:
+          num-worker: 3
+          tp: 2
+          ep: 1
+          dp-attn: false
+      - spec-decoding: "mtp"
+        conc-list: [5120]
+        prefill:
+          num-worker: 7
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p1d-dep4-dep16-c5120.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+      - spec-decoding: "mtp"
+        conc-list: [8192]
+        prefill:
+          num-worker: 7
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p1d-dep4-dep16-c8192.yaml"
+        decode:
+          num-worker: 1
+          tp: 16
+          ep: 16
+          dp-attn: true
+      - spec-decoding: "mtp"
+        conc-list: [12288]
+        prefill:
+          num-worker: 7
+          tp: 4
+          ep: 4
+          dp-attn: true
+          additional-settings:
+          - "CONFIG_FILE=recipes/sglang/qwen3.5/gb200-fp4-v2/8k1k/mtp/7p4d-dep4-dep4-c12288.yaml"
+        decode:
+          num-worker: 4
+          tp: 4
+          ep: 4
+          dp-attn: true
+
 qwen3.5-fp8-gb200-dynamo-sglang-mtp:
   image: lmsysorg/sglang:v0.5.14-cu130@sha256:5027e95bf6ec536856b1b52a91d1f35ff5c564ab83e8a94758a169ff09bb8df3
   model: Qwen/Qwen3.5-397B-A17B-FP8
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 04ba9feb5..deab9c100 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5953,3 +5953,14 @@
     - "Use native EAGLE MTP (3 steps, top-k 1, 4 draft tokens) and golden synthetic acceptance length 2.49 for throughput; eval retains real verification."
     - "Follow the official SGLang DeepSeek-V4 Blackwell recipe, require nonempty SGLang server metrics, keep pooled AgentX connections alive, let AIPerf own HiCache warmup, and reserve transient MoE workspace at DEP8 c512."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2577
+
+- config-keys:
+    - qwen3.5-fp4-gb200-dynamo-sglang-mtp
+  scenario-type:
+    - fixed-seq-len
+  description:
+    - "Add a Qwen3.5 NVFP4 V2 GB200 Dynamo-SGLang disaggregated MTP configuration for the 8k/1k scenario"
+    - "Add nine checked-in srt-slurm recipes covering TP and data-parallel-attention topologies"
+    - "Use chat-formatted benchmark inputs with Mooncake KV transfer"
+    - "Image: lmsysorg/sglang:nightly-dev-cu13-20260709-074bb928"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2594