From 8bd17af62b35dbad2e742cdfe53738ebe63d0687 Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Sat, 3 Oct 2026 14:58:25 +0800
Subject: [PATCH 1/8] perf(agentx): GB200 DSV4 SGLang disagg KV offload via
 Mooncake unified-cache external linker

---
 ...0-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml | 79 ++++++++++++++++--
 ...00-1p1d-dep8-dep16-c128-mtp-kvoffload.yaml | 81 +++++++++++++++++--
 ...200-1p1d-dep8-dep16-c64-mtp-kvoffload.yaml | 81 +++++++++++++++++--
 ...-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml | 79 ++++++++++++++++--
 ...-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml | 79 ++++++++++++++++--
 ...0-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml | 79 ++++++++++++++++--
 6 files changed, 446 insertions(+), 32 deletions(-)

diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml
index 9a2b0265a1..541934f0af 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml
@@ -18,6 +18,11 @@ slurm:
 health_check:
   max_attempts: 1440
   interval_seconds: 10
+# Tachometer's per-node exporters slow decode steps by a few percent and add
+# host memory on the head node; this ladder is a throughput measurement.
+observability:
+  tachometer:
+    enabled: false
 resources:
   gpu_type: gb200
   gpus_per_node: 4
@@ -32,6 +37,54 @@ services:
     node: dedicated
   options:
     max_payload_mb: 32
+- name: mooncake-master
+  type: mooncake-master
+  args:
+    - --eviction_high_watermark_ratio=0.90
+# Decode ranks use a chunk cache, so their host DRAM only joins the store
+# through standalone stores: four per decode node, so segments stay uniform
+# with prefill's per-rank ones instead of one oversized segment per node.
+- &decode-store
+  name: store-decode-0
+  type: mooncake-store
+  placement:
+    node: decode
+  args: ["--port", "8800"]
+  env:
+    MOONCAKE_PROTOCOL: rdma
+    MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+    MOONCAKE_GLOBAL_SEGMENT_SIZE: 180gb
+    MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+    MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+    # Advertise the store transfer engine on the node address instead of the first active
+    # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+    MC_TCP_BIND_ADDRESS: "{node_ip}"
+  preamble: |
+    ulimit -n 1048576
+    ulimit -l unlimited
+  # Cold container starts on a decode node can exceed srtctl's 120 s default;
+  # the <<: merge is shallow, so every store repeats the timeout.
+  readiness:
+    port: 8800
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-1
+  args: ["--port", "8801"]
+  readiness:
+    port: 8801
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-2
+  args: ["--port", "8802"]
+  readiness:
+    port: 8802
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-3
+  args: ["--port", "8803"]
+  readiness:
+    port: 8803
+    timeout_seconds: 600
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -84,6 +137,18 @@ roles:
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
       SGLANG_DSV4_MHC_PREWARM: '1'
+      MOONCAKE_PROTOCOL: rdma
+      # The store only registers host buffers on RDMA; without an explicit list it
+      # falls back to NVLink for same-domain peers and aborts at the warmup put.
+      MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 140gb
+      MOONCAKE_STANDALONE_STORAGE: '0'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      # Advertise the store transfer engine on the node address instead of the first active
+      # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+      MC_TCP_BIND_ADDRESS: "{node}"
+      WITH_NVIDIA_PEERMEM: '0'
     args:
       host: 0.0.0.0
       served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
@@ -117,11 +182,9 @@ roles:
       speculative-num-steps: 1
       speculative-eagle-topk: 1
       speculative-num-draft-tokens: 7
-      enable-hierarchical-cache: true
-      hicache-write-policy: write_back
-      hicache-ratio: 2
-      hicache-io-backend: direct
-      hicache-mem-layout: page_first_direct
+      enable-unified-cache-external-linker: true
+      unified-cache-external-linker-backend: mooncake
+      hicache-storage-backend-extra-config: '{"enable_group_semantics":true}'
       cuda-graph-max-bs-decode: 256
   decode:
     nodes: 8
@@ -157,6 +220,8 @@ roles:
       SGLANG_LOG_MS: '1'
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_DSV4_MHC_PREWARM: '1'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      WITH_NVIDIA_PEERMEM: '0'
     args:
       host: 0.0.0.0
       served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
@@ -199,6 +264,10 @@ srun_options:
 benchmark:
   type: custom
   command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  # The replay client holds ~300 GB of host memory; on the default head node it
+  # shares prefill_0's node and OOMs it. Put it on the etcd/nats node instead.
+  placement:
+    node: dedicated
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
     RESULT_DIR: /logs/agentic
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep8-dep16-c128-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep8-dep16-c128-mtp-kvoffload.yaml
index 9461990907..a44d935fda 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep8-dep16-c128-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep8-dep16-c128-mtp-kvoffload.yaml
@@ -2,7 +2,7 @@ schema: 2
 name: "disagg-gb200-1p1d-dep8-dep16-c128-mtp-kvoffload"
 
 # GB200 AgentX SGLang disaggregated recipe for DeepSeek-V4-Pro-0813
-# (1P x DEP8 / 1D x DEP16, bundled DSpark K=6 + HiCache KV offload), tuned for
+# (1P x DEP8 / 1D x DEP16, bundled DSpark K=6 + Mooncake unified-cache KV offload), tuned for
 # concurrency 128.
 # The DEP8 prefill worker spans two nodes; the DEP16 decode worker spans four.
 
@@ -28,6 +28,11 @@ health_check:
   max_attempts: 1440
   interval_seconds: 10
 
+# Tachometer's per-node exporters slow decode steps by a few percent and add
+# host memory on the head node; this ladder is a throughput measurement.
+observability:
+  tachometer:
+    enabled: false
 resources:
   gpu_type: gb200
   gpus_per_node: 4
@@ -42,6 +47,54 @@ services:
       node: dedicated
     options:
       max_payload_mb: 32
+  - name: mooncake-master
+    type: mooncake-master
+    args:
+      - --eviction_high_watermark_ratio=0.90
+  # Decode ranks use a chunk cache, so their host DRAM only joins the store
+  # through standalone stores: four per decode node, so segments stay uniform
+  # with prefill's per-rank ones instead of one oversized segment per node.
+  - &decode-store
+    name: store-decode-0
+    type: mooncake-store
+    placement:
+      node: decode
+    args: ["--port", "8800"]
+    env:
+      MOONCAKE_PROTOCOL: rdma
+      MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 180gb
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      # Advertise the store transfer engine on the node address instead of the first active
+      # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+      MC_TCP_BIND_ADDRESS: "{node_ip}"
+    preamble: |
+      ulimit -n 1048576
+      ulimit -l unlimited
+    # Cold container starts on a decode node can exceed srtctl's 120 s default;
+    # the <<: merge is shallow, so every store repeats the timeout.
+    readiness:
+      port: 8800
+      timeout_seconds: 600
+  - <<: *decode-store
+    name: store-decode-1
+    args: ["--port", "8801"]
+    readiness:
+      port: 8801
+      timeout_seconds: 600
+  - <<: *decode-store
+    name: store-decode-2
+    args: ["--port", "8802"]
+    readiness:
+      port: 8802
+      timeout_seconds: 600
+  - <<: *decode-store
+    name: store-decode-3
+    args: ["--port", "8803"]
+    readiness:
+      port: 8803
+      timeout_seconds: 600
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -95,6 +148,18 @@ roles:
       SGLANG_LOG_MS: '1'
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
+      MOONCAKE_PROTOCOL: rdma
+      # The store only registers host buffers on RDMA; without an explicit list it
+      # falls back to NVLink for same-domain peers and aborts at the warmup put.
+      MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 140gb
+      MOONCAKE_STANDALONE_STORAGE: '0'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      # Advertise the store transfer engine on the node address instead of the first active
+      # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+      MC_TCP_BIND_ADDRESS: "{node}"
+      WITH_NVIDIA_PEERMEM: '0'
 
     args:
       host: 0.0.0.0
@@ -130,11 +195,9 @@ roles:
       speculative-num-steps: 1
       speculative-eagle-topk: 1
       speculative-num-draft-tokens: 7
-      enable-hierarchical-cache: true
-      hicache-write-policy: write_back
-      hicache-ratio: 8
-      hicache-io-backend: direct
-      hicache-mem-layout: page_first_direct
+      enable-unified-cache-external-linker: true
+      unified-cache-external-linker-backend: mooncake
+      hicache-storage-backend-extra-config: '{"enable_group_semantics":true}'
 
   decode:
     nodes: 4
@@ -169,6 +232,8 @@ roles:
       SGLANG_LOG_FORWARD_ITERS: '1'
       SGLANG_LOG_MS: '1'
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      WITH_NVIDIA_PEERMEM: '0'
 
     args:
       host: 0.0.0.0
@@ -215,6 +280,10 @@ srun_options:
 benchmark:
   type: custom
   command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  # The replay client holds ~300 GB of host memory; on the default head node it
+  # shares prefill_0's node and OOMs it. Put it on the etcd/nats node instead.
+  placement:
+    node: dedicated
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
     RESULT_DIR: /logs/agentic
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep8-dep16-c64-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep8-dep16-c64-mtp-kvoffload.yaml
index d3bcb51688..f0181c16ad 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep8-dep16-c64-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep8-dep16-c64-mtp-kvoffload.yaml
@@ -2,7 +2,7 @@ schema: 2
 name: "disagg-gb200-1p1d-dep8-dep16-c64-mtp-kvoffload"
 
 # GB200 AgentX SGLang disaggregated recipe for DeepSeek-V4-Pro-0813
-# (1P x DEP8 / 1D x DEP16, bundled DSpark K=6 + HiCache KV offload), tuned for
+# (1P x DEP8 / 1D x DEP16, bundled DSpark K=6 + Mooncake unified-cache KV offload), tuned for
 # concurrency 64.
 # The DEP8 prefill worker spans two nodes; the DEP16 decode worker spans four.
 
@@ -28,6 +28,11 @@ health_check:
   max_attempts: 1440
   interval_seconds: 10
 
+# Tachometer's per-node exporters slow decode steps by a few percent and add
+# host memory on the head node; this ladder is a throughput measurement.
+observability:
+  tachometer:
+    enabled: false
 resources:
   gpu_type: gb200
   gpus_per_node: 4
@@ -42,6 +47,54 @@ services:
       node: dedicated
     options:
       max_payload_mb: 32
+  - name: mooncake-master
+    type: mooncake-master
+    args:
+      - --eviction_high_watermark_ratio=0.90
+  # Decode ranks use a chunk cache, so their host DRAM only joins the store
+  # through standalone stores: four per decode node, so segments stay uniform
+  # with prefill's per-rank ones instead of one oversized segment per node.
+  - &decode-store
+    name: store-decode-0
+    type: mooncake-store
+    placement:
+      node: decode
+    args: ["--port", "8800"]
+    env:
+      MOONCAKE_PROTOCOL: rdma
+      MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 180gb
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      # Advertise the store transfer engine on the node address instead of the first active
+      # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+      MC_TCP_BIND_ADDRESS: "{node_ip}"
+    preamble: |
+      ulimit -n 1048576
+      ulimit -l unlimited
+    # Cold container starts on a decode node can exceed srtctl's 120 s default;
+    # the <<: merge is shallow, so every store repeats the timeout.
+    readiness:
+      port: 8800
+      timeout_seconds: 600
+  - <<: *decode-store
+    name: store-decode-1
+    args: ["--port", "8801"]
+    readiness:
+      port: 8801
+      timeout_seconds: 600
+  - <<: *decode-store
+    name: store-decode-2
+    args: ["--port", "8802"]
+    readiness:
+      port: 8802
+      timeout_seconds: 600
+  - <<: *decode-store
+    name: store-decode-3
+    args: ["--port", "8803"]
+    readiness:
+      port: 8803
+      timeout_seconds: 600
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -95,6 +148,18 @@ roles:
       SGLANG_LOG_MS: '1'
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
+      MOONCAKE_PROTOCOL: rdma
+      # The store only registers host buffers on RDMA; without an explicit list it
+      # falls back to NVLink for same-domain peers and aborts at the warmup put.
+      MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 140gb
+      MOONCAKE_STANDALONE_STORAGE: '0'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      # Advertise the store transfer engine on the node address instead of the first active
+      # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+      MC_TCP_BIND_ADDRESS: "{node}"
+      WITH_NVIDIA_PEERMEM: '0'
 
     args:
       host: 0.0.0.0
@@ -130,11 +195,9 @@ roles:
       speculative-num-steps: 1
       speculative-eagle-topk: 1
       speculative-num-draft-tokens: 7
-      enable-hierarchical-cache: true
-      hicache-write-policy: write_back
-      hicache-ratio: 8
-      hicache-io-backend: direct
-      hicache-mem-layout: page_first_direct
+      enable-unified-cache-external-linker: true
+      unified-cache-external-linker-backend: mooncake
+      hicache-storage-backend-extra-config: '{"enable_group_semantics":true}'
 
   decode:
     nodes: 4
@@ -169,6 +232,8 @@ roles:
       SGLANG_LOG_FORWARD_ITERS: '1'
       SGLANG_LOG_MS: '1'
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      WITH_NVIDIA_PEERMEM: '0'
 
     args:
       host: 0.0.0.0
@@ -215,6 +280,10 @@ srun_options:
 benchmark:
   type: custom
   command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  # The replay client holds ~300 GB of host memory; on the default head node it
+  # shares prefill_0's node and OOMs it. Put it on the etcd/nats node instead.
+  placement:
+    node: dedicated
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
     RESULT_DIR: /logs/agentic
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml
index 5a21654eb0..f1672f9e54 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml
@@ -18,6 +18,11 @@ slurm:
 health_check:
   max_attempts: 1440
   interval_seconds: 10
+# Tachometer's per-node exporters slow decode steps by a few percent and add
+# host memory on the head node; this ladder is a throughput measurement.
+observability:
+  tachometer:
+    enabled: false
 resources:
   gpu_type: gb200
   gpus_per_node: 4
@@ -32,6 +37,54 @@ services:
     node: dedicated
   options:
     max_payload_mb: 32
+- name: mooncake-master
+  type: mooncake-master
+  args:
+    - --eviction_high_watermark_ratio=0.90
+# Decode ranks use a chunk cache, so their host DRAM only joins the store
+# through standalone stores: four per decode node, so segments stay uniform
+# with prefill's per-rank ones instead of one oversized segment per node.
+- &decode-store
+  name: store-decode-0
+  type: mooncake-store
+  placement:
+    node: decode
+  args: ["--port", "8800"]
+  env:
+    MOONCAKE_PROTOCOL: rdma
+    MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+    MOONCAKE_GLOBAL_SEGMENT_SIZE: 180gb
+    MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+    MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+    # Advertise the store transfer engine on the node address instead of the first active
+    # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+    MC_TCP_BIND_ADDRESS: "{node_ip}"
+  preamble: |
+    ulimit -n 1048576
+    ulimit -l unlimited
+  # Cold container starts on a decode node can exceed srtctl's 120 s default;
+  # the <<: merge is shallow, so every store repeats the timeout.
+  readiness:
+    port: 8800
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-1
+  args: ["--port", "8801"]
+  readiness:
+    port: 8801
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-2
+  args: ["--port", "8802"]
+  readiness:
+    port: 8802
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-3
+  args: ["--port", "8803"]
+  readiness:
+    port: 8803
+    timeout_seconds: 600
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -84,6 +137,18 @@ roles:
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
       SGLANG_DSV4_MHC_PREWARM: '1'
+      MOONCAKE_PROTOCOL: rdma
+      # The store only registers host buffers on RDMA; without an explicit list it
+      # falls back to NVLink for same-domain peers and aborts at the warmup put.
+      MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 140gb
+      MOONCAKE_STANDALONE_STORAGE: '0'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      # Advertise the store transfer engine on the node address instead of the first active
+      # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+      MC_TCP_BIND_ADDRESS: "{node}"
+      WITH_NVIDIA_PEERMEM: '0'
     args:
       host: 0.0.0.0
       served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
@@ -117,11 +182,9 @@ roles:
       speculative-num-steps: 1
       speculative-eagle-topk: 1
       speculative-num-draft-tokens: 7
-      enable-hierarchical-cache: true
-      hicache-write-policy: write_back
-      hicache-ratio: 2
-      hicache-io-backend: direct
-      hicache-mem-layout: page_first_direct
+      enable-unified-cache-external-linker: true
+      unified-cache-external-linker-backend: mooncake
+      hicache-storage-backend-extra-config: '{"enable_group_semantics":true}'
       cuda-graph-max-bs-decode: 256
   decode:
     nodes: 8
@@ -157,6 +220,8 @@ roles:
       SGLANG_LOG_MS: '1'
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_DSV4_MHC_PREWARM: '1'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      WITH_NVIDIA_PEERMEM: '0'
     args:
       host: 0.0.0.0
       served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
@@ -199,6 +264,10 @@ srun_options:
 benchmark:
   type: custom
   command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  # The replay client holds ~300 GB of host memory; on the default head node it
+  # shares prefill_0's node and OOMs it. Put it on the etcd/nats node instead.
+  placement:
+    node: dedicated
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
     RESULT_DIR: /logs/agentic
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml
index c597addbe6..ac2ed56b08 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml
@@ -18,6 +18,11 @@ slurm:
 health_check:
   max_attempts: 1440
   interval_seconds: 10
+# Tachometer's per-node exporters slow decode steps by a few percent and add
+# host memory on the head node; this ladder is a throughput measurement.
+observability:
+  tachometer:
+    enabled: false
 resources:
   gpu_type: gb200
   gpus_per_node: 4
@@ -32,6 +37,54 @@ services:
     node: dedicated
   options:
     max_payload_mb: 32
+- name: mooncake-master
+  type: mooncake-master
+  args:
+    - --eviction_high_watermark_ratio=0.90
+# Decode ranks use a chunk cache, so their host DRAM only joins the store
+# through standalone stores: four per decode node, so segments stay uniform
+# with prefill's per-rank ones instead of one oversized segment per node.
+- &decode-store
+  name: store-decode-0
+  type: mooncake-store
+  placement:
+    node: decode
+  args: ["--port", "8800"]
+  env:
+    MOONCAKE_PROTOCOL: rdma
+    MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+    MOONCAKE_GLOBAL_SEGMENT_SIZE: 180gb
+    MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+    MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+    # Advertise the store transfer engine on the node address instead of the first active
+    # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+    MC_TCP_BIND_ADDRESS: "{node_ip}"
+  preamble: |
+    ulimit -n 1048576
+    ulimit -l unlimited
+  # Cold container starts on a decode node can exceed srtctl's 120 s default;
+  # the <<: merge is shallow, so every store repeats the timeout.
+  readiness:
+    port: 8800
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-1
+  args: ["--port", "8801"]
+  readiness:
+    port: 8801
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-2
+  args: ["--port", "8802"]
+  readiness:
+    port: 8802
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-3
+  args: ["--port", "8803"]
+  readiness:
+    port: 8803
+    timeout_seconds: 600
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -84,6 +137,18 @@ roles:
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
       SGLANG_DSV4_MHC_PREWARM: '1'
+      MOONCAKE_PROTOCOL: rdma
+      # The store only registers host buffers on RDMA; without an explicit list it
+      # falls back to NVLink for same-domain peers and aborts at the warmup put.
+      MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 140gb
+      MOONCAKE_STANDALONE_STORAGE: '0'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      # Advertise the store transfer engine on the node address instead of the first active
+      # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+      MC_TCP_BIND_ADDRESS: "{node}"
+      WITH_NVIDIA_PEERMEM: '0'
     args:
       host: 0.0.0.0
       served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
@@ -117,11 +182,9 @@ roles:
       speculative-num-steps: 1
       speculative-eagle-topk: 1
       speculative-num-draft-tokens: 7
-      enable-hierarchical-cache: true
-      hicache-write-policy: write_back
-      hicache-ratio: 2
-      hicache-io-backend: direct
-      hicache-mem-layout: page_first_direct
+      enable-unified-cache-external-linker: true
+      unified-cache-external-linker-backend: mooncake
+      hicache-storage-backend-extra-config: '{"enable_group_semantics":true}'
       cuda-graph-max-bs-decode: 256
   decode:
     nodes: 8
@@ -157,6 +220,8 @@ roles:
       SGLANG_LOG_MS: '1'
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_DSV4_MHC_PREWARM: '1'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      WITH_NVIDIA_PEERMEM: '0'
     args:
       host: 0.0.0.0
       served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
@@ -199,6 +264,10 @@ srun_options:
 benchmark:
   type: custom
   command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  # The replay client holds ~300 GB of host memory; on the default head node it
+  # shares prefill_0's node and OOMs it. Put it on the etcd/nats node instead.
+  placement:
+    node: dedicated
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
     RESULT_DIR: /logs/agentic
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml
index 4b7b7ff7b0..ff4cd0a15f 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml
@@ -18,6 +18,11 @@ slurm:
 health_check:
   max_attempts: 1440
   interval_seconds: 10
+# Tachometer's per-node exporters slow decode steps by a few percent and add
+# host memory on the head node; this ladder is a throughput measurement.
+observability:
+  tachometer:
+    enabled: false
 resources:
   gpu_type: gb200
   gpus_per_node: 4
@@ -32,6 +37,54 @@ services:
     node: dedicated
   options:
     max_payload_mb: 32
+- name: mooncake-master
+  type: mooncake-master
+  args:
+    - --eviction_high_watermark_ratio=0.90
+# Decode ranks use a chunk cache, so their host DRAM only joins the store
+# through standalone stores: four per decode node, so segments stay uniform
+# with prefill's per-rank ones instead of one oversized segment per node.
+- &decode-store
+  name: store-decode-0
+  type: mooncake-store
+  placement:
+    node: decode
+  args: ["--port", "8800"]
+  env:
+    MOONCAKE_PROTOCOL: rdma
+    MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+    MOONCAKE_GLOBAL_SEGMENT_SIZE: 180gb
+    MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+    MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+    # Advertise the store transfer engine on the node address instead of the first active
+    # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+    MC_TCP_BIND_ADDRESS: "{node_ip}"
+  preamble: |
+    ulimit -n 1048576
+    ulimit -l unlimited
+  # Cold container starts on a decode node can exceed srtctl's 120 s default;
+  # the <<: merge is shallow, so every store repeats the timeout.
+  readiness:
+    port: 8800
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-1
+  args: ["--port", "8801"]
+  readiness:
+    port: 8801
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-2
+  args: ["--port", "8802"]
+  readiness:
+    port: 8802
+    timeout_seconds: 600
+- <<: *decode-store
+  name: store-decode-3
+  args: ["--port", "8803"]
+  readiness:
+    port: 8803
+    timeout_seconds: 600
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -84,6 +137,18 @@ roles:
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
       SGLANG_DSV4_MHC_PREWARM: '1'
+      MOONCAKE_PROTOCOL: rdma
+      # The store only registers host buffers on RDMA; without an explicit list it
+      # falls back to NVLink for same-domain peers and aborts at the warmup put.
+      MOONCAKE_DEVICE: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 140gb
+      MOONCAKE_STANDALONE_STORAGE: '0'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      # Advertise the store transfer engine on the node address instead of the first active
+      # interface, which can be a BMC Redfish host interface (the same 10.0.1.2 on every node).
+      MC_TCP_BIND_ADDRESS: "{node}"
+      WITH_NVIDIA_PEERMEM: '0'
     args:
       host: 0.0.0.0
       served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
@@ -117,11 +182,9 @@ roles:
       speculative-num-steps: 1
       speculative-eagle-topk: 1
       speculative-num-draft-tokens: 7
-      enable-hierarchical-cache: true
-      hicache-write-policy: write_back
-      hicache-ratio: 2
-      hicache-io-backend: direct
-      hicache-mem-layout: page_first_direct
+      enable-unified-cache-external-linker: true
+      unified-cache-external-linker-backend: mooncake
+      hicache-storage-backend-extra-config: '{"enable_group_semantics":true}'
       cuda-graph-max-bs-decode: 256
   decode:
     nodes: 8
@@ -157,6 +220,8 @@ roles:
       SGLANG_LOG_MS: '1'
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_DSV4_MHC_PREWARM: '1'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      WITH_NVIDIA_PEERMEM: '0'
     args:
       host: 0.0.0.0
       served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
@@ -199,6 +264,10 @@ srun_options:
 benchmark:
   type: custom
   command: bash /infmax-workspace/benchmarks/srt_agentic.sh
+  # The replay client holds ~300 GB of host memory; on the default head node it
+  # shares prefill_0's node and OOMs it. Put it on the etcd/nats node instead.
+  placement:
+    node: dedicated
   env:
     INFMAX_CONTAINER_WORKSPACE: /infmax-workspace
     RESULT_DIR: /logs/agentic

From a1806991be8cdff4e496ccc5f63b62a894b3b202 Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Sat, 3 Oct 2026 17:10:45 +0800
Subject: [PATCH 2/8] config(agentx): record Mooncake as the GB200 DSV4 SGLang
 disagg KV offload backend

---
 inferencex-e2e/configs/nvidia-master.yaml | 12 ++++++------
 1 file changed, 6 insertions(+), 6 deletions(-)

diff --git a/inferencex-e2e/configs/nvidia-master.yaml b/inferencex-e2e/configs/nvidia-master.yaml
index b89ea8e9f7..f3eaf4835d 100644
--- a/inferencex-e2e/configs/nvidia-master.yaml
+++ b/inferencex-e2e/configs/nvidia-master.yaml
@@ -6694,7 +6694,7 @@ dsv4-fp4-gb200-dynamo-sglang-agentic-disagg:
       - spec-decoding: draft_model
         conc-list: [64]
         kv-offloading: dram
-        kv-offload-backend: { name: hicache }
+        kv-offload-backend: { name: mooncake, version: "0.3.13" }
         prefill:
           num-worker: 1
           tp: 8
@@ -6710,7 +6710,7 @@ dsv4-fp4-gb200-dynamo-sglang-agentic-disagg:
       - spec-decoding: draft_model
         conc-list: [128]
         kv-offloading: dram
-        kv-offload-backend: { name: hicache }
+        kv-offload-backend: { name: mooncake, version: "0.3.13" }
         prefill:
           num-worker: 1
           tp: 8
@@ -6726,7 +6726,7 @@ dsv4-fp4-gb200-dynamo-sglang-agentic-disagg:
       - spec-decoding: draft_model
         conc-list: [256]
         kv-offloading: dram
-        kv-offload-backend: { name: hicache }
+        kv-offload-backend: { name: mooncake, version: "0.3.13" }
         prefill:
           num-worker: 1
           tp: 16
@@ -6742,7 +6742,7 @@ dsv4-fp4-gb200-dynamo-sglang-agentic-disagg:
       - spec-decoding: draft_model
         conc-list: [768]
         kv-offloading: dram
-        kv-offload-backend: { name: hicache }
+        kv-offload-backend: { name: mooncake, version: "0.3.13" }
         prefill:
           num-worker: 2
           tp: 16
@@ -6758,7 +6758,7 @@ dsv4-fp4-gb200-dynamo-sglang-agentic-disagg:
       - spec-decoding: draft_model
         conc-list: [1024]
         kv-offloading: dram
-        kv-offload-backend: { name: hicache }
+        kv-offload-backend: { name: mooncake, version: "0.3.13" }
         prefill:
           num-worker: 2
           tp: 16
@@ -6774,7 +6774,7 @@ dsv4-fp4-gb200-dynamo-sglang-agentic-disagg:
       - spec-decoding: draft_model
         conc-list: [1280]
         kv-offloading: dram
-        kv-offload-backend: { name: hicache }
+        kv-offload-backend: { name: mooncake, version: "0.3.13" }
         prefill:
           num-worker: 2
           tp: 16

From f87550bf5825a93922896fbfb9d458207b52009b Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Sat, 3 Oct 2026 20:19:43 +0800
Subject: [PATCH 3/8] docs(changelog): GB200 DSV4 SGLang AgentX Mooncake KV
 offload

---
 inferencex-e2e/perf-changelog.yaml | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index db5662f5b8..5faa51c862 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9292,3 +9292,20 @@
     - "Update the B200 DeepSeek-V4.1-Flash vLLM AgentX image from nightly ddd6fbca to nightly-dev-x86_64-cu130-ac9126e58aa7 and enable FlashInfer autotuning."
     - "Run TP4 at concurrency 1-128 and replace TP2 with DEP2 (TP1 x DP2 + EP2, DeepGEMM MegaMoE) at concurrency 8-32 and DEP4 (TP1 x DP4 + EP4, MegaMoE) at concurrency 64-128, both behind a consistent-hash vLLM Router. All points set gpu-memory-utilization 0.97."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3686
+
+- config-keys:
+    - dsv4-fp4-gb200-dynamo-sglang-agentic-disagg
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Move the GB200 DeepSeek-V4-Pro AgentX disaggregated recipes from hierarchical cache to the Mooncake unified-cache external linker with group semantics, as #3187 did for GB300. MoE settings are unchanged: the recipes keep MegaMoE without W4A4 MXFP4."
+    - "Decode nodes lend host DRAM to the store through srt-slurm mooncake-store services: four 180GB standalone stores per decode node, gated on their readiness ports with a 600-second timeout. Prefill keeps a 140GB segment per rank. GB200 nodes have the same host memory as GB300 nodes, so the sizes carry over."
+    - "The prefill role and the decode stores name their RDMA devices (MOONCAKE_DEVICE=mlx5_0..3, as the GB200 vLLM Mooncake recipes do) and set MC_TCP_BIND_ADDRESS to their node, so each store transfer engine advertises the node address instead of the first active interface."
+    - "The benchmark client runs on the dedicated etcd/nats node, because prefill nodes now also hold Mooncake segments, and Tachometer's per-node exporters are off for this throughput ladder."
+    - "Measured at concurrency 128 (1P DEP8 / 1D DEP16) against the current recipe with the same image, srt-slurm and golden acceptance length: 26,566 vs 23,763 total tok/s/GPU (+11.8%) at the same P90 TPOT (7.31 vs 7.22 ms), with TTFT p50 3.27 vs 4.95 s."
+    - "将 GB200 DeepSeek-V4-Pro AgentX 分离式配方从 hierarchical cache 切换到启用 group semantics 的 Mooncake unified-cache external linker，与 #3187 对 GB300 的改动一致。MoE 设置不变：配方继续使用不启用 W4A4 MXFP4 的 MegaMoE。"
+    - "decode 节点通过 srt-slurm 的 mooncake-store 服务向 store 借出主机 DRAM：每个 decode 节点 4 个 180GB 的独立 store，以就绪端口作为启动检查，超时 600 秒。prefill 每个 rank 保留 140GB segment。GB200 与 GB300 节点的主机内存相同，因此沿用相同大小。"
+    - "prefill 角色与 decode 的 store 显式指定 RDMA 设备（MOONCAKE_DEVICE=mlx5_0..3，与 GB200 vLLM Mooncake 配方一致），并将 MC_TCP_BIND_ADDRESS 设为所在节点，使每个 store 传输引擎公布节点自身地址，而不是第一个活动网卡。"
+    - "压测客户端运行在独立的 etcd/nats 节点上，因为 prefill 节点现在也承载 Mooncake segment；该吞吐阶梯关闭 Tachometer 的逐节点 exporter。"
+    - "在并发 128（1P DEP8 / 1D DEP16）下，与当前配方在相同镜像、srt-slurm 与 golden acceptance length 下对比：单卡总吞吐 26,566 对 23,763 tok/s/GPU（+11.8%），P90 TPOT 相当（7.31 对 7.22 ms），TTFT p50 为 3.27 对 4.95 秒。"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/TBD

From 60d03e8b591bda99234b0e2c80a8cfa9de408e7d Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Sat, 3 Oct 2026 20:20:25 +0800
Subject: [PATCH 4/8] docs(changelog): link #3701

---
 inferencex-e2e/perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 5faa51c862..c441ce4b60 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9308,4 +9308,4 @@
     - "prefill 角色与 decode 的 store 显式指定 RDMA 设备（MOONCAKE_DEVICE=mlx5_0..3，与 GB200 vLLM Mooncake 配方一致），并将 MC_TCP_BIND_ADDRESS 设为所在节点，使每个 store 传输引擎公布节点自身地址，而不是第一个活动网卡。"
     - "压测客户端运行在独立的 etcd/nats 节点上，因为 prefill 节点现在也承载 Mooncake segment；该吞吐阶梯关闭 Tachometer 的逐节点 exporter。"
     - "在并发 128（1P DEP8 / 1D DEP16）下，与当前配方在相同镜像、srt-slurm 与 golden acceptance length 下对比：单卡总吞吐 26,566 对 23,763 tok/s/GPU（+11.8%），P90 TPOT 相当（7.31 对 7.22 ms），TTFT p50 为 3.27 对 4.95 秒。"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/TBD
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3701

From 28b5f47ee758e4b316690dafe7c8b449289d8a7c Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Tue, 6 Oct 2026 18:42:29 +0800
Subject: [PATCH 5/8] perf(agentx): 8k-token prefill chunks per rank on the
 GB200 DSV4 DEP16 prefill recipes

Raise chunked-prefill-size from 65536 to 131072 on the c256/c768/c1024/c1280
DEP16 prefill workers (4k to 8k tokens per DP rank) and lower prefill
mem-fraction-static from 0.85 to 0.80; 8k chunks run out of GPU memory at 0.85.
---
 .../disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml   | 6 ++++--
 .../disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml  | 6 ++++--
 .../disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml  | 6 ++++--
 .../disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml   | 6 ++++--
 4 files changed, 16 insertions(+), 8 deletions(-)

diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml
index 541934f0af..760f4994fd 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml
@@ -169,11 +169,13 @@ roles:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      mem-fraction-static: 0.85
+      # Prefill runs 8k-token chunks per DP rank (chunked-prefill-size 131072 / DEP16),
+      # which run out of GPU memory at 0.85 on GB200.
+      mem-fraction-static: 0.80
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
-      chunked-prefill-size: 65536
+      chunked-prefill-size: 131072
       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"}'
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml
index f1672f9e54..39d118f51e 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml
@@ -169,11 +169,13 @@ roles:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      mem-fraction-static: 0.85
+      # Prefill runs 8k-token chunks per DP rank (chunked-prefill-size 131072 / DEP16),
+      # which run out of GPU memory at 0.85 on GB200.
+      mem-fraction-static: 0.80
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
-      chunked-prefill-size: 65536
+      chunked-prefill-size: 131072
       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"}'
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml
index ac2ed56b08..fcc9007b18 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml
@@ -169,11 +169,13 @@ roles:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      mem-fraction-static: 0.85
+      # Prefill runs 8k-token chunks per DP rank (chunked-prefill-size 131072 / DEP16),
+      # which run out of GPU memory at 0.85 on GB200.
+      mem-fraction-static: 0.80
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
-      chunked-prefill-size: 65536
+      chunked-prefill-size: 131072
       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"}'
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml
index ff4cd0a15f..caa63de7d6 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml
@@ -169,11 +169,13 @@ roles:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      mem-fraction-static: 0.85
+      # Prefill runs 8k-token chunks per DP rank (chunked-prefill-size 131072 / DEP16),
+      # which run out of GPU memory at 0.85 on GB200.
+      mem-fraction-static: 0.80
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
-      chunked-prefill-size: 65536
+      chunked-prefill-size: 131072
       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"}'

From 47f237ca9f5b3c0510c2cad0477550b4265784fe Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Tue, 6 Oct 2026 18:43:05 +0800
Subject: [PATCH 6/8] docs(changelog): GB200 DSV4 agg key, DEP16 prefill
 chunks; English-only entry

List dsv4-fp4-gb200-dynamo-sglang-agentic-agg so the sweep reruns the low-latency
TP8 c1/c4 points, describe the 8k-token DEP16 prefill chunks, and drop the
Chinese bullets now that new changelog entries must be English-only.
---
 inferencex-e2e/perf-changelog.yaml | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)

diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index 0aa77612cc..e3c4e02a3e 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9311,6 +9311,7 @@
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3664
 
 - config-keys:
+    - dsv4-fp4-gb200-dynamo-sglang-agentic-agg
     - dsv4-fp4-gb200-dynamo-sglang-agentic-disagg
   scenario-type:
     - agentic-coding
@@ -9319,10 +9320,7 @@
     - "Decode nodes lend host DRAM to the store through srt-slurm mooncake-store services: four 180GB standalone stores per decode node, gated on their readiness ports with a 600-second timeout. Prefill keeps a 140GB segment per rank. GB200 nodes have the same host memory as GB300 nodes, so the sizes carry over."
     - "The prefill role and the decode stores name their RDMA devices (MOONCAKE_DEVICE=mlx5_0..3, as the GB200 vLLM Mooncake recipes do) and set MC_TCP_BIND_ADDRESS to their node, so each store transfer engine advertises the node address instead of the first active interface."
     - "The benchmark client runs on the dedicated etcd/nats node, because prefill nodes now also hold Mooncake segments, and Tachometer's per-node exporters are off for this throughput ladder."
+    - "The DEP16 prefill workers (c256, c768, c1024, c1280) run 8k-token chunks per DP rank instead of 4k (chunked-prefill-size 65536 to 131072) with prefill mem-fraction-static lowered from 0.85 to 0.80; at 0.85 the 8k chunks run out of GPU memory."
+    - "The aggregate TP8 c1/c4 recipes are unchanged and stay on hierarchical cache; their config key is listed so this sweep covers the whole GB200 Pareto ladder, including the low-latency points."
     - "Measured at concurrency 128 (1P DEP8 / 1D DEP16) against the current recipe with the same image, srt-slurm and golden acceptance length: 26,566 vs 23,763 total tok/s/GPU (+11.8%) at the same P90 TPOT (7.31 vs 7.22 ms), with TTFT p50 3.27 vs 4.95 s."
-    - "将 GB200 DeepSeek-V4-Pro AgentX 分离式配方从 hierarchical cache 切换到启用 group semantics 的 Mooncake unified-cache external linker，与 #3187 对 GB300 的改动一致。MoE 设置不变：配方继续使用不启用 W4A4 MXFP4 的 MegaMoE。"
-    - "decode 节点通过 srt-slurm 的 mooncake-store 服务向 store 借出主机 DRAM：每个 decode 节点 4 个 180GB 的独立 store，以就绪端口作为启动检查，超时 600 秒。prefill 每个 rank 保留 140GB segment。GB200 与 GB300 节点的主机内存相同，因此沿用相同大小。"
-    - "prefill 角色与 decode 的 store 显式指定 RDMA 设备（MOONCAKE_DEVICE=mlx5_0..3，与 GB200 vLLM Mooncake 配方一致），并将 MC_TCP_BIND_ADDRESS 设为所在节点，使每个 store 传输引擎公布节点自身地址，而不是第一个活动网卡。"
-    - "压测客户端运行在独立的 etcd/nats 节点上，因为 prefill 节点现在也承载 Mooncake segment；该吞吐阶梯关闭 Tachometer 的逐节点 exporter。"
-    - "在并发 128（1P DEP8 / 1D DEP16）下，与当前配方在相同镜像、srt-slurm 与 golden acceptance length 下对比：单卡总吞吐 26,566 对 23,763 tok/s/GPU（+11.8%），P90 TPOT 相当（7.31 对 7.22 ms），TTFT p50 为 3.27 对 4.95 秒。"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3701

From e12d06406f46e727d96fbc617189aee448796526 Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Thu, 8 Oct 2026 01:05:31 +0800
Subject: [PATCH 7/8] Revert "perf(agentx): 8k-token prefill chunks per rank on
 the GB200 DSV4 DEP16 prefill recipes"

This reverts commit 28b5f47ee758e4b316690dafe7c8b449289d8a7c.
---
 .../disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml   | 6 ++----
 .../disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml  | 6 ++----
 .../disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml  | 6 ++----
 .../disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml   | 6 ++----
 4 files changed, 8 insertions(+), 16 deletions(-)

diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml
index 760f4994fd..541934f0af 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-1p1d-dep16-dep32-c256-mtp-kvoffload.yaml
@@ -169,13 +169,11 @@ roles:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      # Prefill runs 8k-token chunks per DP rank (chunked-prefill-size 131072 / DEP16),
-      # which run out of GPU memory at 0.85 on GB200.
-      mem-fraction-static: 0.80
+      mem-fraction-static: 0.85
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
-      chunked-prefill-size: 131072
+      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"}'
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml
index 39d118f51e..f1672f9e54 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1024-mtp-kvoffload.yaml
@@ -169,13 +169,11 @@ roles:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      # Prefill runs 8k-token chunks per DP rank (chunked-prefill-size 131072 / DEP16),
-      # which run out of GPU memory at 0.85 on GB200.
-      mem-fraction-static: 0.80
+      mem-fraction-static: 0.85
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
-      chunked-prefill-size: 131072
+      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"}'
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml
index fcc9007b18..ac2ed56b08 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c1280-mtp-kvoffload.yaml
@@ -169,13 +169,11 @@ roles:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      # Prefill runs 8k-token chunks per DP rank (chunked-prefill-size 131072 / DEP16),
-      # which run out of GPU memory at 0.85 on GB200.
-      mem-fraction-static: 0.80
+      mem-fraction-static: 0.85
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
-      chunked-prefill-size: 131072
+      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"}'
diff --git a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml
index caa63de7d6..ff4cd0a15f 100644
--- a/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml
+++ b/inferencex-e2e/benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb200-fp4/agentx/disagg-gb200-2p1d-dep16-dep32-c768-mtp-kvoffload.yaml
@@ -169,13 +169,11 @@ roles:
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: prefill
       load-balance-method: total_tokens
-      # Prefill runs 8k-token chunks per DP rank (chunked-prefill-size 131072 / DEP16),
-      # which run out of GPU memory at 0.85 on GB200.
-      mem-fraction-static: 0.80
+      mem-fraction-static: 0.85
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
-      chunked-prefill-size: 131072
+      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"}'

From b7e858958472d3c5d52454068ebc3addbdd94224 Mon Sep 17 00:00:00 2001
From: weireweire <20922698+weireweire@users.noreply.github.com>
Date: Thu, 8 Oct 2026 01:05:40 +0800
Subject: [PATCH 8/8] docs(changelog): drop the reverted DEP16 prefill chunk
 change from the GB200 DSV4 entry

---
 inferencex-e2e/perf-changelog.yaml | 1 -
 1 file changed, 1 deletion(-)

diff --git a/inferencex-e2e/perf-changelog.yaml b/inferencex-e2e/perf-changelog.yaml
index b5d8df3833..b9b9fde46b 100644
--- a/inferencex-e2e/perf-changelog.yaml
+++ b/inferencex-e2e/perf-changelog.yaml
@@ -9338,7 +9338,6 @@
     - "Decode nodes lend host DRAM to the store through srt-slurm mooncake-store services: four 180GB standalone stores per decode node, gated on their readiness ports with a 600-second timeout. Prefill keeps a 140GB segment per rank. GB200 nodes have the same host memory as GB300 nodes, so the sizes carry over."
     - "The prefill role and the decode stores name their RDMA devices (MOONCAKE_DEVICE=mlx5_0..3, as the GB200 vLLM Mooncake recipes do) and set MC_TCP_BIND_ADDRESS to their node, so each store transfer engine advertises the node address instead of the first active interface."
     - "The benchmark client runs on the dedicated etcd/nats node, because prefill nodes now also hold Mooncake segments, and Tachometer's per-node exporters are off for this throughput ladder."
-    - "The DEP16 prefill workers (c256, c768, c1024, c1280) run 8k-token chunks per DP rank instead of 4k (chunked-prefill-size 65536 to 131072) with prefill mem-fraction-static lowered from 0.85 to 0.80; at 0.85 the 8k chunks run out of GPU memory."
     - "The aggregate TP8 c1/c4 recipes are unchanged and stay on hierarchical cache; their config key is listed so this sweep covers the whole GB200 Pareto ladder, including the low-latency points."
     - "Measured at concurrency 128 (1P DEP8 / 1D DEP16) against the current recipe with the same image, srt-slurm and golden acceptance length: 26,566 vs 23,763 total tok/s/GPU (+11.8%) at the same P90 TPOT (7.31 vs 7.22 ms), with TTFT p50 3.27 vs 4.95 s."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3701