diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4/agentx/agg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4/agentx/agg-variants.yaml
new file mode 100644
index 0000000000..e855c7db4c
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4/agentx/agg-variants.yaml
@@ -0,0 +1,168 @@
+# AgentX dsv41flash sglang gb300-fp4 recipes (Dynamo frontend + SGLang, AGGREGATED
+# topology): shared settings in base, one override per benchmark point. Select one
+# with CONFIG_FILE=recipes/dsv41flash/sglang/gb300-fp4/agentx/agg-variants.yaml:override_<name>.
+#
+# One SGLang worker serves prefill and decode with the checkpoint's bundled DSpark
+# draft (block size 5); the Dynamo frontend routes with the KV-aware router and
+# session affinity. The aggregated arms are the two lowest-latency points of the
+# curve; everything from c16 to c256 comes from the disaggregated recipes in
+# disagg-variants.yaml. Measured on GB300 NVL72 (tokens/s/GPU @ P90 interactivity,
+# P90 TTFT; both GSM8K gates passed):
+#   override_tp4_c1 :  6,215 @ 385.0 tok/s/user, 1.2 s  (4 GPUs; GSM8K 0.9735)
+#   override_tp2_c1 : 12,501 @ 377.1 tok/s/user, 0.9 s  (2 GPUs; GSM8K 0.9735)
+# The aggregated c64 cells (prefill-decode-interval 8: 159,510 @ 62.7; interval 16:
+# 153,542 @ 83.3) were measured as well but are dominated by the disaggregated
+# HiCache cells and are not shipped. The two c1 arms are the highest-interactivity
+# points of the curve (above the
+# published MI355X ATOM curve's last vertex at 336.7 tok/s/user); at c1 the AgentX
+# client replays the trace's idle gaps, so a single lane is client-bound and a second
+# user on the same worker only lowers the P90 (TP2 c2: 12,538 @ 306.8). The harness
+# applies the golden acceptance length (3.51 at K=5) itself.
+
+schema: 2
+
+base:
+  name: dsv41flash-fp4-gb300-dynamo-sglang-agentx-agg
+  model:
+    path: hf:deepseek-ai/DeepSeek-V4.1-Flash
+    container: lmsysorg/sglang:nightly-dev-20260928-81f27fb3@sha256:d9e4917808cfaa4b3be033a0c85a3a73d71eb93c17accbfa2ac3e96f551f3a33
+    precision: fp4
+  identity:
+    model:
+      repo: deepseek-ai/DeepSeek-V4.1-Flash
+  dynamo:
+    install: true
+    source:
+      wheel: "1.6.0.dev20260928"
+  slurm:
+    time_limit: "4:00:00"
+  # Cold weight loads from the shared HF cache plus graph capture take 15-20 min.
+  health_check:
+    max_attempts: 1440
+    interval_seconds: 10
+  resources:
+    gpu_type: gb300
+    gpus_per_node: 4
+  # Dynamo 1.6 uses the TCP request plane; etcd is the only discovery service needed.
+  services:
+    - name: etcd
+      type: etcd
+      placement:
+        node: infra
+  frontend:
+    type: dynamo
+    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"
+  engine: sglang
+  roles:
+    agg:
+      nodes: 1
+      workers: 1
+      gpus: 2
+      env:
+        PIP_BREAK_SYSTEM_PACKAGES: "1"
+        PYTHONNOUSERSITE: "1"
+        PYTHONUNBUFFERED: "1"
+        HF_HUB_CACHE: /hf_hub_cache
+        # Outlast AIPerf's pooled connections past Uvicorn's keep-alive.
+        SGLANG_TIMEOUT_KEEP_ALIVE: "900"
+        # AgentX measures thinking on, the regime of the golden acceptance curve.
+        SGLANG_DEFAULT_THINKING: "1"
+        SGLANG_DSV41_REASONING_EFFORT: high
+        SGLANG_DSPARK_OPT_MARKOV_W2_BF16: "True"
+        # TP2 keeps the row-sharded Engram tables in host DRAM.
+        SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: "1"
+        SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT: per_rank
+      args:
+        served-model-name: deepseek-ai/DeepSeek-V4.1-Flash
+        trust-remote-code: true
+        tensor-parallel-size: 2
+        expert-parallel-size: 2
+        mem-fraction-static: 0.8
+        # Keep active decode requests progressing while long prefixes are queued.
+        prefill-decode-interval: 16
+        # Admission never exceeds the captured decode graph batch.
+        cuda-graph-max-bs-decode: 64
+        # DSpark is the checkpoint's bundled draft; the block size is its only knob.
+        speculative-algorithm: DSPARK
+        speculative-dspark-block-size: 5
+        reasoning-parser: auto
+        tool-call-parser: auto
+        # Draft passes under long-context load outlast the 1800 s default watchdog.
+        watchdog-timeout: 3600
+        enable-metrics: true
+        weight-loader-prefetch-checkpoints: true
+        weight-loader-drop-cache-after-load: false
+        model-loader-extra-config: '{"enable_multithread_load":true}'
+  sbatch_directives:
+    mem: "0"
+    exclusive: ""
+  srun_options:
+    container-remap-root: ""
+  benchmark:
+    type: custom
+    command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+    env:
+      INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+      RESULT_DIR: /logs/agentic
+      PORT: "8000"
+      IS_MULTINODE: "false"
+      MODEL: deepseek-ai/DeepSeek-V4.1-Flash
+      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
+      # Warmup can leave bytes unacknowledged past AIPerf's 30 s default.
+      AIPERF_HTTP_TCP_USER_TIMEOUT: "900000"
+      TP: "2"
+      PP_SIZE: "1"
+      PCP_SIZE: "1"
+
+# Lowest-latency arm: one pure TP4 (EP1) worker on all four GPUs at c1, the standalone
+# recipe's override_tp4_c1 behind the Dynamo frontend: static ragged verify, Engram
+# tables in HBM (host-table path off), admission 2x CONC, 4096-token prefill chunks.
+override_tp4_c1:
+  name: agg-gb300-tp4ep1-c1
+  roles:
+    agg:
+      gpus: 4
+      args:
+        tensor-parallel-size: 4
+        expert-parallel-size: 1
+        chunked-prefill-size: 4096
+        max-running-requests: 2
+      env:
+        SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: "0"
+        SGLANG_RAGGED_VERIFY_MODE: static
+  benchmark:
+    env:
+      CONC: "1"
+      TP: "4"
+      AGENTIC_WARMUP_GRACE_PERIOD: "1800"
+
+# Two-GPU c1 arm: the base TP2/EP2 worker with the Engram tables in host DRAM, static
+# ragged verify, 4096-token prefill chunks and admission 2x CONC: 12,501 tokens/s/GPU
+# @ 377.1 tok/s/user, P90 TTFT 0.90 s - twice the tokens/s/GPU of the TP4 c1 arm at a
+# 2% lower P90 interactivity.
+override_tp2_c1:
+  name: agg-gb300-tp2ep2-c1
+  roles:
+    agg:
+      args:
+        chunked-prefill-size: 4096
+        max-running-requests: 2
+      env:
+        SGLANG_RAGGED_VERIFY_MODE: static
+  benchmark:
+    env:
+      CONC: "1"
+      AGENTIC_WARMUP_GRACE_PERIOD: "1800"
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4/agentx/disagg-variants.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4/agentx/disagg-variants.yaml
new file mode 100644
index 0000000000..dbd3e5d12a
--- /dev/null
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4/agentx/disagg-variants.yaml
@@ -0,0 +1,299 @@
+# AgentX dsv41flash sglang gb300-fp4 recipes (Dynamo frontend + SGLang, DISAGGREGATED
+# topology with Mooncake KV transfer): shared settings in base, one override per
+# benchmark point. Select one with
+# CONFIG_FILE=recipes/dsv41flash/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_<name>.
+#
+# Prefill and decode are separate TP2/EP2 SGLang workers (two GPUs each) behind the
+# Dynamo KV router; with the DSpark draft in PD-disaggregated mode both sides must
+# use the same TP, so capacity is added with more TP2 workers: one prefill worker
+# feeds one or two decode workers (1P1D, 1P2D), which sets the sessions per decode
+# worker and with it the interactivity band, and two prefill workers feed one decode
+# worker (2P1D) at the throughput end. Measured on GB300 NVL72 (tokens/s/GPU @ P90
+# interactivity, P90 TTFT; every override's GSM8K gate passed):
+#   override_1p2d_spread_c16    :  19,280 @ 327.4 tok/s/user,  0.7 s  (6 GPUs, 3 nodes)
+#   override_1p2d_c48           :  59,740 @ 244.3 tok/s/user,  1.4 s  (6 GPUs, 2 nodes)
+#   override_1p2d_c64           :  80,583 @ 207.6 tok/s/user,  2.0 s  (6 GPUs, 2 nodes)
+#   override_1p1d_c96           : 144,448 @ 132.2 tok/s/user,  4.1 s  (4 GPUs, 2 nodes)
+#   override_1p1d_hicache_c160  : 172,192 @  98.8 tok/s/user, 21.7 s  (4 GPUs, 2 nodes)
+#   override_2p1d_hicache_c256  : 174,562 @  68.1 tok/s/user, 18.9 s  (6 GPUs, 3 nodes)
+# Against the published MI355X ATOM curve at matched P90 interactivity: 1P2D spread
+# c16 1.73x, c48 2.11x, c64 2.21x; 1P1D c96 2.72x, HiCache c160 2.40x; 2P1D HiCache
+# c256 1.94x tokens/s/GPU. Other measured cells (1P1D c8/c16/c64, 1P2D c24/c32, 1P4D
+# c64) lie on or under the line through these vertices and are not shipped. The
+# multi-decode cells run the Dynamo frontend with a 1 s session-affinity TTL (per-turn
+# load-aware decode pick); the single-decode cells keep the 3600 s default they were
+# measured with. The harness applies the golden acceptance length (3.51 at K=5) itself.
+
+schema: 2
+
+base:
+  name: dsv41flash-fp4-gb300-dynamo-sglang-agentx-disagg
+  model:
+    path: hf:deepseek-ai/DeepSeek-V4.1-Flash
+    container: lmsysorg/sglang:nightly-dev-20260928-81f27fb3@sha256:d9e4917808cfaa4b3be033a0c85a3a73d71eb93c17accbfa2ac3e96f551f3a33
+    precision: fp4
+  identity:
+    model:
+      repo: deepseek-ai/DeepSeek-V4.1-Flash
+  dynamo:
+    install: true
+    source:
+      wheel: "1.6.0.dev20260928"
+  slurm:
+    # c96 AgentX runs take ~2.5 h including the 60 min profiling phase and the gate.
+    time_limit: "6:00:00"
+  health_check:
+    max_attempts: 1440
+    interval_seconds: 10
+  resources:
+    gpu_type: gb300
+    gpus_per_node: 4
+  # Dynamo 1.6 uses the TCP request plane; etcd is the only discovery service needed
+  # and runs with the frontend on the prefill node (two nodes total for 1P1D).
+  services:
+    - name: etcd
+      type: etcd
+      placement:
+        node: infra
+  frontend:
+    type: dynamo
+    enable_multiple_frontends: false
+    env:
+      PIP_BREAK_SYSTEM_PACKAGES: "1"
+      DYN_NATS_REQUEST_TIMEOUT_SECS: "1800"
+    args:
+      router-mode: kv
+      # Session-affinity TTL as measured for the single-decode 1P1D cells; the
+      # multi-decode overrides below set it to 1 s (per-turn load-aware decode pick).
+      router-session-affinity-ttl-secs: "3600"
+      active-decode-blocks-threshold: "None"
+      active-prefill-tokens-threshold: "None"
+      active-prefill-tokens-threshold-frac: "None"
+  engine: sglang
+  roles:
+    prefill:
+      nodes: 1
+      workers: 1
+      gpus: 2
+      env:
+        PIP_BREAK_SYSTEM_PACKAGES: "1"
+        PYTHONNOUSERSITE: "1"
+        PYTHONUNBUFFERED: "1"
+        HF_HUB_CACHE: /hf_hub_cache
+        SGLANG_TIMEOUT_KEEP_ALIVE: "900"
+        SGLANG_DEFAULT_THINKING: "1"
+        SGLANG_DSV41_REASONING_EFFORT: high
+        SGLANG_DSPARK_OPT_MARKOV_W2_BF16: "True"
+        SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: "1"
+        SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT: per_rank
+        SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+        SGLANG_ENABLE_PREFILL_WAR_READ_DONE: "1"
+        SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+        NCCL_MNNVL_ENABLE: "1"
+        NCCL_CUMEM_ENABLE: "1"
+        MC_FORCE_MNNVL: "1"
+      args:
+        served-model-name: deepseek-ai/DeepSeek-V4.1-Flash
+        trust-remote-code: true
+        tensor-parallel-size: 2
+        expert-parallel-size: 2
+        disaggregation-transfer-backend: mooncake
+        # Long-context first turns (P90 ISL ~250k tokens): one 16k-token chunk per step.
+        chunked-prefill-size: 16384
+        max-prefill-tokens: 16384
+        max-running-requests: 64
+        swa-prefix-tails: 4096
+        mem-fraction-static: 0.85
+        speculative-algorithm: DSPARK
+        speculative-dspark-block-size: 5
+        reasoning-parser: auto
+        tool-call-parser: auto
+        watchdog-timeout: 3600
+        enable-metrics: true
+        weight-loader-prefetch-checkpoints: true
+        weight-loader-drop-cache-after-load: false
+        model-loader-extra-config: '{"enable_multithread_load":true}'
+      # KV events for the Dynamo KV router; srtctl allocates the ZMQ ports per worker.
+      kv_events: true
+    decode:
+      nodes: 1
+      workers: 1
+      gpus: 2
+      env:
+        PIP_BREAK_SYSTEM_PACKAGES: "1"
+        PYTHONNOUSERSITE: "1"
+        PYTHONUNBUFFERED: "1"
+        HF_HUB_CACHE: /hf_hub_cache
+        SGLANG_TIMEOUT_KEEP_ALIVE: "900"
+        SGLANG_DEFAULT_THINKING: "1"
+        SGLANG_DSV41_REASONING_EFFORT: high
+        SGLANG_DSPARK_OPT_MARKOV_W2_BF16: "True"
+        SGLANG_ENABLE_DSV41_ENGRAM_HOST_TABLE: "1"
+        SGLANG_DSV41_ENGRAM_HOST_TABLE_LAYOUT: per_rank
+        SGLANG_DISAGGREGATION_WAITING_TIMEOUT: "900"
+        SGLANG_MOONCAKE_CUSTOM_MEM_POOL: "True"
+        NCCL_MNNVL_ENABLE: "1"
+        NCCL_CUMEM_ENABLE: "1"
+        MC_FORCE_MNNVL: "1"
+      args:
+        served-model-name: deepseek-ai/DeepSeek-V4.1-Flash
+        trust-remote-code: true
+        tensor-parallel-size: 2
+        expert-parallel-size: 2
+        disaggregation-transfer-backend: mooncake
+        max-running-requests: 128
+        cuda-graph-max-bs-decode: 128
+        mem-fraction-static: 0.85
+        speculative-algorithm: DSPARK
+        speculative-dspark-block-size: 5
+        reasoning-parser: auto
+        tool-call-parser: auto
+        watchdog-timeout: 3600
+        enable-metrics: true
+        weight-loader-prefetch-checkpoints: true
+        weight-loader-drop-cache-after-load: false
+        model-loader-extra-config: '{"enable_multithread_load":true}'
+      kv_events: true
+  sbatch_directives:
+    mem: "0"
+    exclusive: ""
+  srun_options:
+    container-remap-root: ""
+  benchmark:
+    type: custom
+    command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+    env:
+      INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
+      RESULT_DIR: /logs/agentic
+      PORT: "8000"
+      IS_MULTINODE: "true"
+      MODEL: deepseek-ai/DeepSeek-V4.1-Flash
+      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
+      AIPERF_HTTP_TCP_USER_TIMEOUT: "900000"
+      AGENTIC_WARMUP_GRACE_PERIOD: "3600"
+
+# 1 prefill TP2/EP2 + 1 decode TP2/EP2 (4 GPUs, 2 nodes). The decode worker alone
+# sets the interactivity (96 sessions give 132 tok/s/user); the single prefill worker
+# holds P90 TTFT under 4.1 s up to c96 and saturates between c96 and c128, where the
+# HiCache override below takes over. 144,448 tokens/s/GPU @ 132.2 tok/s/user with a
+# 4.1 s P90 TTFT.
+override_1p1d_c96:
+  name: disagg-gb300-1p1d-tp2ep2-c96
+  benchmark:
+    env:
+      CONC: "96"
+
+# 1 prefill TP2/EP2 + 2 decode TP2/EP2 (6 GPUs). Two decode workers halve the sessions
+# per decode worker at a given concurrency, which sets the interactivity band. The
+# frontend's session-affinity TTL is 1 s, so every turn takes a load-aware decode pick
+# instead of pinning the session to one decode worker for an hour: +4% P90
+# interactivity at 8 sessions per decode worker, neutral at 12 and more (c24 31,465 @
+# 299.0 vs 31,339 @ 298.1; c32 41,910 @ 276.1 vs 41,768 @ 277.3 at TTL 3600).
+#
+# Same-node layout (both decode workers on one node, 2 nodes total):
+#   c48: 59,740 tokens/s/GPU @ 244.3 tok/s/user, P90 TTFT 1.36 s
+#   c64: 80,583 tokens/s/GPU @ 207.6 tok/s/user, P90 TTFT 2.01 s (GSM8K 0.9742)
+override_1p2d_c48:
+  name: disagg-gb300-1p2d-tp2ep2-c48
+  frontend:
+    args:
+      router-session-affinity-ttl-secs: "1"
+  roles:
+    decode:
+      workers: 2
+      args:
+        max-running-requests: 64
+        cuda-graph-max-bs-decode: 64
+  benchmark:
+    env:
+      CONC: "48"
+
+override_1p2d_c64:
+  name: disagg-gb300-1p2d-tp2ep2-c64
+  frontend:
+    args:
+      router-session-affinity-ttl-secs: "1"
+  roles:
+    decode:
+      workers: 2
+      args:
+        max-running-requests: 64
+        cuda-graph-max-bs-decode: 64
+  benchmark:
+    env:
+      CONC: "64"
+
+# Spread layout (one decode worker per node, 3 nodes total), 8 sessions per decode
+# worker: 19,280 tokens/s/GPU @ 327.4 tok/s/user, P90 TTFT 0.67 s (GSM8K 0.9735) -
+# the highest-interactivity disaggregated point (the same cell on one node at TTL
+# 3600 gave 19,227 @ 308.9).
+override_1p2d_spread_c16:
+  name: disagg-gb300-1p2d-spread-tp2ep2-c16
+  frontend:
+    args:
+      router-session-affinity-ttl-secs: "1"
+  roles:
+    decode:
+      nodes: 2
+      workers: 2
+      args:
+        max-running-requests: 64
+        cuda-graph-max-bs-decode: 64
+  benchmark:
+    env:
+      CONC: "16"
+
+# Throughput end. The prefill worker keeps its KV prefix cache in a host-memory tier
+# (SGLang HiCache: ratio 1 = one host copy of the device pool, write-back, direct I/O),
+# which removes the LRU evictions that pushed P90 TTFT past the 25 s ceiling beyond c96
+# (1P1D c128 without it: 25.4 s). 1P1D c160: 172,192 tokens/s/GPU @ 98.8 tok/s/user,
+# P90 TTFT 21.7 s (GSM8K 0.9727). c192 breaches the ceiling (29.2 s): the single
+# prefill worker is saturated, so the next step is a second prefill worker (below).
+override_1p1d_hicache_c160:
+  name: disagg-gb300-1p1d-hicache-tp2ep2-c160
+  roles:
+    prefill:
+      args:
+        enable-hierarchical-cache: true
+        hicache-ratio: 1
+        hicache-write-policy: write_back
+        hicache-io-backend: direct
+    decode:
+      args:
+        max-running-requests: 192
+        cuda-graph-max-bs-decode: 192
+  benchmark:
+    env:
+      CONC: "160"
+
+# Two HiCache prefill workers (one per node) feeding one decode worker (6 GPUs, 3 nodes),
+# 128 sessions per prefill worker, decode admission and graph batch 256, session-affinity
+# TTL 1 s: 174,562 tokens/s/GPU @ 68.1 tok/s/user, P90 TTFT 18.9 s (GSM8K 0.9712) - the
+# highest tokens/s/GPU of the campaign (the same cell without HiCache: 151,671 @ 70.8
+# with a 36 s P90 TTFT).
+override_2p1d_hicache_c256:
+  name: disagg-gb300-2p1d-hicache-tp2ep2-c256
+  frontend:
+    args:
+      router-session-affinity-ttl-secs: "1"
+  roles:
+    prefill:
+      nodes: 2
+      workers: 2
+      args:
+        max-running-requests: 128
+        enable-hierarchical-cache: true
+        hicache-ratio: 1
+        hicache-write-policy: write_back
+        hicache-io-backend: direct
+    decode:
+      args:
+        max-running-requests: 256
+        cuda-graph-max-bs-decode: 256
+  benchmark:
+    env:
+      CONC: "256"
diff --git a/inferencex-e2e/configs/nvidia-master.yaml b/inferencex-e2e/configs/nvidia-master.yaml
index 63dcf8fff8..8cfd67b436 100644
--- a/inferencex-e2e/configs/nvidia-master.yaml
+++ b/inferencex-e2e/configs/nvidia-master.yaml
@@ -8927,3 +8927,136 @@ dsv41flash-fp4-gb300-sglang-agentic-dspark:
       - { tp: 2, ep: 2, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4-mtp/agentic.yaml }
       - { tp: 4, ep: 1, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4-mtp/agentic.yaml }
       - { tp: 4, ep: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [4, 8, 16, 32, 64, 128], srt-recipe: benchmarks/single_node/srt-slurm-recipes/dsv41flash/sglang/gb300-fp4-mtp/agentic.yaml }
+dsv41flash-fp4-gb300-dynamo-sglang-agentic-agg:
+  image: lmsysorg/sglang:nightly-dev-20260928-81f27fb3@sha256:d9e4917808cfaa4b3be033a0c85a3a73d71eb93c17accbfa2ac3e96f551f3a33
+  model: deepseek-ai/DeepSeek-V4.1-Flash
+  model-prefix: dsv41flash
+  runner: cluster:gb300-nv
+  precision: fp4
+  framework: dynamo-sglang
+  router: { name: dynamo-router, version: "1.6.0.dev20260928" }
+  multinode: true
+  disagg: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - spec-decoding: mtp
+        conc-list: [1]
+        num-nodes: 1
+        worker:
+          num-worker: 1
+          tp: 4
+          ep: 1
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv41flash/sglang/gb300-fp4/agentx/agg-variants.yaml:override_tp4_c1"
+      - spec-decoding: mtp
+        conc-list: [1]
+        num-nodes: 1
+        worker:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv41flash/sglang/gb300-fp4/agentx/agg-variants.yaml:override_tp2_c1"
+dsv41flash-fp4-gb300-dynamo-sglang-agentic-disagg:
+  image: lmsysorg/sglang:nightly-dev-20260928-81f27fb3@sha256:d9e4917808cfaa4b3be033a0c85a3a73d71eb93c17accbfa2ac3e96f551f3a33
+  model: deepseek-ai/DeepSeek-V4.1-Flash
+  model-prefix: dsv41flash
+  runner: cluster:gb300-nv
+  precision: fp4
+  framework: dynamo-sglang
+  router: { name: dynamo-router, version: "1.6.0.dev20260928" }
+  kv-p2p-transfer: mooncake
+  multinode: true
+  disagg: true
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - spec-decoding: mtp
+        conc-list: [16]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv41flash/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_1p2d_spread_c16"
+        decode:
+          num-worker: 2
+          tp: 2
+          ep: 2
+          dp-attn: false
+      - spec-decoding: mtp
+        conc-list: [48]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv41flash/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_1p2d_c48"
+        decode:
+          num-worker: 2
+          tp: 2
+          ep: 2
+          dp-attn: false
+      - spec-decoding: mtp
+        conc-list: [64]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv41flash/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_1p2d_c64"
+        decode:
+          num-worker: 2
+          tp: 2
+          ep: 2
+          dp-attn: false
+      - spec-decoding: mtp
+        conc-list: [96]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv41flash/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_1p1d_c96"
+        decode:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: false
+      - spec-decoding: mtp
+        conc-list: [160]
+        prefill:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv41flash/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_1p1d_hicache_c160"
+        decode:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: false
+      - spec-decoding: mtp
+        conc-list: [256]
+        prefill:
+          num-worker: 2
+          tp: 2
+          ep: 2
+          dp-attn: false
+          additional-settings:
+          - "CONFIG_FILE=recipes/dsv41flash/sglang/gb300-fp4/agentx/disagg-variants.yaml:override_2p1d_hicache_c256"
+        decode:
+          num-worker: 1
+          tp: 2
+          ep: 2
+          dp-attn: false
diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 9b9c48a0ef..df423ce255 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9168,3 +9168,19 @@
     - "Relevant ATOM changes in the range: V4 decode reuses the sparse prefill ASM (ROCm/ATOM#2271), greedy sampler picks via aiter.topk_select (ROCm/ATOM#2244), and FP8 block scales declared scale_fmt ue8m0 are now stored as E8M0 on gfx950 by default (ROCm/ATOM#2419; previously FP32 unless ATOM_FP8_BLOCKSCALE_USE_E8M0_SCALE=1). The upstream DeepSeek-V4 recipes are unchanged across the range, and every recipe flag and choice (all2all-backend rccl, dp-load-balance least_tokens, moe-backend standard) remains valid."
     - "No data-type or precision change to the DeepSeek-V4-Pro-0813 DSpark draft: no online quantization is configured, so it keeps its checkpoint precision; the E8M0 scale storage represents the checkpoint's power-of-two block scales exactly. kv-cache-dtype and index-cache-dtype touch cache storage only."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3605
+
+- config-keys:
+    - dsv41flash-fp4-gb300-dynamo-sglang-agentic-agg
+    - dsv41flash-fp4-gb300-dynamo-sglang-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add DeepSeek-V4.1-Flash FP4 AgentX on GB300 with the Dynamo frontend (KV router, session affinity) and SGLang nightly-dev-20260928-81f27fb3 / ai-dynamo 1.6.0.dev20260928, DSpark block size 5."
+    - "Use the mounted shared Hugging Face cache for every SGLang server role."
+    - "Aggregated arms: the two lowest-latency points of the curve, a pure TP4 (EP1) c1 point (the standalone recipe's TP4 c1 settings, static ragged verify and Engram in HBM, behind Dynamo; 6,215 tok/s/GPU @ 385.0 tok/s/user P90) and a TP2/EP2 c1 point with static ragged verify and Engram in host DRAM (12,501 @ 377.1, P90 TTFT 0.9 s); both passed their GSM8K gates."
+    - "Disaggregated curve with Mooncake KV transfer and TP2/EP2 workers: 1P2D with a 1 s Dynamo session-affinity TTL at c16 (one decode worker per node; 19,280 tok/s/GPU @ 327.4 tok/s/user P90, P90 TTFT 0.7 s), c48 (59,740 @ 244.3, 1.4 s) and c64 (80,583 @ 207.6, 2.0 s); 1P1D at c96 (144,448 @ 132.2, 4.1 s) and, with the SGLang HiCache host-memory prefix-cache tier on the prefill worker, at c160 (172,192 @ 98.8, 21.7 s); 2P1D HiCache at c256 (174,562 @ 68.1, 18.9 s). Every point passed its GSM8K gate; at matched P90 interactivity the curve is 1.7-2.7x the published MI355X ATOM tokens/s/GPU."
+    - "新增 GB300 上 DeepSeek-V4.1-Flash FP4 AgentX 的 Dynamo 前端（KV 路由、会话亲和）+ SGLang nightly-dev-20260928-81f27fb3 / ai-dynamo 1.6.0.dev20260928 配方，DSpark block size 5。"
+    - "所有 SGLang server role 使用已挂载的共享 Hugging Face cache。"
+    - "聚合分支：曲线上延迟最低的两个点——纯 TP4（EP1）c1（沿用单机配方的 TP4 c1 设置、static ragged verify 与 HBM 内 Engram，置于 Dynamo 之后；6,215 tok/s/GPU @ 385.0 tok/s/user P90）与采用 static ragged verify、Engram 置于主机内存的 TP2/EP2 c1（12,501 @ 377.1，P90 TTFT 0.9 s）；两者均通过 GSM8K 验证。"
+    - "分离式曲线（Mooncake KV 传输，TP2/EP2 worker）：1P2D 配合 1 秒 Dynamo 会话亲和 TTL 的 c16（每节点一个 decode worker；19,280 tok/s/GPU @ 327.4 tok/s/user P90，P90 TTFT 0.7 s）、c48（59,740 @ 244.3，1.4 s）与 c64（80,583 @ 207.6，2.0 s）；1P1D 的 c96（144,448 @ 132.2，4.1 s）以及在 prefill worker 上启用 SGLang HiCache 主机内存前缀缓存层后的 c160（172,192 @ 98.8，21.7 s）；2P1D HiCache 的 c256（174,562 @ 68.1，18.9 s）。所有点均通过 GSM8K 验证；在相同 P90 交互性下，该曲线为已发布 MI355X ATOM tokens/s/GPU 的 1.7-2.7 倍。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3598