From c36c2ddc253cbe465a93aaf5167bf9c7d0d47fbf Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Wed, 5 Aug 2026 03:23:51 +0000
Subject: [PATCH 01/18] feat(agentx): add Kimi-K3 DSpark LMCache on MI355X
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add the TP8 AgentX sweep, validated ROCm image, LMCache MP offload recipe, and launcher routing for Kimi-K3 DSpark.

中文：新增 Kimi-K3 DSpark 在 MI355X 上的 TP8 AgentX 扫描、已验证 ROCm 镜像、LMCache MP 卸载配置及启动器路由。

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../kimik3_fp4_mi355x_agentic_dspark.sh       | 196 ++++++++++++++++++
 configs/amd-master.yaml                       |  14 ++
 perf-changelog.yaml                           |  10 +
 runners/launch_mi355x-amds.sh                 |  11 +-
 4 files changed, 230 insertions(+), 1 deletion(-)
 create mode 100755 benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
new file mode 100755
index 0000000000..e390f333b4
--- /dev/null
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -0,0 +1,196 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# Agentic trace replay benchmark for Kimi-K3 MXFP4 on MI355X using vLLM,
+# DSpark speculative decoding, and LMCache MP host-DRAM KV offload.
+#
+# Required env vars:
+#   MODEL, TP, CONC, KV_OFFLOADING, KV_OFFLOAD_BACKEND,
+#   TOTAL_CPU_DRAM_GB, RESULT_DIR, DURATION, EP_SIZE, PORT
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL TP CONC KV_OFFLOADING KV_OFFLOAD_BACKEND TOTAL_CPU_DRAM_GB \
+    RESULT_DIR DURATION EP_SIZE PORT
+
+if [ "$TP" -ne 8 ]; then
+    echo "Error: Kimi-K3 MXFP4 requires TP=8 on MI355X; got TP=$TP." >&2
+    exit 1
+fi
+
+if [ "$EP_SIZE" -gt 1 ]; then
+    echo "Error: this Kimi-K3 DSpark recipe supports pure TP8 only; got EP_SIZE=$EP_SIZE." >&2
+    exit 1
+fi
+
+require_agentic_kv_offload_backend lmcache
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
+fi
+
+if [[ -n "${ROCR_VISIBLE_DEVICES:-}" ]]; then
+    export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
+fi
+
+# The cluster launcher mounts its persistent Hugging Face cache here. MODEL_PATH
+# may instead point at a pre-staged snapshot.
+export HF_HUB_CACHE="${HF_HUB_CACHE:-/models/huggingface_hub}"
+if [[ -n "${MODEL_PATH:-}" ]]; then
+    if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
+        hf download "$MODEL" --local-dir "$MODEL_PATH"
+    fi
+else
+    hf download "$MODEL"
+    export MODEL_PATH="$MODEL"
+fi
+
+rocm-smi || true
+amd-smi || true
+
+resolve_trace_source
+install_agentic_deps
+
+# Environment used by the validated Kimi-K3 + DSpark + LMCache image.
+export PYTHONNOUSERSITE=1
+export PYTHONHASHSEED=42
+export VLLM_ROCM_USE_AITER=1
+export VLLM_ROCM_USE_AITER_MLA=1
+export AITER_SITUV2_A8W4=1
+export AITER_BF16_FP8_MOE_BOUND=0
+export VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION=1
+export VLLM_USE_BREAKABLE_CUDAGRAPH=0
+export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS="${VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS:-1200}"
+export VLLM_ENGINE_READY_TIMEOUT_S="${VLLM_ENGINE_READY_TIMEOUT_S:-7200}"
+export HSA_NO_SCRATCH_RECLAIM=1
+export SAFETENSORS_FAST_GPU=1
+
+SERVER_LOG="$RESULT_DIR/server.log"
+LMCACHE_LOG="$RESULT_DIR/lmcache_server.log"
+mkdir -p "$RESULT_DIR"
+
+SERVER_PID=""
+LMCACHE_PID=""
+
+cleanup_services() {
+    local exit_code=$?
+    trap - EXIT INT TERM
+    set +e
+    stop_background_process_tree "$SERVER_PID" "vLLM server" 60
+    stop_background_process_tree "$LMCACHE_PID" "LMCache server" 30
+    exit "$exit_code"
+}
+trap cleanup_services EXIT
+trap 'exit 130' INT
+trap 'exit 143' TERM
+
+wait_for_lmcache_ready() {
+    { set +x; } 2>/dev/null
+    local attempts="${LMCACHE_READY_ATTEMPTS:-1800}"
+    local paths=(/healthcheck /health /v1/health /status /)
+    local i path
+
+    for ((i = 1; i <= attempts; i++)); do
+        for path in "${paths[@]}"; do
+            if curl --output /dev/null --silent --fail \
+                    "http://127.0.0.1:${LMCACHE_HTTP_PORT}${path}"; then
+                echo "LMCache server healthy after ${i}s (endpoint ${path})"
+                set -x
+                return 0
+            fi
+        done
+
+        if ! kill -0 "$LMCACHE_PID" 2>/dev/null; then
+            echo "LMCache server exited before becoming healthy. Log follows:" >&2
+            tail -200 "$LMCACHE_LOG" >&2 || true
+            exit 1
+        fi
+
+        if ((i % 60 == 0)); then
+            echo "Still waiting for LMCache (${i}s/${attempts}s, L1=${LMCACHE_L1_SIZE_GB}GB)"
+        fi
+        sleep 1
+    done
+
+    echo "Timed out waiting for LMCache on port $LMCACHE_HTTP_PORT. Log follows:" >&2
+    tail -200 "$LMCACHE_LOG" >&2 || true
+    exit 1
+}
+
+# This is the pool size validated on the MI355X host. Keep it overridable for
+# hosts with a different /dev/shm allocation, but never exceed the matrix's DRAM
+# budget.
+LMCACHE_L1_SIZE_GB="${LMCACHE_L1_SIZE_GB:-906}"
+if [ "$LMCACHE_L1_SIZE_GB" -gt "$TOTAL_CPU_DRAM_GB" ]; then
+    echo "Error: LMCache L1=${LMCACHE_L1_SIZE_GB}GB exceeds the generated DRAM budget ${TOTAL_CPU_DRAM_GB}GB." >&2
+    exit 1
+fi
+
+LMCACHE_HTTP_PORT="${LMCACHE_HTTP_PORT:-8080}"
+LMCACHE_CMD=(
+    lmcache server
+    --host 127.0.0.1
+    --port 5555
+    --http-host 127.0.0.1
+    --http-port "$LMCACHE_HTTP_PORT"
+    --l1-size-gb "$LMCACHE_L1_SIZE_GB"
+    --l1-init-size-gb 20
+    --l1-read-ttl-seconds 7200
+    --chunk-size 1536
+    --max-workers 8
+    --eviction-trigger-watermark 0.85
+    --eviction-ratio 0.10
+    --eviction-policy LRU
+)
+printf '%q ' "${LMCACHE_CMD[@]}" > "$RESULT_DIR/lmcache_command.txt"
+printf '\n' >> "$RESULT_DIR/lmcache_command.txt"
+"${LMCACHE_CMD[@]}" > "$LMCACHE_LOG" 2>&1 &
+LMCACHE_PID=$!
+echo "LMCache server PID: $LMCACHE_PID"
+wait_for_lmcache_ready
+
+MAX_NUM_SEQS="${MAX_NUM_SEQS:-$CONC}"
+SPEC_NUM_TOKENS="${SPEC_NUM_TOKENS:-3}"
+SPEC_DRAFT_MODEL="${SPEC_DRAFT_MODEL:-Inferact/Kimi-K3-DSpark}"
+SPEC_CONFIG="{\"model\":\"${SPEC_DRAFT_MODEL}\",\"num_speculative_tokens\":${SPEC_NUM_TOKENS},\"method\":\"dspark\",\"attention_backend\":\"TRITON_MLA\",\"draft_sample_method\":\"probabilistic\",\"rejection_sample_method\":\"block\"}"
+KV_TRANSFER_CONFIG='{"kv_connector":"LMCacheMPConnector","kv_connector_module_path":"lmcache.integration.vllm.lmcache_mp_connector","kv_role":"kv_both","kv_connector_extra_config":{"lmcache.mp.host":"tcp://127.0.0.1","lmcache.mp.port":5555}}'
+
+VLLM_CMD=(
+    vllm serve "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --trust-remote-code
+    --moe-backend auto
+    --tensor-parallel-size "$TP"
+    --load-format auto
+    --gpu-memory-utilization 0.85
+    --mm-encoder-tp-mode data
+    --max-num-seqs "$MAX_NUM_SEQS"
+    --max-num-batched-tokens 3000
+    --enable-auto-tool-choice
+    --tool-call-parser kimi_k3
+    --reasoning-parser kimi_k3
+    --enable-prefix-caching
+    --mamba-cache-mode align
+    --kv-cache-dtype fp8
+    --enforce-eager
+    --speculative-config "$SPEC_CONFIG"
+    --kv-transfer-config "$KV_TRANSFER_CONFIG"
+)
+printf '%q ' "${VLLM_CMD[@]}" | tee "$RESULT_DIR/vllm_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/vllm_command.txt"
+"${VLLM_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+echo "vLLM server PID: $SERVER_PID"
+
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [ "${EVAL_ONLY:-false}" = "true" ]; then
+    run_eval --port "$PORT"
+else
+    build_replay_cmd "$RESULT_DIR"
+    run_agentic_replay_and_write_outputs "$RESULT_DIR"
+fi
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index b8f6f4fc31..875fb54469 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -548,6 +548,20 @@ kimik2.5-fp4-mi355x-vllm-agentic:
       - { tp: 4, kv-offloading: none, conc-list: [16, 24, 32, 40] }
       - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: vllm-native }, conc-list: [16, 24, 32, 40] }
 
+kimik3-fp4-mi355x-vllm-agentic:
+  image: haihub007/hc-gpu:kimik3-dspark-lmcache-working-20260804
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:mi355x-amds
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.3rc2" }, spec-decoding: mtp, conc-list: [1, 4, 8, 16, 32, 64] }
+
 kimik2.5-fp4-mi355x-atom:
   image: rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.4_202607091539
   model: amd/Kimi-K2.5-MXFP4
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 75c8ce68c1..80b53cda52 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5396,3 +5396,13 @@
     - "Bump vLLM nightly image to nightly-5e35a6f4f9bbc217c599692157ca985c894373f7 (fixed-len MiniMax-M3 fix); update B300 FP4 MTP recipe YAML container fields to match (excluding 2p1d-dep2-dep4 used by legacy-dep4)"
     - "Enable VLLM_MINIMAX_M3_MSA_DECODE_BACKEND=cutlass in prefill and decode environments"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2471
+
+- config-keys:
+    - kimik3-fp4-mi355x-vllm-agentic
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-working-20260804 image"
+    - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency 1 through 64"
+    - "Use Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, TRITON_MLA, and block rejection"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
diff --git a/runners/launch_mi355x-amds.sh b/runners/launch_mi355x-amds.sh
index 75a67c1c58..ccdd2d6879 100644
--- a/runners/launch_mi355x-amds.sh
+++ b/runners/launch_mi355x-amds.sh
@@ -306,10 +306,19 @@ else
         export HF_HUB_CACHE_MOUNT="/it-share/hf-hub-cache/"
     fi
 
+    # Kimi-K3's 1.56 TB checkpoint does not fit reliably in the node-local hub
+    # cache, so use the cluster's pre-staged NFS cache.
+    if [[ "$MODEL" == moonshotai/Kimi-K3* ]]; then
+        export HF_HUB_CACHE_MOUNT="/it-share/hf-hub-cache/"
+    fi
+
     SCRIPT_BASE="${EXP_NAME%%_*}_${PRECISION}_mi355x"
+    SCRIPT_AGENTIC_DSPARK="benchmarks/single_node/agentic/${SCRIPT_BASE}_agentic_dspark.sh"
     SCRIPT_FW="benchmarks/single_node/${SCENARIO_SUBDIR:-fixed_seq_len/}${SCRIPT_BASE}_${FRAMEWORK}${SPEC_SUFFIX}.sh"
     SCRIPT_FALLBACK="benchmarks/single_node/${SCENARIO_SUBDIR:-fixed_seq_len/}${SCRIPT_BASE}${FRAMEWORK_SUFFIX}${SPEC_SUFFIX}.sh"
-    if [[ -f "$SCRIPT_FW" ]]; then
+    if [[ "${SCENARIO_SUBDIR:-}" == "agentic/" && "$SPEC_DECODING" == "mtp" && -f "$SCRIPT_AGENTIC_DSPARK" ]]; then
+        BENCHMARK_SCRIPT="$SCRIPT_AGENTIC_DSPARK"
+    elif [[ -f "$SCRIPT_FW" ]]; then
         BENCHMARK_SCRIPT="$SCRIPT_FW"
     else
         BENCHMARK_SCRIPT="$SCRIPT_FALLBACK"

From cbe08fc255cd12e4a1ce51c108b60f5ff1468a22 Mon Sep 17 00:00:00 2001
From: haic0 <149741444+haic0@users.noreply.github.com>
Date: Wed, 5 Aug 2026 14:53:37 +0800
Subject: [PATCH 02/18] Update
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh

Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh  | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index e390f333b4..fa18e2ca69 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -151,7 +151,10 @@ LMCACHE_PID=$!
 echo "LMCache server PID: $LMCACHE_PID"
 wait_for_lmcache_ready
 
-MAX_NUM_SEQS="${MAX_NUM_SEQS:-$CONC}"
+# AgentX concurrency counts live session trees, not individual requests.
+# Subagent fan-out can push instantaneous request concurrency above CONC, so
+# leave 2x headroom rather than clipping those bursts at the scheduler.
+MAX_NUM_SEQS="${MAX_NUM_SEQS:-$((2 * CONC))}"
 SPEC_NUM_TOKENS="${SPEC_NUM_TOKENS:-3}"
 SPEC_DRAFT_MODEL="${SPEC_DRAFT_MODEL:-Inferact/Kimi-K3-DSpark}"
 SPEC_CONFIG="{\"model\":\"${SPEC_DRAFT_MODEL}\",\"num_speculative_tokens\":${SPEC_NUM_TOKENS},\"method\":\"dspark\",\"attention_backend\":\"TRITON_MLA\",\"draft_sample_method\":\"probabilistic\",\"rejection_sample_method\":\"block\"}"

From 5fdafc4f8cd1823fa78f8041a7c71f7ce8ecdb69 Mon Sep 17 00:00:00 2001
From: haic0 <149741444+haic0@users.noreply.github.com>
Date: Wed, 5 Aug 2026 14:54:52 +0800
Subject: [PATCH 03/18] Update
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh

Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index fa18e2ca69..70ef32c0b9 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -157,7 +157,11 @@ wait_for_lmcache_ready
 MAX_NUM_SEQS="${MAX_NUM_SEQS:-$((2 * CONC))}"
 SPEC_NUM_TOKENS="${SPEC_NUM_TOKENS:-3}"
 SPEC_DRAFT_MODEL="${SPEC_DRAFT_MODEL:-Inferact/Kimi-K3-DSpark}"
-SPEC_CONFIG="{\"model\":\"${SPEC_DRAFT_MODEL}\",\"num_speculative_tokens\":${SPEC_NUM_TOKENS},\"method\":\"dspark\",\"attention_backend\":\"TRITON_MLA\",\"draft_sample_method\":\"probabilistic\",\"rejection_sample_method\":\"block\"}"
+SPEC_DRAFT_MODEL_PATH="${SPEC_DRAFT_MODEL_PATH:-$SPEC_DRAFT_MODEL}"
+if [[ "$SPEC_DRAFT_MODEL_PATH" == "$SPEC_DRAFT_MODEL" ]]; then
+    hf download "$SPEC_DRAFT_MODEL"
+fi
+SPEC_CONFIG="{\"model\":\"${SPEC_DRAFT_MODEL_PATH}\",\"num_speculative_tokens\":${SPEC_NUM_TOKENS},\"method\":\"dspark\",\"attention_backend\":\"TRITON_MLA\",\"draft_sample_method\":\"probabilistic\",\"rejection_sample_method\":\"block\"}"
 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.host":"tcp://127.0.0.1","lmcache.mp.port":5555}}'
 
 VLLM_CMD=(

From d70e45fb53206339687ffabf9e7c4ac3d099a74d Mon Sep 17 00:00:00 2001
From: haic0 <149741444+haic0@users.noreply.github.com>
Date: Wed, 5 Aug 2026 14:55:05 +0800
Subject: [PATCH 04/18] Update
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh

Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh     | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index 70ef32c0b9..16c7c3b3cb 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -176,7 +176,7 @@ VLLM_CMD=(
     --gpu-memory-utilization 0.85
     --mm-encoder-tp-mode data
     --max-num-seqs "$MAX_NUM_SEQS"
-    --max-num-batched-tokens 3000
+    --max-num-batched-tokens 32768
     --enable-auto-tool-choice
     --tool-call-parser kimi_k3
     --reasoning-parser kimi_k3

From 78d73ae1ce5372ef0cf285490a7060044e7bbb07 Mon Sep 17 00:00:00 2001
From: haic0 <149741444+haic0@users.noreply.github.com>
Date: Wed, 5 Aug 2026 14:58:21 +0800
Subject: [PATCH 05/18] Update
 benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh

Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh     | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index 16c7c3b3cb..7927e55ffd 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -162,7 +162,7 @@ if [[ "$SPEC_DRAFT_MODEL_PATH" == "$SPEC_DRAFT_MODEL" ]]; then
     hf download "$SPEC_DRAFT_MODEL"
 fi
 SPEC_CONFIG="{\"model\":\"${SPEC_DRAFT_MODEL_PATH}\",\"num_speculative_tokens\":${SPEC_NUM_TOKENS},\"method\":\"dspark\",\"attention_backend\":\"TRITON_MLA\",\"draft_sample_method\":\"probabilistic\",\"rejection_sample_method\":\"block\"}"
-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.host":"tcp://127.0.0.1","lmcache.mp.port":5555}}'
+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.host":"tcp://127.0.0.1","lmcache.mp.port":5555,"lmcache.mp.mq_timeout":6000.0}}'
 
 VLLM_CMD=(
     vllm serve "$MODEL_PATH"

From 13ef7f93bea5cd7ceb5204789d9f3df8414943d9 Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Wed, 5 Aug 2026 07:03:21 +0000
Subject: [PATCH 06/18] fix(agentx): wait for clean MI355X VRAM before Kimi-K3
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Invoke the GPU-drain gate added by PR #2490 before downloading or loading Kimi-K3, preventing overlap with memory still held by a prior Slurm job.

中文：在下载或加载 Kimi-K3 前调用 PR #2490 新增的 GPU 清理门禁，避免与上一 Slurm 任务尚未释放的显存重叠。

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh   | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index 7927e55ffd..295fa7b0b9 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -35,6 +35,10 @@ if [[ -n "${ROCR_VISIBLE_DEVICES:-}" ]]; then
     export HIP_VISIBLE_DEVICES="$ROCR_VISIBLE_DEVICES"
 fi
 
+# Do not start the large Kimi-K3 load while a previous Slurm job is still
+# releasing VRAM on the allocated MI355X GPUs.
+wait_for_amd_gpu_clean
+
 # The cluster launcher mounts its persistent Hugging Face cache here. MODEL_PATH
 # may instead point at a pre-staged snapshot.
 export HF_HUB_CACHE="${HF_HUB_CACHE:-/models/huggingface_hub}"

From a2df37faa7ab36ccc48ecfa45d852d8237001743 Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Wed, 5 Aug 2026 07:32:54 +0000
Subject: [PATCH 07/18] fix(agentx): tune Kimi-K3 DSpark serving limits

Cap batched tokens for stable agentic scheduling and align the validated LMCache and speculative-decoding settings.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../agentic/kimik3_fp4_mi355x_agentic_dspark.sh        | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index 295fa7b0b9..d7d0590186 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -133,10 +133,11 @@ if [ "$LMCACHE_L1_SIZE_GB" -gt "$TOTAL_CPU_DRAM_GB" ]; then
 fi
 
 LMCACHE_HTTP_PORT="${LMCACHE_HTTP_PORT:-8080}"
+LMCACHE_MP_PORT="${LMCACHE_MP_PORT:-6000}"
 LMCACHE_CMD=(
     lmcache server
     --host 127.0.0.1
-    --port 5555
+    --port "$LMCACHE_MP_PORT"
     --http-host 127.0.0.1
     --http-port "$LMCACHE_HTTP_PORT"
     --l1-size-gb "$LMCACHE_L1_SIZE_GB"
@@ -159,14 +160,14 @@ wait_for_lmcache_ready
 # Subagent fan-out can push instantaneous request concurrency above CONC, so
 # leave 2x headroom rather than clipping those bursts at the scheduler.
 MAX_NUM_SEQS="${MAX_NUM_SEQS:-$((2 * CONC))}"
-SPEC_NUM_TOKENS="${SPEC_NUM_TOKENS:-3}"
+SPEC_NUM_TOKENS="${SPEC_NUM_TOKENS:-4}"
 SPEC_DRAFT_MODEL="${SPEC_DRAFT_MODEL:-Inferact/Kimi-K3-DSpark}"
 SPEC_DRAFT_MODEL_PATH="${SPEC_DRAFT_MODEL_PATH:-$SPEC_DRAFT_MODEL}"
 if [[ "$SPEC_DRAFT_MODEL_PATH" == "$SPEC_DRAFT_MODEL" ]]; then
     hf download "$SPEC_DRAFT_MODEL"
 fi
 SPEC_CONFIG="{\"model\":\"${SPEC_DRAFT_MODEL_PATH}\",\"num_speculative_tokens\":${SPEC_NUM_TOKENS},\"method\":\"dspark\",\"attention_backend\":\"TRITON_MLA\",\"draft_sample_method\":\"probabilistic\",\"rejection_sample_method\":\"block\"}"
-KV_TRANSFER_CONFIG='{"kv_connector":"LMCacheMPConnector","kv_connector_module_path":"lmcache.integration.vllm.lmcache_mp_connector","kv_role":"kv_both","kv_connector_extra_config":{"lmcache.mp.host":"tcp://127.0.0.1","lmcache.mp.port":5555,"lmcache.mp.mq_timeout":6000.0}}'
+KV_TRANSFER_CONFIG="{\"kv_connector\":\"LMCacheMPConnector\",\"kv_connector_module_path\":\"lmcache.integration.vllm.lmcache_mp_connector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"lmcache.mp.host\":\"tcp://127.0.0.1\",\"lmcache.mp.port\":${LMCACHE_MP_PORT}}}"
 
 VLLM_CMD=(
     vllm serve "$MODEL_PATH"
@@ -180,14 +181,13 @@ VLLM_CMD=(
     --gpu-memory-utilization 0.85
     --mm-encoder-tp-mode data
     --max-num-seqs "$MAX_NUM_SEQS"
-    --max-num-batched-tokens 32768
+    --max-num-batched-tokens 3000
     --enable-auto-tool-choice
     --tool-call-parser kimi_k3
     --reasoning-parser kimi_k3
     --enable-prefix-caching
     --mamba-cache-mode align
     --kv-cache-dtype fp8
-    --enforce-eager
     --speculative-config "$SPEC_CONFIG"
     --kv-transfer-config "$KV_TRANSFER_CONFIG"
 )

From 1ecd3c07cd6c47f4ffbb0216fee9d0f8e4cdce34 Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Wed, 5 Aug 2026 07:34:06 +0000
Subject: [PATCH 08/18] docs(agentx): align Kimi-K3 benchmark settings

Keep the performance changelog synchronized with the validated scheduler and speculative-decoding configuration.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 perf-changelog.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 7ed3a5841f..76cdef47af 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5457,5 +5457,5 @@
   description:
     - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-working-20260804 image"
     - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency 1 through 64"
-    - "Use Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, TRITON_MLA, and block rejection"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX
+    - "Use Inferact/Kimi-K3-DSpark with four speculative tokens, probabilistic drafting, TRITON_MLA, block rejection, and a 3000-token scheduler cap"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2496

From 040546012c8306bbfea39207cf9d6d0dcacadc29 Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Wed, 5 Aug 2026 08:50:17 +0000
Subject: [PATCH 09/18] fix(agentx): patch Kimi-K3 DSpark MLA verification

Apply the pinned upstream FP8 verification fix, reduce speculative depth, and use eager execution so the MI355X sweep can start reliably.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../kimik3_fp4_mi355x_agentic_dspark.sh       | 45 ++++++++++++++++++-
 docs/waiver/2496.md                           | 33 ++++++++++++++
 perf-changelog.yaml                           |  3 +-
 3 files changed, 78 insertions(+), 3 deletions(-)
 create mode 100644 docs/waiver/2496.md

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index d7d0590186..1bcf8efa36 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -57,6 +57,46 @@ amd-smi || true
 resolve_trace_source
 install_agentic_deps
 
+# Apply the upstream Kimi-K3 DSpark FP8 MLA verification fix until the pinned
+# image includes vLLM PR #50619. The immutable compare URL and checksum make
+# this reproducible; git-apply checks fail loudly if the image has drifted.
+apply_vllm_pr_50619() {
+    local base_sha="38a466e7b6e087d67c35e7f924c04c245423c99f"
+    local head_sha="a9f790851885589ea41b2b99ee750e044573bebd"
+    local patch_sha256="fa7200baae57af574ad018a92fff91430e864bd5609772a89e75fd9743c5abce"
+    local patch_file vllm_site_parent
+
+    patch_file="$(mktemp)"
+    curl --fail --location --retry 3 --silent --show-error \
+        "https://github.com/vllm-project/vllm/compare/${base_sha}...${head_sha}.patch" \
+        --output "$patch_file"
+    echo "${patch_sha256}  ${patch_file}" | sha256sum --check --status
+
+    vllm_site_parent="$(python3 - <<'PY'
+import importlib.util
+from pathlib import Path
+
+spec = importlib.util.find_spec("vllm")
+if spec is None or not spec.submodule_search_locations:
+    raise SystemExit("Unable to locate the installed vLLM package")
+print(Path(next(iter(spec.submodule_search_locations))).resolve().parent)
+PY
+)"
+
+    if git -C "$vllm_site_parent" apply --reverse --check \
+            --include='vllm/**' "$patch_file" 2>/dev/null; then
+        echo "vLLM PR #50619 is already applied"
+    else
+        git -C "$vllm_site_parent" apply --check \
+            --include='vllm/**' "$patch_file"
+        git -C "$vllm_site_parent" apply \
+            --include='vllm/**' "$patch_file"
+        echo "Applied vLLM PR #50619 at ${head_sha}"
+    fi
+    rm -f "$patch_file"
+}
+apply_vllm_pr_50619
+
 # Environment used by the validated Kimi-K3 + DSpark + LMCache image.
 export PYTHONNOUSERSITE=1
 export PYTHONHASHSEED=42
@@ -65,7 +105,6 @@ export VLLM_ROCM_USE_AITER_MLA=1
 export AITER_SITUV2_A8W4=1
 export AITER_BF16_FP8_MOE_BOUND=0
 export VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION=1
-export VLLM_USE_BREAKABLE_CUDAGRAPH=0
 export VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS="${VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS:-1200}"
 export VLLM_ENGINE_READY_TIMEOUT_S="${VLLM_ENGINE_READY_TIMEOUT_S:-7200}"
 export HSA_NO_SCRATCH_RECLAIM=1
@@ -160,7 +199,7 @@ wait_for_lmcache_ready
 # Subagent fan-out can push instantaneous request concurrency above CONC, so
 # leave 2x headroom rather than clipping those bursts at the scheduler.
 MAX_NUM_SEQS="${MAX_NUM_SEQS:-$((2 * CONC))}"
-SPEC_NUM_TOKENS="${SPEC_NUM_TOKENS:-4}"
+SPEC_NUM_TOKENS="${SPEC_NUM_TOKENS:-3}"
 SPEC_DRAFT_MODEL="${SPEC_DRAFT_MODEL:-Inferact/Kimi-K3-DSpark}"
 SPEC_DRAFT_MODEL_PATH="${SPEC_DRAFT_MODEL_PATH:-$SPEC_DRAFT_MODEL}"
 if [[ "$SPEC_DRAFT_MODEL_PATH" == "$SPEC_DRAFT_MODEL" ]]; then
@@ -188,6 +227,8 @@ VLLM_CMD=(
     --enable-prefix-caching
     --mamba-cache-mode align
     --kv-cache-dtype fp8
+    # Keep eager execution until the patched graph path is validated in this image.
+    --enforce-eager
     --speculative-config "$SPEC_CONFIG"
     --kv-transfer-config "$KV_TRANSFER_CONFIG"
 )
diff --git a/docs/waiver/2496.md b/docs/waiver/2496.md
new file mode 100644
index 0000000000..5ac4c51723
--- /dev/null
+++ b/docs/waiver/2496.md
@@ -0,0 +1,33 @@
+# Serving-stack patch waiver for PR #2496
+
+## What is patched
+
+The Kimi-K3 MI355X AgentX launcher applies the production-file portion of
+[vLLM PR #50619](https://github.com/vllm-project/vllm/pull/50619) to the vLLM
+installation in `haihub007/hc-gpu:kimik3-dspark-lmcache-working-20260804`.
+The patch is pinned to base `38a466e7b6e087d67c35e7f924c04c245423c99f`,
+head `a9f790851885589ea41b2b99ee750e044573bebd`, and a SHA-256 checksum. It fixes
+Kimi-K3 TP8 DSpark FP8 MLA verification by preserving native causal
+multi-token Gluon queries and the target model's CUDA-graph policy.
+
+## Why the unmodified image cannot run this benchmark
+
+The image's current vLLM expands each DSpark verification token into an
+independent Gluon MLA row. During engine warmup, 16 requests with the
+four-token draft configuration produced 144 rows and exceeded the kernel's
+maximum batch size of 128, preventing EngineCore from starting. Reducing the
+draft to three tokens and temporarily enabling eager execution lowers risk,
+but neither is a complete serving-stack correction for multi-token FP8 MLA
+verification.
+
+The failure is recorded in
+[run 30985536172](https://github.com/SemiAnalysisAI/InferenceX/actions/runs/30985536172/job/92239885715).
+The upstream Kimi-K3 ROCm roadmap is tracked in
+[vLLM issue #50682](https://github.com/vllm-project/vllm/issues/50682).
+
+## Removal plan
+
+Remove the runtime patch and this waiver after PR #50619, or its upstream
+replacement, is merged and included in an upstream `vllm/vllm-openai-rocm`
+image validated for Kimi-K3 TP8 DSpark with FP8 KV cache and LMCache. Replace
+the personal image with that upstream image in the same update.
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 76cdef47af..9d1974c6b3 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5457,5 +5457,6 @@
   description:
     - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-working-20260804 image"
     - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency 1 through 64"
-    - "Use Inferact/Kimi-K3-DSpark with four speculative tokens, probabilistic drafting, TRITON_MLA, block rejection, and a 3000-token scheduler cap"
+    - "Use Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, TRITON_MLA, block rejection, eager execution, and a 3000-token scheduler cap"
+    - "Apply the pinned vLLM PR #50619 fix for native causal Kimi-K3 DSpark FP8 MLA verification"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2496

From 5bd0fe7fa0f20b8e101bf28bb51dec0ff01c404b Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Wed, 5 Aug 2026 10:12:19 +0000
Subject: [PATCH 10/18] fix(agentx): use Kimi-K3 image with DSpark fix

Switch to the image with vLLM PR #50619 baked in so runners no longer depend on a fragile runtime patch.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../kimik3_fp4_mi355x_agentic_dspark.sh       | 40 -------------------
 configs/amd-master.yaml                       |  2 +-
 docs/waiver/2496.md                           | 23 ++++++-----
 perf-changelog.yaml                           |  4 +-
 4 files changed, 15 insertions(+), 54 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index 1bcf8efa36..be47e680e6 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -57,46 +57,6 @@ amd-smi || true
 resolve_trace_source
 install_agentic_deps
 
-# Apply the upstream Kimi-K3 DSpark FP8 MLA verification fix until the pinned
-# image includes vLLM PR #50619. The immutable compare URL and checksum make
-# this reproducible; git-apply checks fail loudly if the image has drifted.
-apply_vllm_pr_50619() {
-    local base_sha="38a466e7b6e087d67c35e7f924c04c245423c99f"
-    local head_sha="a9f790851885589ea41b2b99ee750e044573bebd"
-    local patch_sha256="fa7200baae57af574ad018a92fff91430e864bd5609772a89e75fd9743c5abce"
-    local patch_file vllm_site_parent
-
-    patch_file="$(mktemp)"
-    curl --fail --location --retry 3 --silent --show-error \
-        "https://github.com/vllm-project/vllm/compare/${base_sha}...${head_sha}.patch" \
-        --output "$patch_file"
-    echo "${patch_sha256}  ${patch_file}" | sha256sum --check --status
-
-    vllm_site_parent="$(python3 - <<'PY'
-import importlib.util
-from pathlib import Path
-
-spec = importlib.util.find_spec("vllm")
-if spec is None or not spec.submodule_search_locations:
-    raise SystemExit("Unable to locate the installed vLLM package")
-print(Path(next(iter(spec.submodule_search_locations))).resolve().parent)
-PY
-)"
-
-    if git -C "$vllm_site_parent" apply --reverse --check \
-            --include='vllm/**' "$patch_file" 2>/dev/null; then
-        echo "vLLM PR #50619 is already applied"
-    else
-        git -C "$vllm_site_parent" apply --check \
-            --include='vllm/**' "$patch_file"
-        git -C "$vllm_site_parent" apply \
-            --include='vllm/**' "$patch_file"
-        echo "Applied vLLM PR #50619 at ${head_sha}"
-    fi
-    rm -f "$patch_file"
-}
-apply_vllm_pr_50619
-
 # Environment used by the validated Kimi-K3 + DSpark + LMCache image.
 export PYTHONNOUSERSITE=1
 export PYTHONHASHSEED=42
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index f34ee0bd62..7ffc7cd43b 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -465,7 +465,7 @@ kimik2.5-fp4-mi355x-vllm:
       - { tp: 4, conc-start: 4, conc-end: 128 }
 
 kimik3-fp4-mi355x-vllm-agentic:
-  image: haihub007/hc-gpu:kimik3-dspark-lmcache-working-20260804
+  image: haihub007/hc-gpu:kimik3-dspark-lmcache-pr50619-20260805
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:mi355x-amds
diff --git a/docs/waiver/2496.md b/docs/waiver/2496.md
index 5ac4c51723..47b81279a9 100644
--- a/docs/waiver/2496.md
+++ b/docs/waiver/2496.md
@@ -2,13 +2,14 @@
 
 ## What is patched
 
-The Kimi-K3 MI355X AgentX launcher applies the production-file portion of
-[vLLM PR #50619](https://github.com/vllm-project/vllm/pull/50619) to the vLLM
-installation in `haihub007/hc-gpu:kimik3-dspark-lmcache-working-20260804`.
-The patch is pinned to base `38a466e7b6e087d67c35e7f924c04c245423c99f`,
-head `a9f790851885589ea41b2b99ee750e044573bebd`, and a SHA-256 checksum. It fixes
-Kimi-K3 TP8 DSpark FP8 MLA verification by preserving native causal
-multi-token Gluon queries and the target model's CUDA-graph policy.
+The Kimi-K3 MI355X AgentX configuration uses
+`haihub007/hc-gpu:kimik3-dspark-lmcache-pr50619-20260805`, which contains the
+production-file portion of
+[vLLM PR #50619](https://github.com/vllm-project/vllm/pull/50619) rebased onto
+the image's vLLM revision `124154a88`. It fixes Kimi-K3 TP8 DSpark FP8 MLA
+verification by preserving native causal multi-token Gluon queries and the
+target model's CUDA-graph policy. The published image digest is
+`sha256:3b5b0131f89be08d34d431f2c4e45e32606f9dff0e3aace79272489e788e6eda`.
 
 ## Why the unmodified image cannot run this benchmark
 
@@ -27,7 +28,7 @@ The upstream Kimi-K3 ROCm roadmap is tracked in
 
 ## Removal plan
 
-Remove the runtime patch and this waiver after PR #50619, or its upstream
-replacement, is merged and included in an upstream `vllm/vllm-openai-rocm`
-image validated for Kimi-K3 TP8 DSpark with FP8 KV cache and LMCache. Replace
-the personal image with that upstream image in the same update.
+Remove this waiver after PR #50619, or its upstream replacement, is merged and
+included in an upstream `vllm/vllm-openai-rocm` image validated for Kimi-K3
+TP8 DSpark with FP8 KV cache and LMCache. Replace the personal image with that
+upstream image in the same update.
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 9d1974c6b3..143caf7c02 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5455,8 +5455,8 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-working-20260804 image"
+    - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-pr50619-20260805 image"
     - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency 1 through 64"
     - "Use Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, TRITON_MLA, block rejection, eager execution, and a 3000-token scheduler cap"
-    - "Apply the pinned vLLM PR #50619 fix for native causal Kimi-K3 DSpark FP8 MLA verification"
+    - "Use vLLM PR #50619 rebased onto the image's vLLM revision for native causal Kimi-K3 DSpark FP8 MLA verification"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2496

From 091b396b1ec8a67a547b279a83e193e3a15c78bc Mon Sep 17 00:00:00 2001
From: haic0 <149741444+haic0@users.noreply.github.com>
Date: Wed, 5 Aug 2026 22:07:34 +0800
Subject: [PATCH 11/18] Update amd-master.yaml

---
 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 7ffc7cd43b..d82d763dce 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -476,7 +476,7 @@ kimik3-fp4-mi355x-vllm-agentic:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.3rc2" }, spec-decoding: mtp, conc-list: [1, 4, 8, 16, 32, 64] }
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.3rc2" }, spec-decoding: mtp, conc-list: [1, 4, 8] }
 
 kimik2.5-fp4-mi355x-atom:
   image: rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.4_202607091539

From 65bd3c4030f7b800da57e48a51106a4c31b4887c Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Thu, 6 Aug 2026 02:22:31 +0000
Subject: [PATCH 12/18] fix(agentx): pin v1.0.0 and skip fast-run evals

Pin AgentX to the requested v1.0.0 release and keep agentx-fast sweeps throughput-only so short validation runs do not dispatch a separate GSM8K evaluation.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .github/workflows/run-sweep.yml | 1 +
 utils/aiperf                    | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/.github/workflows/run-sweep.yml b/.github/workflows/run-sweep.yml
index b46c4c5901..89a7ed5a83 100644
--- a/.github/workflows/run-sweep.yml
+++ b/.github/workflows/run-sweep.yml
@@ -747,6 +747,7 @@ jobs:
         if: >-
             ${{
               !cancelled() &&
+              !contains(github.event.pull_request.labels.*.name, 'agentx-fast') &&
               needs.setup.result == 'success' &&
               needs.setup.outputs.reuse-enabled != 'true' &&
               (needs.canary-sweep.result == 'success' || needs.canary-sweep.result == 'skipped') &&
diff --git a/utils/aiperf b/utils/aiperf
index b7b16cf851..818c3a5a29 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit b7b16cf851885567988a643282266bce74e34437
+Subproject commit 818c3a5a2922c535af6271ff296ed374e292b8e4

From 8244d9d1d65b212834a8628b10a88aba74b91746 Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Thu, 6 Aug 2026 03:57:18 +0000
Subject: [PATCH 13/18] fix(agentx): use patched Kimi-K3 ROCm image

Move the MI355X sweep to the image containing the skinny-GEMM, FP8 MLA prefill, DSpark verification, and large-KV fixes, while disabling the unsafe skinny-GEMM path as a defensive fallback.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh     | 1 +
 configs/amd-master.yaml                                         | 2 +-
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index be47e680e6..3f4cbb8401 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -62,6 +62,7 @@ export PYTHONNOUSERSITE=1
 export PYTHONHASHSEED=42
 export VLLM_ROCM_USE_AITER=1
 export VLLM_ROCM_USE_AITER_MLA=1
+export VLLM_ROCM_USE_SKINNY_GEMM=0
 export AITER_SITUV2_A8W4=1
 export AITER_BF16_FP8_MOE_BOUND=0
 export VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION=1
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index d82d763dce..d4410fe0da 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -465,7 +465,7 @@ kimik2.5-fp4-mi355x-vllm:
       - { tp: 4, conc-start: 4, conc-end: 128 }
 
 kimik3-fp4-mi355x-vllm-agentic:
-  image: haihub007/hc-gpu:kimik3-dspark-lmcache-pr50619-20260805
+  image: haihub007/hc-gpu:kimik3-dspark-lmcache-pr50618-pr51040-aiter4474-20260806
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:mi355x-amds

From 23bff1a4a1eda71a36c682105ca6cc18149f5b93 Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Thu, 6 Aug 2026 05:51:39 +0000
Subject: [PATCH 14/18] fix(agentx): rerun patched Kimi-K3 at c2

Use a single concurrency-2 AgentX validation with the rebuilt combined ROCm image and keep the changelog aligned with its integrated fixes.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 configs/amd-master.yaml | 2 +-
 perf-changelog.yaml     | 6 +++---
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index d4410fe0da..92619c5c1f 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -476,7 +476,7 @@ kimik3-fp4-mi355x-vllm-agentic:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.3rc2" }, spec-decoding: mtp, conc-list: [1, 4, 8] }
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.3rc2" }, spec-decoding: mtp, conc-list: [2] }
 
 kimik2.5-fp4-mi355x-atom:
   image: rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.4_202607091539
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index af54eb32fd..0c56d3417f 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5465,8 +5465,8 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-pr50619-20260805 image"
-    - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency 1 through 64"
+    - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-pr50618-pr51040-aiter4474-20260806 image"
+    - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency 2"
     - "Use Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, TRITON_MLA, block rejection, eager execution, and a 3000-token scheduler cap"
-    - "Use vLLM PR #50619 rebased onto the image's vLLM revision for native causal Kimi-K3 DSpark FP8 MLA verification"
+    - "Combine vLLM PRs #50618, #50619, and #51040 with AITER PR #4474, remove the integration import shadow, and disable skinny GEMM"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2496

From 0ccacf664a724435a5a39b23e7e1fd4d1ae0097d Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Thu, 6 Aug 2026 06:23:10 +0000
Subject: [PATCH 15/18] fix(agentx): use immutable Kimi-K3 shadowfix image

Move the c2 validation to a fresh image tag so MI355X squashfs caching cannot reuse the pre-shadow-fix container.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 configs/amd-master.yaml | 2 +-
 perf-changelog.yaml     | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 92619c5c1f..7c4465277d 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -465,7 +465,7 @@ kimik2.5-fp4-mi355x-vllm:
       - { tp: 4, conc-start: 4, conc-end: 128 }
 
 kimik3-fp4-mi355x-vllm-agentic:
-  image: haihub007/hc-gpu:kimik3-dspark-lmcache-pr50618-pr51040-aiter4474-20260806
+  image: haihub007/hc-gpu:kimik3-dspark-lmcache-pr50618-pr51040-aiter4474-20260806-shadowfix1
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:mi355x-amds
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0c56d3417f..40f51672ed 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5465,7 +5465,7 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-pr50618-pr51040-aiter4474-20260806 image"
+    - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-pr50618-pr51040-aiter4474-20260806-shadowfix1 image"
     - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency 2"
     - "Use Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, TRITON_MLA, block rejection, eager execution, and a 3000-token scheduler cap"
     - "Combine vLLM PRs #50618, #50619, and #51040 with AITER PR #4474, remove the integration import shadow, and disable skinny GEMM"

From a53c228028fa8873f066ddeef9ca5355f53340e8 Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Thu, 6 Aug 2026 07:28:46 +0000
Subject: [PATCH 16/18] fix(agentx): use Kimi-K3 MLA fallback image

Route the c2 validation through the image that keeps 12-head single-token decode on Gluon, avoiding the unavailable padded AITER ASM kernel.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 configs/amd-master.yaml | 2 +-
 perf-changelog.yaml     | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 7c4465277d..eef2e01fc2 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -465,7 +465,7 @@ kimik2.5-fp4-mi355x-vllm:
       - { tp: 4, conc-start: 4, conc-end: 128 }
 
 kimik3-fp4-mi355x-vllm-agentic:
-  image: haihub007/hc-gpu:kimik3-dspark-lmcache-pr50618-pr51040-aiter4474-20260806-shadowfix1
+  image: haihub007/hc-gpu:kimik3-dspark-lmcache-fix2
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:mi355x-amds
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 40f51672ed..131fa0862b 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5465,7 +5465,7 @@
   scenario-type:
     - agentic-coding
   description:
-    - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-pr50618-pr51040-aiter4474-20260806-shadowfix1 image"
+    - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-fix2 image"
     - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency 2"
     - "Use Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, TRITON_MLA, block rejection, eager execution, and a 3000-token scheduler cap"
     - "Combine vLLM PRs #50618, #50619, and #51040 with AITER PR #4474, remove the integration import shadow, and disable skinny GEMM"

From c3f12c3bc919bf52d051632a222e6852c01fff7a Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Thu, 6 Aug 2026 11:54:56 +0000
Subject: [PATCH 17/18] perf(agentx): expand Kimi-K3 concurrency sweep

Exercise graph execution with prefix caching across concurrency 2 through 32 now that the MLA fallback image completes the c2 run.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh   | 2 --
 configs/amd-master.yaml                                       | 2 +-
 perf-changelog.yaml                                           | 4 ++--
 3 files changed, 3 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index 3f4cbb8401..83e2121f29 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -188,8 +188,6 @@ VLLM_CMD=(
     --enable-prefix-caching
     --mamba-cache-mode align
     --kv-cache-dtype fp8
-    # Keep eager execution until the patched graph path is validated in this image.
-    --enforce-eager
     --speculative-config "$SPEC_CONFIG"
     --kv-transfer-config "$KV_TRANSFER_CONFIG"
 )
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index eef2e01fc2..1f6a6f5b60 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -476,7 +476,7 @@ kimik3-fp4-mi355x-vllm-agentic:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.3rc2" }, spec-decoding: mtp, conc-list: [2] }
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.3rc2" }, spec-decoding: mtp, conc-list: [2, 4, 8, 16, 32] }
 
 kimik2.5-fp4-mi355x-atom:
   image: rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.4_202607091539
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 131fa0862b..0e996567d7 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5466,7 +5466,7 @@
     - agentic-coding
   description:
     - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-fix2 image"
-    - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency 2"
-    - "Use Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, TRITON_MLA, block rejection, eager execution, and a 3000-token scheduler cap"
+    - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency sweep [2, 4, 8, 16, 32]"
+    - "Use Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, TRITON_MLA, block rejection, default graph execution, and a 3000-token scheduler cap"
     - "Combine vLLM PRs #50618, #50619, and #51040 with AITER PR #4474, remove the integration import shadow, and disable skinny GEMM"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2496

From 22f2aa523b8f10be7ec6b70dcb9082e4ab5ab357 Mon Sep 17 00:00:00 2001
From: "haichzha@amd.com" <haichzha@amd.com@mia1-p01-g18.mia.tensorwave.lan>
Date: Thu, 6 Aug 2026 14:51:35 +0000
Subject: [PATCH 18/18] fix(agentx): isolate Kimi-K3 piecewise graph run

Use TRITON_MLA with piecewise graph capture and fused RMS norm gating at concurrency 4 to narrow the prior full-graph sampling failure.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh   | 2 ++
 configs/amd-master.yaml                                       | 2 +-
 perf-changelog.yaml                                           | 4 ++--
 3 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
index 83e2121f29..f86234bf2d 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_mi355x_agentic_dspark.sh
@@ -188,6 +188,8 @@ VLLM_CMD=(
     --enable-prefix-caching
     --mamba-cache-mode align
     --kv-cache-dtype fp8
+    --attention-backend TRITON_MLA
+    --compilation-config '{"cudagraph_mode":"PIECEWISE","custom_ops":["+fused_rms_norm_gated"]}'
     --speculative-config "$SPEC_CONFIG"
     --kv-transfer-config "$KV_TRANSFER_CONFIG"
 )
diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml
index 1f6a6f5b60..e529c5e04c 100644
--- a/configs/amd-master.yaml
+++ b/configs/amd-master.yaml
@@ -476,7 +476,7 @@ kimik3-fp4-mi355x-vllm-agentic:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.3rc2" }, spec-decoding: mtp, conc-list: [2, 4, 8, 16, 32] }
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.3rc2" }, spec-decoding: mtp, conc-list: [4] }
 
 kimik2.5-fp4-mi355x-atom:
   image: rocm/atom:rocm7.2.4_ubuntu24.04_py3.12_pytorch_release_2.10.0_atom0.1.4_202607091539
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 0e996567d7..1fb34de760 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5466,7 +5466,7 @@
     - agentic-coding
   description:
     - "Add Kimi-K3 MXFP4 DSpark speculative decoding on TP8 MI355X with the validated haihub007/hc-gpu:kimik3-dspark-lmcache-fix2 image"
-    - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency sweep [2, 4, 8, 16, 32]"
-    - "Use Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, TRITON_MLA, block rejection, default graph execution, and a 3000-token scheduler cap"
+    - "Run LMCache 0.5.3rc2 MP host-DRAM offload with a 906 GB L1 pool, FP8 KV cache, Mamba alignment, and concurrency 4"
+    - "Use TRITON_MLA with PIECEWISE CUDA graphs, fused RMS norm gating, and Inferact/Kimi-K3-DSpark with three speculative tokens, probabilistic drafting, block rejection, and a 3000-token scheduler cap"
     - "Combine vLLM PRs #50618, #50619, and #51040 with AITER PR #4474, remove the integration import shadow, and disable skinny GEMM"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2496