[NV] Move GB300 DeepSeek-V4-Pro AgentX disagg to the Mooncake external linker / 将 GB300 DeepSeek-V4-Pro AgentX 分离式配置切换到 Mooncake external linker

benchmarks/multi_node/srt-slurm-recipes/configs/dsv4-gb300-sglang-mooncake-opt.sh

@@ -0,0 +1,51 @@
+#!/usr/bin/env bash
+# TEMPORARY: verify the SGLang source tree carrying the DSV4 Mooncake
+# external-linker optimizations before any server starts.
+#
+# The recipes put ${SGLANG_SRC}/python on PYTHONPATH so the servers import the
+# reviewed branch instead of the container's installed SGLang, while still
+# using the container's compiled kernels. If the runner failed to clone the
+# tree, PYTHONPATH silently resolves to the stock package and the job would
+# benchmark the wrong code, so fail loudly here instead.
+#
+# Delete this script, its runner wiring, and the recipes' setup_script and
+# environment keys once the optimizations ship in the pinned image.
+set -euo pipefail
+
+SGLANG_SRC="${SGLANG_MOONCAKE_OPT_SRC:-/configs/sglang-mooncake-opt}"
+INIT="${SGLANG_SRC}/python/sglang/__init__.py"
+
+if [ ! -f "${INIT}" ]; then
+    echo "ERROR: SGLang optimization source missing at ${INIT}." >&2
+    echo "The runner must clone it before submitting; refusing to run against" >&2
+    echo "the container's stock SGLang." >&2
+    exit 1
+fi
+
+python3 - "${SGLANG_SRC}" <<'PYEOF'
+import sys
+from pathlib import Path
+
+root = Path(sys.argv[1]) / "python"
+required = {
+    "SGLANG_EXTERNAL_LINKER_SWA_RETENTION_INTERVAL": "sglang/srt/environ.py",
+    "SGLANG_MOONCAKE_STORE_CONTRIBUTOR": "sglang/srt/environ.py",
+}
+for token, relative in sorted(required.items()):
+    path = root / relative
+    if token not in path.read_text():
+        raise SystemExit(f"ERROR: {token} not found in {path}; wrong revision")
+print(f"[sglang-mooncake-opt] verified source tree at {root}")
+PYEOF
+
+COMPAT_SHIM="${SGLANG_MOONCAKE_OPT_COMPAT:-/configs/sglang-server-args-compat}/sitecustomize.py"
+if [ ! -f "${COMPAT_SHIM}" ]; then
+    echo "ERROR: Dynamo compatibility shim missing at ${COMPAT_SHIM}." >&2
+    exit 1
+fi
+
+resolved=$(PYTHONPATH="${SGLANG_SRC}/python:${PYTHONPATH:-}" python3 -c 'import sglang; print(sglang.__file__)')
+case "${resolved}" in
+    "${SGLANG_SRC}"/*) echo "[sglang-mooncake-opt] sglang resolves to ${resolved}" ;;
+    *) echo "ERROR: sglang resolves to ${resolved}, not ${SGLANG_SRC}" >&2; exit 1 ;;
+esac

benchmarks/multi_node/srt-slurm-recipes/configs/sglang-server-args-compat/sitecustomize.py

@@ -0,0 +1,27 @@
+"""TEMPORARY: restore ServerArgs.get_model_config for the pinned Dynamo wheel.
+
+The reviewed SGLang branch these recipes import moved model-config resolution
+out of ServerArgs, but the pinned Dynamo wheel still calls
+``server_args.get_model_config()`` while parsing worker arguments. Re-attach
+the accessor so the wheel keeps working against the newer source tree.
+
+This file is picked up because its directory is on PYTHONPATH, so it is
+imported by every interpreter in the job, including ones that never import
+SGLang. Failing to import SGLang there is expected and must stay silent.
+
+Remove this directory, its PYTHONPATH entry, and the rest of the temporary
+source override once the optimizations ship in the pinned image.
+"""
+
+try:
+    from sglang.srt.arg_groups.model_override_base import model_config_of
+    from sglang.srt.server_args import ServerArgs
+except Exception:  # noqa: BLE001 - non-SGLang interpreters legitimately fail here
+    pass
+else:
+    if not hasattr(ServerArgs, "get_model_config"):
+
+        def get_model_config(self):
+            return model_config_of(self)
+
+        ServerArgs.get_model_config = get_model_config

benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb300-fp4/agentx/disagg-1p1d-dep8-dep16-c480-mtp-kvoffload.yaml

@@ -18,7 +18,7 @@ identity:
   model:
     repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
   container:
-    image: "lmsysorg/sglang:nightly-dev-cu13-20260829-89816a21"
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260916-c9a8fba9"
 
 dynamo:
   install: true
@@ -32,20 +32,35 @@ health_check:
   max_attempts: 1440
   interval_seconds: 10
 
+# TEMPORARY: the Mooncake external-linker optimizations this Pareto point
+# depends on are not in a released SGLang image yet. The gb300-nv runner clones
+# the reviewed branch into the srt-slurm configs directory and the servers
+# import it from there. The setup script aborts the job when that tree is
+# missing, so a run can never silently fall back to the container's SGLang.
+# Remove both keys once the optimizations ship in the pinned image.
+setup_script: dsv4-gb300-sglang-mooncake-opt.sh
+
+environment:
+  PYTHONPATH: /configs/sglang-server-args-compat:/configs/sglang-mooncake-opt/python
+
 resources:
   gpu_type: gb300
   gpus_per_node: 4
 services:
-  - name: etcd
-    type: etcd
-    placement:
-      node: dedicated
-  - name: nats
-    type: nats
-    placement:
-      node: dedicated
-    options:
-      max_payload_mb: 32
+- name: etcd
+  type: etcd
+  placement:
+    node: dedicated
+- name: nats
+  type: nats
+  placement:
+    node: dedicated
+  options:
+    max_payload_mb: 32
+- name: mooncake-master
+  type: mooncake-master
+  args:
+  - --eviction_high_watermark_ratio=0.90
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -100,6 +115,15 @@ roles:
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
 
+      MOONCAKE_PROTOCOL: rdma
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 140gb
+      MOONCAKE_STANDALONE_STORAGE: '0'
+      MC_STORE_CLIENT_METRIC: '1'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      WITH_NVIDIA_PEERMEM: '0'
+      SGLANG_EXTERNAL_LINKER_WRITE_THROUGH_THRESHOLD: '1'
+      SGLANG_EXTERNAL_LINKER_SWA_RETENTION_INTERVAL: '32768'
     args:
       host: 0.0.0.0
       served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
@@ -125,7 +149,7 @@ roles:
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
-      cuda-graph-max-bs: 256
+      cuda-graph-max-bs-decode: 256
       chunked-prefill-size: 65536
       disable-flashinfer-autotune: true
       model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
@@ -135,11 +159,11 @@ 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: 1
-      hicache-io-backend: direct
-
+      prefill-chunk-size-per-request: 2048
+      disaggregation-defer-partial-kv-transfer: true
+      enable-unified-cache-external-linker: true
+      unified-cache-external-linker-backend: mooncake
+      hicache-storage-backend-extra-config: '{"enable_group_semantics":true}'
   decode:
     nodes: 4
     workers: 1
@@ -175,6 +199,13 @@ roles:
       SGLANG_LOG_MS: '1'
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
 
+      MOONCAKE_PROTOCOL: rdma
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 140gb
+      MOONCAKE_STANDALONE_STORAGE: '0'
+      MC_STORE_CLIENT_METRIC: '1'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      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
@@ -195,12 +226,13 @@ roles:
       enable-w4a4-mxfp4-megamoe: true
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: decode
+      mooncake-store-contributor: true
       load-balance-method: total_tokens
       mem-fraction-static: 0.9
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 3072
-      cuda-graph-max-bs: 256
+      cuda-graph-max-bs-decode: 256
       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"}'

benchmarks/multi_node/srt-slurm-recipes/dsv4/sglang/gb300-fp4/agentx/disagg-2p1d-dep8-dep16-c960-mtp-kvoffload.yaml

@@ -18,7 +18,7 @@ identity:
   model:
     repo: "deepseek-ai/DeepSeek-V4-Pro-0813"
   container:
-    image: "lmsysorg/sglang:nightly-dev-cu13-20260829-89816a21"
+    image: "lmsysorg/sglang:nightly-dev-cu13-20260916-c9a8fba9"
 
 dynamo:
   install: true
@@ -32,20 +32,35 @@ health_check:
   max_attempts: 1440
   interval_seconds: 10
 
+# TEMPORARY: the Mooncake external-linker optimizations this Pareto point
+# depends on are not in a released SGLang image yet. The gb300-nv runner clones
+# the reviewed branch into the srt-slurm configs directory and the servers
+# import it from there. The setup script aborts the job when that tree is
+# missing, so a run can never silently fall back to the container's SGLang.
+# Remove both keys once the optimizations ship in the pinned image.
+setup_script: dsv4-gb300-sglang-mooncake-opt.sh
+
+environment:
+  PYTHONPATH: /configs/sglang-server-args-compat:/configs/sglang-mooncake-opt/python
+
 resources:
   gpu_type: gb300
   gpus_per_node: 4
 services:
-  - name: etcd
-    type: etcd
-    placement:
-      node: dedicated
-  - name: nats
-    type: nats
-    placement:
-      node: dedicated
-    options:
-      max_payload_mb: 32
+- name: etcd
+  type: etcd
+  placement:
+    node: dedicated
+- name: nats
+  type: nats
+  placement:
+    node: dedicated
+  options:
+    max_payload_mb: 32
+- name: mooncake-master
+  type: mooncake-master
+  args:
+  - --eviction_high_watermark_ratio=0.90
 frontend:
   type: dynamo
   nginx_session_affinity: true
@@ -100,6 +115,15 @@ roles:
       SGLANG_NCCL_ALL_GATHER_IN_OVERLAP_SCHEDULER_SYNC_BATCH: '1'
       SGLANG_DSV4_MHC_PREWARM: '1'
 
+      MOONCAKE_PROTOCOL: rdma
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 140gb
+      MOONCAKE_STANDALONE_STORAGE: '0'
+      MC_STORE_CLIENT_METRIC: '1'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      MC_ENABLE_DEST_DEVICE_AFFINITY: '1'
+      WITH_NVIDIA_PEERMEM: '0'
+      SGLANG_EXTERNAL_LINKER_WRITE_THROUGH_THRESHOLD: '1'
+      SGLANG_EXTERNAL_LINKER_SWA_RETENTION_INTERVAL: '32768'
     args:
       host: 0.0.0.0
       served-model-name: deepseek-ai/DeepSeek-V4-Pro-0813
@@ -125,7 +149,7 @@ roles:
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 256
-      cuda-graph-max-bs: 256
+      cuda-graph-max-bs-decode: 256
       chunked-prefill-size: 65536
       disable-flashinfer-autotune: true
       model-loader-extra-config: '{"enable_multithread_load":true,"num_threads":8}'
@@ -135,11 +159,11 @@ 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: 1
-      hicache-io-backend: direct
-
+      prefill-chunk-size-per-request: 2048
+      disaggregation-defer-partial-kv-transfer: true
+      enable-unified-cache-external-linker: true
+      unified-cache-external-linker-backend: mooncake
+      hicache-storage-backend-extra-config: '{"enable_group_semantics":true}'
   decode:
     nodes: 4
     workers: 1
@@ -177,6 +201,13 @@ roles:
       SGLANG_REQUEST_STATE_WAIT_TIMEOUT: '60'
       SGLANG_DSV4_MHC_PREWARM: '1'
 
+      MOONCAKE_PROTOCOL: rdma
+      MOONCAKE_GLOBAL_SEGMENT_SIZE: 140gb
+      MOONCAKE_STANDALONE_STORAGE: '0'
+      MC_STORE_CLIENT_METRIC: '1'
+      MC_STORE_CLIENT_METRIC_INTERVAL: '5'
+      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
@@ -197,12 +228,13 @@ roles:
       enable-w4a4-mxfp4-megamoe: true
       disaggregation-transfer-backend: mooncake
       disaggregation-mode: decode
+      mooncake-store-contributor: true
       load-balance-method: total_tokens
       mem-fraction-static: 0.9
       page-size: 256
       swa-full-tokens-ratio: 0.02
       max-running-requests: 3072
-      cuda-graph-max-bs: 256
+      cuda-graph-max-bs-decode: 256
       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"}'