From 055dfc2ddcd77841d127ef29e77adb04c7802ddd Mon Sep 17 00:00:00 2001
From: Samuel Shen <slshen@tensormesh.ai>
Date: Wed, 12 Aug 2026 19:58:31 +0000
Subject: [PATCH 01/12] kimik3-fp4-mi355x-vllm-agentic-mtp: add LMCache DRAM
 KV-offload arm

Add an lmcache kv-offload-backend point at TP8 conc 10 on top of the
existing DSpark MTP serving stack, mirroring the vllm-simple offload arm
for a direct backend comparison. The benchmark script gains an lmcache
case arm that installs the LMCache 0.5.4rc1 ROCm wheel (torch/ROCm stack
untouched), starts one MP server per the Kimi-K3 recipe (chunk size 768
= K3 unified block size at 8 GPUs, --separate-object-groups for the
hybrid KDA/MLA two-group KV layout, --enable-extra-logging,
--max-cpu-workers 8 --max-gpu-workers 1), and wires vLLM to it via
LMCacheMPConnector.
---
 .../agentic/kimik3_fp4_mi355x_mtp.sh          | 75 +++++++++++++++++++
 configs/amd-master.yaml                       |  4 +
 2 files changed, 79 insertions(+)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index ec2eaf3c3..0f1355dfb 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -109,12 +109,14 @@ SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
 
 SERVER_PID=""
+LMCACHE_PID=""
 
 cleanup_agentic_services() {
     local exit_code=$?
     trap - EXIT INT TERM
     set +e
     stop_background_process_tree "$SERVER_PID" "vLLM server" 60
+    stop_background_process_tree "$LMCACHE_PID" "LMCache server"
     exit "$exit_code"
 }
 trap cleanup_agentic_services EXIT
@@ -143,6 +145,79 @@ case "${KV_OFFLOAD_BACKEND:-}" in
     )
     echo "SimpleCPUOffloadConnector: ${CPU_BYTES_PER_RANK} B/rank x ${TP} ranks, lazy_offload=$SIMPLE_LAZY_OFFLOAD"
     ;;
+  lmcache)
+    require_agentic_kv_offload_backend "$KV_OFFLOAD_BACKEND"
+
+    # Keep the image's tested torch/ROCm stack and install only LMCache's
+    # missing runtime dependencies, same as the MiniMax-M3 lmcache arm.
+    LMCACHE_VERSION="0.5.4rc1"
+    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/v${LMCACHE_VERSION}-rocm"
+    agentic_pip_install --quiet --no-cache-dir --no-deps \
+        "sortedcontainers==2.4.0" \
+        "opentelemetry-exporter-prometheus==0.61b0" \
+        "cupy-rocm-7-0==14.1.1" \
+        "lmcache==${LMCACHE_VERSION}" --find-links "$LMCACHE_ROCM_INDEX"
+    python3 -c \
+        "import cupy; import lmcache.integration.vllm.lmcache_mp_connector; import opentelemetry.exporter.prometheus" \
+        >/dev/null
+
+    # One MP server for the node, per the Kimi-K3 recipe
+    # (docs.lmcache.ai/recipes/kimi_k3.html). --chunk-size must equal the
+    # K3 unified block size N=768 at 8 GPUs, and the hybrid KDA/MLA layout
+    # (two KV-cache groups under MTP) requires one object group per
+    # sliding-window size: --separate-object-groups.
+    LMCACHE_PORT=6555
+    LMCACHE_HTTP_PORT=8090
+    LMCACHE_LOG="$RESULT_DIR/lmcache_server.log"
+
+    # The whole generated node-DRAM budget backs the single server's L1,
+    # which lives in /dev/shm; fail early if it cannot fit.
+    LMCACHE_L1_SIZE_GB="$TOTAL_CPU_DRAM_GB"
+    SHM_FREE_GB=$(df -BG --output=avail /dev/shm 2>/dev/null | tail -1 | tr -dc '0-9')
+    if [ -n "$SHM_FREE_GB" ] && [ "$SHM_FREE_GB" -gt 0 ]; then
+        SHM_CAP_GB=$((SHM_FREE_GB * 90 / 100))
+        if [ "$LMCACHE_L1_SIZE_GB" -gt "$SHM_CAP_GB" ]; then
+            echo "Error: LMCache L1 ${LMCACHE_L1_SIZE_GB} GB exceeds 90% of free /dev/shm (${SHM_CAP_GB} GB)." >&2
+            exit 1
+        fi
+    fi
+
+    LMCACHE_CMD=(
+        lmcache server
+        --host 127.0.0.1
+        --port "$LMCACHE_PORT"
+        --http-host 127.0.0.1
+        --http-port "$LMCACHE_HTTP_PORT"
+        --l1-size-gb "$LMCACHE_L1_SIZE_GB"
+        --l1-init-size-gb 10
+        --chunk-size 768
+        --separate-object-groups
+        --enable-extra-logging
+        --max-cpu-workers 8
+        --max-gpu-workers 1
+        --eviction-policy LRU
+    )
+    append_command "$RESULT_DIR/lmcache_command.txt" "${LMCACHE_CMD[@]}"
+    "${LMCACHE_CMD[@]}" > "$LMCACHE_LOG" 2>&1 &
+    LMCACHE_PID=$!
+    wait_for_ready \
+        --endpoint "http://127.0.0.1:${LMCACHE_HTTP_PORT}/healthcheck" \
+        --log "$LMCACHE_LOG" \
+        --pid "$LMCACHE_PID" \
+        --sleep-interval 1 \
+        --timeout 600
+
+    # 100k-330k-token agentic prefixes make single retrieves large; use the
+    # same MQ timeout headroom as the MiniMax-M3 arm.
+    OFFLOAD_ARGS=(
+        --kv-transfer-config
+        "{\"kv_connector\":\"LMCacheMPConnector\",\"kv_connector_module_path\":\"lmcache.integration.vllm.lmcache_mp_connector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"lmcache.mp.port\":$LMCACHE_PORT,\"lmcache.mp.mq_timeout\":6000.0}}"
+    )
+    ;;
+  *)
+    echo "Error: unsupported KV_OFFLOAD_BACKEND='$KV_OFFLOAD_BACKEND' (expected vllm-simple or lmcache)" >&2
+    exit 1
+    ;;
 esac
 fi
 
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index c94029484..3aa757c0a 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -650,6 +650,10 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       search-space:
       - { tp: 8, kv-offloading: none, conc-list: [1, 4, 8] , spec-decoding: mtp}
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [10], spec-decoding: mtp }
+      # LMCache MP-server DRAM offload on top of the same DSpark MTP serving
+      # stack, at the same concurrency as the vllm-simple arm for a direct
+      # offload-backend comparison.
+      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.4rc1" }, conc-list: [10], spec-decoding: mtp }
 
 dsr1-fp4-mi355x-sglang-disagg:
   image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519

From 328836b6f1aea4da24ac60651cdf4b7614920a25 Mon Sep 17 00:00:00 2001
From: Samuel Shen <slshen@tensormesh.ai>
Date: Wed, 12 Aug 2026 19:59:05 +0000
Subject: [PATCH 02/12] perf-changelog: select
 kimik3-fp4-mi355x-vllm-agentic-mtp LMCache arm

---
 perf-changelog.yaml | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index f118b99aa..da67c0138 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5909,3 +5909,12 @@
     - "Increase MTP speculative steps from 3 to 5 (num-draft-tokens 4→6) and raise SGLANG_SIMULATE_ACC_LEN from 2.99 to 3.61 to reflect higher measured acceptance rate, targeting ~15% throughput improvement"
     - "Replace the TP8/EP8 low-concurrency arm (conc [1,2,4]) hicache offload with kv-offloading: none to reduce per-request latency at low load; conc [1,2,4] are now tested on both tp=4+hicache and tp=8+no-offload so SemiAnalysis can select the Pareto-optimal point per concurrency"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2570
+
+- config-keys:
+    - kimik3-fp4-mi355x-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add an LMCache 0.5.4rc1 DRAM KV-offload arm at TP8 conc 10 on top of the DSpark MTP stack, mirroring the vllm-simple offload point for a direct backend comparison."
+    - "Run one LMCache MP server per node with chunk size 768 (the K3 unified block size at 8 GPUs) and --separate-object-groups for the hybrid KDA/MLA two-group KV layout."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2583

From 4584d37ab75aa38fd8af0060b3d6a429a6b82154 Mon Sep 17 00:00:00 2001
From: Samuel Shen <slshen@tensormesh.ai>
Date: Wed, 12 Aug 2026 21:20:39 +0000
Subject: [PATCH 03/12] Move the LMCache arm to a dedicated config key at conc
 4/8/16

A separate kimik3-fp4-mi355x-vllm-agentic-mtp-lmcache key lets the
changelog select only the LMCache points instead of re-running the
resident and vllm-simple arms of the base key. The base key returns to
its upstream shape.
---
 configs/amd-master.yaml | 22 ++++++++++++++++++----
 perf-changelog.yaml     |  4 ++--
 2 files changed, 20 insertions(+), 6 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 3aa757c0a..5afd30620 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -650,10 +650,24 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       search-space:
       - { tp: 8, kv-offloading: none, conc-list: [1, 4, 8] , spec-decoding: mtp}
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [10], spec-decoding: mtp }
-      # LMCache MP-server DRAM offload on top of the same DSpark MTP serving
-      # stack, at the same concurrency as the vllm-simple arm for a direct
-      # offload-backend comparison.
-      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.4rc1" }, conc-list: [10], spec-decoding: mtp }
+
+# LMCache MP-server DRAM offload on top of the same DSpark MTP serving stack as
+# kimik3-fp4-mi355x-vllm-agentic-mtp (same image, script, and topology). A
+# dedicated key so LMCache points can be selected and swept without re-running
+# the resident and vllm-simple arms of the base key.
+kimik3-fp4-mi355x-vllm-agentic-mtp-lmcache:
+  image: vllm/vllm-openai-rocm:nightly-cb8104839c141609d99f1254459ef3a4f1bd4263
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:mi355x-amds
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.50
+      search-space:
+      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.4rc1" }, conc-list: [4, 8, 16], spec-decoding: mtp }
 
 dsr1-fp4-mi355x-sglang-disagg:
   image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index da67c0138..b5e9172a3 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5911,10 +5911,10 @@
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2570
 
 - config-keys:
-    - kimik3-fp4-mi355x-vllm-agentic-mtp
+    - kimik3-fp4-mi355x-vllm-agentic-mtp-lmcache
   scenario-type:
     - agentic-coding
   description:
-    - "Add an LMCache 0.5.4rc1 DRAM KV-offload arm at TP8 conc 10 on top of the DSpark MTP stack, mirroring the vllm-simple offload point for a direct backend comparison."
+    - "Add a dedicated LMCache 0.5.4rc1 DRAM KV-offload key at TP8 conc 4/8/16 on top of the unchanged kimik3-fp4-mi355x-vllm-agentic-mtp DSpark MTP stack."
     - "Run one LMCache MP server per node with chunk size 768 (the K3 unified block size at 8 GPUs) and --separate-object-groups for the hybrid KDA/MLA two-group KV layout."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2583

From aba14d1b1f300be44daaf97ecf0693e4f5dcdd1e Mon Sep 17 00:00:00 2001
From: Samuel Shen <slshen@tensormesh.ai>
Date: Wed, 12 Aug 2026 21:57:33 +0000
Subject: [PATCH 04/12] Lower the lmcache key's dram-utilization to fit the
 shm-backed L1

The LMCache MP server's L1 lives in /dev/shm and the script rejects
budgets above 90% of free shm. mi355x-amds nodes mount ~1.5 TB of shm
(cap ~1360 GB), so 0.50's 1499 GB budget failed the check in run
31644286169. 0.40 generates ~1199 GB, which fits with margin.
---
 configs/amd-master.yaml | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 5afd30620..96d3b2986 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -665,7 +665,11 @@ kimik3-fp4-mi355x-vllm-agentic-mtp-lmcache:
   multinode: false
   scenarios:
     agentic-coding:
-    - dram-utilization: 0.50
+    # 0.40, not the base key's 0.50: the LMCache L1 is /dev/shm-backed and the
+    # script refuses budgets above 90% of free shm. mi355x-amds nodes mount
+    # ~1.5 TB of shm (cap ~1360 GB), so 0.50's 1499 GB budget fails the check
+    # (run 31644286169); 0.40 -> ~1199 GB fits with margin.
+    - dram-utilization: 0.40
       search-space:
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.4rc1" }, conc-list: [4, 8, 16], spec-decoding: mtp }
 

From 6b95b7b830beec89a2417cc2d2478cc5a233443e Mon Sep 17 00:00:00 2001
From: Samuel Shen <slshen@tensormesh.ai>
Date: Wed, 12 Aug 2026 22:10:05 +0000
Subject: [PATCH 05/12] Match the LMCache chunk size to this stack's 1536-token
 block size

vLLM sizes the K3 unified attention block to 1536 tokens on the MI355X
fp8-KV TRITON_MLA path (attention page >= mamba page), and the MP
connector asserts chunk %% block == 0, so the recipe's CUDA-path 768
fails engine init (run 31644990546).
---
 .../single_node/agentic/kimik3_fp4_mi355x_mtp.sh     | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index 0f1355dfb..eba514b26 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -163,9 +163,13 @@ case "${KV_OFFLOAD_BACKEND:-}" in
 
     # One MP server for the node, per the Kimi-K3 recipe
     # (docs.lmcache.ai/recipes/kimi_k3.html). --chunk-size must equal the
-    # K3 unified block size N=768 at 8 GPUs, and the hybrid KDA/MLA layout
-    # (two KV-cache groups under MTP) requires one object group per
-    # sliding-window size: --separate-object-groups.
+    # unified attention block size, which THIS stack (fp8 KV, TP8, no
+    # mamba-cache-mode align) sets to 1536 -- "Setting attention block size
+    # to 1536 tokens to ensure that attention page size is >= mamba page
+    # size" (run 31644990546); the recipe's 768 is the CUDA-path value and
+    # fails the connector's chunk %% block == 0 assert here. The hybrid
+    # KDA/MLA layout (two KV-cache groups under MTP) requires one object
+    # group per sliding-window size: --separate-object-groups.
     LMCACHE_PORT=6555
     LMCACHE_HTTP_PORT=8090
     LMCACHE_LOG="$RESULT_DIR/lmcache_server.log"
@@ -190,7 +194,7 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --http-port "$LMCACHE_HTTP_PORT"
         --l1-size-gb "$LMCACHE_L1_SIZE_GB"
         --l1-init-size-gb 10
-        --chunk-size 768
+        --chunk-size 1536
         --separate-object-groups
         --enable-extra-logging
         --max-cpu-workers 8

From a8cda1afc0cfa50c85515fb7632dd5b7f027662c Mon Sep 17 00:00:00 2001
From: Samuel Shen <slshen@tensormesh.ai>
Date: Wed, 12 Aug 2026 22:21:13 +0000
Subject: [PATCH 06/12] Raise the LMCache chunk size to 3072 for the KDA state
 group

The connector requires the chunk to be a multiple of every engine KV
group's tokens_per_block. On this stack the hybrid layout registers
attention groups at 1536 and a KDA state group at 3072 (run
31645828378), so 1536 fails registration; 3072 is the minimum valid
chunk.
---
 .../agentic/kimik3_fp4_mi355x_mtp.sh          | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index eba514b26..b9e58ee03 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -162,14 +162,15 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         >/dev/null
 
     # One MP server for the node, per the Kimi-K3 recipe
-    # (docs.lmcache.ai/recipes/kimi_k3.html). --chunk-size must equal the
-    # unified attention block size, which THIS stack (fp8 KV, TP8, no
-    # mamba-cache-mode align) sets to 1536 -- "Setting attention block size
-    # to 1536 tokens to ensure that attention page size is >= mamba page
-    # size" (run 31644990546); the recipe's 768 is the CUDA-path value and
-    # fails the connector's chunk %% block == 0 assert here. The hybrid
-    # KDA/MLA layout (two KV-cache groups under MTP) requires one object
-    # group per sliding-window size: --separate-object-groups.
+    # (docs.lmcache.ai/recipes/kimi_k3.html), with --chunk-size sized for
+    # THIS stack rather than the recipe's CUDA-path 768: the connector
+    # requires the chunk to be a multiple of every engine KV group's
+    # tokens_per_block, and the hybrid KDA/MLA layout here registers
+    # attention groups at 1536 ("Setting attention block size to 1536",
+    # run 31644990546) plus a KDA state group at 3072 (run 31645828378),
+    # so 3072 is the minimum valid chunk. The multi-group layout also
+    # requires one object group per sliding-window size:
+    # --separate-object-groups.
     LMCACHE_PORT=6555
     LMCACHE_HTTP_PORT=8090
     LMCACHE_LOG="$RESULT_DIR/lmcache_server.log"
@@ -194,7 +195,7 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --http-port "$LMCACHE_HTTP_PORT"
         --l1-size-gb "$LMCACHE_L1_SIZE_GB"
         --l1-init-size-gb 10
-        --chunk-size 1536
+        --chunk-size 3072
         --separate-object-groups
         --enable-extra-logging
         --max-cpu-workers 8

From afee680555d81c8783ee45cfd44047f0a434fadf Mon Sep 17 00:00:00 2001
From: Samuel Shen <slshen@tensormesh.ai>
Date: Wed, 12 Aug 2026 22:44:35 +0000
Subject: [PATCH 07/12] Pin the LMCache MP server to the lmcache_driven
 transfer path

Auto mode loads both transfer paths; pin server-driven STORE/RETRIEVE
(as the MiniMax-M3 arm does) so the benchmark measures one
deterministic path. The L1 stays shm-backed either way, so the /dev/shm
capacity check still applies.
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index b9e58ee03..aa2a5ae89 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -201,6 +201,12 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --max-cpu-workers 8
         --max-gpu-workers 1
         --eviction-policy LRU
+        # Pin the server-driven STORE/RETRIEVE path (same as the MiniMax-M3
+        # arm) so the benchmark measures one deterministic transfer path
+        # instead of the auto-mode pair. The L1 stays /dev/shm-backed either
+        # way (shm_name defaults on), which is why the capacity check above
+        # applies in this mode too.
+        --supported-transfer-mode lmcache_driven
     )
     append_command "$RESULT_DIR/lmcache_command.txt" "${LMCACHE_CMD[@]}"
     "${LMCACHE_CMD[@]}" > "$LMCACHE_LOG" 2>&1 &

From 20b4fdae81f816d7b9741f2c03d66b801f53b7dc Mon Sep 17 00:00:00 2001
From: Samuel Shen <slshen@tensormesh.ai>
Date: Wed, 12 Aug 2026 23:21:43 +0000
Subject: [PATCH 08/12] Hold LMCache L1 read locks for the job duration

The default 300s read-lock TTL expires under a single GPU worker
serializing huge K3 transfers: run 31648224111 logged 57k
finish-read-on-non-read-locked-key warnings starting exactly at
warmup+300s, followed by a GPU illegal-access crash mid-profile.
Match the MiniMax-M3 arm's 7200s read TTL.
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index aa2a5ae89..55944729b 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -195,6 +195,13 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --http-port "$LMCACHE_HTTP_PORT"
         --l1-size-gb "$LMCACHE_L1_SIZE_GB"
         --l1-init-size-gb 10
+        # Read locks default to a 300s TTL, and with a single GPU worker
+        # serializing 100k-300k-token K3 transfers, queued reads routinely
+        # outlive it: run 31648224111 logged 57k "finish read on
+        # non-read-locked key" warnings starting exactly warmup+300s before
+        # a GPU illegal-access crash. Outlast the job like the MiniMax-M3
+        # arm does.
+        --l1-read-ttl-seconds 7200
         --chunk-size 3072
         --separate-object-groups
         --enable-extra-logging

From 3c1908b33c48ba9a37a6efd7e031a672b5a7ae33 Mon Sep 17 00:00:00 2001
From: Samuel Shen <slshen@tensormesh.ai>
Date: Wed, 12 Aug 2026 23:36:49 +0000
Subject: [PATCH 09/12] Revert "Hold LMCache L1 read locks for the job
 duration"

This reverts commit 20b4fdae81f816d7b9741f2c03d66b801f53b7dc.
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 7 -------
 1 file changed, 7 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index 55944729b..aa2a5ae89 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -195,13 +195,6 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --http-port "$LMCACHE_HTTP_PORT"
         --l1-size-gb "$LMCACHE_L1_SIZE_GB"
         --l1-init-size-gb 10
-        # Read locks default to a 300s TTL, and with a single GPU worker
-        # serializing 100k-300k-token K3 transfers, queued reads routinely
-        # outlive it: run 31648224111 logged 57k "finish read on
-        # non-read-locked key" warnings starting exactly warmup+300s before
-        # a GPU illegal-access crash. Outlast the job like the MiniMax-M3
-        # arm does.
-        --l1-read-ttl-seconds 7200
         --chunk-size 3072
         --separate-object-groups
         --enable-extra-logging

From 0811fb0c6962785d07607a9b1cffb621fe01473c Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Thu, 13 Aug 2026 08:50:20 -0700
Subject: [PATCH 10/12] trigger lmcache 0.5.4rc2

Signed-off-by: ApostaC <yihua98@uchicago.edu>
---
 configs/amd-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 96d3b2986..4b95cd413 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -671,7 +671,7 @@ kimik3-fp4-mi355x-vllm-agentic-mtp-lmcache:
     # (run 31644286169); 0.40 -> ~1199 GB fits with margin.
     - dram-utilization: 0.40
       search-space:
-      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.4rc1" }, conc-list: [4, 8, 16], spec-decoding: mtp }
+      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.4rc2" }, conc-list: [4, 8, 16], spec-decoding: mtp }
 
 dsr1-fp4-mi355x-sglang-disagg:
   image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519

From ac6ad7d514a4a5346293e62a5378204d19e32f55 Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Thu, 13 Aug 2026 11:02:46 -0700
Subject: [PATCH 11/12] update lmcache configus

Signed-off-by: ApostaC <yihua98@uchicago.edu>
---
 .../agentic/kimik3_fp4_mi355x_mtp.sh          | 19 +++----------------
 configs/amd-master.yaml                       | 10 +++++-----
 perf-changelog.yaml                           |  4 ++--
 3 files changed, 10 insertions(+), 23 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index aa2a5ae89..0a3fd4318 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -150,7 +150,7 @@ case "${KV_OFFLOAD_BACKEND:-}" in
 
     # Keep the image's tested torch/ROCm stack and install only LMCache's
     # missing runtime dependencies, same as the MiniMax-M3 lmcache arm.
-    LMCACHE_VERSION="0.5.4rc1"
+    LMCACHE_VERSION="0.5.4rc2"
     LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/v${LMCACHE_VERSION}-rocm"
     agentic_pip_install --quiet --no-cache-dir --no-deps \
         "sortedcontainers==2.4.0" \
@@ -175,17 +175,7 @@ case "${KV_OFFLOAD_BACKEND:-}" in
     LMCACHE_HTTP_PORT=8090
     LMCACHE_LOG="$RESULT_DIR/lmcache_server.log"
 
-    # The whole generated node-DRAM budget backs the single server's L1,
-    # which lives in /dev/shm; fail early if it cannot fit.
     LMCACHE_L1_SIZE_GB="$TOTAL_CPU_DRAM_GB"
-    SHM_FREE_GB=$(df -BG --output=avail /dev/shm 2>/dev/null | tail -1 | tr -dc '0-9')
-    if [ -n "$SHM_FREE_GB" ] && [ "$SHM_FREE_GB" -gt 0 ]; then
-        SHM_CAP_GB=$((SHM_FREE_GB * 90 / 100))
-        if [ "$LMCACHE_L1_SIZE_GB" -gt "$SHM_CAP_GB" ]; then
-            echo "Error: LMCache L1 ${LMCACHE_L1_SIZE_GB} GB exceeds 90% of free /dev/shm (${SHM_CAP_GB} GB)." >&2
-            exit 1
-        fi
-    fi
 
     LMCACHE_CMD=(
         lmcache server
@@ -198,15 +188,12 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --chunk-size 3072
         --separate-object-groups
         --enable-extra-logging
+        --extra-logging-interval 30
         --max-cpu-workers 8
         --max-gpu-workers 1
         --eviction-policy LRU
-        # Pin the server-driven STORE/RETRIEVE path (same as the MiniMax-M3
-        # arm) so the benchmark measures one deterministic transfer path
-        # instead of the auto-mode pair. The L1 stays /dev/shm-backed either
-        # way (shm_name defaults on), which is why the capacity check above
-        # applies in this mode too.
         --supported-transfer-mode lmcache_driven
+        --shm-name ""
     )
     append_command "$RESULT_DIR/lmcache_command.txt" "${LMCACHE_CMD[@]}"
     "${LMCACHE_CMD[@]}" > "$LMCACHE_LOG" 2>&1 &
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 4b95cd413..64b156590 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -665,11 +665,11 @@ kimik3-fp4-mi355x-vllm-agentic-mtp-lmcache:
   multinode: false
   scenarios:
     agentic-coding:
-    # 0.40, not the base key's 0.50: the LMCache L1 is /dev/shm-backed and the
-    # script refuses budgets above 90% of free shm. mi355x-amds nodes mount
-    # ~1.5 TB of shm (cap ~1360 GB), so 0.50's 1499 GB budget fails the check
-    # (run 31644286169); 0.40 -> ~1199 GB fits with margin.
-    - dram-utilization: 0.40
+    # 0.50 matches the base key: the LMCache server runs with --shm-name ""
+    # so its L1 lives in regular process memory instead of /dev/shm, and the
+    # budget is no longer capped by the ~1.5 TB shm mount (which forced 0.40
+    # before, run 31644286169).
+    - dram-utilization: 0.50
       search-space:
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.4rc2" }, conc-list: [4, 8, 16], spec-decoding: mtp }
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index d8619dec9..9b0682f01 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5924,6 +5924,6 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add a dedicated LMCache 0.5.4rc1 DRAM KV-offload key at TP8 conc 4/8/16 on top of the unchanged kimik3-fp4-mi355x-vllm-agentic-mtp DSpark MTP stack."
-    - "Run one LMCache MP server per node with chunk size 768 (the K3 unified block size at 8 GPUs) and --separate-object-groups for the hybrid KDA/MLA two-group KV layout."
+    - "Add a dedicated LMCache 0.5.4rc2 DRAM KV-offload key at TP8 conc 4/8/16 on top of the unchanged kimik3-fp4-mi355x-vllm-agentic-mtp DSpark MTP stack, with the version pinned in the master config and consumed by the script via KV_OFFLOAD_BACKEND_METADATA."
+    - "Run one LMCache MP server per node with chunk size 3072 (the minimum multiple of the hybrid KDA/MLA group block sizes) and --separate-object-groups, keeping the L1 in process memory (--shm-name \"\") so the DRAM budget is not capped by /dev/shm."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2583

From e890a2b44a770a2cfec61a9b025fbad6fcbe8e69 Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Thu, 13 Aug 2026 11:25:39 -0700
Subject: [PATCH 12/12] update conc

Signed-off-by: ApostaC <yihua98@uchicago.edu>
---
 configs/amd-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 64b156590..44075c764 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -671,7 +671,7 @@ kimik3-fp4-mi355x-vllm-agentic-mtp-lmcache:
     # before, run 31644286169).
     - dram-utilization: 0.50
       search-space:
-      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.4rc2" }, conc-list: [4, 8, 16], spec-decoding: mtp }
+      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.4rc2" }, conc-list: [4, 8, 10, 12], spec-decoding: mtp }
 
 dsr1-fp4-mi355x-sglang-disagg:
   image: lmsysorg/sglang-rocm:v0.5.12-rocm720-mi35x-20260519