From 71acdd405f13dc1c61a1d051801bd32cfe60ee9a Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 09:10:47 +0900
Subject: [PATCH 01/24] perf(agentx): refresh K3 MI355X vLLM recipe with DCP8
 MTP arm

Bump the kimik3-fp4-mi355x-vllm-agentic-mtp image to vLLM ROCm nightly
7c5dc571 and LMCache to 0.5.5rc3, and add a TP8/DCP8 spec-mtp DRAM-offload
arm (conc 44/48/52). Retune the LMCache tier (chunk-size 12288, max-gpu-workers
8) and the no-spec decode profile (GMU 0.9, 8192 batched tokens), and switch
the DCP attention backend to ROCM_AITER_MLA.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../agentic/kimik3_fp4_mi355x_mtp.sh           | 18 +++++++++---------
 configs/amd-master.yaml                        |  5 +++--
 2 files changed, 12 insertions(+), 11 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index a8f2c42c1c..95ada0bb38 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -35,10 +35,10 @@ set -x
 #   GPU_MEM_UTIL             0.95   (reference)
 #   MAX_NUM_BATCHED_TOKENS   8192   (default)
 #   AITER_A8W4               1      (reference; 0 = aiter a16w4 MoE path)
-#   LANGUAGE_MODEL_ONLY      true   
+#   LANGUAGE_MODEL_ONLY      true
 #   KV_CACHE_DTYPE           fp8    (default for every arm; =auto for a bf16 A/B)
 #   KV_BLOCK_SIZE            unset  (unset -> vLLM sizes the page; 128 under fp8)
-#   MAX_MODEL_LEN            1M     
+#   MAX_MODEL_LEN            1M
 #   SPEC_DECODE              true   (this is the _mtp DSpark recipe; =false for a no-spec A/B)
 #   SPEC_NUM_TOKENS          2      (DSpark draft length; validated by the _mtp config)
 
@@ -151,7 +151,7 @@ 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="0.5.5.dev60+rocm7.2"
+    LMCACHE_VERSION="0.5.5rc3"
     LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/nightly-rocm"
     agentic_pip_install --quiet --no-cache-dir --no-deps \
         "sortedcontainers==2.4.0" \
@@ -205,12 +205,12 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --http-port "$LMCACHE_HTTP_PORT"
         --l1-size-gb "$LMCACHE_L1_SIZE_GB"
         --l1-init-size-gb 10
-        --chunk-size 3072
+        --chunk-size 12288
         --separate-object-groups
         --enable-extra-logging
         --extra-logging-interval 30
         --max-cpu-workers 8
-        --max-gpu-workers 1
+        --max-gpu-workers 8
         --eviction-policy LRU
         --supported-transfer-mode lmcache_driven
         --shm-name ""
@@ -264,8 +264,8 @@ case "$CONC" in
         ;;
     *)
         SPEC_NUM_TOKENS=0
-        GPU_MEM_UTIL=0.85
-        MAX_NUM_BATCHED_TOKENS=4096
+        GPU_MEM_UTIL=0.9
+        MAX_NUM_BATCHED_TOKENS=8192
         ;;
 esac
 
@@ -306,7 +306,7 @@ 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)
+    ATTN_BE_ARGS+=(--attention-backend ROCM_AITER_MLA)
 fi
 export VLLM_USE_DIRECT_DCP_A2A=0
 export VLLM_USE_DIRECT_DCP_Q_GATHER=0
@@ -352,4 +352,4 @@ if [ "${EVAL_ONLY}" = "true" ]; then
 else
     build_replay_cmd "$RESULT_DIR"
     run_agentic_replay_and_write_outputs "$RESULT_DIR"
-fi
+fi
\ No newline at end of file
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 0d380154a0..feae382240 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
@@ -637,7 +637,8 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
     - 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" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48, 52], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From c281dbece1fb8919e2204e273ed5cca49480b1c7 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 09:11:31 +0900
Subject: [PATCH 02/24] docs(changelog): link K3 MI355X vLLM DCP8 MTP refresh
 PR

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 perf-changelog.yaml | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 53a2ba7e75..e02c093848 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -6831,3 +6831,13 @@
     - "Use a 169 GB/rank HiCache target pool at c12 and c16 while retaining ratio mode for lower concurrencies."
     - "Isolate SGLang runtime caches per Slurm allocation to prevent concurrent sweep cells from sharing per-rank cache files."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2808
+
+- config-keys:
+    - kimik3-fp4-mi355x-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Bump the MI355X K3 vLLM recipe to vLLM ROCm nightly 7c5dc571 and LMCache 0.5.5rc3."
+    - "Add a TP8/DCP8 spec-mtp DRAM-offload arm at concurrency 44/48/52."
+    - "Retune the LMCache tier (12288-token chunk, 8 GPU workers) and the no-spec decode profile (gpu-memory-utilization 0.9, 8192 batched tokens); switch the DCP attention backend to ROCM_AITER_MLA."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2810

From 7b934687b32e7dfe5ff1f5ac72f1385b7fb67ccf Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 09:13:10 +0900
Subject: [PATCH 03/24] style(agentx): add trailing newline to K3 MI355X
 benchmark script

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 2 +-
 1 file changed, 1 insertion(+), 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 95ada0bb38..4ba204c355 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -352,4 +352,4 @@ if [ "${EVAL_ONLY}" = "true" ]; then
 else
     build_replay_cmd "$RESULT_DIR"
     run_agentic_replay_and_write_outputs "$RESULT_DIR"
-fi
\ No newline at end of file
+fi

From db7732d3c7eda62e92adcc2f13c196677d91d3cb Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 10:03:20 +0900
Subject: [PATCH 04/24] perf(agentx): drop c52 from K3 MI355X DCP8 arm

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 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 feae382240..6313ffbd97 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -638,7 +638,7 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       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: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
-      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48, 52], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From 0a32492dcff178adf5b49c6ab24afcd0d57ab242 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 10:44:52 +0900
Subject: [PATCH 05/24] perf(agentx): gate K3 LMCache GPU workers on DCP; trim
 sweep to c14/c48

Set --max-gpu-workers to 8 only when DCP_SIZE>1 (else 1), and reduce the
kimik3-fp4-mi355x-vllm-agentic-mtp search-space to a c14 (non-DCP) and c48
(DCP8) smoke test.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_mtp.sh       | 10 +++++++++-
 configs/amd-master.yaml                                |  9 ++++++---
 2 files changed, 15 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 4ba204c355..5978b74122 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -197,6 +197,14 @@ case "${KV_OFFLOAD_BACKEND:-}" in
 
     LMCACHE_L1_SIZE_GB="$TOTAL_CPU_DRAM_GB"
 
+    # DCP shards decode KV across the TP ranks, so the LMCache GPU transfer
+    # pool needs one worker per rank; a non-DCP arm only needs a single worker.
+    if [ "${DCP_SIZE:-1}" -gt 1 ]; then
+        LMCACHE_MAX_GPU_WORKERS=8
+    else
+        LMCACHE_MAX_GPU_WORKERS=1
+    fi
+
     LMCACHE_CMD=(
         lmcache server
         --host 127.0.0.1
@@ -210,7 +218,7 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --enable-extra-logging
         --extra-logging-interval 30
         --max-cpu-workers 8
-        --max-gpu-workers 8
+        --max-gpu-workers "$LMCACHE_MAX_GPU_WORKERS"
         --eviction-policy LRU
         --supported-transfer-mode lmcache_driven
         --shm-name ""
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 6313ffbd97..ee77661a66 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -636,9 +636,12 @@ 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: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
-      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
+      #- { tp: 8, kv-offloading: none, conc-list: [1] , spec-decoding: mtp}
+      #- { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
+      #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
+
+      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [14], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From 1701d427b46655842cca2e794aab7f32dc44af31 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 11:10:26 +0900
Subject: [PATCH 06/24] perf(agentx): lower K3 c12/c14 GMU to 0.88; run
 c14-only smoke test

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 8 +++++++-
 configs/amd-master.yaml                                 | 2 +-
 2 files changed, 8 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index 5978b74122..e60a874102 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -264,12 +264,18 @@ case "$CONC" in
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=16384
         ;;
-    2|4|8|10|12|14)
+    2|4|8)
         SYNTHETIC_ACCEPT_LEN=3.00
         SPEC_NUM_TOKENS=3
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
+    12|14)
+        SYNTHETIC_ACCEPT_LEN=3.00
+        SPEC_NUM_TOKENS=3
+        GPU_MEM_UTIL=0.88
+        MAX_NUM_BATCHED_TOKENS=8192
+        ;;
     *)
         SPEC_NUM_TOKENS=0
         GPU_MEM_UTIL=0.9
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index ee77661a66..7b661185ab 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -641,7 +641,7 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
       - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [14], spec-decoding: mtp }
-      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
+      #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From aefefb37cd75f180ec96a0eb87459b1cc4e081ec Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 11:30:11 +0900
Subject: [PATCH 07/24] perf(agentx): run K3 c12/c14 GMU 0.88 smoke test

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 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 7b661185ab..fd8058b606 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -640,7 +640,7 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       #- { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
-      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [14], spec-decoding: mtp }
+      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [12, 14], spec-decoding: mtp }
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark

From 5cebe65d23cd1e89e4e144828ca1a41e2adeac81 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 12:28:04 +0900
Subject: [PATCH 08/24] perf(agentx): lower K3 c12/c14 GMU to 0.85

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 2 +-
 1 file changed, 1 insertion(+), 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 e60a874102..6211c674b1 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -273,7 +273,7 @@ case "$CONC" in
     12|14)
         SYNTHETIC_ACCEPT_LEN=3.00
         SPEC_NUM_TOKENS=3
-        GPU_MEM_UTIL=0.88
+        GPU_MEM_UTIL=0.85
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
     *)

From 1b2469300a9ed6b7a3378b2289814a5a9b0a3191 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 12:53:44 +0900
Subject: [PATCH 09/24] perf(agentx): test K3 c14 GPU-resident (no KV offload),
 GMU 0.9

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 4 ++--
 configs/amd-master.yaml                                 | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index 6211c674b1..7a6483cb34 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -271,9 +271,9 @@ case "$CONC" in
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
     12|14)
-        SYNTHETIC_ACCEPT_LEN=3.00
+        SYNTHETIC_ACCEPT_LEN=0
         SPEC_NUM_TOKENS=3
-        GPU_MEM_UTIL=0.85
+        GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
     *)
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index fd8058b606..63e825c289 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -640,7 +640,7 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       #- { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
-      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [12, 14], spec-decoding: mtp }
+      - { tp: 8, ep: 1, kv-offloading: none, conc-list: [14], spec-decoding: mtp }
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark

From ffc4d74e0ebe74104e956ee591577173ce56a6a3 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 13:25:53 +0900
Subject: [PATCH 10/24] perf(agentx): bump K3 image to nightly-73029d42; c14
 no-spec GPU-resident

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 2 +-
 configs/amd-master.yaml                                 | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index 7a6483cb34..0f96fe01af 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -272,7 +272,7 @@ case "$CONC" in
         ;;
     12|14)
         SYNTHETIC_ACCEPT_LEN=0
-        SPEC_NUM_TOKENS=3
+        SPEC_NUM_TOKENS=0
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 63e825c289..32438c89cb 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-7c5dc571cbd1064ecc8a9b1045637ff647aa22cb
+  image: vllm/vllm-openai-rocm:nightly-73029d42441321b631779db3475031f5ec26dd6c
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:mi355x-amds

From 99d06715f060dd8699dfd0b523391dcb5fa229d5 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 13:51:25 +0900
Subject: [PATCH 11/24] perf(agentx): run K3 c48 DCP8 LMCache smoke test on
 nightly-73029d42

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 configs/amd-master.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 32438c89cb..87ee30eefb 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -640,8 +640,8 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       #- { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
-      - { tp: 8, ep: 1, kv-offloading: none, conc-list: [14], spec-decoding: mtp }
-      #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
+      #- { tp: 8, ep: 1, kv-offloading: none, conc-list: [14], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From 34d93c9c451e27636b911eaf2dec6c1715f8cdc1 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 13:55:38 +0900
Subject: [PATCH 12/24] perf(agentx): run K3 c14 GPU-resident with MTP spec on
 nightly-73029d42

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 4 ++--
 configs/amd-master.yaml                                 | 4 ++--
 2 files changed, 4 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 0f96fe01af..8e95e4c5e3 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -271,8 +271,8 @@ case "$CONC" in
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
     12|14)
-        SYNTHETIC_ACCEPT_LEN=0
-        SPEC_NUM_TOKENS=0
+        SYNTHETIC_ACCEPT_LEN=3.00
+        SPEC_NUM_TOKENS=3
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 87ee30eefb..32438c89cb 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -640,8 +640,8 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       #- { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
-      #- { tp: 8, ep: 1, kv-offloading: none, conc-list: [14], spec-decoding: mtp }
-      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
+      - { tp: 8, ep: 1, kv-offloading: none, conc-list: [14], spec-decoding: mtp }
+      #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From 883d6e37d57c38d6cfd312b2e11ffbe57c1661d6 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 14:18:10 +0900
Subject: [PATCH 13/24] perf(agentx): drop K3 high-conc max-num-batched-tokens
 to 4096 (OOM fix)

Widen prefill activation headroom for the c48 DCP8 arm, which repeatedly
OOM'd in the attention-residual forward at 8192 batched tokens under GMU 0.9.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 2 +-
 1 file changed, 1 insertion(+), 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 8e95e4c5e3..cee20214a1 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -279,7 +279,7 @@ case "$CONC" in
     *)
         SPEC_NUM_TOKENS=0
         GPU_MEM_UTIL=0.9
-        MAX_NUM_BATCHED_TOKENS=8192
+        MAX_NUM_BATCHED_TOKENS=4096
         ;;
 esac
 

From 60acfb363fdc6f563830463a61a2dc575fcfbc31 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 14:21:57 +0900
Subject: [PATCH 14/24] perf(agentx): activate K3 c48 DCP8 arm for OOM-fix
 validation

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 configs/amd-master.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 32438c89cb..87ee30eefb 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -640,8 +640,8 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       #- { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
-      - { tp: 8, ep: 1, kv-offloading: none, conc-list: [14], spec-decoding: mtp }
-      #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
+      #- { tp: 8, ep: 1, kv-offloading: none, conc-list: [14], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From 47745bf424afa9e92bbea6ae0ca35466a0cb38ce Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 14:32:39 +0900
Subject: [PATCH 15/24] perf(agentx): gate K3 LMCache chunk-size on DCP; pin
 dev60 wheel; run c14

Set --chunk-size 12288 only when DCP_SIZE>1 (else 3072), install the
0.5.5.dev60+rocm7.2 wheel from the 0.5.5rc3 release index, and activate the
c14 LMCache-DRAM arm.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_mtp.sh      | 11 ++++++++---
 configs/amd-master.yaml                               |  4 ++--
 2 files changed, 10 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index cee20214a1..4e54866e91 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -151,8 +151,9 @@ 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="0.5.5rc3"
-    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/nightly-rocm"
+    LMCACHE_VERSION="0.5.5.dev60+rocm7.2"
+    LMCACHE_RELEASE="0.5.5rc3"
+    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/${LMCACHE_RELEASE}-rocm"
     agentic_pip_install --quiet --no-cache-dir --no-deps \
         "sortedcontainers==2.4.0" \
         "opentelemetry-exporter-prometheus==0.61b0" \
@@ -199,10 +200,14 @@ case "${KV_OFFLOAD_BACKEND:-}" in
 
     # DCP shards decode KV across the TP ranks, so the LMCache GPU transfer
     # pool needs one worker per rank; a non-DCP arm only needs a single worker.
+    # The DCP KV interleave also needs the larger 12288 chunk; a non-DCP arm
+    # uses the 3072 minimum (one KDA state group).
     if [ "${DCP_SIZE:-1}" -gt 1 ]; then
         LMCACHE_MAX_GPU_WORKERS=8
+        LMCACHE_CHUNK_SIZE=12288
     else
         LMCACHE_MAX_GPU_WORKERS=1
+        LMCACHE_CHUNK_SIZE=3072
     fi
 
     LMCACHE_CMD=(
@@ -213,7 +218,7 @@ case "${KV_OFFLOAD_BACKEND:-}" in
         --http-port "$LMCACHE_HTTP_PORT"
         --l1-size-gb "$LMCACHE_L1_SIZE_GB"
         --l1-init-size-gb 10
-        --chunk-size 12288
+        --chunk-size "$LMCACHE_CHUNK_SIZE"
         --separate-object-groups
         --enable-extra-logging
         --extra-logging-interval 30
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 87ee30eefb..b90a16769f 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -640,8 +640,8 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       #- { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
-      #- { tp: 8, ep: 1, kv-offloading: none, conc-list: [14], spec-decoding: mtp }
-      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
+      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [14], spec-decoding: mtp }
+      #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From f48e07a28ed14db9e28af38b9177963189bebd10 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 16:00:46 +0900
Subject: [PATCH 16/24] perf(agentx): gate K3 LMCache wheel on DCP; run c14 on
 dev60

Use 0.5.5rc3 when DCP_SIZE>1 and 0.5.5.dev60+rocm7.2 otherwise, and pin the
active c14 arm to dev60.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_mtp.sh       | 10 +++++++---
 configs/amd-master.yaml                                |  3 +--
 2 files changed, 8 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index 4e54866e91..16be48c0f4 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -151,9 +151,13 @@ 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="0.5.5.dev60+rocm7.2"
-    LMCACHE_RELEASE="0.5.5rc3"
-    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/${LMCACHE_RELEASE}-rocm"
+    # The DCP arm needs the 0.5.5rc3 wheel; the non-DCP arm pins dev60.
+    if [ "${DCP_SIZE:-1}" -gt 1 ]; then
+        LMCACHE_VERSION="0.5.5rc3"
+    else
+        LMCACHE_VERSION="0.5.5.dev60+rocm7.2"
+    fi
+    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/nightly-rocm"
     agentic_pip_install --quiet --no-cache-dir --no-deps \
         "sortedcontainers==2.4.0" \
         "opentelemetry-exporter-prometheus==0.61b0" \
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index b90a16769f..9a6e36c459 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -640,8 +640,7 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       #- { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
-      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [14], spec-decoding: mtp }
-      #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
+      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev60+rocm7.2" }, conc-list: [14], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From ed613b6c3c5cce4b784f1129995a623ce3dd8ba1 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 16:34:09 +0900
Subject: [PATCH 17/24] fix(agentx): define LMCACHE_RELEASE for the K3 index
 URL

The index was templated on ${LMCACHE_RELEASE} but the var was never set,
which aborts the lmcache arm under set -u. Pin it to the nightly-rocm asset
page that hosts both wheels.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 4 +++-
 1 file changed, 3 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 16be48c0f4..ed4edf549a 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -157,7 +157,9 @@ case "${KV_OFFLOAD_BACKEND:-}" in
     else
         LMCACHE_VERSION="0.5.5.dev60+rocm7.2"
     fi
-    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/nightly-rocm"
+    # Both wheels are published on the rolling nightly-rocm asset page.
+    LMCACHE_RELEASE="nightly"
+    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/${LMCACHE_RELEASE}-rocm"
     agentic_pip_install --quiet --no-cache-dir --no-deps \
         "sortedcontainers==2.4.0" \
         "opentelemetry-exporter-prometheus==0.61b0" \

From bbce1d25c7672bec371d8e68f5e28d0d7dcd8c9a Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 16:37:50 +0900
Subject: [PATCH 18/24] perf(agentx): pin K3 LMCache to v0.5.5rc4 from the
 release asset page

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_mtp.sh    | 13 +++----------
 1 file changed, 3 insertions(+), 10 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index ed4edf549a..3d3070adab 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -149,17 +149,10 @@ 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.
-    # The DCP arm needs the 0.5.5rc3 wheel; the non-DCP arm pins dev60.
-    if [ "${DCP_SIZE:-1}" -gt 1 ]; then
-        LMCACHE_VERSION="0.5.5rc3"
-    else
-        LMCACHE_VERSION="0.5.5.dev60+rocm7.2"
-    fi
-    # Both wheels are published on the rolling nightly-rocm asset page.
-    LMCACHE_RELEASE="nightly"
+    LMCACHE_VERSION=v0.5.5rc4
+    LMCACHE_RELEASE=v0.5.5rc4+rocm7.2
     LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/${LMCACHE_RELEASE}-rocm"
+
     agentic_pip_install --quiet --no-cache-dir --no-deps \
         "sortedcontainers==2.4.0" \
         "opentelemetry-exporter-prometheus==0.61b0" \

From 0a3b5728e167750176d15c12efceb9cbe9dffd80 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 19:54:48 +0900
Subject: [PATCH 19/24] perf(agentx): pin K3 LMCache to dev89; run c14 (dcp1)
 and c48 (dcp8)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 5 ++---
 configs/amd-master.yaml                                 | 7 ++++---
 2 files changed, 6 insertions(+), 6 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index 3d3070adab..141d635a07 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -149,9 +149,8 @@ case "${KV_OFFLOAD_BACKEND:-}" in
       lmcache)
     require_agentic_kv_offload_backend "$KV_OFFLOAD_BACKEND"
 
-    LMCACHE_VERSION=v0.5.5rc4
-    LMCACHE_RELEASE=v0.5.5rc4+rocm7.2
-    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/${LMCACHE_RELEASE}-rocm"
+    LMCACHE_VERSION=0.5.5.dev89+rocm7.2
+    LMCACHE_ROCM_INDEX="https://github.com/LMCache/LMCache/releases/expanded_assets/nightly-rocm"
 
     agentic_pip_install --quiet --no-cache-dir --no-deps \
         "sortedcontainers==2.4.0" \
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 9a6e36c459..0a52ebd4e1 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -636,11 +636,12 @@ 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: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
+      #- { tp: 8, ep: 1, dcp-size: 1, kv-offloading: none, conc-list: [1] , spec-decoding: mtp}
+      #- { tp: 8, ep: 1, dcp-size: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
-      - { tp: 8, ep: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev60+rocm7.2" }, conc-list: [14], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev89+rocm7.2" }, conc-list: [14], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From d5e7a1c57f4b5d557def988079aa4606ff789d64 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 21:47:03 +0900
Subject: [PATCH 20/24] perf(agentx): revert K3 image to 7c5dc571; c48 label to
 dev89

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 configs/amd-master.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 0a52ebd4e1..c8140d817a 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-73029d42441321b631779db3475031f5ec26dd6c
+  image: vllm/vllm-openai-rocm:nightly-7c5dc571cbd1064ecc8a9b1045637ff647aa22cb
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:mi355x-amds
@@ -641,7 +641,7 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
 
       - { tp: 8, ep: 1, dcp-size: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev89+rocm7.2" }, conc-list: [14], spec-decoding: mtp }
-      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [48], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev89+rocm7.2" }, conc-list: [48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From 05c3f3f62fbdb95b00ebde76c8067ab91b3fa74c Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 22:41:29 +0900
Subject: [PATCH 21/24] perf(agentx): bump K3 image to 73029d42; run full conc
 sweep on dev89

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_mtp.sh      |  9 ++++-----
 configs/amd-master.yaml                               | 11 ++++-------
 2 files changed, 8 insertions(+), 12 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index 141d635a07..774dd0b278 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -267,22 +267,21 @@ case "$CONC" in
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=16384
         ;;
-    2|4|8)
+    4|8|10|12|14)
         SYNTHETIC_ACCEPT_LEN=3.00
         SPEC_NUM_TOKENS=3
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
-    12|14)
-        SYNTHETIC_ACCEPT_LEN=3.00
-        SPEC_NUM_TOKENS=3
+    44|48)
+        SPEC_NUM_TOKENS=0
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
     *)
         SPEC_NUM_TOKENS=0
         GPU_MEM_UTIL=0.9
-        MAX_NUM_BATCHED_TOKENS=4096
+        MAX_NUM_BATCHED_TOKENS=8192
         ;;
 esac
 
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index c8140d817a..af3fe7a820 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-7c5dc571cbd1064ecc8a9b1045637ff647aa22cb
+  image: vllm/vllm-openai-rocm:nightly-73029d42441321b631779db3475031f5ec26dd6c
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:mi355x-amds
@@ -636,12 +636,9 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.60
       search-space:
-      #- { tp: 8, ep: 1, dcp-size: 1, kv-offloading: none, conc-list: [1] , spec-decoding: mtp}
-      #- { tp: 8, ep: 1, dcp-size: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
-      #- { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5rc3" }, conc-list: [44, 48], spec-decoding: mtp }
-
-      - { tp: 8, ep: 1, dcp-size: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev89+rocm7.2" }, conc-list: [14], spec-decoding: mtp }
-      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev89+rocm7.2" }, conc-list: [48], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 1, kv-offloading: none, conc-list: [1] , spec-decoding: mtp}
+      - { tp: 8, ep: 1, dcp-size: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev89+rocm7.2" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev89+rocm7.2" }, conc-list: [44, 48], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From 932280a154bf065f5799b5db1445d28e2d08f57b Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 23:13:35 +0900
Subject: [PATCH 22/24] perf(agentx): add conc 52 to K3 dcp8 sweep

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh | 2 +-
 configs/amd-master.yaml                                 | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
index 774dd0b278..f0416e4af0 100644
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh
@@ -273,7 +273,7 @@ case "$CONC" in
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=8192
         ;;
-    44|48)
+    44|48|52)
         SPEC_NUM_TOKENS=0
         GPU_MEM_UTIL=0.9
         MAX_NUM_BATCHED_TOKENS=8192
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index af3fe7a820..6ab04839d8 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -638,7 +638,7 @@ kimik3-fp4-mi355x-vllm-agentic-mtp:
       search-space:
       - { tp: 8, ep: 1, dcp-size: 1, kv-offloading: none, conc-list: [1] , spec-decoding: mtp}
       - { tp: 8, ep: 1, dcp-size: 1, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev89+rocm7.2" }, conc-list: [4, 8, 10, 12, 14], spec-decoding: mtp }
-      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev89+rocm7.2" }, conc-list: [44, 48], spec-decoding: mtp }
+      - { tp: 8, ep: 1, dcp-size: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.5.dev89+rocm7.2" }, conc-list: [44, 48, 52], spec-decoding: mtp }
 
 # Kimi-K3 MXFP4 agentic-coding benchmark on MI355X via ATOM with DSpark
 # speculative decoding. Acceptance is pinned to the committed golden curve in

From d9f68559b1a393fbbdbdd347427a14237b77b9ec Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Thu, 3 Sep 2026 23:31:15 +0900
Subject: [PATCH 23/24] perf(agentx): revert K3 image to 7c5dc571

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 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 6ab04839d8..0269ddfa05 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-73029d42441321b631779db3475031f5ec26dd6c
+  image: vllm/vllm-openai-rocm:nightly-7c5dc571cbd1064ecc8a9b1045637ff647aa22cb
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:mi355x-amds

From cbf74f5ab34d28448660ed37da7fc1d7a5a5d171 Mon Sep 17 00:00:00 2001
From: seungrokj <seungrok.jung@amd.com>
Date: Fri, 4 Sep 2026 00:35:44 +0900
Subject: [PATCH 24/24] fix(ci): pin pydantic>=2 for perf-changelog validation

The validator imports field_validator/model_validator/ConfigDict (pydantic
v2), but the unpinned pip install was resolving pydantic 1.10.26, crashing
check-changelog on ImportError before validating anything.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .github/workflows/run-sweep.yml           | 2 +-
 .github/workflows/test-changelog-gate.yml | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/.github/workflows/run-sweep.yml b/.github/workflows/run-sweep.yml
index c2d1445b24..9be5c651b6 100644
--- a/.github/workflows/run-sweep.yml
+++ b/.github/workflows/run-sweep.yml
@@ -109,7 +109,7 @@ jobs:
                   ALL_EVALS: ${{ contains(github.event.pull_request.labels.*.name, 'all-evals') }}
                   EVALS_ONLY: ${{ contains(github.event.pull_request.labels.*.name, 'evals-only') }}
               run: |
-                  pip install pydantic pyyaml
+                  pip install "pydantic>=2" pyyaml
                   CMD=(
                       python3 utils/validate_perf_changelog.py
                       --changelog-file perf-changelog.yaml
diff --git a/.github/workflows/test-changelog-gate.yml b/.github/workflows/test-changelog-gate.yml
index c1f11f34f8..65bd1eebc5 100644
--- a/.github/workflows/test-changelog-gate.yml
+++ b/.github/workflows/test-changelog-gate.yml
@@ -68,7 +68,7 @@ jobs:
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
-          python -m pip install pytest pydantic pyyaml tabulate
+          python -m pip install pytest "pydantic>=2" pyyaml tabulate
 
       - name: Run validator, reuse, and gating tests
         run: |