diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml
deleted file mode 100644
index 7bb82d2b5..000000000
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml
+++ /dev/null
@@ -1,156 +0,0 @@
-name: "agg-gb300-tp4-mtp-kvoffload"
-
-# Agentic-coding SGLang aggregated recipe for DeepSeek-V4-Pro on GB300
-# (single aggregated worker, TP4, MTP + hierarchical-cache KV offload).
-#
-# Uses the flat single-variant schema the agentic CI flow expects. Concurrency is NOT
-# a recipe field here: the GHA matrix fans out one job per concurrency from
-# the master-config conc-list, exporting CONC into agentic_srt.sh. Modeled on
-# the flat vllm/deepseek-v4/agentic recipes (agentic infra + benchmark wiring)
-# and the flat sglang/deepseek-v4/8k1k recipes (sglang backend schema).
-
-model:
-  path: "deepseek-v4-pro"
-  container: "dynamo-sglang"
-  precision: "fp4"
-
-dynamo:
-  install: true
-  wheel: "1.3.0.dev20260718"
-
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 1440
-  interval_seconds: 10
-
-resources:
-  gpu_type: "gb300"
-  gpus_per_node: 4
-  agg_nodes: 1
-  agg_workers: 1
-  gpus_per_agg: 4
-
-infra:
-  # Dedicated etcd/nats node — matches the vllm agentic recipes; under the
-  # dynamo router + multiple frontends the infra services need their own node.
-  etcd_nats_dedicated_node: true
-  # cc-traces prompts (~125K tokens) serialize to ~2.8MB; the 1 MiB NATS
-  # default drops them. 32 MiB matches the agentic vllm recipes (~10x headroom
-  # over the largest observed payload); schema recommends 24+ for long ISL.
-  nats_max_payload_mb: 32
-
-frontend:
-  type: dynamo
-  nginx_session_affinity: true
-  nginx_session_affinity_header: X-Dynamo-Session-ID
-  enable_multiple_frontends: true
-  num_additional_frontends: 4
-  env:
-    # The sglang image is PEP 668 externally-managed; the runtime dynamo
-    # install (dynamo_wheels.py / source build) runs pip and fails with
-    # "externally-managed-environment" without this. Lets pip install into
-    # the system env. Applies to both dynamo.hash (source) and dynamo.wheel.
-    PIP_BREAK_SYSTEM_PACKAGES: "1"
-  args:
-    router-mode: "kv"
-    router-session-affinity-ttl-secs: "3600"
-    active-decode-blocks-threshold: "None"
-    active-prefill-tokens-threshold: "None"
-    active-prefill-tokens-threshold-frac: "None"
-
-backend:
-  type: sglang
-
-  aggregated_environment:
-    SGLANG_DEFAULT_THINKING: '1'
-    # TP4/bs128 can spend over an hour capturing the DSV4 EAGLE draft-extend
-    # graph. Keep the target/decode graphs and use the upstream eager fallback.
-    SGLANG_DISABLE_DRAFT_EXTEND_CUDA_GRAPH: "1"
-    SGLANG_DSV4_REASONING_EFFORT: high
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
-    # Lets the runtime dynamo install's pip bypass PEP 668 on the sglang image.
-    PIP_BREAK_SYSTEM_PACKAGES: "1"
-    # from submission for B300
-    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
-    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
-    SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
-    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
-    SGLANG_OPT_USE_ONLINE_COMPRESS: "0"
-    SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
-    SGLANG_OPT_USE_JIT_NORM: "1"
-    SGLANG_OPT_USE_TOPK_V2: "True"
-    # for kvcache offload
-    SGLANG_ENABLE_UNIFIED_RADIX_TREE: "1"
-
-  sglang_config:
-    aggregated:
-      served-model-name: "deepseek-ai/DeepSeek-V4-Pro"
-      enable-metrics: true
-      trust-remote-code: true
-      stream-interval: 50
-      watchdog-timeout: 1000000
-      mem-fraction-static: 0.90
-      chunked-prefill-size: 8192  # can only support 11000 for TP
-      moe-runner-backend: "flashinfer_mxfp4"
-      disable-flashinfer-autotune: true
-      swa-full-tokens-ratio: 0.1
-      max-running-requests: 128
-      cuda-graph-max-bs: 128
-
-      scheduler-recv-interval: 30
-      dp-size: 1
-      tp-size: 4
-      ep-size: 1
-      # kv-cache offload
-      enable-hierarchical-cache: true
-      hicache-ratio: 6
-      hicache-write-policy: write_through
-      hicache-io-backend: kernel
-      # mtp section
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
-      speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
-
-sbatch_directives:
-  mem: "0"
-  cpus-per-task: "144"
-
-srun_options:
-  mem: "0"
-  # gb300-nv: pyxis maps the calling user into the container as a non-root
-  # uid; agentic_srt.sh's apt-get install git step needs EUID 0. Remap to
-  # uid 0 inside the container. srt-slurm renders empty-string values as
-  # flag-only srun args.
-  container-remap-root: ""
-
-benchmark:
-  type: custom
-  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
-  env:
-    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
-    RESULT_DIR: /logs/agentic
-    PORT: "8000"
-    # Aggregated: one worker serves both prefill and decode, so GPU accounting
-    # is the single-worker form num_gpus = TP (not the disagg prefill+decode
-    # sum). The multinode post-processing path assumes disagg and would divide
-    # throughput by prefill_gpus + decode_gpus; force the single-node code path
-    # in process_agentic_result.py instead. TP must match sglang_config tp-size.
-    IS_MULTINODE: "false"
-    TP: "4"
-    # Stamp X-Dynamo-Session-ID on every turn so both nginx and Dynamo's
-    # router keep the session on the worker that owns its KV prefix.
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
-    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
-    # Container-side path of the aiperf mmap dataset cache; host-side mount
-    # wired via launch_gb300-*.sh srtslurm.yaml default_mounts. Without it,
-    # aiperf re-tokenizes + re-writes the dataset mmap on every run.
-    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
-    # Persistent HF hub cache (also via default_mounts) so the trace dataset
-    # isn't re-downloaded each run; overrides the workflow-level HF_HUB_CACHE.
-    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
new file mode 100644
index 000000000..5f125e330
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml
@@ -0,0 +1,120 @@
+name: "agg-gb300-tp4-mtp-lowlatency"
+
+# Low-latency AgentX aggregate topology: one TP4 worker occupies one
+# four-GPU GB300 node and serves both prefill and decode with DSpark K=6.
+
+model:
+  path: "deepseek-v4-pro-0813"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260816-4a6dc267"
+
+dynamo:
+  install: true
+  hash: "7a4e47ead90ec6b479e397996f6479ce7d251510"
+
+slurm:
+  time_limit: "4:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  agg_nodes: 1
+  agg_workers: 1
+  gpus_per_agg: 4
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: false
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    DYN_NATS_REQUEST_TIMEOUT_SECS: "1800"
+  args:
+    router-mode: "kv"
+    router-session-affinity-ttl-secs: "3600"
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+backend:
+  type: sglang
+  aggregated_environment:
+    SGLANG_RAGGED_VERIFY_MODE: "static"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: high
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
+    # Avoid the TP4 DSpark startup stall in the shared-read-done fastpath.
+    SGLANG_FORCE_COARSE_WAR_BARRIER: "1"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+    SGLANG_OPT_USE_ONLINE_COMPRESS: "0"
+    SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
+    SGLANG_OPT_USE_JIT_NORM: "1"
+    SGLANG_OPT_USE_TOPK_V2: "True"
+
+  sglang_config:
+    aggregated:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813"
+      enable-metrics: true
+      enable-cache-report: true
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      stream-interval: 10
+      watchdog-timeout: 1000000
+      mem-fraction-static: 0.94
+      chunked-prefill-size: 8192
+      max-prefill-tokens: 8192
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+      swa-full-tokens-ratio: 0.1
+      max-running-requests: 32
+      cuda-graph-max-bs-decode: 32
+      scheduler-recv-interval: 30
+      dp-size: 1
+      tp-size: 4
+      ep-size: 1
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "144"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "false"
+    TP: "4"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml
new file mode 100644
index 000000000..da765fd1e
--- /dev/null
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml
@@ -0,0 +1,120 @@
+name: "agg-gb300-tp8-mtp-lowlatency"
+
+# Low-latency AgentX aggregate topology: one TP8 worker spans two
+# four-GPU GB300 nodes and serves both prefill and decode with DSpark K=6.
+
+model:
+  path: "deepseek-v4-pro-0813"
+  container: "dynamo-sglang"
+  precision: "fp4"
+
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260816-4a6dc267"
+
+dynamo:
+  install: true
+  hash: "7a4e47ead90ec6b479e397996f6479ce7d251510"
+
+slurm:
+  time_limit: "4:00:00"
+
+health_check:
+  max_attempts: 1440
+  interval_seconds: 10
+
+resources:
+  gpu_type: "gb300"
+  gpus_per_node: 4
+  agg_nodes: 2
+  agg_workers: 1
+  gpus_per_agg: 8
+
+infra:
+  etcd_nats_dedicated_node: false
+  nats_max_payload_mb: 32
+
+frontend:
+  type: dynamo
+  nginx_session_affinity: true
+  nginx_session_affinity_header: X-Dynamo-Session-ID
+  enable_multiple_frontends: false
+  env:
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    DYN_NATS_REQUEST_TIMEOUT_SECS: "1800"
+  args:
+    router-mode: "kv"
+    router-session-affinity-ttl-secs: "3600"
+    active-decode-blocks-threshold: "None"
+    active-prefill-tokens-threshold: "None"
+    active-prefill-tokens-threshold-frac: "None"
+
+backend:
+  type: sglang
+  aggregated_environment:
+    SGLANG_RAGGED_VERIFY_MODE: "static"
+    # Avoid aggregate DSpark startup stalls in the shared-read-done fastpath.
+    SGLANG_FORCE_COARSE_WAR_BARRIER: "1"
+    SGLANG_DEFAULT_THINKING: "1"
+    SGLANG_DSV4_REASONING_EFFORT: high
+    PIP_BREAK_SYSTEM_PACKAGES: "1"
+    SGLANG_JIT_DEEPGEMM_PRECOMPILE: "1"
+    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: "1"
+    SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
+    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: "1"
+    SGLANG_OPT_USE_ONLINE_COMPRESS: "0"
+    SGLANG_OPT_USE_JIT_INDEXER_METADATA: "1"
+    SGLANG_OPT_USE_JIT_NORM: "1"
+    SGLANG_OPT_USE_TOPK_V2: "True"
+
+  sglang_config:
+    aggregated:
+      served-model-name: "deepseek-ai/DeepSeek-V4-Pro-0813"
+      enable-metrics: true
+      enable-cache-report: true
+      trust-remote-code: true
+      weight-loader-prefetch-checkpoints: true
+      stream-interval: 10
+      watchdog-timeout: 1000000
+      mem-fraction-static: 0.94
+      chunked-prefill-size: 8192
+      max-prefill-tokens: 8192
+      moe-runner-backend: "flashinfer_mxfp4"
+      disable-flashinfer-autotune: true
+      swa-full-tokens-ratio: 0.1
+      max-running-requests: 4
+      cuda-graph-max-bs-decode: 4
+      scheduler-recv-interval: 30
+      dp-size: 1
+      tp-size: 8
+      ep-size: 1
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
+      speculative-eagle-topk: 1
+      speculative-num-draft-tokens: 7
+
+sbatch_directives:
+  mem: "0"
+  cpus-per-task: "144"
+
+srun_options:
+  mem: "0"
+  container-remap-root: ""
+
+benchmark:
+  type: custom
+  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
+  env:
+    INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
+    RESULT_DIR: "/logs/agentic"
+    PORT: "8000"
+    IS_MULTINODE: "false"
+    TP: "8"
+    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
+    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
+    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
+    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml
deleted file mode 100644
index e41438765..000000000
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml
+++ /dev/null
@@ -1,216 +0,0 @@
-name: "disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload"
-
-# Agentic-coding SGLang disaggregated Pareto recipe for DeepSeek-V4-Pro on GB300
-# (2P x DEP8 / 1D x TP4, MTP + hierarchical-cache KV offload), tuned for concurrency 80.
-#
-# Uses the flat single-variant srtctl schema the agentic CI flow expects;
-# resources + backend (prefill/decode env + sglang_config) are normalized
-# from the Pareto run.
-# Concurrency is exported into agentic_srt.sh from the master-config conc-list.
-
-model:
-  path: "deepseek-v4-pro"
-  container: "dynamo-sglang"
-  precision: "fp4"
-
-dynamo:
-  install: true
-  wheel: "1.3.0.dev20260718"
-
-slurm:
-  time_limit: "8:00:00"
-
-health_check:
-  max_attempts: 1440
-  interval_seconds: 10
-
-resources:
-  gpu_type: gb300
-  gpus_per_node: 4
-  prefill_nodes: 2
-  decode_nodes: 1
-  prefill_workers: 1
-  decode_workers: 1
-  gpus_per_prefill: 8
-  gpus_per_decode: 4
-
-infra:
-  etcd_nats_dedicated_node: true
-  nats_max_payload_mb: 32
-
-frontend:
-  type: dynamo
-  nginx_session_affinity: true
-  nginx_session_affinity_header: X-Dynamo-Session-ID
-  enable_multiple_frontends: true
-  num_additional_frontends: 4
-  env:
-    PIP_BREAK_SYSTEM_PACKAGES: "1"
-  args:
-    router-mode: "kv"
-    router-session-affinity-ttl-secs: "3600"
-    active-decode-blocks-threshold: "None"
-    active-prefill-tokens-threshold: "None"
-    active-prefill-tokens-threshold-frac: "None"
-
-backend:
-  type: sglang
-
-  prefill_environment:
-    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
-    SGLANG_DSV4_MHC_PREWARM: '1'
-    PIP_BREAK_SYSTEM_PACKAGES: '1'
-    PYTHONUNBUFFERED: '1'
-    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
-    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
-    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '1'
-    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: '1'
-    SGLANG_DEFAULT_THINKING: '1'
-    SGLANG_DSV4_REASONING_EFFORT: high
-    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
-    SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
-    SGLANG_OPT_USE_CUSTOM_ALL_REDUCE_V2: '1'
-    SGLANG_OPT_USE_DEEPGEMM_MEGA_MOE: '1'
-    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: '1'
-    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: '1'
-    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '9216'
-    SGLANG_OPT_USE_ONLINE_COMPRESS: '0'
-    NCCL_MNNVL_ENABLE: '1'
-    NCCL_CUMEM_ENABLE: '1'
-    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
-    MC_FORCE_MNNVL: '1'
-    NCCL_TIMEOUT: '100000'
-    NCCL_MNNVL_UUID_QUERY_TIMEOUT: '100000'
-    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
-    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
-    DYN_LOG: info,dynamo_runtime::pipeline::network::ingress::push_handler=warn
-    SGLANG_LOG_FORWARD_ITERS: '1'
-    SGLANG_LOG_MS: '1'
-    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
-    SGLANG_ENABLE_UNIFIED_RADIX_TREE: '1'
-
-  decode_environment:
-    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
-    # Match the TP4 aggregate worker: avoid the pathological DSV4 EAGLE
-    # draft-extend graph capture while retaining the target/decode graphs.
-    SGLANG_DISABLE_DRAFT_EXTEND_CUDA_GRAPH: "1"
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
-    PIP_BREAK_SYSTEM_PACKAGES: '1'
-    PYTHONUNBUFFERED: '1'
-    SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
-    SGLANG_DG_CACHE_DIR: /configs/deepgemm_cache
-    SGLANG_JIT_DEEPGEMM_PRECOMPILE: '1'
-    SGLANG_JIT_DEEPGEMM_FAST_WARMUP: '1'
-    SGLANG_DEFAULT_THINKING: '1'
-    SGLANG_DSV4_REASONING_EFFORT: high
-    SGLANG_OPT_SWA_SPLIT_LEAF_ON_INSERT: '1'
-    SGLANG_OPT_UNIFIED_CACHE_FREE_OUT_OF_WINDOW_SLOTS: "1"
-    SGLANG_OPT_DEEPGEMM_MEGA_MOE: '1'
-    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS: '1'
-    SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_MXF4_KIND: '1'
-    SGLANG_OPT_DEEPGEMM_MEGA_MOE_NUM_MAX_TOKENS_PER_RANK: '2048'
-    SGLANG_OPT_USE_ONLINE_COMPRESS: '0'
-    NCCL_MNNVL_ENABLE: '1'
-    NCCL_CUMEM_ENABLE: '1'
-    SGLANG_MOONCAKE_CUSTOM_MEM_POOL: 'True'
-    MC_FORCE_MNNVL: '1'
-    NCCL_TIMEOUT: '100000'
-    NCCL_MNNVL_UUID_QUERY_TIMEOUT: '100000'
-    SGLANG_OPT_SWA_RELEASE_LEAF_LOCK_AFTER_WINDOW: '1'
-    SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION: '8'
-    DYN_SKIP_SGLANG_LOG_FORMATTING: '1'
-    DYN_LOG: info,dynamo_runtime::pipeline::network::ingress::push_handler=warn
-    SGLANG_LOG_FORWARD_ITERS: '1'
-    SGLANG_LOG_MS: '1'
-    SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
-    SGLANG_ENABLE_UNIFIED_RADIX_TREE: '1'
-
-  sglang_config:
-    prefill:
-      host: 0.0.0.0
-      served-model-name: deepseek-ai/DeepSeek-V4-Pro
-      enable-metrics: true
-      model-path: /model/
-      trust-remote-code: true
-      watchdog-timeout: 86400
-      stream-interval: 60
-      tp-size: 8
-      dp-size: 8
-      ep-size: 8
-      enable-dp-attention: true
-      enable-dp-lm-head: true
-      moe-dense-tp-size: 1
-      moe-a2a-backend: megamoe
-      disaggregation-transfer-backend: mooncake
-      disaggregation-mode: prefill
-      load-balance-method: total_tokens
-      mem-fraction-static: 0.9
-      swa-full-tokens-ratio: 0.02
-      max-running-requests: 128
-      cuda-graph-max-bs: 128
-      chunked-prefill-size: 32768
-      disable-flashinfer-autotune: true
-      model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
-      kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}'
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
-      speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
-      enable-hierarchical-cache: true
-      hicache-write-policy: write_back
-      hicache-ratio: 1
-      hicache-io-backend: direct
-
-    decode:
-      host: 0.0.0.0
-      served-model-name: deepseek-ai/DeepSeek-V4-Pro
-      enable-metrics: true
-      model-path: /model/
-      trust-remote-code: true
-      watchdog-timeout: 86400
-      stream-interval: 60
-      tp-size: 4
-      dp-size: 1
-      ep-size: 1
-      enable-dp-attention: false
-      enable-dp-lm-head: false
-      moe-runner-backend: flashinfer_mxfp4
-      disaggregation-transfer-backend: mooncake
-      disaggregation-mode: decode
-      load-balance-method: total_tokens
-      mem-fraction-static: 0.9
-      swa-full-tokens-ratio: 0.02
-      max-running-requests: 128
-      cuda-graph-max-bs: 128
-      disable-flashinfer-autotune: true
-      model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
-      kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5567"}'
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
-      speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
-
-
-sbatch_directives:
-  mem: "0"
-  cpus-per-task: "144"
-
-srun_options:
-  mem: "0"
-  container-remap-root: ""
-
-benchmark:
-  type: custom
-  command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
-  env:
-    INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
-    RESULT_DIR: /logs/agentic
-    PORT: "8000"
-    IS_MULTINODE: "true"
-    AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
-    AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
-    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
-    AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
-    HF_HUB_CACHE: "/hf_hub_cache"
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c192-mtp-kvoffload.yaml
similarity index 86%
rename from benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c192-mtp-kvoffload.yaml
index 74b371e94..eb1ef5313 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c192-mtp-kvoffload.yaml
@@ -1,7 +1,7 @@
-name: "disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload"
+name: "disagg-gb300-2p4d-dep8-dep16-c192-mtp-kvoffload"
 
 # Agentic-coding SGLang disaggregated Pareto recipe for DeepSeek-V4-Pro on GB300
-# (2P x DEP8 / 4D x DEP16, MTP + hierarchical-cache KV offload), tuned for concurrency 256.
+# (2P x DEP8 / 4D x DEP16, DSpark K=6 + hierarchical-cache KV offload), tuned for concurrency 192.
 #
 # Uses the flat single-variant srtctl schema the agentic CI flow expects;
 # resources + backend (prefill/decode env + sglang_config) are normalized
@@ -9,13 +9,19 @@ name: "disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload"
 # Concurrency is exported into agentic_srt.sh from the master-config conc-list.
 
 model:
-  path: "deepseek-v4-pro"
+  path: "deepseek-v4-pro-0813"
   container: "dynamo-sglang"
   precision: "fp4"
 
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260816-4a6dc267"
+
 dynamo:
   install: true
-  wheel: "1.3.0.dev20260718"
+  hash: "7a4e47ead90ec6b479e397996f6479ce7d251510"
 
 slurm:
   time_limit: "8:00:00"
@@ -57,6 +63,7 @@ backend:
   type: sglang
 
   prefill_environment:
+    SGLANG_RAGGED_VERIFY_MODE: "static"
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
     SGLANG_DSV4_MHC_PREWARM: '1'
     PIP_BREAK_SYSTEM_PACKAGES: '1'
@@ -90,10 +97,8 @@ backend:
     SGLANG_ENABLE_UNIFIED_RADIX_TREE: '1'
 
   decode_environment:
+    SGLANG_RAGGED_VERIFY_MODE: "static"
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
@@ -127,8 +132,9 @@ backend:
   sglang_config:
     prefill:
       host: 0.0.0.0
-      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
       enable-metrics: true
+      enable-cache-report: true
       model-path: /model/
       trust-remote-code: true
       watchdog-timeout: 86400
@@ -143,18 +149,19 @@ backend:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      mem-fraction-static: 0.9
+      mem-fraction-static: 0.85
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
       cuda-graph-max-bs: 256
-      chunked-prefill-size: 32768
+      chunked-prefill-size: 65536
       disable-flashinfer-autotune: true
       model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
       kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}'
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
       speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
+      speculative-num-draft-tokens: 7
       enable-hierarchical-cache: true
       hicache-write-policy: write_back
       hicache-ratio: 1
@@ -162,8 +169,9 @@ backend:
 
     decode:
       host: 0.0.0.0
-      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
       enable-metrics: true
+      enable-cache-report: true
       model-path: /model/
       trust-remote-code: true
       watchdog-timeout: 86400
@@ -185,10 +193,11 @@ backend:
       disable-flashinfer-autotune: true
       model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
       kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5567"}'
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
       speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
+      speculative-num-draft-tokens: 7
 
 
 sbatch_directives:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c400-mtp-kvoffload.yaml
similarity index 86%
rename from benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c400-mtp-kvoffload.yaml
index 6b585fba8..df487dd82 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c400-mtp-kvoffload.yaml
@@ -1,7 +1,7 @@
-name: "disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload"
+name: "disagg-gb300-4p4d-dep8-dep16-c400-mtp-kvoffload"
 
 # Agentic-coding SGLang disaggregated recipe for DeepSeek-V4-Pro on GB300
-# (4P x DEP8 / 4D x DEP16, MTP + hierarchical-cache KV offload), tuned for concurrency 512.
+# (4P x DEP8 / 4D x DEP16, DSpark K=6 + hierarchical-cache KV offload), tuned for concurrency 400.
 #
 # DEP8-prefill variant aligned with the measured Pareto point: prefill uses
 # tp/dp/ep 8, four nodes, and SGLANG_DSV4_MHC_PREWARM=1.
@@ -9,13 +9,19 @@ name: "disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload"
 # from the master-config conc-list.
 
 model:
-  path: "deepseek-v4-pro"
+  path: "deepseek-v4-pro-0813"
   container: "dynamo-sglang"
   precision: "fp4"
 
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260816-4a6dc267"
+
 dynamo:
   install: true
-  wheel: "1.3.0.dev20260718"
+  hash: "7a4e47ead90ec6b479e397996f6479ce7d251510"
 
 slurm:
   time_limit: "8:00:00"
@@ -56,6 +62,7 @@ backend:
   type: sglang
 
   prefill_environment:
+    SGLANG_RAGGED_VERIFY_MODE: "static"
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
@@ -89,10 +96,8 @@ backend:
     SGLANG_DSV4_MHC_PREWARM: '1'
 
   decode_environment:
+    SGLANG_RAGGED_VERIFY_MODE: "static"
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
@@ -127,8 +132,9 @@ backend:
   sglang_config:
     prefill:
       host: 0.0.0.0
-      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
       enable-metrics: true
+      enable-cache-report: true
       model-path: /model/
       trust-remote-code: true
       watchdog-timeout: 86400
@@ -143,18 +149,19 @@ backend:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      mem-fraction-static: 0.9
+      mem-fraction-static: 0.85
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
       cuda-graph-max-bs: 256
-      chunked-prefill-size: 32768
+      chunked-prefill-size: 65536
       disable-flashinfer-autotune: true
       model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
       kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}'
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
       speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
+      speculative-num-draft-tokens: 7
       enable-hierarchical-cache: true
       hicache-write-policy: write_back
       hicache-ratio: 1
@@ -162,8 +169,9 @@ backend:
 
     decode:
       host: 0.0.0.0
-      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
       enable-metrics: true
+      enable-cache-report: true
       model-path: /model/
       trust-remote-code: true
       watchdog-timeout: 86400
@@ -185,10 +193,11 @@ backend:
       disable-flashinfer-autotune: true
       model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
       kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5567"}'
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
       speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
+      speculative-num-draft-tokens: 7
 
 
 sbatch_directives:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c640-mtp-kvoffload.yaml
similarity index 86%
rename from benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c640-mtp-kvoffload.yaml
index b2cac21c7..d78f36153 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c640-mtp-kvoffload.yaml
@@ -1,7 +1,7 @@
-name: "disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload"
+name: "disagg-gb300-6p4d-dep8-dep16-c640-mtp-kvoffload"
 
 # Agentic-coding SGLang disaggregated Pareto recipe for DeepSeek-V4-Pro on GB300
-# (6P x DEP8 / 4D x DEP16, MTP + hierarchical-cache KV offload), tuned for concurrency 768.
+# (6P x DEP8 / 4D x DEP16, DSpark K=6 + hierarchical-cache KV offload), tuned for concurrency 640.
 #
 # Uses the flat single-variant srtctl schema the agentic CI flow expects;
 # resources + backend (prefill/decode env + sglang_config) are normalized
@@ -9,13 +9,19 @@ name: "disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload"
 # Concurrency is exported into agentic_srt.sh from the master-config conc-list.
 
 model:
-  path: "deepseek-v4-pro"
+  path: "deepseek-v4-pro-0813"
   container: "dynamo-sglang"
   precision: "fp4"
 
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260816-4a6dc267"
+
 dynamo:
   install: true
-  wheel: "1.3.0.dev20260718"
+  hash: "7a4e47ead90ec6b479e397996f6479ce7d251510"
 
 slurm:
   time_limit: "8:00:00"
@@ -57,6 +63,7 @@ backend:
   type: sglang
 
   prefill_environment:
+    SGLANG_RAGGED_VERIFY_MODE: "static"
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
     SGLANG_DSV4_MHC_PREWARM: '1'
     PIP_BREAK_SYSTEM_PACKAGES: '1'
@@ -90,10 +97,8 @@ backend:
     SGLANG_ENABLE_UNIFIED_RADIX_TREE: '1'
 
   decode_environment:
+    SGLANG_RAGGED_VERIFY_MODE: "static"
     SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
@@ -127,8 +132,9 @@ backend:
   sglang_config:
     prefill:
       host: 0.0.0.0
-      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
       enable-metrics: true
+      enable-cache-report: true
       model-path: /model/
       trust-remote-code: true
       watchdog-timeout: 86400
@@ -143,18 +149,19 @@ backend:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      mem-fraction-static: 0.9
+      mem-fraction-static: 0.85
       swa-full-tokens-ratio: 0.02
       max-running-requests: 512
       cuda-graph-max-bs: 512
-      chunked-prefill-size: 32768
+      chunked-prefill-size: 65536
       disable-flashinfer-autotune: true
       model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
       kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}'
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
       speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
+      speculative-num-draft-tokens: 7
       enable-hierarchical-cache: true
       hicache-write-policy: write_back
       hicache-ratio: 1
@@ -162,8 +169,9 @@ backend:
 
     decode:
       host: 0.0.0.0
-      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
       enable-metrics: true
+      enable-cache-report: true
       model-path: /model/
       trust-remote-code: true
       watchdog-timeout: 86400
@@ -185,10 +193,11 @@ backend:
       disable-flashinfer-autotune: true
       model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
       kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5567"}'
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
       speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
+      speculative-num-draft-tokens: 7
 
 
 sbatch_directives:
diff --git a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-8p4d-dep8-dep16-c1152-mtp-kvoffload.yaml
similarity index 83%
rename from benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml
rename to benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-8p4d-dep8-dep16-c1152-mtp-kvoffload.yaml
index 709962936..798d96345 100644
--- a/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml
+++ b/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic/disagg-gb300-8p4d-dep8-dep16-c1152-mtp-kvoffload.yaml
@@ -1,7 +1,7 @@
-name: "disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload"
+name: "disagg-gb300-8p4d-dep8-dep16-c1152-mtp-kvoffload"
 
 # Agentic-coding SGLang disaggregated Pareto recipe for DeepSeek-V4-Pro on GB300
-# (12P x DEP8 / 4D x DEP16, MTP + hierarchical-cache KV offload), tuned for concurrency 1536.
+# (8P x DEP8 / 4D x DEP16, DSpark K=6 + hierarchical-cache KV offload), tuned for concurrency 1152.
 #
 # Uses the flat single-variant srtctl schema the agentic CI flow expects;
 # resources + backend (prefill/decode env + sglang_config) are normalized
@@ -9,13 +9,19 @@ name: "disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload"
 # Concurrency is exported into agentic_srt.sh from the master-config conc-list.
 
 model:
-  path: "deepseek-v4-pro"
+  path: "deepseek-v4-pro-0813"
   container: "dynamo-sglang"
   precision: "fp4"
 
+identity:
+  model:
+    repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
+  container:
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260816-4a6dc267"
+
 dynamo:
   install: true
-  wheel: "1.3.0.dev20260718"
+  hash: "7a4e47ead90ec6b479e397996f6479ce7d251510"
 
 slurm:
   time_limit: "8:00:00"
@@ -27,9 +33,9 @@ health_check:
 resources:
   gpu_type: gb300
   gpus_per_node: 4
-  prefill_nodes: 12
+  prefill_nodes: 8
   decode_nodes: 4
-  prefill_workers: 6
+  prefill_workers: 4
   decode_workers: 1
   gpus_per_prefill: 8
   gpus_per_decode: 16
@@ -42,6 +48,7 @@ frontend:
   type: dynamo
   nginx_session_affinity: true
   nginx_session_affinity_header: X-Dynamo-Session-ID
+  nginx_keepalive_timeout: "900s"
   enable_multiple_frontends: true
   num_additional_frontends: 4
   env:
@@ -60,7 +67,8 @@ backend:
   type: sglang
 
   prefill_environment:
-    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+    SGLANG_RAGGED_VERIFY_MODE: "static"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "3600"
     SGLANG_DSV4_MHC_PREWARM: '1'
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
@@ -93,10 +101,8 @@ backend:
     SGLANG_ENABLE_UNIFIED_RADIX_TREE: '1'
 
   decode_environment:
-    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
-    SGLANG_SIMULATE_ACC_LEN: '2.49'
-    SGLANG_SIMULATE_ACC_METHOD: match-expected
-    SGLANG_SIMULATE_ACC_TOKEN_MODE: "real-draft-token"
+    SGLANG_RAGGED_VERIFY_MODE: "static"
+    SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "3600"
     PIP_BREAK_SYSTEM_PACKAGES: '1'
     PYTHONUNBUFFERED: '1'
     SGLANG_DISABLE_TP_MEMORY_INBALANCE_CHECK: '1'
@@ -130,8 +136,9 @@ backend:
   sglang_config:
     prefill:
       host: 0.0.0.0
-      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
       enable-metrics: true
+      enable-cache-report: true
       model-path: /model/
       trust-remote-code: true
       watchdog-timeout: 86400
@@ -146,18 +153,19 @@ backend:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      mem-fraction-static: 0.9
+      mem-fraction-static: 0.85
       swa-full-tokens-ratio: 0.02
       max-running-requests: 1024
       cuda-graph-max-bs: 1024
-      chunked-prefill-size: 32768
+      chunked-prefill-size: 65536
       disable-flashinfer-autotune: true
       model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
       kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5557"}'
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
       speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
+      speculative-num-draft-tokens: 7
       enable-hierarchical-cache: true
       hicache-write-policy: write_back
       hicache-ratio: 1
@@ -165,8 +173,9 @@ backend:
 
     decode:
       host: 0.0.0.0
-      served-model-name: deepseek-ai/DeepSeek-V4-Pro
+      served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
       enable-metrics: true
+      enable-cache-report: true
       model-path: /model/
       trust-remote-code: true
       watchdog-timeout: 86400
@@ -184,14 +193,15 @@ backend:
       mem-fraction-static: 0.9
       swa-full-tokens-ratio: 0.02
       max-running-requests: 3072
-      cuda-graph-max-bs: 1024
+      cuda-graph-max-bs: 192
       disable-flashinfer-autotune: true
       model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
       kv-events-config: '{"publisher":"zmq","topic":"kv-events","endpoint":"tcp://*:5567"}'
-      speculative-algorithm: EAGLE
-      speculative-num-steps: 3
+      speculative-algorithm: DSPARK
+      speculative-dspark-block-size: 6
+      speculative-num-steps: 1
       speculative-eagle-topk: 1
-      speculative-num-draft-tokens: 4
+      speculative-num-draft-tokens: 7
 
 
 sbatch_directives:
@@ -211,6 +221,7 @@ benchmark:
     PORT: "8000"
     IS_MULTINODE: "true"
     AIPERF_HTTP_X_DYNAMO_SESSION_ID_FROM_CORRELATION_ID: "true"
+    AIPERF_HTTP_TCP_USER_TIMEOUT: "900000"
     AIPERF_USE_DYNAMO_CONV_AWARE_ROUTING: "0"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "sglang:"
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 2f73b1c64..7be3444cc 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7784,33 +7784,46 @@ kimik3-fp4-gb200-dynamo-vllm-agentic:
           dp-attn: true
 
 dsv4-fp4-gb300-dynamo-sglang-agentic-agg:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260719-99f5a6f4
-  model: deepseek-ai/DeepSeek-V4-Pro
+  image: lmsysorg/sglang:nightly-dev-cu13-20260816-4a6dc267
+  model: deepseek-ai/DeepSeek-V4-Pro-0813
   model-prefix: dsv4
   runner: cluster:gb300-nv
   precision: fp4
   framework: dynamo-sglang
-  router: { name: dynamo-router, version: "1.3.0.dev20260718" }
+  router: { name: dynamo-router, version: "7a4e47ead90ec6b479e397996f6479ce7d251510" }
   multinode: true
   disagg: false
   scenarios:
     agentic-coding:
-    - dram-utilization: 0.80
-      search-space:
-      - spec-decoding: mtp
-        conc-list: [2, 4, 8, 16]
-        kv-offloading: dram
-        kv-offload-backend: { name: hicache }
+    - search-space:
+      - spec-decoding: draft_model
+        conc-list: [1, 4]
+        prefill:
+          num-worker: 1
+          tp: 8
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/agg-gb300-tp8-mtp-lowlatency.yaml"
+        decode:
+          num-worker: 0
+          tp: 8
+          ep: 1
+          dp-attn: false
+    - search-space:
+      - spec-decoding: draft_model
+        conc-list: [8]
         prefill:
           num-worker: 1
           tp: 4
           ep: 1
           dp-attn: false
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-kvoffload.yaml"
-        # Aggregated worker: prefill's GPUs also serve decode, so decode has no
-        # separate allocation (num-worker: 0). Deployment is recipe-driven
-        # (agg_workers: 1); this only makes GPU accounting count 4, not 4+4.
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/agg-gb300-tp4-mtp-lowlatency.yaml"
         decode:
           num-worker: 0
           tp: 4
@@ -7818,13 +7831,13 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-agg:
           dp-attn: false
 
 dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260719-99f5a6f4
-  model: deepseek-ai/DeepSeek-V4-Pro
+  image: lmsysorg/sglang:nightly-dev-cu13-20260816-4a6dc267
+  model: deepseek-ai/DeepSeek-V4-Pro-0813
   model-prefix: dsv4
   runner: cluster:gb300-nv
   precision: fp4
   framework: dynamo-sglang
-  router: { name: dynamo-router, version: "1.3.0.dev20260718" }
+  router: { name: dynamo-router, version: "7a4e47ead90ec6b479e397996f6479ce7d251510" }
   kv-p2p-transfer: mooncake
   multinode: true
   disagg: true
@@ -7832,24 +7845,8 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - spec-decoding: mtp
-        conc-list: [80]
-        kv-offloading: dram
-        kv-offload-backend: { name: hicache }
-        prefill:
-          num-worker: 1
-          tp: 8
-          ep: 8
-          dp-attn: true
-          additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p1d-dep8-tp4-c80-mtp-kvoffload.yaml"
-        decode:
-          num-worker: 1
-          tp: 4
-          ep: 1
-          dp-attn: false
-      - spec-decoding: mtp
-        conc-list: [256]
+      - spec-decoding: draft_model
+        conc-list: [192]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
         prefill:
@@ -7858,14 +7855,16 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c256-mtp-kvoffload.yaml"
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-2p4d-dep8-dep16-c192-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
           tp: 16
           ep: 16
           dp-attn: true
-      - spec-decoding: mtp
-        conc-list: [512]
+      - spec-decoding: draft_model
+        conc-list: [400]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
         prefill:
@@ -7874,14 +7873,16 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c512-mtp-kvoffload.yaml"
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-4p4d-dep8-dep16-c400-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
           tp: 16
           ep: 16
           dp-attn: true
-      - spec-decoding: mtp
-        conc-list: [768]
+      - spec-decoding: draft_model
+        conc-list: [640]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
         prefill:
@@ -7890,23 +7891,27 @@ dsv4-fp4-gb300-dynamo-sglang-agentic-disagg:
           ep: 8
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c768-mtp-kvoffload.yaml"
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-6p4d-dep8-dep16-c640-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
           tp: 16
           ep: 16
           dp-attn: true
-      - spec-decoding: mtp
-        conc-list: [1536]
+      - spec-decoding: draft_model
+        conc-list: [1152]
         kv-offloading: dram
         kv-offload-backend: { name: hicache }
         prefill:
-          num-worker: 6
+          num-worker: 4
           tp: 8
           ep: 8
           dp-attn: true
           additional-settings:
-          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-12p4d-dep8-dep16-c1536-mtp-kvoffload.yaml"
+          - "SYNTHETIC_ACCEPTANCE=true"
+          - "SYNTHETIC_ACCEPTANCE_LENGTH=3.77"
+          - "CONFIG_FILE=recipes/sglang/deepseek-v4/agentic/disagg-gb300-8p4d-dep8-dep16-c1152-mtp-kvoffload.yaml"
         decode:
           num-worker: 1
           tp: 16
diff --git a/golden_al_distribution/README.md b/golden_al_distribution/README.md
index 95e4a5622..2993f9c1b 100644
--- a/golden_al_distribution/README.md
+++ b/golden_al_distribution/README.md
@@ -109,6 +109,8 @@ Before accepting an updated curve, reviewers should verify:
 | Model | Method | Golden YAML | Source run |
 | --- | --- | --- | --- |
 | DeepSeek V4 Pro | MTP | [`dsv4_mtp.yaml`](dsv4_mtp.yaml) | [27180633016](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27180633016) |
+| DeepSeek V4 Pro DSpark | DSpark (probabilistic drafting) | [`dsv4_dspark_probabilistic.yaml`](dsv4_dspark_probabilistic.yaml) | [31279568355](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/31279568355) |
+| DeepSeek V4 Pro 0813 | DSpark (probabilistic drafting) | [`dsv4_0813_dspark_probabilistic.yaml`](dsv4_0813_dspark_probabilistic.yaml) | Provided 2026-08-17; same SpeedBench AL method |
 | Qwen3.5 397B-A17B | MTP | [`qwen3.5_mtp.yaml`](qwen3.5_mtp.yaml) | [27317114007](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27317114007) |
 | Kimi K2.5 | EAGLE3 | [`kimik2.5_eagle3.yaml`](kimik2.5_eagle3.yaml) | [28122195822](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28122195822) |
 | MiniMax-M3 | EAGLE3 | [`minimaxm3_eagle3.yaml`](minimaxm3_eagle3.yaml) | [28061204145](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28061204145) |
diff --git a/golden_al_distribution/README_zh.md b/golden_al_distribution/README_zh.md
index a56a005f2..18b38be80 100644
--- a/golden_al_distribution/README_zh.md
+++ b/golden_al_distribution/README_zh.md
@@ -109,6 +109,8 @@ gh workflow run speedbench-al.yml \
 | 模型 | 方法 | 黄金 YAML | 源 run |
 | --- | --- | --- | --- |
 | DeepSeek V4 Pro | MTP | [`dsv4_mtp.yaml`](dsv4_mtp.yaml) | [27180633016](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27180633016) |
+| DeepSeek V4 Pro DSpark | DSpark（概率式草稿采样） | [`dsv4_dspark_probabilistic.yaml`](dsv4_dspark_probabilistic.yaml) | [31279568355](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/31279568355) |
+| DeepSeek V4 Pro 0813 | DSpark（概率式草稿采样） | [`dsv4_0813_dspark_probabilistic.yaml`](dsv4_0813_dspark_probabilistic.yaml) | 2026-08-17 提供；使用相同 SpeedBench AL 测法 |
 | Qwen3.5 397B-A17B | MTP | [`qwen3.5_mtp.yaml`](qwen3.5_mtp.yaml) | [27317114007](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/27317114007) |
 | Kimi K2.5 | EAGLE3 | [`kimik2.5_eagle3.yaml`](kimik2.5_eagle3.yaml) | [28122195822](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28122195822) |
 | MiniMax-M3 | EAGLE3 | [`minimaxm3_eagle3.yaml`](minimaxm3_eagle3.yaml) | [28061204145](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/28061204145) |
diff --git a/golden_al_distribution/dsv4_0813_dspark_probabilistic.yaml b/golden_al_distribution/dsv4_0813_dspark_probabilistic.yaml
new file mode 100644
index 000000000..1333b4de8
--- /dev/null
+++ b/golden_al_distribution/dsv4_0813_dspark_probabilistic.yaml
@@ -0,0 +1,24 @@
+# Acceptance Length (AL) reference values measured with the same SpeedBench
+# setup as the existing DeepSeek-V4-Pro DSpark probabilistic curve, using the
+# DeepSeek-V4-Pro-0813 model. Values provided on 2026-08-17.
+#
+# key = num_speculative_tokens (DSpark level); value = golden AL
+deepseek-v4-pro-0813:
+  thinking_on:
+    1: 1.84
+    2: 2.51
+    3: 3.01
+    4: 3.36
+    5: 3.61
+    6: 3.77
+    7: 3.73
+    8: 3.47
+  thinking_off:
+    1: 1.94
+    2: 2.81
+    3: 3.58
+    4: 4.17
+    5: 4.74
+    6: 5.12
+    7: 5.24
+    8: 4.95
diff --git a/golden_al_distribution/dsv4_dspark_probabilistic.yaml b/golden_al_distribution/dsv4_dspark_probabilistic.yaml
new file mode 100644
index 000000000..516aeb2ab
--- /dev/null
+++ b/golden_al_distribution/dsv4_dspark_probabilistic.yaml
@@ -0,0 +1,28 @@
+# Source GitHub Actions run: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/31279568355
+# Acceptance Length (AL) reference values measured with SPEED-Bench.
+# dataset: coding | temperature: 1.0 | output_len: 4096
+# thinking_on chat_template_kwargs: {"thinking": true, "reasoning_effort": "high"}
+# speculative-config: method=dspark | draft_sample_method=probabilistic
+# Measured on deepseek-v4-pro-dspark (B300, vLLM DSpark), per num_speculative_tokens.
+# Auto-generated by benchmarks/single_node/speedbench/dsv4dspark_fp4_b300_vllm.sh (speedbench-al.yml).
+#
+# key = num_speculative_tokens (DSpark level); value = golden AL
+deepseek-v4-pro-dspark:
+  thinking_on:
+    1: 1.86
+    2: 2.55
+    3: 3.08
+    4: 3.43
+    5: 3.74
+    6: 3.82
+    7: 3.86
+    8: 3.75
+  thinking_off:
+    1: 1.93
+    2: 2.76
+    3: 3.47
+    4: 4.10
+    5: 4.52
+    6: 4.87
+    7: 4.94
+    8: 4.94
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 085e9cb88..b8ea19f20 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6039,7 +6039,6 @@
     - "EVAL_ONLY runs switch to real block verification by removing synthetic acceptance (TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS stripped from all agentx recipes before srtctl)."
     - "Add SERVED_MODEL_NAME: Qwen3.5-397B-A17B-NVFP4-V2 to all 6 agentx recipe env blocks and make build_replay_cmd prefer SERVED_MODEL_NAME over $MODEL for the aiperf --model flag."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2612
-
 - config-keys:
     - glm5.2-fp4-gb200-dynamo-sglang-agentic-agg
     - glm5.2-fp4-gb200-dynamo-sglang-agentic-disagg
@@ -6049,3 +6048,35 @@
     - "Add GB200 GLM-5.2 FP4 Dynamo-SGLang AgentX with EAGLE MTP, HiCache DRAM offload, and KV-aware correlation-ID affinity."
     - "Measure the TP8 aggregate curve and tuned 1P1D TP4 HiCache c10/c12 disaggregated frontier with every logical SGLang metrics endpoint."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2620
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-agg
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Replace the GB300 DeepSeek-V4-Pro AgentX EAGLE/MTP serving path with DSpark block size 6 across two aggregate and four disaggregated recipes"
+    - "Use the official DeepSeek-V4-Pro-0813 checkpoint with SGLANG_RAGGED_VERIFY_MODE=static, MegaMoE for DEP workers, and throughput-only thinking-on acceptance length 3.77 from the committed DeepSeek-V4-Pro-0813 probabilistic DSpark curve (golden_al_distribution/dsv4_0813_dspark_probabilistic.yaml); eval keeps real DSpark verification"
+    - "Refresh SGLang to nightly-dev-cu13-20260816-4a6dc267, which already contains the required DSpark performance and MegaMoE support"
+    - "Retune the Pareto ladder to concurrency 1/4/8, 192, 400, 640, and 1536; the 8P4D concurrency-1536 point is the throughput endpoint"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2623
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Restore concurrency 1152 as the validated 8P4D DSpark throughput endpoint after the concurrency-1536 formal AgentX warmup exceeded prefill activation-memory headroom"
+    - "Keep the 8P4D topology, DSpark block size 6, prefill mem-fraction-static 0.85, decode CUDA Graph batch size 192, and all other serving parameters unchanged"
+    - "Use the coarse WAR barrier only for the TP4 aggregate point to avoid the shared-read-done startup stall observed with the 20260816 nightly"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2623
+
+- config-keys:
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-agg
+    - dsv4-fp4-gb300-dynamo-sglang-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Enable SGLang cache reporting on every GB300 DeepSeek-V4-Pro DSpark worker and prefetch checkpoint shards on the two pure-TP aggregate recipes"
+    - "Move the DSv4 AgentX runner from srt-slurm v1.0.38 to v1.0.40; set the concurrency-1152 recipe's nginx client/upstream keepalive to 900 seconds and AIPerf TCP user timeout to 900 seconds so a pooled multi-turn connection is not closed between turns"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2623
diff --git a/runners/launch_gb300-nv.sh b/runners/launch_gb300-nv.sh
index 88b964a8a..bea9f742b 100644
--- a/runners/launch_gb300-nv.sh
+++ b/runners/launch_gb300-nv.sh
@@ -44,6 +44,11 @@ elif [[ $MODEL_PREFIX == "dsr1" && $PRECISION == "fp8" ]]; then
     export SERVED_MODEL_NAME="deepseek-r1-fp8"
     export MODEL_PATH=/scratch/models/DeepSeek-R1-0528
     export SRT_SLURM_MODEL_PREFIX="dsr1-fp8"
+elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" && $MODEL == "deepseek-ai/DeepSeek-V4-Pro-0813" ]]; then
+    # The official 0813 checkpoint, including its attached DSpark weights, is
+    # staged on every GB300 node in batch_1.
+    export MODEL_PATH="/scratch/models/DeepSeek-V4-Pro-0813"
+    export SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro-0813"
 elif [[ $MODEL_PREFIX == "dsv4" && $PRECISION == "fp4" ]]; then
     # Use the node-local /scratch SSD for the 806 GB DSv4-Pro
     # checkpoint. Faster than the Vast NFS path, but this dir only
@@ -224,15 +229,16 @@ elif [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX =
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/qwen3.5" \
         recipes/sglang/qwen3.5
 elif [[ "$IS_AGENTIC" == "1" && $FRAMEWORK == "dynamo-sglang" && $MODEL_PREFIX == "dsv4" ]]; then
-    # DSv4 GB300 SGLang agentic uses NVIDIA/srt-slurm v1.0.38. In addition to
+    # DSv4 GB300 SGLang agentic uses NVIDIA/srt-slurm v1.0.40. In addition to
     # the nginx body-size fix, session-affinity frontend, and custom benchmark
     # schema required by these recipes, this release injects every logical
-    # SGLang worker leader's /metrics URL into AIPERF_SERVER_METRICS_URLS.
+    # SGLang worker leader's /metrics URL into AIPERF_SERVER_METRICS_URLS and
+    # supports long-lived nginx keepalive for multi-turn AgentX replay.
     # AgentX forwards that list to aiperf's --server-metrics argument so its
     # trace artifacts include backend metrics for every engine.
     git clone https://github.com/NVIDIA/srt-slurm.git "$SRT_REPO_DIR"
     cd "$SRT_REPO_DIR"
-    git checkout v1.0.38
+    git checkout v1.0.40
     mkdir -p recipes/sglang/deepseek-v4/agentic
     cp -rT "$GITHUB_WORKSPACE/benchmarks/multi_node/srt-slurm-recipes/sglang/deepseek-v4/agentic" \
         recipes/sglang/deepseek-v4/agentic
diff --git a/runners/test_synthetic_acceptance.py b/runners/test_synthetic_acceptance.py
new file mode 100644
index 000000000..6a5b910a6
--- /dev/null
+++ b/runners/test_synthetic_acceptance.py
@@ -0,0 +1,136 @@
+import os
+import subprocess
+import sys
+from pathlib import Path
+
+import yaml
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+INJECTOR = REPO_ROOT / "runners" / "inject_synthetic_acceptance.py"
+LAUNCHER = REPO_ROOT / "runners" / "launch_gb300-nv.sh"
+MASTER = REPO_ROOT / "configs" / "nvidia-master.yaml"
+RECIPES = REPO_ROOT / "benchmarks" / "multi_node" / "srt-slurm-recipes"
+GOLDEN_AL = REPO_ROOT / "golden_al_distribution" / "dsv4_0813_dspark_probabilistic.yaml"
+CONFIG_KEYS = (
+    "dsv4-fp4-gb300-dynamo-sglang-agentic-agg",
+    "dsv4-fp4-gb300-dynamo-sglang-agentic-disagg",
+)
+
+
+def run_injector(
+    recipe: Path, *, eval_only: bool, acceptance_length: str = "3.77"
+) -> subprocess.CompletedProcess[str]:
+    env = {
+        **os.environ,
+        "EVAL_ONLY": str(eval_only).lower(),
+        "SYNTHETIC_ACCEPTANCE": "true",
+        "SYNTHETIC_ACCEPTANCE_LENGTH": acceptance_length,
+    }
+    return subprocess.run(
+        [sys.executable, str(INJECTOR), str(recipe), "dynamo-sglang"],
+        check=False,
+        capture_output=True,
+        text=True,
+        env=env,
+    )
+
+
+def _configured_recipe_paths(master: dict) -> set[Path]:
+    recipe_paths = set()
+    for config_key in CONFIG_KEYS:
+        assert master[config_key]["model"] == "deepseek-ai/DeepSeek-V4-Pro-0813"
+        for group in master[config_key]["scenarios"]["agentic-coding"]:
+            for point in group["search-space"]:
+                assert point["spec-decoding"] == "draft_model"
+                settings = point["prefill"]["additional-settings"]
+                assert "SYNTHETIC_ACCEPTANCE=true" in settings
+                assert "SYNTHETIC_ACCEPTANCE_LENGTH=3.77" in settings
+                config = next(
+                    item for item in settings if item.startswith("CONFIG_FILE=")
+                )
+                recipe_paths.add(RECIPES / config.removeprefix("CONFIG_FILE=recipes/"))
+    return recipe_paths
+
+
+def test_sglang_injection_is_throughput_only(tmp_path: Path) -> None:
+    original = "  prefill_environment:\n    A: B\n  decode_environment:\n    C: D\n"
+
+    throughput_recipe = tmp_path / "throughput.yaml"
+    throughput_recipe.write_text(original)
+    result = run_injector(throughput_recipe, eval_only=False)
+    assert result.returncode == 0, result.stderr
+    assert throughput_recipe.read_text().count("SGLANG_SIMULATE_ACC_LEN") == 2
+
+    eval_recipe = tmp_path / "eval.yaml"
+    eval_recipe.write_text(original)
+    result = run_injector(eval_recipe, eval_only=True)
+    assert result.returncode == 0, result.stderr
+    assert eval_recipe.read_text() == original
+
+
+def test_gb300_dsv4_configs_inject_only_at_launch() -> None:
+    master = yaml.safe_load(MASTER.read_text())
+    recipe_paths = _configured_recipe_paths(master)
+
+    assert len(recipe_paths) == 6
+    for recipe in recipe_paths:
+        text = recipe.read_text()
+        parsed = yaml.safe_load(text)
+        worker_count = 1 if recipe.name.startswith("agg-") else 2
+        assert "SGLANG_SIMULATE_ACC_" not in text
+        assert text.count('SGLANG_RAGGED_VERIFY_MODE: "static"') == worker_count
+        assert text.count("speculative-algorithm: DSPARK") == worker_count
+        assert text.count("speculative-dspark-block-size: 6") == worker_count
+        assert text.count("speculative-num-draft-tokens: 7") == worker_count
+        assert "speculative-algorithm: EAGLE" not in text
+        assert 'path: "deepseek-v4-pro-0813"' in text
+        assert "AIPERF_TOKENIZER" not in parsed["benchmark"]["env"]
+
+    launcher = LAUNCHER.read_text()
+    assert 'MODEL == "deepseek-ai/DeepSeek-V4-Pro-0813"' in launcher
+    assert 'MODEL_PATH="/scratch/models/DeepSeek-V4-Pro-0813"' in launcher
+    assert 'SRT_SLURM_MODEL_PREFIX="deepseek-v4-pro-0813"' in launcher
+    source = 'source "$(dirname "${BASH_SOURCE[0]}")/slurm_utils.sh"'
+    inject = 'inject_synthetic_acceptance "$CONFIG_PATH" "$FRAMEWORK" || exit 1'
+    apply = "SRTCTL_OUTPUT=$(srtctl apply"
+    assert launcher.index(source) < launcher.index(inject) < launcher.index(apply)
+
+
+def test_gb300_dsv4_dspark6_matches_committed_golden_curve() -> None:
+    golden = yaml.safe_load(GOLDEN_AL.read_text())
+    assert golden["deepseek-v4-pro-0813"]["thinking_on"][6] == 3.77
+
+    master = yaml.safe_load(MASTER.read_text())
+    for config_key in CONFIG_KEYS:
+        for group in master[config_key]["scenarios"]["agentic-coding"]:
+            for point in group["search-space"]:
+                settings = point["prefill"]["additional-settings"]
+                assert "SYNTHETIC_ACCEPTANCE_LENGTH=3.77" in settings
+
+
+def test_gb300_dsv4_dspark6_uses_golden_acceptance_only_for_throughput(
+    tmp_path: Path,
+) -> None:
+    master = yaml.safe_load(MASTER.read_text())
+    for recipe in _configured_recipe_paths(master):
+        original = recipe.read_text()
+        worker_count = 1 if recipe.name.startswith("agg-") else 2
+
+        throughput_recipe = tmp_path / recipe.name
+        throughput_recipe.write_text(original)
+        result = run_injector(throughput_recipe, eval_only=False)
+        assert result.returncode == 0, result.stderr
+        throughput_text = throughput_recipe.read_text()
+        assert throughput_text.count('SGLANG_SIMULATE_ACC_LEN: "3.77"') == worker_count
+        assert (
+            throughput_text.count('SGLANG_RAGGED_VERIFY_MODE: "static"') == worker_count
+        )
+
+        eval_recipe = tmp_path / f"eval-{recipe.name}"
+        eval_recipe.write_text(original)
+        result = run_injector(eval_recipe, eval_only=True)
+        assert result.returncode == 0, result.stderr
+        eval_text = eval_recipe.read_text()
+        assert eval_text == original
+        assert "SGLANG_SIMULATE_ACC_" not in eval_text
+        assert eval_text.count('SGLANG_RAGGED_VERIFY_MODE: "static"') == worker_count