From 36515f7d008a8d692b13411969ae74da860f658d Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Wed, 15 Jul 2026 17:39:52 -0700
Subject: [PATCH 1/9] feat(agentic): add LMCache arm to tuned DSV4 FP4 B300
 vLLM AgentX recipe
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Combine the tuned B300 recipe from PR #2225 (nightly image, sparse DSV4
FlashInfer attention, FULL_DECODE_ONLY CUDA graphs, AMXF4 mega-MoE) with
the LMCache 0.5.1 KV-offload backend from PR #2153, mirroring the B200
arm in PR #2231. The lmcache arm keeps the stock PyTorch caching
allocator instead of PYTORCH_ALLOC_CONF=expandable_segments:True
(expandable-segment cuMem/VMM allocations cannot be CUDA-IPC-exported to
the LMCache MP server, the same failure mode as --enable-cumem-allocator
on B200) and runs otherwise identical serving flags, so backends are
directly comparable. Adds a standalone dsv4-fp4-b300-vllm-agentic-lmcache
config section mirroring the vllm-simple DEP4/DEP8 concurrency ladders
and a perf-changelog entry triggering it.

中文：将 PR #2225 的 B300 调优配方（nightly 镜像、稀疏 DSV4 FlashInfer
注意力、FULL_DECODE_ONLY CUDA graph、AMXF4 mega-MoE）与 PR #2153 的
LMCache 0.5.1 KV 卸载后端合并，对应 B200 的 PR #2231。lmcache 分支仅以
默认 PyTorch 缓存分配器替代 PYTORCH_ALLOC_CONF=expandable_segments:True
（expandable segments 的 cuMem/VMM 分配无法通过 CUDA IPC 导出给 LMCache
MP server，与 B200 上 --enable-cumem-allocator 的失败模式相同），其余
serving 参数与其他分支保持一致，便于卸载后端间直接对比。新增独立的
dsv4-fp4-b300-vllm-agentic-lmcache 配置（测试点与 vllm-simple 的
DEP4/DEP8 阶梯对齐）及触发它的 perf-changelog 条目。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 .../single_node/agentic/dsv4_fp4_b300_vllm.sh | 306 +++++++++++++-----
 configs/nvidia-master.yaml                    |  40 ++-
 perf-changelog.yaml                           |  11 +
 3 files changed, 260 insertions(+), 97 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh b/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh
index 3a5be7b578..30c6a76836 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh
@@ -1,60 +1,47 @@
 #!/usr/bin/env bash
-set -euo pipefail
+set -eo pipefail
 set -x
 
 # Agentic trace replay benchmark for DeepSeek-V4-Pro FP4 on B300 using vLLM.
-# Mirrors the fixed-seq-len parallelism options (pure TP and DEP) so the
-# agentic sweep can probe both interactivity and throughput regimes:
-#   pure TP (DP_ATTENTION=false, EP_SIZE=1):  attention TP-sharded across
-#       all $TP GPUs in a single engine. Lower TPOT, lower batch.
-#   TP+EP   (DP_ATTENTION=false, EP_SIZE>1):  attention TP-sharded, MoE
-#       experts EP-sharded within the TP group.
-#   DEP     (DP_ATTENTION=true, EP_SIZE>1):   per-DP-rank attention with
-#       experts EP-sharded across DP ranks (per the vLLM blog recipe).
-#       Highest aggregate throughput at large CONC.
+# v4pro-b300.yaml TP4 and DEP4 recipe. SimpleCPUOffload / MooncakeStore /
+# LMCache
 #
-# Image is vllm/vllm-openai:v0.20.0-cu130. block_size=256, kv-cache-dtype=fp8,
-# FP4 indexer cache enabled, FULL_AND_PIECEWISE cudagraph capture with
-# custom_ops=all (per the vLLM blog recipe at https://vllm.ai/blog/deepseek-v4).
+# Image is configured in nvidia-master.yaml. The recipe uses FP8 KV cache,
+# sparse DeepSeek-V4 FlashInfer attention with an FP4 indexer cache, mega-MoE,
+# and FULL_DECODE_ONLY CUDA graphs with every batch size captured explicitly.
 #
 # Required env vars:
 #   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR
 #
-# KV_OFFLOADING=dram requires KV_OFFLOAD_BACKEND=mooncake.
+# KV_OFFLOADING=none is used by TP4. DEP4 uses KV_OFFLOADING=dram with
+# KV_OFFLOAD_BACKEND=vllm-simple, mooncake, or lmcache.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
 check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION EP_SIZE DP_ATTENTION
 
-DCP_SIZE="${DCP_SIZE:-1}"
-PCP_SIZE="${PCP_SIZE:-1}"
-VLLM_CP_ARGS=()
-if [ "$DCP_SIZE" -gt 1 ]; then
-    VLLM_CP_ARGS+=(--decode-context-parallel-size "$DCP_SIZE")
-fi
-if [ "$PCP_SIZE" -gt 1 ]; then
-    VLLM_CP_ARGS+=(--prefill-context-parallel-size "$PCP_SIZE")
-fi
-
-GPU_COUNT="${GPU_COUNT:-$((TP * PCP_SIZE))}"
+GPU_COUNT=$TP
 if [[ ! "$GPU_COUNT" =~ ^[1-9][0-9]*$ ]]; then
     echo "Error: GPU_COUNT must be a positive integer, got '$GPU_COUNT'" >&2
     exit 1
 fi
 export GPU_COUNT
 
-if declare -p SLURM_JOB_ID >/dev/null 2>&1 && [ -n "$SLURM_JOB_ID" ]; then
-    SLURM_NODE=unknown
-    if declare -p SLURMD_NODENAME >/dev/null 2>&1 && [ -n "$SLURMD_NODENAME" ]; then
-        SLURM_NODE="$SLURMD_NODENAME"
-    fi
-    echo "JOB $SLURM_JOB_ID running on $SLURM_NODE"
+# Under DP-attention the DP world size equals TP, and the DEP recipe sizes
+# per-rank batch as MAX_NUM_SEQS = 2*CONC/TP, which must be an integer.
+if [ "$DP_ATTENTION" = "true" ] && [ $((2 * CONC % TP)) -ne 0 ]; then
+    echo "Error: DEP requires 2*CONC divisible by TP, got CONC='$CONC' and TP='$TP'" >&2
+    exit 1
+fi
+
+if [[ -n "$SLURM_JOB_ID" ]]; then
+    echo "JOB $SLURM_JOB_ID running on $SLURMD_NODENAME"
 fi
 
 # `hf download` creates the target dir if missing and is itself idempotent.
-# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
+# When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE.
 # Either way, MODEL_PATH is what the server is launched with.
-if declare -p MODEL_PATH >/dev/null 2>&1 && [ -n "$MODEL_PATH" ]; then
+if [[ -n "$MODEL_PATH" ]]; then
     if [[ ! -d "$MODEL_PATH" || -z "$(ls -A "$MODEL_PATH" 2>/dev/null)" ]]; then
         hf download "$MODEL" --local-dir "$MODEL_PATH"
     fi
@@ -68,17 +55,9 @@ nvidia-smi
 resolve_trace_source
 install_agentic_deps
 
-# vLLM v0.22.1 can ship CUTLASS DSL 4.5.2 with stale native MLIR bindings,
-# which fails DSV4 indexer compilation with mlir_global_dtors(..., data).
-# Reinstall the matching native wheel until NVIDIA/cutlass#3259 is resolved.
-agentic_pip_install --quiet --force-reinstall --no-deps \
-    'nvidia-cutlass-dsl-libs-cu13==4.5.2'
-
 # vllm-project/router expands the one HTTP backend into one logical worker per
-# DP rank and sends X-data-parallel-rank on forwarded requests. aiperf's
-# X-Correlation-ID is stable for every turn of a conversation; alias it to the
-# router's preferred X-Session-ID header. This also keeps affinity correct when
-# testing older wheels that prioritize per-request X-Request-ID.
+# DP rank. Bind every turn of a conversation to the same rank by mapping
+# AIPerf's stable correlation ID to the router's X-Session-ID header.
 USE_VLLM_ROUTER=false
 VLLM_BACKEND_PORT="$PORT"
 if [ "$DP_ATTENTION" = "true" ]; then
@@ -91,31 +70,60 @@ if [ "$DP_ATTENTION" = "true" ]; then
     agentic_pip_install --quiet "vllm-router==$VLLM_ROUTER_VERSION"
 fi
 
-# DeepSeek-V4-Pro weights are large; engine startup can exceed default 600s.
+# Match the environment used by v4pro-b300.yaml.
+export VLLM_USE_V2_MODEL_RUNNER=1
 export VLLM_ENGINE_READY_TIMEOUT_S=3600
-
-# vllm-project/vllm#43447 keeps local SWA prefix-cache tails sparsely, while
-# vllm-project/vllm#44774 applies the same reachability policy to Mooncake's
-# store mask. 32k matches the trace-replay tuning validated for this workload.
 export VLLM_PREFIX_CACHE_RETENTION_INTERVAL=32768
+export VLLM_DSV4_MEGA_FP8_COMBINE=1
+export NCCL_NVLS_ENABLE=1
+export VLLM_USE_RUST_FRONTEND=1
 
 # ---- Server config ----------------------------------------------------------
 SERVER_LOG="$RESULT_DIR/server.log"
 ROUTER_LOG="$RESULT_DIR/router.log"
 MOONCAKE_MASTER_LOG="$RESULT_DIR/mooncake_master.log"
+LMCACHE_SERVER_LOG="$RESULT_DIR/lmcache_server.log"
 mkdir -p "$RESULT_DIR"
 
 SERVER_PID=""
 ROUTER_PID=""
 MOONCAKE_MASTER_PID=""
+LMCACHE_SERVER_PID=""
 
+# The generated TOTAL_CPU_DRAM_GB budget is proportional to allocated GPUs.
+# On cluster:b300-nv, dram-utilization=0.80 and DEP4 resolve to roughly the
+# source recipe's 280 GiB per DP rank. TP4 remains GPU-resident.
 OFFLOAD_ARGS=()
-if require_agentic_kv_offload_backend mooncake; then
-        # Mooncake embedded mode contributes one global segment per GPU rank to
-        # a shared distributed store. Pre-divide the aggregate host budget
-        # across those rank-contributed segments.
+case "$KV_OFFLOAD_BACKEND" in
+    "")
+        require_agentic_kv_offload_none
+        ;;
+    vllm-simple)
+        require_agentic_kv_offload_backend vllm-simple
+        CPU_BYTES_PER_RANK=$(( TOTAL_CPU_DRAM_GB * 1000 * 1000 * 1000 / GPU_COUNT ))
+        # Identical prefixes must hash to identical block keys across DP ranks.
+        export PYTHONHASHSEED=42
+        OFFLOAD_CONFIG=$(cat <<EOF
+{
+  "kv_connector": "SimpleCPUOffloadConnector",
+  "kv_role": "kv_both",
+  "kv_connector_extra_config": {
+    "cpu_bytes_to_use": ${CPU_BYTES_PER_RANK},
+    "enable_cross_layers_blocks": "true"
+  }
+}
+EOF
+)
+        OFFLOAD_ARGS=(
+            --kv-transfer-config
+            "$OFFLOAD_CONFIG"
+        )
+        ;;
+    mooncake)
+        require_agentic_kv_offload_backend mooncake
+        # Embedded mode contributes one global segment per DP rank to the
+        # shared store, so divide the aggregate host budget across ranks.
         PER_RANK_GB=$((TOTAL_CPU_DRAM_GB / GPU_COUNT))
-
         MOONCAKE_VERSION=0.3.11.post1
         agentic_pip_install --quiet --no-cache-dir --no-deps \
             --force-reinstall "mooncake-transfer-engine-cuda13==$MOONCAKE_VERSION"
@@ -139,9 +147,7 @@ EOF
         export MC_ENABLE_DEST_DEVICE_AFFINITY=1
         # Identical prefixes must hash to identical store keys across DP ranks.
         export PYTHONHASHSEED=0
-        # Large agentic KV writes can exceed Mooncake Store's fixed 60-second
-        # transfer deadline at the default 64 KiB RDMA slice size. Reduce
-        # per-transfer bookkeeping and give the shared RNIC more workers.
+        export WITH_NVIDIA_PEERMEM=0
         export MC_SLICE_SIZE=1048576
         export MC_WORKERS_PER_CTX=4
 
@@ -164,55 +170,185 @@ EOF
             exit 1
         fi
 
+        unset VLLM_USE_SIMPLE_KV_OFFLOAD
+        OFFLOAD_CONFIG='{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
+        OFFLOAD_ARGS=(--kv-transfer-config "$OFFLOAD_CONFIG")
+        ;;
+    lmcache)
+        require_agentic_kv_offload_backend lmcache
+        # The LMCache MP server owns the host-DRAM KV pool as one shared
+        # tier; vLLM ranks attach via LMCacheMPConnector, so the aggregate
+        # host budget is passed through undivided (unlike Mooncake's
+        # per-rank segments). Follows the LMCache DeepSeek-V4 recipe
+        # (docs.lmcache.ai/recipes/deepseek_v4_flash.html); LMCache handles
+        # DSV4's Sparse-MLA hybrid KV geometries automatically.
+        LMCACHE_VERSION=0.5.1
+        agentic_pip_install --quiet --no-cache-dir "lmcache==$LMCACHE_VERSION"
+        python3 -c "import lmcache.integration.vllm.lmcache_mp_connector" >/dev/null
+
+        LMCACHE_HOST=127.0.0.1
+        LMCACHE_PORT=$((PORT + 12000))
+        LMCACHE_HTTP_PORT=$((PORT + 13000))
+        # LMCacheMPConnector concatenates lmcache.mp.host and port into the
+        # ZMQ endpoint. Bind the server to a raw host, but pass the connector
+        # a ZMQ-style host string.
+        LMCACHE_CONNECT_HOST="tcp://$LMCACHE_HOST"
+        # Pool target derated to 75% of the aggregate budget: pinned host
+        # memory is unswappable and also consumes GPU-side mapping
+        # resources, so leave headroom for vLLM host buffers and the OS.
+        # Full-budget targets OOM-killed the node (host OOM-killer or
+        # cudaErrorMemoryAllocation) as the cache filled past ~2 TB during
+        # PR #2153 bring-up.
+        LMCACHE_L1_SIZE_GB=$((TOTAL_CPU_DRAM_GB * 3 / 4))
+        # The pool grows lazily from the initial allocation, so the full
+        # --l1-size-gb target is not pinned at startup.
+        LMCACHE_L1_INIT_SIZE_GB=20
+        LMCACHE_MQ_TIMEOUT=300
+        # Identical prefixes must hash to identical cache keys across DP ranks.
+        export PYTHONHASHSEED=0
+        # Per-engine scheduler stats every 5s, to diagnose per-DP-rank KV
+        # cache imbalance under the session-sticky router.
+        export VLLM_LOG_STATS_INTERVAL=5
+
+        echo "Starting LMCache MP server on port $LMCACHE_PORT..."
+        # One GPU-side transfer worker avoids concurrent-GPU-transfer stalls
+        # under heavy async-load pressure; CPU-side workers stay at 8.
+        lmcache server \
+            --host "$LMCACHE_HOST" \
+            --port "$LMCACHE_PORT" \
+            --http-host "$LMCACHE_HOST" \
+            --http-port "$LMCACHE_HTTP_PORT" \
+            --l1-size-gb "$LMCACHE_L1_SIZE_GB" \
+            --l1-init-size-gb "$LMCACHE_L1_INIT_SIZE_GB" \
+            --max-gpu-workers 1 \
+            --max-cpu-workers 8 \
+            --chunk-size 1024 \
+            --l1-align-bytes 16384 \
+            --eviction-trigger-watermark 0.85 \
+            --eviction-ratio 0.10 \
+            --eviction-policy LRU \
+            --supported-transfer-mode lmcache_driven \
+            --no-separate-object-groups \
+            > "$LMCACHE_SERVER_LOG" 2>&1 &
+        LMCACHE_SERVER_PID=$!
+        LMCACHE_READY=0
+        for _ in $(seq 1 60); do
+            if ! kill -0 "$LMCACHE_SERVER_PID" 2>/dev/null; then
+                echo "LMCache server died during startup." >&2
+                cat "$LMCACHE_SERVER_LOG" >&2
+                exit 1
+            fi
+            if curl --output /dev/null --silent --fail \
+                "http://127.0.0.1:$LMCACHE_HTTP_PORT/healthcheck"; then
+                LMCACHE_READY=1
+                break
+            fi
+            sleep 2
+        done
+        if [ "$LMCACHE_READY" -ne 1 ]; then
+            echo "LMCache server did not become healthy in time." >&2
+            cat "$LMCACHE_SERVER_LOG" >&2
+            exit 1
+        fi
+
         unset VLLM_USE_SIMPLE_KV_OFFLOAD
         OFFLOAD_ARGS=(
             --kv-transfer-config
-            '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_connector_extra_config":{"load_async":true}}'
+            "{\"kv_connector\":\"LMCacheMPConnector\",\"kv_role\":\"kv_both\",\"kv_connector_extra_config\":{\"lmcache.mp.host\":\"$LMCACHE_CONNECT_HOST\",\"lmcache.mp.port\":$LMCACHE_PORT,\"lmcache.mp.mq_timeout\":$LMCACHE_MQ_TIMEOUT}}"
         )
-fi
+        ;;
+    *)
+        echo "Error: unsupported B300 KV_OFFLOAD_BACKEND='$KV_OFFLOAD_BACKEND'" >&2
+        exit 1
+        ;;
+esac
 
 PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
 if [ "$DP_ATTENTION" = "true" ]; then
     PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
 fi
 
-EP_ARGS=()
+TP_ARGS=()
+if [ "$DP_ATTENTION" = "true" ]; then
+    # LMCacheMPConnector exports the KV cache to the LMCache server through
+    # legacy CUDA IPC handles, and expandable-segment (cuMem/VMM) allocations
+    # cannot be exported that way (register_kv_caches fails with
+    # cudaErrorInvalidValue, same failure mode as --enable-cumem-allocator on
+    # the B200 lmcache arm in PR #2231), so the lmcache arm keeps the stock
+    # caching allocator.
+    if [ "$KV_OFFLOAD_BACKEND" != "lmcache" ]; then
+        export PYTORCH_ALLOC_CONF=expandable_segments:True
+    fi
+else
+    export VLLM_ALLREDUCE_USE_FLASHINFER=1
+    export VLLM_FLASHINFER_ALLREDUCE_BACKEND=auto
+    TP_ARGS+=(--disable-custom-all-reduce)
+fi
+
+MODE_ARGS=()
 if [ "$EP_SIZE" -gt 1 ]; then
-    EP_ARGS=(--enable-expert-parallel)
+    MODE_ARGS+=(
+        --enable-expert-parallel
+        --enable-ep-weight-filter
+        --moe-backend deep_gemm_amxf4_mega_moe
+    )
+fi
+if [ "$DP_ATTENTION" = "true" ]; then
+    MODE_ARGS+=(
+        --prefill-schedule-interval 8
+        --max-num-batched-tokens 8192
+    )
 fi
 
-# 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=$((2 * CONC))
-if [ "$MAX_NUM_SEQS" -eq 128 ]; then
-    MAX_NUM_SEQS=136
+if [ "$DP_ATTENTION" = "true" ]; then
+    # The DEP source recipe enforces 2*CONC = DP_WORLD_SIZE*MAX_NUM_SEQS.
+    MAX_NUM_SEQS=$((2 * CONC / TP))
+else
+    # Preserve the previous TP4 scheduler headroom for agentic fan-out.
+    MAX_NUM_SEQS=$((2 * CONC))
 fi
+CUDA_GRAPH_CAPTURE_SIZES=""
+for ((capture_size = 1; capture_size <= MAX_NUM_SEQS; capture_size++)); do
+    if [ -n "$CUDA_GRAPH_CAPTURE_SIZES" ]; then
+        CUDA_GRAPH_CAPTURE_SIZES+=","
+    fi
+    CUDA_GRAPH_CAPTURE_SIZES+="$capture_size"
+done
+COMPILATION_CONFIG="{\"cudagraph_mode\":\"FULL_DECODE_ONLY\",\"cudagraph_capture_sizes\":[${CUDA_GRAPH_CAPTURE_SIZES}],\"mode\":0}"
 
 echo "Starting vllm server..."
 export TORCH_CUDA_ARCH_LIST="10.0"
 export PYTHONNOUSERSITE=1
 export VLLM_FLOAT32_MATMUL_PRECISION=high
 
-vllm serve "$MODEL_PATH" --served-model-name "$MODEL" \
---host 0.0.0.0 \
---port "$VLLM_BACKEND_PORT" \
---trust-remote-code \
---kv-cache-dtype fp8 \
---block-size 256 \
-"${PARALLEL_ARGS[@]}" \
-"${VLLM_CP_ARGS[@]}" \
-"${EP_ARGS[@]}" \
---compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \
---attention_config.use_fp4_indexer_cache=True \
---tokenizer-mode deepseek_v4 \
---tool-call-parser deepseek_v4 \
---enable-auto-tool-choice \
---reasoning-parser deepseek_v4 \
---enable-prefix-caching \
---no-disable-hybrid-kv-cache-manager \
---max-num-seqs "$MAX_NUM_SEQS" \
-"${OFFLOAD_ARGS[@]}" > "$SERVER_LOG" 2>&1 &
+{ set +x; } 2>/dev/null
+VLLM_CMD=(
+    vllm serve "$MODEL_PATH" --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$VLLM_BACKEND_PORT"
+    --gpu-memory-utilization 0.96
+    --trust-remote-code
+    --no-enable-flashinfer-autotune
+    --no-disable-hybrid-kv-cache-manager
+    --max-num-seqs "$MAX_NUM_SEQS"
+    --kv-cache-dtype fp8
+    --block-size 256
+    --max-model-len 1048576
+    --attention-config '{"use_fp4_indexer_cache":true,"backend":"FLASHINFER_MLA_SPARSE_DSV4","use_prefill_query_quantization":true}'
+    --disable-uvicorn-access-log
+    --tokenizer-mode deepseek_v4
+    --tool-call-parser deepseek_v4
+    --enable-auto-tool-choice
+    --reasoning-parser deepseek_v4
+    --compilation-config "$COMPILATION_CONFIG"
+    "${PARALLEL_ARGS[@]}"
+    "${TP_ARGS[@]}"
+    "${MODE_ARGS[@]}"
+    "${OFFLOAD_ARGS[@]}"
+)
+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 "Server PID: $SERVER_PID"
 
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index e4e0a7a75f..569adc78d1 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -3346,7 +3346,7 @@ dsr1-fp8-b200-trt:
       osl: 1024
       search-space:
       - { tp: 8, ep: 1, conc-start: 64, conc-end: 128 }
-      - { tp: 4, ep: 1, conc-start: 8, conc-end: 16 } 
+      - { tp: 4, ep: 1, conc-start: 8, conc-end: 16 }
       - { tp: 8, ep: 1, conc-start: 4, conc-end: 8 }
     - isl: 8192
       osl: 1024
@@ -3560,7 +3560,31 @@ dsv4-fp4-b300-vllm:
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 2048, conc-end: 2048 }
 
 dsv4-fp4-b300-vllm-agentic:
-  image: vllm/vllm-openai:v0.23.0
+  image: vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec
+  model: deepseek-ai/DeepSeek-V4-Pro
+  model-prefix: dsv4
+  runner: cluster:b300-nv
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      # Preserve the previous GPU-resident TP4 search space on the new image.
+      - { tp: 8, kv-offloading: none, conc-list: [1, 2, 4, 6, 8, 16] }
+      - { tp: 4, kv-offloading: none, conc-list: [1, 2, 4, 6, 8, 16] }
+      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: vllm-simple, version: "904e4ec" }, conc-list: [8, 16, 24, 32, 40, 48, 56, 64, 72], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [12, 20, 28, 36, 44, 52, 60, 68, 76], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: vllm-simple, version: "904e4ec" }, conc-list: [32, 64, 96, 128, 160, 192, 224, 228], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [40, 72, 104, 136, 168, 200], router: { name: vllm-router, version: "0.1.14" } }
+
+# LMCache CPU-offload arm of dsv4-fp4-b300-vllm-agentic, split into its own
+# config so it can be triggered/tested independently of the vllm-simple and
+# Mooncake curves. Points mirror the vllm-simple offload ladders above on the
+# same tuned image, so the backends are directly comparable.
+dsv4-fp4-b300-vllm-agentic-lmcache:
+  image: vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:b300-nv
@@ -3571,16 +3595,8 @@ dsv4-fp4-b300-vllm-agentic:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      # Compare native GPU-cache and MooncakeStore CPU-offload cliffs.
-      - { tp: 4, kv-offloading: none,  conc-list: [1, 2, 4, 6, 8, 16] }
-      - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [16, 18, 20, 24] }
-      # TP8 remains cache-resident through conc 52.
-      - { tp: 8, kv-offloading: none,  conc-list: [1, 2, 4, 6, 8, 16, 32, 40, 48, 52] }
-      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [52] }
-      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: none,  conc-list: [8, 16], router: { name: vllm-router, version: "0.1.14" } }
-      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: mooncake, version: "0.3.11.post1" }, conc-list: [32], router: { name: vllm-router, version: "0.1.14" } }
-      # TP8 DEP retains representative low, peak, and transition points.
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none,  conc-list: [52, 72, 100, 128, 144], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [8, 16, 24, 32, 40, 48, 56, 64, 72], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [32, 64, 96, 128, 160, 192, 224, 228], router: { name: vllm-router, version: "0.1.14" } }
 
 dsv4-fp4-b300-trt:
   image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-c185066
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 44b403ae9f..3cd43afd48 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4815,3 +4815,14 @@
     - "Image: lmsysorg/sglang:v0.5.13.post1-cu130"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2123
   
+- config-keys:
+    - dsv4-fp4-b300-vllm-agentic-lmcache
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add LMCache 0.5.1 DRAM KV-offload arm on the tuned B300 AgentX recipe from PR #2225 (sparse DSV4 FlashInfer attention, AMXF4 mega-MoE, FULL_DECODE_ONLY CUDA graphs) with image vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec"
+    - "LMCache MP server (lmcache_driven transfer mode) + LMCacheMPConnector per PR #2153/#2231; L1 pool derated to 75% of TOTAL_CPU_DRAM_GB; PYTORCH_ALLOC_CONF=expandable_segments:True dropped on the lmcache arm only (VMM allocations cannot be CUDA-IPC-exported to the LMCache server, same failure mode as cuMem on B200)"
+    - "Points mirror the vllm-simple ladders: TP4-DEP4 conc [8, 16, 24, 32, 40, 48, 56, 64, 72]; TP8-DEP8 conc [32, 64, 96, 128, 160, 192, 224, 228]"
+    - "Parent dsv4-fp4-b300-vllm-agentic updated in the same change to the PR #2225 recipe and search space (image bump, vllm-simple + Mooncake arms) so the shared script stays consistent; the parent points are benchmarked by PR #2225 and not re-triggered here"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX
+

From 8dec80cedfe8e7e08e6df4109903dc5530fce80a Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Wed, 15 Jul 2026 17:40:53 -0700
Subject: [PATCH 2/9] chore: fill in perf-changelog pr-link with PR #2232
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将 perf-changelog 条目的 pr-link 填写为 PR #2232。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 3cd43afd48..732a4feccd 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4824,5 +4824,5 @@
     - "LMCache MP server (lmcache_driven transfer mode) + LMCacheMPConnector per PR #2153/#2231; L1 pool derated to 75% of TOTAL_CPU_DRAM_GB; PYTORCH_ALLOC_CONF=expandable_segments:True dropped on the lmcache arm only (VMM allocations cannot be CUDA-IPC-exported to the LMCache server, same failure mode as cuMem on B200)"
     - "Points mirror the vllm-simple ladders: TP4-DEP4 conc [8, 16, 24, 32, 40, 48, 56, 64, 72]; TP8-DEP8 conc [32, 64, 96, 128, 160, 192, 224, 228]"
     - "Parent dsv4-fp4-b300-vllm-agentic updated in the same change to the PR #2225 recipe and search space (image bump, vllm-simple + Mooncake arms) so the shared script stays consistent; the parent points are benchmarked by PR #2225 and not re-triggered here"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXXX
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2232
 

From 54c612d571b7d701a4e570dcf04692251797acf0 Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Thu, 16 Jul 2026 14:12:34 -0700
Subject: [PATCH 3/9] feat(agentic): tier lmcache GPU mem util and add pure-TP
 lmcache arms
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Bring-up sweep on PR #2232 (run 29463061871) showed the lmcache arm has
<1 GiB GPU headroom at the recipe's gpu-memory-utilization 0.96 (LMCache
MP server GPU worker + DeepGEMM/TileLang JIT driver allocations + no
expandable segments): every DEP8 point OOMed deterministically growing
the torch pool, and 2/9 DEP4 points failed at the margin (JIT module
load driver OOM at startup; cuBLAS workspace failure mid-run). Derate
DEP8 to the B200-proven 0.92 and DEP4/pure-TP to 0.94.

Also add pure-TP LMCache arms (TP8 and TP4, conc [8, 12, 16]) mirroring
the upper end of the parent's GPU-resident TP ladders and the B200
lmcache TP8 ladder, for a direct GPU-cache vs LMCache-offload
comparison.

中文：PR #2232 的调试扫描（run 29463061871）显示 lmcache 分支在配方的
gpu-memory-utilization 0.96 下 GPU 余量不足 1 GiB（LMCache MP server 的
GPU worker + DeepGEMM/TileLang JIT 驱动层分配 + 无 expandable
segments）：所有 DEP8 测试点在扩展 torch 内存池时必然 OOM，DEP4 有 2/9
测试点在边缘失败（启动时 JIT 模块加载驱动 OOM；运行中 cuBLAS workspace
失败）。将 DEP8 降额至 B200 验证过的 0.92，DEP4 与纯 TP 降额至 0.94。
另新增纯 TP 的 LMCache 分支（TP8 与 TP4，并发 [8, 12, 16]），与父配置
GPU 驻留 TP 阶梯的高段及 B200 lmcache TP8 阶梯对齐，便于 GPU 缓存与
LMCache 卸载的直接对比。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 .../single_node/agentic/dsv4_fp4_b300_vllm.sh | 24 +++++++++++++++++--
 configs/nvidia-master.yaml                    |  5 ++++
 perf-changelog.yaml                           |  4 ++--
 3 files changed, 29 insertions(+), 4 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh b/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh
index 30c6a76836..5dc8419fe9 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh
@@ -13,7 +13,8 @@ set -x
 # Required env vars:
 #   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR
 #
-# KV_OFFLOADING=none is used by TP4. DEP4 uses KV_OFFLOADING=dram with
+# KV_OFFLOADING=none is used by the GPU-resident TP arms. DRAM offload arms
+# (DEP, and pure-TP on the lmcache config) use KV_OFFLOADING=dram with
 # KV_OFFLOAD_BACKEND=vllm-simple, mooncake, or lmcache.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
@@ -94,6 +95,8 @@ LMCACHE_SERVER_PID=""
 # On cluster:b300-nv, dram-utilization=0.80 and DEP4 resolve to roughly the
 # source recipe's 280 GiB per DP rank. TP4 remains GPU-resident.
 OFFLOAD_ARGS=()
+# Recipe default; the lmcache arm derates it (see the lmcache case below).
+GPU_MEM_UTIL=0.96
 case "$KV_OFFLOAD_BACKEND" in
     "")
         require_agentic_kv_offload_none
@@ -209,6 +212,23 @@ EOF
         # Per-engine scheduler stats every 5s, to diagnose per-DP-rank KV
         # cache imbalance under the session-sticky router.
         export VLLM_LOG_STATS_INTERVAL=5
+        # The LMCache MP server keeps a GPU worker (~0.8 GiB CUDA context +
+        # staging buffers) on every GPU, and this arm cannot run expandable
+        # segments (the KV cache must be legacy-CUDA-IPC-exportable), so the
+        # recipe's 0.96 leaves <1 GiB free after vLLM sizes its pool. In the
+        # PR #2232 bring-up sweep every DEP8 point OOMed deterministically
+        # growing the torch pool for a 1 GiB DeepGEMM workspace, so DEP8
+        # gets the B200 lmcache arm's proven 0.92. DEP4 was only marginal
+        # (7/9 passed; the failures were a DeepGEMM JIT module load hitting
+        # driver CUDA_ERROR_OUT_OF_MEMORY at startup and a cuBLAS workspace
+        # failure mid-run), so DEP4 and the pure-TP arms take a lighter 0.94
+        # to keep more GPU KV capacity. KV overflows to the DRAM pool by
+        # design, so the derate costs little.
+        if [ "$DP_ATTENTION" = "true" ] && [ "$TP" -eq 8 ]; then
+            GPU_MEM_UTIL=0.92
+        else
+            GPU_MEM_UTIL=0.94
+        fi
 
         echo "Starting LMCache MP server on port $LMCACHE_PORT..."
         # One GPU-side transfer worker avoids concurrent-GPU-transfer stalls
@@ -326,7 +346,7 @@ VLLM_CMD=(
     vllm serve "$MODEL_PATH" --served-model-name "$MODEL"
     --host 0.0.0.0
     --port "$VLLM_BACKEND_PORT"
-    --gpu-memory-utilization 0.96
+    --gpu-memory-utilization "$GPU_MEM_UTIL"
     --trust-remote-code
     --no-enable-flashinfer-autotune
     --no-disable-hybrid-kv-cache-manager
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 569adc78d1..6ae7f426c7 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -3595,6 +3595,11 @@ dsv4-fp4-b300-vllm-agentic-lmcache:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
+      # Pure-TP LMCache arms: upper end of the parent's GPU-resident TP
+      # ladders, matching the B200 lmcache TP8 conc ladder, for a direct
+      # GPU-cache vs LMCache-offload comparison at the same points.
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [8, 12, 16] }
+      - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [8, 12, 16] }
       - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [8, 16, 24, 32, 40, 48, 56, 64, 72], router: { name: vllm-router, version: "0.1.14" } }
       - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [32, 64, 96, 128, 160, 192, 224, 228], router: { name: vllm-router, version: "0.1.14" } }
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 732a4feccd..0b49c5583b 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4821,8 +4821,8 @@
     - agentic-coding
   description:
     - "Add LMCache 0.5.1 DRAM KV-offload arm on the tuned B300 AgentX recipe from PR #2225 (sparse DSV4 FlashInfer attention, AMXF4 mega-MoE, FULL_DECODE_ONLY CUDA graphs) with image vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec"
-    - "LMCache MP server (lmcache_driven transfer mode) + LMCacheMPConnector per PR #2153/#2231; L1 pool derated to 75% of TOTAL_CPU_DRAM_GB; PYTORCH_ALLOC_CONF=expandable_segments:True dropped on the lmcache arm only (VMM allocations cannot be CUDA-IPC-exported to the LMCache server, same failure mode as cuMem on B200)"
-    - "Points mirror the vllm-simple ladders: TP4-DEP4 conc [8, 16, 24, 32, 40, 48, 56, 64, 72]; TP8-DEP8 conc [32, 64, 96, 128, 160, 192, 224, 228]"
+    - "LMCache MP server (lmcache_driven transfer mode) + LMCacheMPConnector per PR #2153/#2231; L1 pool derated to 75% of TOTAL_CPU_DRAM_GB; PYTORCH_ALLOC_CONF=expandable_segments:True dropped on the lmcache arm only (VMM allocations cannot be CUDA-IPC-exported to the LMCache server, same failure mode as cuMem on B200); gpu-memory-utilization derated 0.96 -> 0.92 on DEP8 and 0.94 on DEP4/pure-TP lmcache points (LMCache's per-GPU server worker + no expandable segments left <1 GiB headroom at 0.96: every DEP8 point OOMed growing the torch pool, DEP4 marginally failed via DeepGEMM JIT driver OOM and a cuBLAS workspace failure in the PR #2232 bring-up sweep)"
+    - "Points mirror the vllm-simple ladders plus pure-TP arms matching the B200 lmcache TP8 ladder: TP8 and TP4 conc [8, 12, 16]; TP4-DEP4 conc [8, 16, 24, 32, 40, 48, 56, 64, 72]; TP8-DEP8 conc [32, 64, 96, 128, 160, 192, 224, 228]"
     - "Parent dsv4-fp4-b300-vllm-agentic updated in the same change to the PR #2225 recipe and search space (image bump, vllm-simple + Mooncake arms) so the shared script stays consistent; the parent points are benchmarked by PR #2225 and not re-triggered here"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2232
 

From bdc33baba9b8c4008aba6e6e60d7bad4724f258d Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Fri, 17 Jul 2026 12:55:49 -0700
Subject: [PATCH 4/9] fix(matrix): accept agentic SWE-bench eval rows in the
 changelog matrix schema
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

PR #1947 made process_changelog's eval pass select single-node agentic
(SWE-bench) eval rows via mark_eval_entries --evals-only, but
ChangelogMatrixEntry.evals still only accepted SingleNodeMatrixEntry, so
any changelog entry touching an agentic-coding config now fails
validation (missing isl/osl, extra kv-offloading/scenario-type/duration).
Accept Union[SingleNodeMatrixEntry, SingleNodeAgenticMatrixEntry] in
evals and add run-eval/eval-only (default false) to the agentic entry
model, with a regression test.

中文：PR #1947 使 process_changelog 的评估生成步骤通过 mark_eval_entries
--evals-only 选择单节点 agentic（SWE-bench）评估行，但
ChangelogMatrixEntry.evals 仍只接受 SingleNodeMatrixEntry，导致任何涉及
agentic-coding 配置的变更日志条目都无法通过校验（缺少 isl/osl，多出
kv-offloading/scenario-type/duration 字段）。将 evals 改为接受
Union[SingleNodeMatrixEntry, SingleNodeAgenticMatrixEntry]，并为 agentic
条目模型添加 run-eval/eval-only（默认 false）字段，附回归测试。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 utils/matrix_logic/test_validation.py | 48 +++++++++++++++++++++++++++
 utils/matrix_logic/validation.py      |  9 ++++-
 2 files changed, 56 insertions(+), 1 deletion(-)

diff --git a/utils/matrix_logic/test_validation.py b/utils/matrix_logic/test_validation.py
index 9925adb626..6a23c68f21 100644
--- a/utils/matrix_logic/test_validation.py
+++ b/utils/matrix_logic/test_validation.py
@@ -18,6 +18,7 @@
     SingleNodeMasterConfigEntry,
     MultiNodeMasterConfigEntry,
     ChangelogEntry,
+    ChangelogMatrixEntry,
     validate_matrix_entry,
     validate_master_config,
     validate_runner_config,
@@ -405,6 +406,53 @@ def test_arbitrary_backend_is_valid_for_single_node_agentic_entry(self):
         assert entry.kv_offload_backend.name == "future-backend"
         assert entry.kv_offload_backend.version is None
 
+    def test_agentic_eval_row_is_valid_in_changelog_matrix_evals(self):
+        """Agentic SWE-bench eval rows (mark_eval_entries with --evals-only)
+        carry run-eval/eval-only plus agentic-only fields and must validate
+        inside ChangelogMatrixEntry.evals (regression test for the schema gap
+        left by the PR #1947 SWE-bench eval selection)."""
+        agentic_eval_row = {
+            "image": "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec",
+            "model": "deepseek-ai/DeepSeek-V4-Pro",
+            "model-prefix": "dsv4",
+            "precision": "fp4",
+            "framework": "vllm",
+            "runner": "cluster:b300-nv",
+            "tp": 8,
+            "pp": 1,
+            "dcp-size": 1,
+            "pcp-size": 1,
+            "ep": 8,
+            "dp-attn": True,
+            "spec-decoding": "none",
+            "conc": 224,
+            "kv-offloading": "none",
+            "total-cpu-dram-gb": 0,
+            "duration": 3600,
+            "exp-name": "dsv4_tp8_conc224_kvnone",
+            "scenario-type": "agentic-coding",
+            "router": {"name": "vllm-router", "version": "0.1.14"},
+            "run-eval": True,
+            "eval-only": True,
+        }
+        validated = ChangelogMatrixEntry.model_validate({
+            "single_node": {},
+            "multi_node": {},
+            "evals": [agentic_eval_row],
+            "multinode_evals": [],
+            "changelog_metadata": {
+                "base_ref": "origin/main",
+                "head_ref": "abc123",
+                "entries": [{
+                    "config-keys": ["dsv4-fp4-b300-vllm-agentic"],
+                    "description": ["test"],
+                    "pr-link": "https://github.com/SemiAnalysisAI/InferenceX/pull/2232",
+                }],
+            },
+        })
+        assert isinstance(validated.evals[0], SingleNodeAgenticMatrixEntry)
+        assert validated.evals[0].eval_only is True
+
     def test_arbitrary_backend_is_valid_for_agentic_search_space(self):
         entry = AgenticCodingSearchSpaceEntry(**{
             "tp": 8,
diff --git a/utils/matrix_logic/validation.py b/utils/matrix_logic/validation.py
index 6b7c671289..60e26e58f8 100644
--- a/utils/matrix_logic/validation.py
+++ b/utils/matrix_logic/validation.py
@@ -291,6 +291,10 @@ class SingleNodeAgenticMatrixEntry(BaseModel):
     duration: int = Field(alias=Fields.DURATION.value)
     exp_name: str = Field(alias=Fields.EXP_NAME.value)
     scenario_type: str = Field(alias=Fields.SCENARIO_TYPE.value)
+    # Present on agentic eval rows (SWE-bench selection via mark_eval_entries);
+    # benchmark rows are generated with --no-evals and omit both fields.
+    run_eval: bool = Field(alias=Fields.RUN_EVAL.value, default=False)
+    eval_only: bool = Field(alias=Fields.EVAL_ONLY.value, default=False)
 
     @model_validator(mode='after')
     def validate_kv_offload_fields(self):
@@ -889,7 +893,10 @@ class ChangelogMatrixEntry(BaseModel):
                       ] = Field(default_factory=dict)
     multi_node: dict[str, list[Union[MultiNodeMatrixEntry, MultiNodeAgenticMatrixEntry]]
                      ] = Field(default_factory=dict)
-    evals: list[SingleNodeMatrixEntry] = Field(default_factory=list)
+    # Fixed-seq-len evals plus single-node agentic (SWE-bench) eval rows,
+    # selected by mark_eval_entries since the SWE-bench Lite eval (PR #1947).
+    evals: list[Union[SingleNodeMatrixEntry, SingleNodeAgenticMatrixEntry]] = Field(
+        default_factory=list)
     multinode_evals: list[MultiNodeMatrixEntry] = Field(default_factory=list)
     changelog_metadata: ChangelogMetadata
 

From 189aed4882da6ada675deb8edfeac5477613b884 Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Fri, 17 Jul 2026 12:59:45 -0700
Subject: [PATCH 5/9] Revert "fix(matrix): accept agentic SWE-bench eval rows
 in the changelog matrix schema"
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

This reverts commit bdc33bab6a72a0847c26be0dbe4e10bd8fe4a2eb to keep this
PR scoped to the B300 LMCache recipe. The PR #1947 schema regression
(agentic changelog entries fail matrix validation) needs a standalone fix
PR; until that merges, check-changelog on this PR is expected to fail.

中文：回退 bdc33bab，使本 PR 仅保留 B300 LMCache 配方相关改动。PR #1947
造成的 schema 回归（agentic 变更日志条目无法通过矩阵校验）需通过独立 PR
修复；在该修复合并前，本 PR 的 check-changelog 预期会失败。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 utils/matrix_logic/test_validation.py | 48 ---------------------------
 utils/matrix_logic/validation.py      |  9 +----
 2 files changed, 1 insertion(+), 56 deletions(-)

diff --git a/utils/matrix_logic/test_validation.py b/utils/matrix_logic/test_validation.py
index 6a23c68f21..9925adb626 100644
--- a/utils/matrix_logic/test_validation.py
+++ b/utils/matrix_logic/test_validation.py
@@ -18,7 +18,6 @@
     SingleNodeMasterConfigEntry,
     MultiNodeMasterConfigEntry,
     ChangelogEntry,
-    ChangelogMatrixEntry,
     validate_matrix_entry,
     validate_master_config,
     validate_runner_config,
@@ -406,53 +405,6 @@ def test_arbitrary_backend_is_valid_for_single_node_agentic_entry(self):
         assert entry.kv_offload_backend.name == "future-backend"
         assert entry.kv_offload_backend.version is None
 
-    def test_agentic_eval_row_is_valid_in_changelog_matrix_evals(self):
-        """Agentic SWE-bench eval rows (mark_eval_entries with --evals-only)
-        carry run-eval/eval-only plus agentic-only fields and must validate
-        inside ChangelogMatrixEntry.evals (regression test for the schema gap
-        left by the PR #1947 SWE-bench eval selection)."""
-        agentic_eval_row = {
-            "image": "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec",
-            "model": "deepseek-ai/DeepSeek-V4-Pro",
-            "model-prefix": "dsv4",
-            "precision": "fp4",
-            "framework": "vllm",
-            "runner": "cluster:b300-nv",
-            "tp": 8,
-            "pp": 1,
-            "dcp-size": 1,
-            "pcp-size": 1,
-            "ep": 8,
-            "dp-attn": True,
-            "spec-decoding": "none",
-            "conc": 224,
-            "kv-offloading": "none",
-            "total-cpu-dram-gb": 0,
-            "duration": 3600,
-            "exp-name": "dsv4_tp8_conc224_kvnone",
-            "scenario-type": "agentic-coding",
-            "router": {"name": "vllm-router", "version": "0.1.14"},
-            "run-eval": True,
-            "eval-only": True,
-        }
-        validated = ChangelogMatrixEntry.model_validate({
-            "single_node": {},
-            "multi_node": {},
-            "evals": [agentic_eval_row],
-            "multinode_evals": [],
-            "changelog_metadata": {
-                "base_ref": "origin/main",
-                "head_ref": "abc123",
-                "entries": [{
-                    "config-keys": ["dsv4-fp4-b300-vllm-agentic"],
-                    "description": ["test"],
-                    "pr-link": "https://github.com/SemiAnalysisAI/InferenceX/pull/2232",
-                }],
-            },
-        })
-        assert isinstance(validated.evals[0], SingleNodeAgenticMatrixEntry)
-        assert validated.evals[0].eval_only is True
-
     def test_arbitrary_backend_is_valid_for_agentic_search_space(self):
         entry = AgenticCodingSearchSpaceEntry(**{
             "tp": 8,
diff --git a/utils/matrix_logic/validation.py b/utils/matrix_logic/validation.py
index 60e26e58f8..6b7c671289 100644
--- a/utils/matrix_logic/validation.py
+++ b/utils/matrix_logic/validation.py
@@ -291,10 +291,6 @@ class SingleNodeAgenticMatrixEntry(BaseModel):
     duration: int = Field(alias=Fields.DURATION.value)
     exp_name: str = Field(alias=Fields.EXP_NAME.value)
     scenario_type: str = Field(alias=Fields.SCENARIO_TYPE.value)
-    # Present on agentic eval rows (SWE-bench selection via mark_eval_entries);
-    # benchmark rows are generated with --no-evals and omit both fields.
-    run_eval: bool = Field(alias=Fields.RUN_EVAL.value, default=False)
-    eval_only: bool = Field(alias=Fields.EVAL_ONLY.value, default=False)
 
     @model_validator(mode='after')
     def validate_kv_offload_fields(self):
@@ -893,10 +889,7 @@ class ChangelogMatrixEntry(BaseModel):
                       ] = Field(default_factory=dict)
     multi_node: dict[str, list[Union[MultiNodeMatrixEntry, MultiNodeAgenticMatrixEntry]]
                      ] = Field(default_factory=dict)
-    # Fixed-seq-len evals plus single-node agentic (SWE-bench) eval rows,
-    # selected by mark_eval_entries since the SWE-bench Lite eval (PR #1947).
-    evals: list[Union[SingleNodeMatrixEntry, SingleNodeAgenticMatrixEntry]] = Field(
-        default_factory=list)
+    evals: list[SingleNodeMatrixEntry] = Field(default_factory=list)
     multinode_evals: list[MultiNodeMatrixEntry] = Field(default_factory=list)
     changelog_metadata: ChangelogMetadata
 

From 7c8bd63716143a3887ff2a7929dc9cca7c776d75 Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Fri, 17 Jul 2026 13:05:10 -0700
Subject: [PATCH 6/9] refactor(agentic): fold B300 lmcache arms into the
 official agentic config
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Per review: run the LMCache sweep together with the official
dsv4-fp4-b300-vllm-agentic config instead of a separate section. The
standalone dsv4-fp4-b300-vllm-agentic-lmcache config is removed; the
parent now carries one lmcache arm per official arm at a +4 conc offset
(TP4 [32, 36, 40, 44]; DEP4 [36, 44, 52, 60, 68, 76]; DEP8
[68, 100, 116, 132, 148, 164, 180, 196, 228]), all inside the ranges
validated in the PR #2232 bring-up sweeps. The changelog entry now
triggers only the parent config.

中文：按评审意见，LMCache 与官方 dsv4-fp4-b300-vllm-agentic 配置在同一
扫描中运行，不再单列配置段。移除独立的
dsv4-fp4-b300-vllm-agentic-lmcache 配置；父配置为每个官方分支各增加一个
+4 并发偏移的 lmcache 分支（TP4 [32, 36, 40, 44]；DEP4
[36, 44, 52, 60, 68, 76]；DEP8 [68, 100, 116, 132, 148, 164, 180, 196,
228]），全部处于 PR #2232 调试扫描验证过的范围内。变更日志条目现仅触发
父配置。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 .../single_node/agentic/dsv4_fp4_b300_vllm.sh |  6 ++--
 configs/nvidia-master.yaml                    | 34 +++++--------------
 perf-changelog.yaml                           |  8 ++---
 3 files changed, 14 insertions(+), 34 deletions(-)

diff --git a/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh b/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh
index 7ca0fd4410..5f0048e8ce 100755
--- a/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh
+++ b/benchmarks/single_node/agentic/dsv4_fp4_b300_vllm.sh
@@ -13,9 +13,9 @@ set -x
 # Required env vars:
 #   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR
 #
-# GPU-resident arms (TP4 and DEP8 in the parent config) use
-# KV_OFFLOADING=none. DRAM offload arms (TP and DEP) use KV_OFFLOADING=dram
-# with KV_OFFLOAD_BACKEND=vllm-simple, mooncake, or lmcache.
+# GPU-resident arms (TP4 and DEP8) use KV_OFFLOADING=none. DRAM offload
+# arms (TP and DEP) use KV_OFFLOADING=dram with
+# KV_OFFLOAD_BACKEND=vllm-simple, mooncake, or lmcache.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 33750cf874..12a9486be7 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -3962,34 +3962,16 @@ dsv4-fp4-b300-vllm-agentic:
       - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: vllm-simple, version: "904e4ec" }, conc-list: [32, 40, 48, 56, 64, 72], router: { name: vllm-router, version: "0.1.14" } }
       # DEP8 SimpleCPU
       - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, conc-list: [64, 96, 112, 128, 144, 160, 176, 192, 224], router: { name: vllm-router, version: "0.1.14" } }
-      # TP4 and DEP4 LMCache: interleave the SimpleCPU ladders above at a
-      # +4 conc offset, so the curves sample distinct points and the conc
-      # values stay disjoint from dsv4-fp4-b300-vllm-agentic-lmcache below
-      # (identical exp-names across configs would collide).
+      # LMCache arms: interleave each official ladder above at a +4 conc
+      # offset so the backends sample distinct points on the same recipe.
+      # Bring-up validated in the PR #2232 sweeps (runs 29463061871 and
+      # 29535333851): TP/DEP4 at gpu-memory-utilization 0.94, DEP8 at 0.92.
+      # TP4 LMCache (vs TP4 SimpleCPU)
       - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [32, 36, 40, 44] }
+      # DEP4 LMCache (vs DEP4 SimpleCPU)
       - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [36, 44, 52, 60, 68, 76], router: { name: vllm-router, version: "0.1.14" } }
-
-# LMCache CPU-offload arms of dsv4-fp4-b300-vllm-agentic, split into their
-# own config so they can be triggered/tested independently. Conc ladders were
-# validated in the PR #2232 bring-up sweeps (all 23 points green) and stay
-# disjoint from the parent's lmcache arms above.
-dsv4-fp4-b300-vllm-agentic-lmcache:
-  image: vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec
-  model: deepseek-ai/DeepSeek-V4-Pro
-  model-prefix: dsv4
-  runner: cluster:b300-nv
-  precision: fp4
-  framework: vllm
-  multinode: false
-  scenarios:
-    agentic-coding:
-    - dram-utilization: 0.80
-      search-space:
-      # Pure-TP LMCache arms matching the B200 lmcache TP8 conc ladder.
-      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [8, 12, 16] }
-      - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [8, 12, 16] }
-      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [8, 16, 24, 32, 40, 48, 56, 64, 72], router: { name: vllm-router, version: "0.1.14" } }
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [32, 64, 96, 128, 160, 192, 224, 228], router: { name: vllm-router, version: "0.1.14" } }
+      # DEP8 LMCache (vs GPU-resident DEP8)
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [68, 100, 116, 132, 148, 164, 180, 196, 228], router: { name: vllm-router, version: "0.1.14" } }
 
 dsv4-fp4-b300-trt:
   image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-c185066
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index ea3c1ce69e..82ab4b6cf7 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4929,14 +4929,12 @@
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2094
 
 - config-keys:
-    - dsv4-fp4-b300-vllm-agentic-lmcache
     - dsv4-fp4-b300-vllm-agentic
   scenario-type:
     - agentic-coding
   description:
-    - "Add LMCache 0.5.1 DRAM KV-offload arms on the official B300 AgentX recipe from PR #2241 (sparse DSV4 FlashInfer attention, AMXF4 mega-MoE, FULL_DECODE_ONLY CUDA graphs, DEP8 large prefill budget + long-prefill chunking) with image vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec"
-    - "LMCache MP server (lmcache_driven transfer mode) + LMCacheMPConnector per PR #2153/#2231; L1 pool derated to 75% of TOTAL_CPU_DRAM_GB; PYTORCH_ALLOC_CONF=expandable_segments:True dropped on the lmcache arms only (VMM allocations cannot be CUDA-IPC-exported to the LMCache server, same failure mode as cuMem on B200); gpu-memory-utilization 0.92 on lmcache DEP8 (matches the official DEP8 derate) and 0.94 on lmcache TP/DEP4 (at 0.96 the PR #2232 bring-up sweep hit torch-pool, DeepGEMM-JIT, and cuBLAS-workspace OOMs; all 23 lmcache points green after the derate)"
-    - "Standalone dsv4-fp4-b300-vllm-agentic-lmcache points (validated in the PR #2232 bring-up sweeps): TP8 and TP4 conc [8, 12, 16]; TP4-DEP4 conc [8, 16, 24, 32, 40, 48, 56, 64, 72]; TP8-DEP8 conc [32, 64, 96, 128, 160, 192, 224, 228]"
-    - "Parent dsv4-fp4-b300-vllm-agentic gains lmcache arms interleaving the SimpleCPU ladders at a +4 conc offset (TP4 conc [32, 36, 40, 44]; DEP4 conc [36, 44, 52, 60, 68, 76]), so the lmcache backend is tracked in the main B300 AgentX sweep on distinct points"
+    - "Add LMCache 0.5.1 DRAM KV-offload arms to the official B300 AgentX config from PR #2241 (sparse DSV4 FlashInfer attention, AMXF4 mega-MoE, FULL_DECODE_ONLY CUDA graphs, image vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec), so LMCache is benchmarked in the same sweep as the SimpleCPU and GPU-resident arms"
+    - "LMCache MP server (lmcache_driven transfer mode) + LMCacheMPConnector per PR #2153/#2231; L1 pool derated to 75% of TOTAL_CPU_DRAM_GB; PYTORCH_ALLOC_CONF=expandable_segments:True dropped on the lmcache arms only (VMM allocations cannot be CUDA-IPC-exported to the LMCache server, same failure mode as cuMem on B200); gpu-memory-utilization 0.92 on lmcache DEP8 (matches the official DEP8 derate) and 0.94 on lmcache TP4/DEP4 (at 0.96 the PR #2232 bring-up sweeps hit torch-pool, DeepGEMM-JIT, and cuBLAS-workspace OOMs; 23 lmcache points validated green in runs 29463061871/29535333851 after the derate)"
+    - "LMCache ladders interleave the official arms at a +4 conc offset: TP4 [32, 36, 40, 44] vs SimpleCPU [28, 32, 36, 40]; DEP4 [36, 44, 52, 60, 68, 76] vs SimpleCPU [32, 40, 48, 56, 64, 72]; DEP8 [68, 100, 116, 132, 148, 164, 180, 196, 228] vs GPU-resident [64, 96, 112, 128, 144, 160, 176, 192, 224]"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2232
 

From 690b4f102f1e5872f367c4a7d55202343c0d7d9a Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Fri, 17 Jul 2026 13:08:13 -0700
Subject: [PATCH 7/9] chore(agentic): DEP8 lmcache ladder at +16 conc offset
 capped at 208
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Per review: offset the DEP8 lmcache arm by +16 conc from the official
GPU-resident DEP8 ladder and cap the maximum at 208, giving
[80, 112, 128, 144, 160, 176, 192, 208]. TP4/DEP4 keep the +4 offset.

中文：按评审意见，DEP8 lmcache 分支相对官方 GPU 驻留 DEP8 阶梯偏移 +16
并发，最大并发限制为 208，即 [80, 112, 128, 144, 160, 176, 192, 208]。
TP4/DEP4 保持 +4 偏移。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 configs/nvidia-master.yaml | 11 ++++++-----
 perf-changelog.yaml        |  2 +-
 2 files changed, 7 insertions(+), 6 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 12a9486be7..24dc102b98 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -3962,16 +3962,17 @@ dsv4-fp4-b300-vllm-agentic:
       - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: vllm-simple, version: "904e4ec" }, conc-list: [32, 40, 48, 56, 64, 72], router: { name: vllm-router, version: "0.1.14" } }
       # DEP8 SimpleCPU
       - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, conc-list: [64, 96, 112, 128, 144, 160, 176, 192, 224], router: { name: vllm-router, version: "0.1.14" } }
-      # LMCache arms: interleave each official ladder above at a +4 conc
-      # offset so the backends sample distinct points on the same recipe.
-      # Bring-up validated in the PR #2232 sweeps (runs 29463061871 and
-      # 29535333851): TP/DEP4 at gpu-memory-utilization 0.94, DEP8 at 0.92.
+      # LMCache arms: offset each official ladder above so the backends
+      # sample distinct points on the same recipe — TP4/DEP4 at +4 conc,
+      # DEP8 at +16 conc capped at 208. Bring-up validated in the PR #2232
+      # sweeps (runs 29463061871 and 29535333851): TP/DEP4 at
+      # gpu-memory-utilization 0.94, DEP8 at 0.92.
       # TP4 LMCache (vs TP4 SimpleCPU)
       - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [32, 36, 40, 44] }
       # DEP4 LMCache (vs DEP4 SimpleCPU)
       - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [36, 44, 52, 60, 68, 76], router: { name: vllm-router, version: "0.1.14" } }
       # DEP8 LMCache (vs GPU-resident DEP8)
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [68, 100, 116, 132, 148, 164, 180, 196, 228], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [80, 112, 128, 144, 160, 176, 192, 208], router: { name: vllm-router, version: "0.1.14" } }
 
 dsv4-fp4-b300-trt:
   image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-c185066
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 82ab4b6cf7..bfee94c7fd 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4935,6 +4935,6 @@
   description:
     - "Add LMCache 0.5.1 DRAM KV-offload arms to the official B300 AgentX config from PR #2241 (sparse DSV4 FlashInfer attention, AMXF4 mega-MoE, FULL_DECODE_ONLY CUDA graphs, image vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec), so LMCache is benchmarked in the same sweep as the SimpleCPU and GPU-resident arms"
     - "LMCache MP server (lmcache_driven transfer mode) + LMCacheMPConnector per PR #2153/#2231; L1 pool derated to 75% of TOTAL_CPU_DRAM_GB; PYTORCH_ALLOC_CONF=expandable_segments:True dropped on the lmcache arms only (VMM allocations cannot be CUDA-IPC-exported to the LMCache server, same failure mode as cuMem on B200); gpu-memory-utilization 0.92 on lmcache DEP8 (matches the official DEP8 derate) and 0.94 on lmcache TP4/DEP4 (at 0.96 the PR #2232 bring-up sweeps hit torch-pool, DeepGEMM-JIT, and cuBLAS-workspace OOMs; 23 lmcache points validated green in runs 29463061871/29535333851 after the derate)"
-    - "LMCache ladders interleave the official arms at a +4 conc offset: TP4 [32, 36, 40, 44] vs SimpleCPU [28, 32, 36, 40]; DEP4 [36, 44, 52, 60, 68, 76] vs SimpleCPU [32, 40, 48, 56, 64, 72]; DEP8 [68, 100, 116, 132, 148, 164, 180, 196, 228] vs GPU-resident [64, 96, 112, 128, 144, 160, 176, 192, 224]"
+    - "LMCache ladders offset the official arms: TP4 +4 conc [32, 36, 40, 44] vs SimpleCPU [28, 32, 36, 40]; DEP4 +4 conc [36, 44, 52, 60, 68, 76] vs SimpleCPU [32, 40, 48, 56, 64, 72]; DEP8 +16 conc capped at 208 [80, 112, 128, 144, 160, 176, 192, 208] vs GPU-resident [64, 96, 112, 128, 144, 160, 176, 192, 224]"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2232
 

From 8a88a78627b2310bb4c6ad1dd1c426aa1269a30c Mon Sep 17 00:00:00 2001
From: ApostaC <yihua98@uchicago.edu>
Date: Fri, 17 Jul 2026 13:09:54 -0700
Subject: [PATCH 8/9] chore(agentic): DEP8 lmcache ladder at +8 conc offset
 capped at 208
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Correct the DEP8 lmcache offset from +16 to +8 (still capped at 208),
giving [72, 104, 120, 136, 152, 168, 184, 200].

中文：将 DEP8 lmcache 分支的并发偏移从 +16 更正为 +8（最大并发仍限制为
208），即 [72, 104, 120, 136, 152, 168, 184, 200]。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 configs/nvidia-master.yaml | 4 ++--
 perf-changelog.yaml        | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 24dc102b98..34c0809aa4 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -3964,7 +3964,7 @@ dsv4-fp4-b300-vllm-agentic:
       - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, conc-list: [64, 96, 112, 128, 144, 160, 176, 192, 224], router: { name: vllm-router, version: "0.1.14" } }
       # LMCache arms: offset each official ladder above so the backends
       # sample distinct points on the same recipe — TP4/DEP4 at +4 conc,
-      # DEP8 at +16 conc capped at 208. Bring-up validated in the PR #2232
+      # DEP8 at +8 conc capped at 208. Bring-up validated in the PR #2232
       # sweeps (runs 29463061871 and 29535333851): TP/DEP4 at
       # gpu-memory-utilization 0.94, DEP8 at 0.92.
       # TP4 LMCache (vs TP4 SimpleCPU)
@@ -3972,7 +3972,7 @@ dsv4-fp4-b300-vllm-agentic:
       # DEP4 LMCache (vs DEP4 SimpleCPU)
       - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [36, 44, 52, 60, 68, 76], router: { name: vllm-router, version: "0.1.14" } }
       # DEP8 LMCache (vs GPU-resident DEP8)
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [80, 112, 128, 144, 160, 176, 192, 208], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: lmcache, version: "0.5.1" }, conc-list: [72, 104, 120, 136, 152, 168, 184, 200], router: { name: vllm-router, version: "0.1.14" } }
 
 dsv4-fp4-b300-trt:
   image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-c185066
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index bfee94c7fd..97ba12ef34 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4935,6 +4935,6 @@
   description:
     - "Add LMCache 0.5.1 DRAM KV-offload arms to the official B300 AgentX config from PR #2241 (sparse DSV4 FlashInfer attention, AMXF4 mega-MoE, FULL_DECODE_ONLY CUDA graphs, image vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec), so LMCache is benchmarked in the same sweep as the SimpleCPU and GPU-resident arms"
     - "LMCache MP server (lmcache_driven transfer mode) + LMCacheMPConnector per PR #2153/#2231; L1 pool derated to 75% of TOTAL_CPU_DRAM_GB; PYTORCH_ALLOC_CONF=expandable_segments:True dropped on the lmcache arms only (VMM allocations cannot be CUDA-IPC-exported to the LMCache server, same failure mode as cuMem on B200); gpu-memory-utilization 0.92 on lmcache DEP8 (matches the official DEP8 derate) and 0.94 on lmcache TP4/DEP4 (at 0.96 the PR #2232 bring-up sweeps hit torch-pool, DeepGEMM-JIT, and cuBLAS-workspace OOMs; 23 lmcache points validated green in runs 29463061871/29535333851 after the derate)"
-    - "LMCache ladders offset the official arms: TP4 +4 conc [32, 36, 40, 44] vs SimpleCPU [28, 32, 36, 40]; DEP4 +4 conc [36, 44, 52, 60, 68, 76] vs SimpleCPU [32, 40, 48, 56, 64, 72]; DEP8 +16 conc capped at 208 [80, 112, 128, 144, 160, 176, 192, 208] vs GPU-resident [64, 96, 112, 128, 144, 160, 176, 192, 224]"
+    - "LMCache ladders offset the official arms: TP4 +4 conc [32, 36, 40, 44] vs SimpleCPU [28, 32, 36, 40]; DEP4 +4 conc [36, 44, 52, 60, 68, 76] vs SimpleCPU [32, 40, 48, 56, 64, 72]; DEP8 +8 conc capped at 208 [72, 104, 120, 136, 152, 168, 184, 200] vs GPU-resident [64, 96, 112, 128, 144, 160, 176, 192, 224]"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2232
 

From 26bf7c79a61ba6b647056a0e5a118be0cb863a39 Mon Sep 17 00:00:00 2001
From: Cameron Quilici <cjquilici@gmail.com>
Date: Tue, 21 Jul 2026 10:18:34 -0500
Subject: [PATCH 9/9] Update perf-changelog.yaml

---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 7cc4b998e8..8162941c39 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -4992,4 +4992,4 @@
     - "LMCache MP server (lmcache_driven transfer mode) + LMCacheMPConnector per PR #2153/#2231; L1 pool derated to 75% of TOTAL_CPU_DRAM_GB; PYTORCH_ALLOC_CONF=expandable_segments:True dropped on the lmcache arms only (VMM allocations cannot be CUDA-IPC-exported to the LMCache server, same failure mode as cuMem on B200); gpu-memory-utilization 0.92 on lmcache DEP8 (matches the official DEP8 derate) and 0.94 on lmcache TP4/DEP4 (at 0.96 the PR #2232 bring-up sweeps hit torch-pool, DeepGEMM-JIT, and cuBLAS-workspace OOMs; 23 lmcache points validated green in runs 29463061871/29535333851 after the derate)"
     - "LMCache ladders offset the official arms: TP4 +4 conc [32, 36, 40, 44] vs SimpleCPU [28, 32, 36, 40]; DEP4 +4 conc [36, 44, 52, 60, 68, 76] vs SimpleCPU [32, 40, 48, 56, 64, 72]; DEP8 +8 conc capped at 208 [72, 104, 120, 136, 152, 168, 184, 200] vs GPU-resident [64, 96, 112, 128, 144, 160, 176, 192, 224]"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2232
-  
\ No newline at end of file
+