SemiAnalysisAI · cquil11 · Jul 13, 2026 · Jul 13, 2026
diff --git a/benchmarks/single_node/agentic/README.md b/benchmarks/single_node/agentic/README.md
@@ -20,7 +20,7 @@ currently either `none` or `dram`; when it is `dram`, the backend must be set:
 ```yaml
 - dram-utilization: 0.80
   search-space:
-  - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: native }, conc-list: [16, 32] }
+  - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: vllm-native }, conc-list: [16, 32] }
   - { tp: 8, kv-offloading: none, conc-list: [16, 32] }
 ```
 

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm.sh b/benchmarks/single_node/agentic/dsv4_fp4_mi355x_vllm.sh
@@ -22,7 +22,7 @@ set -x
 #   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR
 #
 # KV_OFFLOADING=dram requires one of these. 
-#   KV_OFFLOAD_BACKEND=native.
+#   KV_OFFLOAD_BACKEND=vllm-native.
 #   KV_OFFLOAD_BACKEND=mooncake.
 #   KV_OFFLOAD_BACKEND=lmcache.
 #   KV_OFFLOAD_BACKEND=hicache.
@@ -102,16 +102,16 @@ OFFLOAD_ARGS=()
 
 if agentic_kv_offload_enabled; then
 case "${KV_OFFLOAD_BACKEND:-}" in
-  native)
-    require_agentic_kv_offload_backend native
+  vllm-native)
+    require_agentic_kv_offload_backend vllm-native
     # ---- vLLM native config ----------------------------------------------------------
     unset VLLM_USE_SIMPLE_KV_OFFLOAD
     # MI355X nodes have ~2.7 TiB of host DRAM available for offload;
     # reserve 2.5 TB for the offload pool (leaves ~200 GB headroom for
     # worker RSS / page cache / slurm cgroup).
     TOTAL_CPU_DRAM_PARTITION_GB="$((TOTAL_CPU_DRAM_GB / (8 / TP)))"
     # Use vLLM's regular native KV-offload path (OffloadingConnector),
-    # NOT the SimpleCPUOffloadConnector. The "native" backend resolves to
+    # NOT the SimpleCPUOffloadConnector. The "vllm-native" backend resolves to
     # OffloadingConnector by default; setting VLLM_USE_SIMPLE_KV_OFFLOAD=1
     # would switch it to SimpleCPUOffloadConnector. We intentionally leave
     # that env var UNSET here so the regular OffloadingConnector path is
@@ -277,7 +277,7 @@ EOF
         git clone https://github.com/LMCache/LMCache.git
         cd LMCache
         # https://github.com/LMCache/LMCache/pull/3853
-        git checkout dev
+        git checkout 9229067cec0b3a63bb8a39368d101db7ac0bc3c1
         pip install -r requirements/build.txt
         pip install grpcio==1.78.0
         CXX=hipcc BUILD_WITH_HIP=1 pip install -e .   --no-build-isolation
@@ -345,7 +345,7 @@ EOF
         )
     ;;
   *)
-    echo "Error: unsupported KV_OFFLOAD_BACKEND '${KV_OFFLOAD_BACKEND:-}' (expected: native, mooncake, lmcache)" >&2
+    echo "Error: unsupported KV_OFFLOAD_BACKEND '${KV_OFFLOAD_BACKEND:-}' (expected: vllm-native, mooncake, lmcache)" >&2
     exit 1
     ;;
 esac

diff --git a/benchmarks/single_node/agentic/kimik2.5_fp4_b200.sh b/benchmarks/single_node/agentic/kimik2.5_fp4_b200.sh
@@ -96,7 +96,7 @@ if require_agentic_kv_offload_backend lmcache; then
         { set +x; } 2>/dev/null
         unset VLLM_USE_SIMPLE_KV_OFFLOAD
 
-        agentic_pip_install --quiet --no-cache-dir lmcache
+        agentic_pip_install --quiet --no-cache-dir 'lmcache==0.5.1'
         python3 -c "import lmcache.integration.vllm.lmcache_mp_connector" >/dev/null
 
         # MP mode owns the configured CPU pool in the external LMCache

diff --git a/benchmarks/single_node/agentic/kimik2.5_fp4_b300.sh b/benchmarks/single_node/agentic/kimik2.5_fp4_b300.sh
@@ -7,7 +7,7 @@ set -x
 # Required env vars:
 #   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR
 #
-# KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=native.
+# KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=vllm-simple.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
@@ -42,7 +42,7 @@ mkdir -p "$RESULT_DIR"
 OFFLOAD_ARGS=()
 PREFIX_CACHE_ARGS=()
 
-if require_agentic_kv_offload_backend native; then
+if require_agentic_kv_offload_backend vllm-simple; then
     export VLLM_USE_SIMPLE_KV_OFFLOAD=1
     OFFLOAD_ARGS=(
         --kv_offloading_backend native

diff --git a/benchmarks/single_node/agentic/kimik2.5_fp4_mi355x.sh b/benchmarks/single_node/agentic/kimik2.5_fp4_mi355x.sh
@@ -7,7 +7,7 @@ set -x
 # Required env vars:
 #   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR
 #
-# KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=native.
+# KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=vllm-native.
 
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
@@ -64,10 +64,10 @@ mkdir -p "$RESULT_DIR"
 OFFLOAD_ARGS=()
 PREFIX_CACHE_ARGS=()
 
-if require_agentic_kv_offload_backend native; then
+if require_agentic_kv_offload_backend vllm-native; then
     unset VLLM_USE_SIMPLE_KV_OFFLOAD
     # Use vLLM's regular native KV-offload path (OffloadingConnector),
-    # NOT the SimpleCPUOffloadConnector. The "native" backend resolves to
+    # NOT the SimpleCPUOffloadConnector. The "vllm-native" backend resolves to
     # OffloadingConnector by default; setting VLLM_USE_SIMPLE_KV_OFFLOAD=1
     # would switch it to SimpleCPUOffloadConnector. We intentionally leave
     # that env var UNSET here so the regular OffloadingConnector path is

diff --git a/benchmarks/single_node/agentic/kimik2.5_int4_b200.sh b/benchmarks/single_node/agentic/kimik2.5_int4_b200.sh
@@ -38,7 +38,7 @@ SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
 
 OFFLOAD_ARGS=""
-if require_agentic_kv_offload_backend native; then
+if require_agentic_kv_offload_backend vllm-simple; then
     export VLLM_USE_SIMPLE_KV_OFFLOAD=1
     OFFLOAD_ARGS="--kv_offloading_backend native --kv_offloading_size $TOTAL_CPU_DRAM_GB --disable-hybrid-kv-cache-manager"
 fi

diff --git a/benchmarks/single_node/agentic/kimik2.5_int4_h100.sh b/benchmarks/single_node/agentic/kimik2.5_int4_h100.sh
@@ -38,7 +38,7 @@ SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
 
 OFFLOAD_ARGS=""
-if require_agentic_kv_offload_backend native; then
+if require_agentic_kv_offload_backend vllm-simple; then
     export VLLM_USE_SIMPLE_KV_OFFLOAD=1
     OFFLOAD_ARGS="--kv_offloading_backend native --kv_offloading_size $TOTAL_CPU_DRAM_GB --disable-hybrid-kv-cache-manager"
 fi

diff --git a/benchmarks/single_node/agentic/kimik2.5_int4_h200.sh b/benchmarks/single_node/agentic/kimik2.5_int4_h200.sh
@@ -38,7 +38,7 @@ SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
 
 OFFLOAD_ARGS=""
-if require_agentic_kv_offload_backend native; then
+if require_agentic_kv_offload_backend vllm-simple; then
     # Kimi K2.5 is pure TP (no DP-attn): single engine, world_size=TP.
     # SimpleCPUOffloadConnector internally divides cpu_bytes_to_use by
     # world_size, so pass the full TOTAL_CPU_DRAM_GB; TP-shared mmap

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
@@ -355,6 +355,7 @@ qwen3.5-fp8-mi355x-sglang-disagg:
   runner: mi355x-disagg
   precision: fp8
   framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
   kv-p2p-transfer: mori
   multinode: true
   disagg: true
@@ -483,6 +484,7 @@ qwen3.5-fp4-mi355x-sglang-disagg:
   runner: mi355x-disagg
   precision: fp4
   framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
   kv-p2p-transfer: mori
   multinode: true
   disagg: true
@@ -597,6 +599,7 @@ glm5-fp8-mi355x-sglang-disagg:
   runner: mi355x-disagg
   precision: fp8
   framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
   kv-p2p-transfer: mori
   multinode: true
   disagg: true
@@ -803,13 +806,13 @@ kimik2.5-fp4-mi355x-vllm-agentic:
       # DRAM offload only above the KV cliff. Lower concurrencies fit
       # entirely on-GPU, so paying the offload-path overhead there would
       # just slow them down without measuring anything new.
-      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: native }, conc-list: [32, 40, 48, 56] }
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: vllm-native }, conc-list: [32, 40, 48, 56] }
       # TP=4 probe: half-node layout doubles per-GPU weight footprint
       # (~62 GB on MI355X's 288 GB HBM, plenty of headroom). Restrict to
       # cliff-region concurrencies on both offload modes so we can directly
       # compare TP=4 vs TP=8 at the same conc points.
       - { tp: 4, kv-offloading: none, conc-list: [16, 24, 32, 40] }
-      - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: native }, conc-list: [16, 24, 32, 40] }
+      - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: vllm-native }, conc-list: [16, 24, 32, 40] }
 
 kimik2.5-fp4-mi355x-atom:
   image: rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.4_202607091539
@@ -876,6 +879,7 @@ dsr1-fp8-mi355x-sglang-disagg:
   runner: mi355x-disagg
   precision: fp8
   framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
   kv-p2p-transfer: mori
   multinode: true
   disagg: true
@@ -1031,6 +1035,7 @@ dsr1-fp8-mi355x-sglang-disagg-mtp:
   runner: mi355x-disagg
   precision: fp8
   framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
   kv-p2p-transfer: mori
   multinode: true
   disagg: true
@@ -1186,6 +1191,7 @@ kimik2.5-fp4-mi355x-vllm-disagg:
   runner: mi355x-disagg
   precision: fp4
   framework: vllm-disagg
+  router: { name: vllm-router, version: "0.1.14" }
   kv-p2p-transfer: moriio
   multinode: true
   disagg: true
@@ -1239,6 +1245,7 @@ dsr1-fp4-mi355x-sglang-disagg:
   runner: mi355x-disagg
   precision: fp4
   framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
   kv-p2p-transfer: mori
   multinode: true
   disagg: true
@@ -1468,6 +1475,7 @@ dsr1-fp4-mi355x-sglang-disagg-1k1k-mtp:
   runner: mi355x-disagg
   precision: fp4
   framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
   kv-p2p-transfer: mori
   multinode: true
   disagg: true
@@ -1579,6 +1587,7 @@ dsr1-fp4-mi355x-sglang-disagg-8k1k-mtp:
   runner: mi355x-disagg
   precision: fp4
   framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
   kv-p2p-transfer: mori
   multinode: true
   disagg: true
@@ -1748,6 +1757,7 @@ dsv4-fp4-mi355x-sglang-disagg:
   runner: mi355x-disagg
   precision: fp4
   framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
   kv-p2p-transfer: mori
   multinode: true
   disagg: true
@@ -2093,8 +2103,8 @@ dsv4-fp4-mi355x-vllm-agentic:
     - dram-utilization: 0.60
       search-space:
       - { tp: 8, ep: 1, dp-attn: false, kv-offloading: none, conc-list: [1, 4, 8, 16, 32, 40, 48] }
-      - { tp: 8, ep: 1, dp-attn: true, kv-offloading: none, conc-list: [64, 72] }
-      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: lmcache }, conc-list: [32, 40, 48] }
+      - { tp: 8, ep: 1, dp-attn: true, kv-offloading: none, conc-list: [64, 72], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 1, dp-attn: false, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "9229067cec0b3a63bb8a39368d101db7ac0bc3c1" }, conc-list: [32, 40, 48] }
 
 # DSv4-Pro FP4 on MI355X via SGLang. Uses a rocm720 mi35x image built off the
 # amd/deepseek_v4 branch in sgl-project/sglang; the SHA is encoded in the
@@ -2108,6 +2118,7 @@ dsr1-fp4-mi355x-sglang-disagg-mtp:
   runner: mi355x-disagg
   precision: fp4
   framework: sglang-disagg
+  router: { name: sglang-router, version: "0.3.2" }
   kv-p2p-transfer: mori
   multinode: true
   disagg: true
@@ -2359,6 +2370,7 @@ dsv4-fp4-mi355x-atom-disagg:
   runner: mi355x
   precision: fp4
   framework: atom-disagg
+  router: { name: atomesh, version: "087b82d9c1f630e79149ba37e6213257ec9a76f8" }
   kv-p2p-transfer: mooncake
   multinode: true
   disagg: true
@@ -2479,6 +2491,7 @@ minimaxm3-fp4-mi355x-vllm-disagg:
   runner: mi355x-disagg
   precision: fp4
   framework: vllm-disagg
+  router: { name: vllm-router, version: "0.1.14" }
   kv-p2p-transfer: moriio
   multinode: true
   disagg: true
@@ -2668,6 +2681,7 @@ minimaxm3-fp8-mi355x-atom-disagg:
   runner: mi355x-disagg
   precision: fp8
   framework: atom-disagg
+  router: { name: atomesh, version: "04b120d3040e0dd1a6915e88a3a3c1f588e6684a" }
   kv-p2p-transfer: mooncake
   multinode: true
   disagg: true
@@ -2739,6 +2753,7 @@ minimaxm3-fp4-mi355x-atom-disagg:
   runner: mi355x-disagg
   precision: fp4
   framework: atom-disagg
+  router: { name: atomesh, version: "04b120d3040e0dd1a6915e88a3a3c1f588e6684a" }
   kv-p2p-transfer: mooncake
   multinode: true
   disagg: true
@@ -2943,6 +2958,7 @@ minimaxm3-fp8-mi355x-vllm-disagg:
   runner: mi355x-disagg
   precision: fp8
   framework: vllm-disagg
+  router: { name: vllm-router, version: "0.1.14" }
   kv-p2p-transfer: moriio
   multinode: true
   disagg: true
@@ -3050,8 +3066,8 @@ minimaxm3-fp8-mi300x-vllm-agentic:
       search-space:
       - { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
       - { tp: 8, ep: 8, kv-offloading: none, conc-list: [2, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
-      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
-      - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
+      - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 16, 18, 20] }
 
 minimaxm3-fp8-mi325x-vllm-agentic:
   image: vllm/vllm-openai-rocm:nightly-04c2a8deac44fdb1ca3e2b5ec3e6bf16f3f6a914
@@ -3074,7 +3090,7 @@ minimaxm3-fp8-mi325x-vllm-agentic:
       - { tp: 4, ep: 4, kv-offloading: none, conc-list: [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 14, 16] }
       - { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 32] }
       - { tp: 8, ep: 8, kv-offloading: none, conc-list: [2, 4, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 32] }
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 24, 28, 32, 36, 40, 44, 48, 52, 56, 60, 64, 72, 80] }
-      - { tp: 4, ep: 4, kv-offloading: dram, kv-offload-backend: { name: native }, conc-list: [3, 4, 5, 6, 7, 8, 9, 10, 12, 14, 16] }
-      - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: native }, conc-list: [10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 32] }
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: native }, conc-list: [24, 32, 36, 40, 44, 48, 52, 56, 60, 64, 72, 80, 96] }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, conc-list: [16, 24, 28, 32, 36, 40, 44, 48, 52, 56, 60, 64, 72, 80], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 4, ep: 4, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [3, 4, 5, 6, 7, 8, 9, 10, 12, 14, 16] }
+      - { tp: 8, ep: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 32] }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [24, 32, 36, 40, 44, 48, 52, 56, 60, 64, 72, 80, 96], router: { name: vllm-router, version: "0.1.14" } }