From 2b68f986791c82cd978f25c8dda14da53f4a0b9d Mon Sep 17 00:00:00 2001
From: hyukjlee <hyukjlee@amd.com>
Date: Wed, 2 Sep 2026 18:24:07 +0900
Subject: [PATCH 1/6] perf(agentx): refresh K3 MI355X LMCache curve

Use the merged Kimi-K3 recipe with vLLM nightly 7c5dc571 and LMCache 0.5.5rc3 for C1, C8, C14, and TP8/DCP8 C40. Keep the validated C40 profile with GPU memory utilization reduced to 0.88.

Assisted-by: OpenAI Codex <codex@openai.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x.sh  |  5 ++
 .../agentic/kimik3_fp4_mi355x_mtp.sh          | 84 ++++++++++++-------
 configs/amd-master.yaml                       |  6 +-
 perf-changelog.yaml                           | 10 +++
 4 files changed, 71 insertions(+), 34 deletions(-)
 create mode 100755 benchmarks/single_node/agentic/kimik3_fp4_mi355x.sh

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x.sh
new file mode 100755
index 0000000000..eff8ce9408
--- /dev/null
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x.sh
@@ -0,0 +1,5 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+export SPEC_DECODING=none
+exec bash "$(dirname "$0")/kimik3_fp4_mi355x_mtp.sh"
diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index a8f2c42c1c..8e954240c7 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -149,25 +149,18 @@ case "${KV_OFFLOAD_BACKEND:-}" in
       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="0.5.5.dev60+rocm7.2"
-    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/nightly-rocm"
+    # Keep the stock image's torch/ROCm stack and install the LMCache runtime
+    # dependencies used by the validated reference run.
+    LMCACHE_VERSION="0.5.5rc3+rocm7.2"
+    export KV_OFFLOAD_BACKEND_METADATA="{\"name\":\"lmcache\",\"version\":\"${LMCACHE_VERSION}\"}"
+    LMCACHE_RELEASE="v0.5.5rc3"
+    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/${LMCACHE_RELEASE}-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"
 
-    # LMCache 0.5.5's transfer-channel layer eagerly imports the Mooncake
-    # backend (mooncake_te_impl.py -> `from mooncake.engine import
-    # TransferEngine`), whose native .so resolves all of its DT_NEEDED libs at
-    # import. The vLLM ROCm image ships none of them, so the import sanity
-    # check below (and the LMCache server) would otherwise fail with
-    # "ImportError: lib*.so: cannot open shared object file" (first libglog,
-    # then libjsoncpp, ...). Provision Mooncake's full runtime lib set from the
-    # distro before importing. apt-get install is idempotent, so run it
-    # whenever any of the libs is still missing rather than gating on one.
     LMCACHE_NATIVE_LIBS=(libglog.so.0 libjsoncpp.so.25 libibverbs.so.1 librdmacm.so.1 libnuma.so.1)
     for lib in "${LMCACHE_NATIVE_LIBS[@]}"; do
         if ! ldconfig -p | grep -q "$lib"; then
@@ -178,17 +171,16 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         fi
     done
     python3 -c \
-        "import cupy; import lmcache.integration.vllm.lmcache_mp_connector; import opentelemetry.exporter.prometheus" \
-        >/dev/null
+        "import cupy; import opentelemetry.exporter.prometheus; from lmcache.v1.multiprocess.http_server import run_http_server"
 
     # One MP server for the node, per the Kimi-K3 recipe
     # (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
+    # tokens_per_block. The hybrid KDA/MLA layout registers attention groups
+    # at 1536 tokens and a KDA state group at 3072. Use 12288 for every point
+    # so it is divisible by both group sizes and matches the tested LMCache
+    # scheduler geometry. The multi-group layout also
     # requires one object group per sliding-window size:
     # --separate-object-groups.
     LMCACHE_PORT=6555
@@ -205,7 +197,7 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --http-port "$LMCACHE_HTTP_PORT"
         --l1-size-gb "$LMCACHE_L1_SIZE_GB"
         --l1-init-size-gb 10
-        --chunk-size 3072
+        --chunk-size 12288
         --separate-object-groups
         --enable-extra-logging
         --extra-logging-interval 30
@@ -248,20 +240,24 @@ if [ "$EP_SIZE" -gt 1 ]; then
 fi
 
 # ---- Speculative / Util------------------------------------------------------
-case "$CONC" in
-    # No KV offload; the working set fits in HBM.
-    1)
+case "${SPEC_DECODING:-mtp}:$CONC" in
+    mtp:1)
         SYNTHETIC_ACCEPT_LEN=3.75
         SPEC_NUM_TOKENS=6
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=16384
         ;;
-    2|4|8|10|12|14)
+    mtp:2|mtp:4|mtp:8|mtp:10|mtp:12|mtp:14)
         SYNTHETIC_ACCEPT_LEN=3.00
         SPEC_NUM_TOKENS=3
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
+    none:40)
+        SPEC_NUM_TOKENS=0
+        GPU_MEM_UTIL=0.88
+        MAX_NUM_BATCHED_TOKENS=16384
+        ;;
     *)
         SPEC_NUM_TOKENS=0
         GPU_MEM_UTIL=0.85
@@ -285,10 +281,24 @@ else
 fi
 
 # ---- HIP graph ------------------------------------------------------------
-MAX_NUM_SEQS=$((2 * CONC))
-MAX_CUDAGRAPH_CAPTURE_SIZE=$((MAX_NUM_SEQS * (1 + SPEC_NUM_TOKENS)))
-CUDAGRAPH_CAPTURE_SIZES="$(seq -s, 2 "$MAX_CUDAGRAPH_CAPTURE_SIZE")"
-COMPILATION_CONFIG_ARGS=(--compilation-config "{\"mode\":3,\"cudagraph_mode\":\"FULL_AND_PIECEWISE\",\"max_cudagraph_capture_size\":$MAX_CUDAGRAPH_CAPTURE_SIZE,\"custom_ops\":[\"+fused_rms_norm_gated\"],\"cudagraph_capture_sizes\":[$CUDAGRAPH_CAPTURE_SIZES]}")
+SERVER_STREAM_ARGS=()
+PREFIX_MATCH_ARGS=()
+ATTENTION_CONFIG='{"mla_prefill_backend":"ROCM_AITER_FA"}'
+COMPILATION_CUSTOM_OPS='["+fused_rms_norm_gated"]'
+if [ "${SPEC_DECODING:-mtp}:$CONC" = "none:40" ]; then
+    MAX_NUM_SEQS=80
+    MAX_CUDAGRAPH_CAPTURE_SIZE=4096
+    CUDAGRAPH_CAPTURE_SIZES="$(seq -s, 1 "$MAX_NUM_SEQS"),128,256,512,1024,2048,4096"
+    SERVER_STREAM_ARGS=(--stream-interval 10)
+    PREFIX_MATCH_ARGS=(--prefix-match-unit 128)
+    ATTENTION_CONFIG='{"mla_prefill_backend":"ROCM_AITER_FA","use_prefill_query_quantization":true}'
+    COMPILATION_CUSTOM_OPS='["+fused_rms_norm_gated","+quant_fp8","+grouped_topk","+sparse_attn_indexer","none"]'
+else
+    MAX_NUM_SEQS=$((2 * CONC))
+    MAX_CUDAGRAPH_CAPTURE_SIZE=$((MAX_NUM_SEQS * (1 + SPEC_NUM_TOKENS)))
+    CUDAGRAPH_CAPTURE_SIZES="$(seq -s, 2 "$MAX_CUDAGRAPH_CAPTURE_SIZE")"
+fi
+COMPILATION_CONFIG_ARGS=(--compilation-config "{\"mode\":3,\"cudagraph_mode\":\"FULL_AND_PIECEWISE\",\"max_cudagraph_capture_size\":$MAX_CUDAGRAPH_CAPTURE_SIZE,\"custom_ops\":$COMPILATION_CUSTOM_OPS,\"cudagraph_capture_sizes\":[$CUDAGRAPH_CAPTURE_SIZES]}")
 
 echo "Starting vllm server..."
 export PYTHONNOUSERSITE=1
@@ -305,8 +315,18 @@ fi
 CP_ARGS=()
 ATTN_BE_ARGS=()
 if [ "$DCP_SIZE" -gt 1 ]; then
-    CP_ARGS+=(--decode-context-parallel-size "$DCP_SIZE" --dcp-comm-backend a2a)
-    ATTN_BE_ARGS+=(--attention-backend TRITON_MLA)
+    CP_KV_CACHE_INTERLEAVE_SIZE=1
+    if [ "${KV_OFFLOAD_BACKEND:-}" = "lmcache" ]; then
+        CP_KV_CACHE_INTERLEAVE_SIZE=1536
+    fi
+    CP_ARGS+=(
+        --decode-context-parallel-size "$DCP_SIZE"
+        --dcp-comm-backend a2a
+        --cp-kv-cache-interleave-size "$CP_KV_CACHE_INTERLEAVE_SIZE"
+    )
+    ATTN_BE_ARGS+=(--attention-backend ROCM_AITER_MLA)
+    export VLLM_ALLOW_DCP_FULL_CUDAGRAPH=1
+    export PREFIX_CACHING_HASH_ALGO=sha256
 fi
 export VLLM_USE_DIRECT_DCP_A2A=0
 export VLLM_USE_DIRECT_DCP_Q_GATHER=0
@@ -329,10 +349,12 @@ VLLM_CMD=(
     --tool-call-parser kimi_k3
     --reasoning-parser kimi_k3
     --max-model-len 1048576
+    "${SERVER_STREAM_ARGS[@]}"
     --enable-prefix-caching
+    "${PREFIX_MATCH_ARGS[@]}"
     --kv-cache-dtype "fp8"
     --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS"
-    --attention-config '{"mla_prefill_backend":"ROCM_AITER_FA"}'
+    --attention-config "$ATTENTION_CONFIG"
     "${ATTN_BE_ARGS[@]}"
     "${COMPILATION_CONFIG_ARGS[@]}"
     "${SPEC_ARGS[@]}"
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 0d380154a0..db1324eb95 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -625,7 +625,7 @@ dsr1-fp8-mi355x-sglang-disagg-mtp:
           - "DECODE_MTP_SIZE=2"
 
 kimik3-fp4-mi355x-vllm-agentic-mtp:
-  image: vllm/vllm-openai-rocm:nightly-1dc464d42681d22f38caf1fdc1eb632dc4421c45
+  image: vllm/vllm-openai-rocm:nightly-7c5dc571cbd1064ecc8a9b1045637ff647aa22cb
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:mi355x-amds
@@ -636,8 +636,8 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.60
       search-space:
-      - { tp: 8, kv-offloading: none, conc-list: [1] , spec-decoding: mtp}
-      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev60+rocm7.2" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
+      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3+rocm7.2" }, conc-list: [1, 8, 14], spec-decoding: mtp }
+      - { tp: 8, dcp-size: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3+rocm7.2" }, conc-list: [40], spec-decoding: none }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 4eb0a2caa6..5dc48fa252 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6821,3 +6821,13 @@
   description:
     - "Refresh to collect TensorRT-LLM server metrics."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2774
+
+- config-keys:
+    - kimik3-fp4-mi355x-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Refresh the Kimi-K3 MI355X AgentX recipe from the merged PR #2787 baseline to vLLM ROCm nightly 7c5dc571 and LMCache 0.5.5rc3+rocm7.2."
+    - "Run C1, C8, and C14 with TP8 DSpark and LMCache DRAM KV offload; run C40 with TP8/DCP8 and no speculative decoding."
+    - "Use a 12288-token LMCache chunk and a 1536-token DCP KV-cache interleave; set C40 gpu-memory-utilization to 0.88 while retaining its 16K batched-token, max-num-seqs 80, and full CUDA-graph profile."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX

From ef281f09d429a16dfb18258b2d3dd63836de5fb7 Mon Sep 17 00:00:00 2001
From: hyukjlee <hyukjlee@amd.com>
Date: Wed, 2 Sep 2026 18:26:38 +0900
Subject: [PATCH 2/6] docs(changelog): link K3 LMCache refresh PR

Assisted-by: OpenAI Codex <codex@openai.com>
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 5dc48fa252..9504a4dc6d 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5595,7 +5595,7 @@
     - "Image ghcr.io/tile-ai/tilert:0.1.5 (tilert 0.1.5.post2 installed at container start); commands aligned to TileRT README Topology A -- NIXL KV transfer, --kv-cache-dtype fp8_ds_mla (prefill) <-> fp8 (decode), max-seq-len 202752; MTP speculative-config wired via spec-decoding=mtp"
     - "Topology: 1 prefill node (TP8) + 1 decode node (TP8), each 8xB200 exclusive; TileRT decode is bs=1 only so conc-list is a single point [1], ISL 1k/8k OSL 1k"
     - "Runner: launch_b200-dgxc.sh tilert early-return branch (zero impact on the dynamo path); tilert_utils/submit.sh issues two srun --ntasks=1, one per role, because prefill and decode need different container images; roles are dispatched by the TILERT_ROLE it exports, and torn down across nodes via a sentinel file on the shared /workspace"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2804
 
 - config-keys:
     - qwen3.5-fp8-b200-sglang-agentic-mtp

From d72e884db8d3cb1609156e95c9c551daac061d6e Mon Sep 17 00:00:00 2001
From: hyukjlee <hyukjlee@amd.com>
Date: Wed, 2 Sep 2026 18:27:37 +0900
Subject: [PATCH 3/6] fix(changelog): link the appended K3 entry

Assisted-by: OpenAI Codex <codex@openai.com>
---
 perf-changelog.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 9504a4dc6d..abb33dd1ed 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5595,7 +5595,7 @@
     - "Image ghcr.io/tile-ai/tilert:0.1.5 (tilert 0.1.5.post2 installed at container start); commands aligned to TileRT README Topology A -- NIXL KV transfer, --kv-cache-dtype fp8_ds_mla (prefill) <-> fp8 (decode), max-seq-len 202752; MTP speculative-config wired via spec-decoding=mtp"
     - "Topology: 1 prefill node (TP8) + 1 decode node (TP8), each 8xB200 exclusive; TileRT decode is bs=1 only so conc-list is a single point [1], ISL 1k/8k OSL 1k"
     - "Runner: launch_b200-dgxc.sh tilert early-return branch (zero impact on the dynamo path); tilert_utils/submit.sh issues two srun --ntasks=1, one per role, because prefill and decode need different container images; roles are dispatched by the TILERT_ROLE it exports, and torn down across nodes via a sentinel file on the shared /workspace"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2804
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
 
 - config-keys:
     - qwen3.5-fp8-b200-sglang-agentic-mtp
@@ -6830,4 +6830,4 @@
     - "Refresh the Kimi-K3 MI355X AgentX recipe from the merged PR #2787 baseline to vLLM ROCm nightly 7c5dc571 and LMCache 0.5.5rc3+rocm7.2."
     - "Run C1, C8, and C14 with TP8 DSpark and LMCache DRAM KV offload; run C40 with TP8/DCP8 and no speculative decoding."
     - "Use a 12288-token LMCache chunk and a 1536-token DCP KV-cache interleave; set C40 gpu-memory-utilization to 0.88 while retaining its 16K batched-token, max-num-seqs 80, and full CUDA-graph profile."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2804

From 6793742ac2460376532d6c81b724bfdb68a8d982 Mon Sep 17 00:00:00 2001
From: hyukjlee <hyukjlee@amd.com>
Date: Wed, 2 Sep 2026 19:17:09 +0900
Subject: [PATCH 4/6] perf(agentx): add K3 C44 and C48 LMCache points
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

为 Kimi-K3 MI355X LMCache 曲线添加 C44 和 C48，并沿用 C40 的 TP8/DCP8、无推测解码和 GMU 0.88 配置。

Assisted-by: OpenAI Codex <codex@openai.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_mtp.sh      | 11 +++++++----
 configs/amd-master.yaml                               |  2 +-
 perf-changelog.yaml                                   |  4 ++--
 3 files changed, 10 insertions(+), 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 8e954240c7..a442ba469e 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -253,7 +253,7 @@ case "${SPEC_DECODING:-mtp}:$CONC" in
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
-    none:40)
+    none:40|none:44|none:48)
         SPEC_NUM_TOKENS=0
         GPU_MEM_UTIL=0.88
         MAX_NUM_BATCHED_TOKENS=16384
@@ -285,7 +285,8 @@ SERVER_STREAM_ARGS=()
 PREFIX_MATCH_ARGS=()
 ATTENTION_CONFIG='{"mla_prefill_backend":"ROCM_AITER_FA"}'
 COMPILATION_CUSTOM_OPS='["+fused_rms_norm_gated"]'
-if [ "${SPEC_DECODING:-mtp}:$CONC" = "none:40" ]; then
+case "${SPEC_DECODING:-mtp}:$CONC" in
+    none:40|none:44|none:48)
     MAX_NUM_SEQS=80
     MAX_CUDAGRAPH_CAPTURE_SIZE=4096
     CUDAGRAPH_CAPTURE_SIZES="$(seq -s, 1 "$MAX_NUM_SEQS"),128,256,512,1024,2048,4096"
@@ -293,11 +294,13 @@ if [ "${SPEC_DECODING:-mtp}:$CONC" = "none:40" ]; then
     PREFIX_MATCH_ARGS=(--prefix-match-unit 128)
     ATTENTION_CONFIG='{"mla_prefill_backend":"ROCM_AITER_FA","use_prefill_query_quantization":true}'
     COMPILATION_CUSTOM_OPS='["+fused_rms_norm_gated","+quant_fp8","+grouped_topk","+sparse_attn_indexer","none"]'
-else
+    ;;
+    *)
     MAX_NUM_SEQS=$((2 * CONC))
     MAX_CUDAGRAPH_CAPTURE_SIZE=$((MAX_NUM_SEQS * (1 + SPEC_NUM_TOKENS)))
     CUDAGRAPH_CAPTURE_SIZES="$(seq -s, 2 "$MAX_CUDAGRAPH_CAPTURE_SIZE")"
-fi
+    ;;
+esac
 COMPILATION_CONFIG_ARGS=(--compilation-config "{\"mode\":3,\"cudagraph_mode\":\"FULL_AND_PIECEWISE\",\"max_cudagraph_capture_size\":$MAX_CUDAGRAPH_CAPTURE_SIZE,\"custom_ops\":$COMPILATION_CUSTOM_OPS,\"cudagraph_capture_sizes\":[$CUDAGRAPH_CAPTURE_SIZES]}")
 
 echo "Starting vllm server..."
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index db1324eb95..502572a78d 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -637,7 +637,7 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
     - dram-utilization: 0.60
       search-space:
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3+rocm7.2" }, conc-list: [1, 8, 14], spec-decoding: mtp }
-      - { tp: 8, dcp-size: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3+rocm7.2" }, conc-list: [40], spec-decoding: none }
+      - { tp: 8, dcp-size: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3+rocm7.2" }, conc-list: [40, 44, 48], spec-decoding: none }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index abb33dd1ed..dad68b8648 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6828,6 +6828,6 @@
     - agentic-coding
   description:
     - "Refresh the Kimi-K3 MI355X AgentX recipe from the merged PR #2787 baseline to vLLM ROCm nightly 7c5dc571 and LMCache 0.5.5rc3+rocm7.2."
-    - "Run C1, C8, and C14 with TP8 DSpark and LMCache DRAM KV offload; run C40 with TP8/DCP8 and no speculative decoding."
-    - "Use a 12288-token LMCache chunk and a 1536-token DCP KV-cache interleave; set C40 gpu-memory-utilization to 0.88 while retaining its 16K batched-token, max-num-seqs 80, and full CUDA-graph profile."
+    - "Run C1, C8, and C14 with TP8 DSpark and LMCache DRAM KV offload; run C40, C44, and C48 with TP8/DCP8 and no speculative decoding."
+    - "Use a 12288-token LMCache chunk and a 1536-token DCP KV-cache interleave; set C40/C44/C48 gpu-memory-utilization to 0.88 while retaining the 16K batched-token, max-num-seqs 80, and full CUDA-graph profile."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2804

From c7b685c4a08fddb0f60a49a0c2ee55abc41c590c Mon Sep 17 00:00:00 2001
From: hyukjlee <hyukjlee@amd.com>
Date: Wed, 2 Sep 2026 21:31:31 +0900
Subject: [PATCH 5/6] fix(agentx): tolerate K3 LMCache transfer stalls
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将 LMCache MP 心跳超时提高到 90 秒，并将 worker 回收超时提高到 300 秒，避免 C14 大批量传输期间的短暂阻塞触发错误恢复路径。

Assisted-by: OpenAI Codex <codex@openai.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 5 ++++-
 perf-changelog.yaml                                     | 1 +
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index a442ba469e..abbfc79558 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -186,6 +186,8 @@ case "${KV_OFFLOAD_BACKEND:-}" in
     LMCACHE_PORT=6555
     LMCACHE_HTTP_PORT=8090
     LMCACHE_LOG="$RESULT_DIR/lmcache_server.log"
+    LMCACHE_HEARTBEAT_INTERVAL=90.0
+    LMCACHE_WORKER_REAP_TIMEOUT=300
 
     LMCACHE_L1_SIZE_GB="$TOTAL_CPU_DRAM_GB"
 
@@ -205,6 +207,7 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --max-gpu-workers 1
         --eviction-policy LRU
         --supported-transfer-mode lmcache_driven
+        --worker-reap-timeout-seconds "$LMCACHE_WORKER_REAP_TIMEOUT"
         --shm-name ""
     )
     append_command "$RESULT_DIR/lmcache_command.txt" "${LMCACHE_CMD[@]}"
@@ -221,7 +224,7 @@ case "${KV_OFFLOAD_BACKEND:-}" in
     # same MQ timeout headroom as the MiniMax-M3 arm.
     OFFLOAD_ARGS=(
         --kv-transfer-config
-        "{\"kv_connector\":\"LMCacheMPConnector\",\"kv_connector_module_path\":\"lmcache.integration.vllm.lmcache_mp_connector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"lmcache.mp.port\":$LMCACHE_PORT,\"lmcache.mp.mq_timeout\":6000.0}}"
+        "{\"kv_connector\":\"LMCacheMPConnector\",\"kv_connector_module_path\":\"lmcache.integration.vllm.lmcache_mp_connector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"lmcache.mp.port\":$LMCACHE_PORT,\"lmcache.mp.mq_timeout\":6000.0,\"lmcache.mp.heartbeat_interval\":$LMCACHE_HEARTBEAT_INTERVAL}}"
     )
     ;;
     *)
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index dad68b8648..f87a57a934 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6830,4 +6830,5 @@
     - "Refresh the Kimi-K3 MI355X AgentX recipe from the merged PR #2787 baseline to vLLM ROCm nightly 7c5dc571 and LMCache 0.5.5rc3+rocm7.2."
     - "Run C1, C8, and C14 with TP8 DSpark and LMCache DRAM KV offload; run C40, C44, and C48 with TP8/DCP8 and no speculative decoding."
     - "Use a 12288-token LMCache chunk and a 1536-token DCP KV-cache interleave; set C40/C44/C48 gpu-memory-utilization to 0.88 while retaining the 16K batched-token, max-num-seqs 80, and full CUDA-graph profile."
+    - "Allow 90 seconds for LMCache MP heartbeats and 300 seconds before reaping workers so transient C14 transfer stalls do not trigger invalid-block recovery."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2804

From a720a04fc480a4d8fd2640870c366c2971a41d4e Mon Sep 17 00:00:00 2001
From: hyukjlee <hyukjlee@amd.com>
Date: Wed, 2 Sep 2026 21:39:46 +0900
Subject: [PATCH 6/6] perf(agentx): set K3 sweep GMU to 0.88
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

将 C1、C8 和 C14 的 gpu-memory-utilization 从 0.90 降至 0.88，使全部六个测试点使用相同的 GMU。移除上一轮未证实的 heartbeat workaround，以便单独验证显存利用率变化。

Assisted-by: OpenAI Codex <codex@openai.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 9 +++------
 perf-changelog.yaml                                     | 3 +--
 2 files changed, 4 insertions(+), 8 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index abbfc79558..81b3b29915 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -186,8 +186,6 @@ case "${KV_OFFLOAD_BACKEND:-}" in
     LMCACHE_PORT=6555
     LMCACHE_HTTP_PORT=8090
     LMCACHE_LOG="$RESULT_DIR/lmcache_server.log"
-    LMCACHE_HEARTBEAT_INTERVAL=90.0
-    LMCACHE_WORKER_REAP_TIMEOUT=300
 
     LMCACHE_L1_SIZE_GB="$TOTAL_CPU_DRAM_GB"
 
@@ -207,7 +205,6 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --max-gpu-workers 1
         --eviction-policy LRU
         --supported-transfer-mode lmcache_driven
-        --worker-reap-timeout-seconds "$LMCACHE_WORKER_REAP_TIMEOUT"
         --shm-name ""
     )
     append_command "$RESULT_DIR/lmcache_command.txt" "${LMCACHE_CMD[@]}"
@@ -224,7 +221,7 @@ case "${KV_OFFLOAD_BACKEND:-}" in
     # same MQ timeout headroom as the MiniMax-M3 arm.
     OFFLOAD_ARGS=(
         --kv-transfer-config
-        "{\"kv_connector\":\"LMCacheMPConnector\",\"kv_connector_module_path\":\"lmcache.integration.vllm.lmcache_mp_connector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"lmcache.mp.port\":$LMCACHE_PORT,\"lmcache.mp.mq_timeout\":6000.0,\"lmcache.mp.heartbeat_interval\":$LMCACHE_HEARTBEAT_INTERVAL}}"
+        "{\"kv_connector\":\"LMCacheMPConnector\",\"kv_connector_module_path\":\"lmcache.integration.vllm.lmcache_mp_connector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"lmcache.mp.port\":$LMCACHE_PORT,\"lmcache.mp.mq_timeout\":6000.0}}"
     )
     ;;
     *)
@@ -247,13 +244,13 @@ case "${SPEC_DECODING:-mtp}:$CONC" in
     mtp:1)
         SYNTHETIC_ACCEPT_LEN=3.75
         SPEC_NUM_TOKENS=6
-        GPU_MEM_UTIL=0.9
+        GPU_MEM_UTIL=0.88
         MAX_NUM_BATCHED_TOKENS=16384
         ;;
     mtp:2|mtp:4|mtp:8|mtp:10|mtp:12|mtp:14)
         SYNTHETIC_ACCEPT_LEN=3.00
         SPEC_NUM_TOKENS=3
-        GPU_MEM_UTIL=0.9
+        GPU_MEM_UTIL=0.88
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
     none:40|none:44|none:48)
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index f87a57a934..a369acd371 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6829,6 +6829,5 @@
   description:
     - "Refresh the Kimi-K3 MI355X AgentX recipe from the merged PR #2787 baseline to vLLM ROCm nightly 7c5dc571 and LMCache 0.5.5rc3+rocm7.2."
     - "Run C1, C8, and C14 with TP8 DSpark and LMCache DRAM KV offload; run C40, C44, and C48 with TP8/DCP8 and no speculative decoding."
-    - "Use a 12288-token LMCache chunk and a 1536-token DCP KV-cache interleave; set C40/C44/C48 gpu-memory-utilization to 0.88 while retaining the 16K batched-token, max-num-seqs 80, and full CUDA-graph profile."
-    - "Allow 90 seconds for LMCache MP heartbeats and 300 seconds before reaping workers so transient C14 transfer stalls do not trigger invalid-block recovery."
+    - "Use gpu-memory-utilization 0.88 for all six points, a 12288-token LMCache chunk, and a 1536-token DCP KV-cache interleave; C40/C44/C48 retain the 16K batched-token, max-num-seqs 80, and full CUDA-graph profile."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2804