From 9726490ad4c092465c3046c907df7250a8632b32 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 29 Jul 2026 23:48:20 -0500
Subject: [PATCH 01/30] feat: add Qwen3.5 FP4 B300 AgentX MTP
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add the initial SGLang NEXTN AgentX recipe with golden synthetic acceptance and 256k traces.\n\n中文：新增 Qwen3.5 FP4 B300 AgentX MTP 初始配置，使用 SGLang NEXTN、黄金合成接受长度和 256k 轨迹数据集。
---
 .../agentic/qwen3.5_fp4_b300_sglang_mtp.sh    | 118 ++++++++++++++++++
 configs/nvidia-master.yaml                    |  15 +++
 2 files changed, 133 insertions(+)
 create mode 100755 benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
new file mode 100755
index 0000000000..a622a42500
--- /dev/null
+++ b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
@@ -0,0 +1,118 @@
+#!/usr/bin/env bash
+set -euo pipefail
+set -x
+
+# AgentX trace replay for Qwen3.5-397B-A17B NVFP4 on B300 with SGLang
+# native NEXTN MTP. Throughput uses the committed golden synthetic AL; evals
+# retain real target-model verification.
+
+source "$(dirname "$0")/../../benchmark_lib.sh"
+
+check_env_vars \
+    MODEL TP CONC EP_SIZE KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
+
+SCHEDULER_RECV_INTERVAL=${SCHEDULER_RECV_INTERVAL:-10}
+
+if [[ -n "${SLURM_JOB_ID:-}" ]]; then
+    echo "JOB $SLURM_JOB_ID running on ${SLURMD_NODENAME:-unknown}"
+fi
+
+if [[ -n "${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="$MODEL"
+fi
+nvidia-smi
+
+export WEKA_LOADER_OVERRIDE=semianalysis_cc_traces_weka_062126_256k
+resolve_trace_source
+install_agentic_deps
+
+SERVER_LOG="$RESULT_DIR/server.log"
+mkdir -p "$RESULT_DIR"
+
+CACHE_ARGS=()
+if require_agentic_kv_offload_backend hicache; then
+    HICACHE_HOST_POOL_COUNT="${HICACHE_HOST_POOL_COUNT:-2}"
+    MAX_HICACHE_SIZE_GB=$((TOTAL_CPU_DRAM_GB / TP / HICACHE_HOST_POOL_COUNT))
+    HICACHE_SIZE_GB="${HICACHE_SIZE_GB:-$MAX_HICACHE_SIZE_GB}"
+    if [ "$HICACHE_SIZE_GB" -lt 1 ] || [ "$HICACHE_SIZE_GB" -gt "$MAX_HICACHE_SIZE_GB" ]; then
+        echo "Error: HICACHE_SIZE_GB=$HICACHE_SIZE_GB outside 1..$MAX_HICACHE_SIZE_GB" >&2
+        exit 1
+    fi
+    CACHE_ARGS=(
+        --page-size 64
+        --enable-hierarchical-cache
+        --hicache-size "$HICACHE_SIZE_GB"
+        --hicache-io-backend kernel
+        --hicache-mem-layout page_first
+        --hicache-write-policy write_through_selective
+    )
+fi
+
+export TORCH_CUDA_ARCH_LIST="10.0"
+export PYTHONNOUSERSITE=1
+export NCCL_NVLS_ENABLE=1
+export SGL_ENABLE_JIT_DEEPGEMM=false
+export SGLANG_ENABLE_FLASHINFER_GEMM=true
+export SGLANG_ENABLE_SPEC_V2=1
+
+if [ "${EVAL_ONLY:-false}" != "true" ]; then
+    export SGLANG_SIMULATE_ACC_LEN=3.39
+    export SGLANG_SIMULATE_ACC_METHOD=match-expected
+    export SGLANG_SIMULATE_ACC_TOKEN_MODE=real-draft-token
+fi
+
+SGLANG_CMD=(
+    python3 -m sglang.launch_server
+    --model-path "$MODEL_PATH"
+    --served-model-name "$MODEL"
+    --host 0.0.0.0
+    --port "$PORT"
+    --trust-remote-code
+    --tensor-parallel-size "$TP"
+    --data-parallel-size 1
+    --expert-parallel-size "$EP_SIZE"
+    --enable-symm-mem
+    --quantization modelopt_fp4
+    --fp4-gemm-backend flashinfer_cutlass
+    --kv-cache-dtype fp8_e4m3
+    --mamba-ssm-dtype bfloat16
+    --attention-backend trtllm_mha
+    --moe-runner-backend flashinfer_trtllm
+    --cuda-graph-max-bs "$CONC"
+    --max-running-requests "$CONC"
+    --max-prefill-tokens 16384
+    --chunked-prefill-size 16384
+    --mem-fraction-static 0.80
+    --stream-interval 50
+    --scheduler-recv-interval "$SCHEDULER_RECV_INTERVAL"
+    --tokenizer-worker-num 6
+    --tokenizer-path "$MODEL"
+    --reasoning-parser qwen3
+    --tool-call-parser qwen3_coder
+    --enable-auto-tool-choice
+    --speculative-algorithm NEXTN
+    --speculative-num-steps 3
+    --speculative-eagle-topk 1
+    --speculative-num-draft-tokens 4
+    --enable-metrics
+    "${CACHE_ARGS[@]}"
+)
+
+printf '%q ' "${SGLANG_CMD[@]}" | tee "$RESULT_DIR/sglang_command.txt"
+printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
+"${SGLANG_CMD[@]}" > "$SERVER_LOG" 2>&1 &
+SERVER_PID=$!
+
+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/nvidia-master.yaml b/configs/nvidia-master.yaml
index 8ee037c8c2..eca7b7bbf4 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7140,6 +7140,21 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
       - { tp: 4, ep: 1, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
       - { tp: 4, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
+qwen3.5-fp4-b300-sglang-agentic-mtp:
+  image: lmsysorg/sglang:nightly-dev-cu13-20260724-433429b1
+  model: nvidia/Qwen3.5-397B-A17B-NVFP4
+  model-prefix: qwen3.5
+  runner: cluster:b300-nv
+  precision: fp4
+  framework: sglang
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 8, 16] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 16, 32] }
+
 kimik2.5-fp4-b200-vllm-agentic-lmcache:
   image: vllm/vllm-openai:v0.22.0
   model: nvidia/Kimi-K2.5-NVFP4

From c8dd6602ee8ea6001a9df4bbac740b46c795b528 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 29 Jul 2026 23:48:49 -0500
Subject: [PATCH 02/30] chore: trigger B300 AgentX sweep
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Append the Qwen3.5 FP4 B300 AgentX MTP benchmark entry.\n\n中文：追加 Qwen3.5 FP4 B300 AgentX MTP 基准测试触发记录。
---
 perf-changelog.yaml | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index c1736ae420..55520b4d95 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5311,3 +5311,10 @@
     - "Search space mirrors the non-MTP entry's KV arms -- TP8 GPU-resident and TP8 host-DRAM offload -- so the spec-decoding delta is readable at equal concurrency, but stops at conc 16 rather than 24. The non-MTP bring-up sweep (run 30326393603) showed the GPU-resident arm already thrashing at conc >= 16 (prefix cache hit rate 2.7%, TTFT p50 86-191s) because GPU KV holds only ~3.1 max-length requests, so conc 24 would spend a full job per arm re-measuring that regime; conc 16 still exercises the DRAM tier meaningfully (62% external prefix cache hit rate). TP8-only for the same memory reason: a ~1.5 TB MXFP4 checkpoint needs ~188 GB/GPU across 8 B300s and does not fit below 8 GPUs."
     - "Sets VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION=1, which the upstream recipe requires on both its blackwell and nvidia paths and which this repo had never set. It defaults to 0 and is threaded into LatentMoERunner as runner_args={\"enable_k3_latent_moe_tail_fusion\": ...} at vllm/models/kimi_k3/nvidia/model.py:549 -- the same runner whose shared-experts output buffer asserted (fused_moe/runner/shared_experts.py:165, all 8 TP ranks at once) when the wider flag alignment was attempted, so every K3 run here so far has been exercising a MoE tail path upstream does not use. Enabled on its own, ahead of re-attempting gpu-memory-utilization 0.95, max-num-seqs 32, --no-enable-flashinfer-autotune or VLLM_USE_V2_MODEL_RUNNER=1, to isolate its effect."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2418
+
+- config-keys:
+    - qwen3.5-fp4-b300-sglang-agentic-mtp
+  description:
+    - "Add Qwen3.5-397B-A17B NVFP4 AgentX benchmark on B300 with SGLang native NEXTN MTP"
+    - "Use the 256k trace dataset and golden synthetic acceptance length 3.39"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2421

From 85521aad83b874d10475da69077e3146bbe2a589 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 29 Jul 2026 23:50:42 -0500
Subject: [PATCH 03/30] perf: expand B300 AgentX preflight range

---
 configs/nvidia-master.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index eca7b7bbf4..679bf95b82 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7152,8 +7152,8 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 8, 16] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 16, 32] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 16, 32, 64, 96, 128] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 8, 16, 32, 64, 96, 128, 192, 256] }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:
   image: vllm/vllm-openai:v0.22.0

From ad6e2d3424996fbc87063fa27d3be9fbe97346d4 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 29 Jul 2026 23:54:43 -0500
Subject: [PATCH 04/30] chore: use SGLang v0.5.16 for Qwen B300

---
 configs/nvidia-master.yaml | 2 +-
 perf-changelog.yaml        | 1 +
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 679bf95b82..d3011ccd6e 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7141,7 +7141,7 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
       - { tp: 4, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
 qwen3.5-fp4-b300-sglang-agentic-mtp:
-  image: lmsysorg/sglang:nightly-dev-cu13-20260724-433429b1
+  image: lmsysorg/sglang:v0.5.16-cu130
   model: nvidia/Qwen3.5-397B-A17B-NVFP4
   model-prefix: qwen3.5
   runner: cluster:b300-nv
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 55520b4d95..4b2e033d4d 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5317,4 +5317,5 @@
   description:
     - "Add Qwen3.5-397B-A17B NVFP4 AgentX benchmark on B300 with SGLang native NEXTN MTP"
     - "Use the 256k trace dataset and golden synthetic acceptance length 3.39"
+    - "Image: lmsysorg/sglang:v0.5.16-cu130"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2421

From a58dcef9a990fe60c8fa46947159116ec14bf10d Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 29 Jul 2026 23:55:38 -0500
Subject: [PATCH 05/30] chore: collect SGLang cache metrics on B300

---
 benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh | 1 +
 1 file changed, 1 insertion(+)

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
index a622a42500..703546064d 100755
--- a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
@@ -114,5 +114,6 @@ if [ "${EVAL_ONLY:-false}" = "true" ]; then
     run_eval --port "$PORT"
 else
     build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --server-metrics http://localhost:$PORT/metrics"
     run_agentic_replay_and_write_outputs "$RESULT_DIR"
 fi

From 8ffac86310384bade5a43c3d14acde1904992049 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Wed, 29 Jul 2026 23:59:58 -0500
Subject: [PATCH 06/30] fix: remove unsupported SGLang tool-choice flag on B300

---
 benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh | 1 -
 1 file changed, 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
index 703546064d..93692c6fb2 100755
--- a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
@@ -94,7 +94,6 @@ SGLANG_CMD=(
     --tokenizer-path "$MODEL"
     --reasoning-parser qwen3
     --tool-call-parser qwen3_coder
-    --enable-auto-tool-choice
     --speculative-algorithm NEXTN
     --speculative-num-steps 3
     --speculative-eagle-topk 1

From 87bf252239418965960489b534329982877bfe61 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 00:59:57 -0500
Subject: [PATCH 07/30] perf: prepare Qwen B300 cache and DEP tuning
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add strict two-pool HiCache accounting, cache reporting, session-sticky DEP8 routing, AgentX request/graph limits, and measured-cache guidance for the B300 frontier.\n\n中文：为 B300 前沿调优加入严格的双池 HiCache 容量核算、缓存统计、会话粘性的 DEP8 路由、AgentX 请求与 CUDA Graph 上限，并记录实测缓存容量依据。
---
 .../agentic/qwen3.5_fp4_b300_sglang_mtp.sh    | 101 +++++++++++++++---
 configs/nvidia-master.yaml                    |   7 ++
 2 files changed, 96 insertions(+), 12 deletions(-)

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
index 93692c6fb2..3ce3af6a12 100755
--- a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
@@ -9,7 +9,8 @@ set -x
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
 check_env_vars \
-    MODEL TP CONC EP_SIZE KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
+    MODEL TP CONC EP_SIZE DP_ATTENTION KV_OFFLOADING \
+    TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
 
 SCHEDULER_RECV_INTERVAL=${SCHEDULER_RECV_INTERVAL:-10}
 
@@ -36,6 +37,12 @@ mkdir -p "$RESULT_DIR"
 
 CACHE_ARGS=()
 if require_agentic_kv_offload_backend hicache; then
+    REQUESTED_HICACHE_TOTAL_GB="${HICACHE_TOTAL_CPU_DRAM_GB:-$TOTAL_CPU_DRAM_GB}"
+    if [ "$REQUESTED_HICACHE_TOTAL_GB" -gt "$TOTAL_CPU_DRAM_GB" ]; then
+        echo "Error: requested HiCache pool ${REQUESTED_HICACHE_TOTAL_GB} GB exceeds configured capacity ${TOTAL_CPU_DRAM_GB} GB" >&2
+        exit 1
+    fi
+    TOTAL_CPU_DRAM_GB="$REQUESTED_HICACHE_TOTAL_GB"
     HICACHE_HOST_POOL_COUNT="${HICACHE_HOST_POOL_COUNT:-2}"
     MAX_HICACHE_SIZE_GB=$((TOTAL_CPU_DRAM_GB / TP / HICACHE_HOST_POOL_COUNT))
     HICACHE_SIZE_GB="${HICACHE_SIZE_GB:-$MAX_HICACHE_SIZE_GB}"
@@ -43,6 +50,7 @@ if require_agentic_kv_offload_backend hicache; then
         echo "Error: HICACHE_SIZE_GB=$HICACHE_SIZE_GB outside 1..$MAX_HICACHE_SIZE_GB" >&2
         exit 1
     fi
+    echo "HiCache CPU pools: ${HICACHE_SIZE_GB} GB per rank per pool across TP=${TP}, pools=${HICACHE_HOST_POOL_COUNT}; node total <= ${TOTAL_CPU_DRAM_GB} GB"
     CACHE_ARGS=(
         --page-size 64
         --enable-hierarchical-cache
@@ -53,12 +61,51 @@ if require_agentic_kv_offload_backend hicache; then
     )
 fi
 
+USE_SGLANG_ROUTER=false
+SGLANG_BACKEND_PORT="$PORT"
+ROUTER_LOG="$RESULT_DIR/router.log"
+if [ "$DP_ATTENTION" = "true" ]; then
+    USE_SGLANG_ROUTER=true
+    export AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true
+    SGLANG_BACKEND_PORT=$((PORT + 1))
+    SGLANG_ROUTER_METRICS_PORT=$((PORT + 10000))
+    SGLANG_ROUTER_CMD=(python3 -m sglang_router.launch_router)
+fi
+
+PARALLEL_ARGS=(
+    --tp "$TP"
+    --dp 1
+    --ep-size "$EP_SIZE"
+)
+TOKENIZER_WORKERS=6
+STREAM_INTERVAL=50
+if [ "$DP_ATTENTION" = "true" ]; then
+    PARALLEL_ARGS=(
+        --tp "$TP"
+        --dp "$TP"
+        --ep-size "$EP_SIZE"
+        --enable-dp-attention
+        --enable-dp-attention-local-control-broadcast
+        --enable-dp-lm-head
+        --dist-init-addr "127.0.0.1:$((PORT + 2000))"
+        --incremental-streaming-output
+    )
+    TOKENIZER_WORKERS="$TP"
+    STREAM_INTERVAL=20
+fi
+
+# AgentX concurrency counts live session trees rather than individual HTTP
+# requests. Leave room for subagent fan-out and avoid spending HBM on graphs
+# above the batch sizes that remain useful for this long-context workload.
+MAX_RUNNING_REQUESTS=$((2 * CONC))
+CUDA_GRAPH_MAX_BS="$CONC"
+[ "$CUDA_GRAPH_MAX_BS" -gt 64 ] && CUDA_GRAPH_MAX_BS=64
+
 export TORCH_CUDA_ARCH_LIST="10.0"
 export PYTHONNOUSERSITE=1
 export NCCL_NVLS_ENABLE=1
 export SGL_ENABLE_JIT_DEEPGEMM=false
 export SGLANG_ENABLE_FLASHINFER_GEMM=true
-export SGLANG_ENABLE_SPEC_V2=1
 
 if [ "${EVAL_ONLY:-false}" != "true" ]; then
     export SGLANG_SIMULATE_ACC_LEN=3.39
@@ -71,11 +118,9 @@ SGLANG_CMD=(
     --model-path "$MODEL_PATH"
     --served-model-name "$MODEL"
     --host 0.0.0.0
-    --port "$PORT"
+    --port "$SGLANG_BACKEND_PORT"
     --trust-remote-code
-    --tensor-parallel-size "$TP"
-    --data-parallel-size 1
-    --expert-parallel-size "$EP_SIZE"
+    "${PARALLEL_ARGS[@]}"
     --enable-symm-mem
     --quantization modelopt_fp4
     --fp4-gemm-backend flashinfer_cutlass
@@ -83,14 +128,14 @@ SGLANG_CMD=(
     --mamba-ssm-dtype bfloat16
     --attention-backend trtllm_mha
     --moe-runner-backend flashinfer_trtllm
-    --cuda-graph-max-bs "$CONC"
-    --max-running-requests "$CONC"
+    --cuda-graph-max-bs "$CUDA_GRAPH_MAX_BS"
+    --max-running-requests "$MAX_RUNNING_REQUESTS"
     --max-prefill-tokens 16384
     --chunked-prefill-size 16384
     --mem-fraction-static 0.80
-    --stream-interval 50
+    --stream-interval "$STREAM_INTERVAL"
     --scheduler-recv-interval "$SCHEDULER_RECV_INTERVAL"
-    --tokenizer-worker-num 6
+    --tokenizer-worker-num "$TOKENIZER_WORKERS"
     --tokenizer-path "$MODEL"
     --reasoning-parser qwen3
     --tool-call-parser qwen3_coder
@@ -99,6 +144,7 @@ SGLANG_CMD=(
     --speculative-eagle-topk 1
     --speculative-num-draft-tokens 4
     --enable-metrics
+    --enable-cache-report
     "${CACHE_ARGS[@]}"
 )
 
@@ -107,12 +153,43 @@ printf '\n' | tee -a "$RESULT_DIR/sglang_command.txt"
 "${SGLANG_CMD[@]}" > "$SERVER_LOG" 2>&1 &
 SERVER_PID=$!
 
-wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+capture_cache_metrics() {
+    {
+        echo "=== SGLang cache metrics snapshot $(date --iso-8601=seconds) ==="
+        curl -fsS "http://localhost:$SGLANG_BACKEND_PORT/metrics" 2>/dev/null \
+            | grep -E '^(sglang:(cache_hit_rate|cached_tokens_total|prompt_tokens_total|hicache_host_used_tokens|hicache_host_total_tokens|token_usage|num_requests_running|num_requests_waiting))' \
+            || true
+        echo "============================================================"
+    } >> "$SERVER_LOG"
+}
+
+wait_for_server_ready --port "$SGLANG_BACKEND_PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
+
+if [ "$USE_SGLANG_ROUTER" = "true" ]; then
+    "${SGLANG_ROUTER_CMD[@]}" \
+        --worker-urls "http://localhost:$SGLANG_BACKEND_PORT" \
+        --policy consistent_hashing \
+        --request-id-headers x-correlation-id \
+        --dp-aware \
+        --host 0.0.0.0 \
+        --port "$PORT" \
+        --prometheus-host 127.0.0.1 \
+        --prometheus-port "$SGLANG_ROUTER_METRICS_PORT" \
+        --connect-timeout-secs 900 \
+        --request-timeout-secs 14400 \
+        --disable-health-check \
+        --disable-retries > "$ROUTER_LOG" 2>&1 &
+    ROUTER_PID=$!
+    wait_for_server_ready --port "$PORT" --server-log "$ROUTER_LOG" --server-pid "$ROUTER_PID"
+fi
+
+capture_cache_metrics
+trap capture_cache_metrics EXIT
 
 if [ "${EVAL_ONLY:-false}" = "true" ]; then
     run_eval --port "$PORT"
 else
     build_replay_cmd "$RESULT_DIR"
-    REPLAY_CMD+=" --server-metrics http://localhost:$PORT/metrics"
+    REPLAY_CMD+=" --server-metrics http://localhost:$SGLANG_BACKEND_PORT/metrics"
     run_agentic_replay_and_write_outputs "$RESULT_DIR"
 fi
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index d3011ccd6e..dd348211df 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7140,6 +7140,12 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
       - { tp: 4, ep: 1, kv-offloading: none, conc-list: [1, 2, 4, 8, 16, 32] }
       - { tp: 4, ep: 1, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [16, 32, 48, 64] }
 
+# Qwen3.5 target FP8 KV is 15,360 bytes/token before the native MTP layer.
+# TP4 replicates the model's two KV heads and measured 30.5 KB/token node-wide;
+# TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
+# p75 ISL, the measured 9.56M-token TP2 pool should bend around conc 56-80,
+# while the 14.88M-token TP4 pool should bend around conc 88-128. DEP8 covers
+# the throughput end with session-sticky routing across attention-DP ranks.
 qwen3.5-fp4-b300-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
   model: nvidia/Qwen3.5-397B-A17B-NVFP4
@@ -7154,6 +7160,7 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
       search-space:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 16, 32, 64, 96, 128] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 8, 16, 32, 64, 96, 128, 192, 256] }
+      - { tp: 8, ep: 8, dp-attn: true, spec-decoding: mtp, kv-offloading: none, conc-list: [64, 128, 256], router: { name: sglang-router, version: "0.3.2" } }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:
   image: vllm/vllm-openai:v0.22.0

From 064b790099880be3e6a0721ce6cab68805316fd5 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 01:36:40 -0500
Subject: [PATCH 08/30] perf: add B300 HiCache cliff probes
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add isolated TP2 c72 and TP4 c80 DRAM HiCache points so each cache-cliff recovery topology can be measured independently.

中文：添加独立的 TP2 c72 与 TP4 c80 DRAM HiCache 测试点，以便分别测量各拓扑在缓存工作集拐点后的恢复效果。
---
 configs/nvidia-master.yaml | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index dd348211df..da5ac30a78 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7159,7 +7159,9 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
     - dram-utilization: 0.80
       search-space:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 16, 32, 64, 96, 128] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [80] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 8, 16, 32, 64, 96, 128, 192, 256] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [72] }
       - { tp: 8, ep: 8, dp-attn: true, spec-decoding: mtp, kv-offloading: none, conc-list: [64, 128, 256], router: { name: sglang-router, version: "0.3.2" } }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:

From 8570002a05d5874f3002f71dc79f9845a8a473bd Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 01:45:59 -0500
Subject: [PATCH 09/30] fix: enforce Qwen HiCache DRAM total
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Account for SGLang target KV, Mamba, and native NEXTN draft host pools as H * 31/15 per rank, with page-alignment reserve, so projected use stays within runner-generated DRAM.

中文：按每个 rank 的 H * 31/15 计入 SGLang 目标 KV、Mamba 与原生 NEXTN 草稿主机缓存，并预留页对齐空间，确保预计用量不超过 runner 生成的 DRAM 上限。
---
 .../agentic/qwen3.5_fp4_b300_sglang_mtp.sh    | 20 ++++++++++++++++---
 1 file changed, 17 insertions(+), 3 deletions(-)

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
index 3ce3af6a12..6552e75cd8 100755
--- a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
@@ -43,14 +43,28 @@ if require_agentic_kv_offload_backend hicache; then
         exit 1
     fi
     TOTAL_CPU_DRAM_GB="$REQUESTED_HICACHE_TOTAL_GB"
-    HICACHE_HOST_POOL_COUNT="${HICACHE_HOST_POOL_COUNT:-2}"
-    MAX_HICACHE_SIZE_GB=$((TOTAL_CPU_DRAM_GB / TP / HICACHE_HOST_POOL_COUNT))
+    # SGLang applies --hicache-size independently to Qwen's target KV and
+    # Mamba pools. Native NEXTN also creates a draft KV pool with the same
+    # slot count; its one attention layer adds 1/15 of the target KV bytes.
+    # Reserve 1 GB/rank for page alignment and enforce H * 31/15 per rank.
+    HICACHE_ALIGNMENT_RESERVE_GB=$TP
+    HICACHE_USABLE_TOTAL_GB=$((TOTAL_CPU_DRAM_GB - HICACHE_ALIGNMENT_RESERVE_GB))
+    if [ "$HICACHE_USABLE_TOTAL_GB" -lt 1 ]; then
+        echo "Error: insufficient DRAM after HiCache alignment reserve" >&2
+        exit 1
+    fi
+    MAX_HICACHE_SIZE_GB=$((HICACHE_USABLE_TOTAL_GB * 15 / TP / 31))
     HICACHE_SIZE_GB="${HICACHE_SIZE_GB:-$MAX_HICACHE_SIZE_GB}"
     if [ "$HICACHE_SIZE_GB" -lt 1 ] || [ "$HICACHE_SIZE_GB" -gt "$MAX_HICACHE_SIZE_GB" ]; then
         echo "Error: HICACHE_SIZE_GB=$HICACHE_SIZE_GB outside 1..$MAX_HICACHE_SIZE_GB" >&2
         exit 1
     fi
-    echo "HiCache CPU pools: ${HICACHE_SIZE_GB} GB per rank per pool across TP=${TP}, pools=${HICACHE_HOST_POOL_COUNT}; node total <= ${TOTAL_CPU_DRAM_GB} GB"
+    PROJECTED_HICACHE_TOTAL_GB=$(((HICACHE_SIZE_GB * TP * 31 + 14) / 15 + HICACHE_ALIGNMENT_RESERVE_GB))
+    if [ "$PROJECTED_HICACHE_TOTAL_GB" -gt "$TOTAL_CPU_DRAM_GB" ]; then
+        echo "Error: projected HiCache use ${PROJECTED_HICACHE_TOTAL_GB} GB exceeds configured capacity ${TOTAL_CPU_DRAM_GB} GB" >&2
+        exit 1
+    fi
+    echo "HiCache CPU pools: ${HICACHE_SIZE_GB} GB target + Mamba + 1/15 draft per rank across TP=${TP}; projected node total ${PROJECTED_HICACHE_TOTAL_GB} GB <= ${TOTAL_CPU_DRAM_GB} GB"
     CACHE_ARGS=(
         --page-size 64
         --enable-hierarchical-cache

From 40649797a4654e432bff0ea79787e9cadee2a008 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:16:05 -0500
Subject: [PATCH 10/30] perf: densify B300 Qwen cache crossover
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add TP4 concurrency points around the measured HiCache crossover and extend the DRAM-offload curve beyond it.\n\n中文：在实测 HiCache 切换区间细化 B300 Qwen TP4 并发点，并将 DRAM 卸载曲线扩展到拐点之后。
---
 configs/nvidia-master.yaml | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 298df88657..04dabea8f5 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7179,9 +7179,10 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # Qwen3.5 target FP8 KV is 15,360 bytes/token before the native MTP layer.
 # TP4 replicates the model's two KV heads and measured 30.5 KB/token node-wide;
 # TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
-# p75 ISL, the measured 9.56M-token TP2 pool should bend around conc 56-80,
-# while the 14.88M-token TP4 pool should bend around conc 88-128. DEP8 covers
-# the throughput end with session-sticky routing across attention-DP ranks.
+# p75 ISL, the measured 9.56M-token TP2 pool should bend around conc 56-80.
+# TP4 remains healthy at c64; HiCache sustains 87% prefix hits and 234k input
+# tok/s at c80, so resolve the crossover densely and carry offload past it.
+# DEP8 probes the throughput end with session-sticky attention-DP routing.
 qwen3.5-fp4-b300-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
   model: nvidia/Qwen3.5-397B-A17B-NVFP4
@@ -7194,8 +7195,8 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 16, 32, 64, 96, 128] }
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [80] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72, 80, 88, 96] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 72, 80, 88, 96, 104, 112] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 8, 16, 32, 64, 96, 128, 192, 256] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [72] }
       - { tp: 8, ep: 8, dp-attn: true, spec-decoding: mtp, kv-offloading: none, conc-list: [64, 128, 256], router: { name: sglang-router, version: "0.3.2" } }

From 05d97cfb07b4988a49bb4b11afead847011a5d9a Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:17:16 -0500
Subject: [PATCH 11/30] perf: bracket B300 TP2 offload frontier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Remove the measured no-offload thrashing tail and add dense TP2 points across the expected HiCache crossover.\n\n中文：移除实测发生抖动的无卸载尾部配置，并在预期 HiCache 切换区间细化 B300 TP2 并发点。
---
 configs/nvidia-master.yaml | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 04dabea8f5..38704e96e7 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7179,7 +7179,8 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # Qwen3.5 target FP8 KV is 15,360 bytes/token before the native MTP layer.
 # TP4 replicates the model's two KV heads and measured 30.5 KB/token node-wide;
 # TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
-# p75 ISL, the measured 9.56M-token TP2 pool should bend around conc 56-80.
+# p75 ISL, TP2 is healthy through c32 but reaches 97% GPU KV with zero prefix
+# hits by c96, placing its useful no-offload/offload crossover below c64.
 # TP4 remains healthy at c64; HiCache sustains 87% prefix hits and 234k input
 # tok/s at c80, so resolve the crossover densely and carry offload past it.
 # DEP8 probes the throughput end with session-sticky attention-DP routing.
@@ -7197,8 +7198,8 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
       search-space:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72, 80, 88, 96] }
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 72, 80, 88, 96, 104, 112] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 8, 16, 32, 64, 96, 128, 192, 256] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [72] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72] }
       - { tp: 8, ep: 8, dp-attn: true, spec-decoding: mtp, kv-offloading: none, conc-list: [64, 128, 256], router: { name: sglang-router, version: "0.3.2" } }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:

From 0b5465ce47a74fea022c1c0bef121606f4d282bc Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:18:55 -0500
Subject: [PATCH 12/30] perf: remove dominated B300 Qwen DEP topology
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Drop the attention-DP arm after live AgentX profiling showed fragmented prefix caches and throughput below the TP frontier. Remove the now-unused router path.\n\n中文：AgentX 实测表明注意力 DP 会导致前缀缓存碎片化且吞吐低于 TP 前沿，因此移除该配置及不再使用的路由路径。
---
 .../agentic/qwen3.5_fp4_b300_sglang_mtp.sh    | 53 ++-----------------
 configs/nvidia-master.yaml                    |  3 +-
 2 files changed, 6 insertions(+), 50 deletions(-)

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
index 6552e75cd8..b0aa51657e 100755
--- a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
@@ -9,7 +9,7 @@ set -x
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
 check_env_vars \
-    MODEL TP CONC EP_SIZE DP_ATTENTION KV_OFFLOADING \
+    MODEL TP CONC EP_SIZE KV_OFFLOADING \
     TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
 
 SCHEDULER_RECV_INTERVAL=${SCHEDULER_RECV_INTERVAL:-10}
@@ -75,17 +75,6 @@ if require_agentic_kv_offload_backend hicache; then
     )
 fi
 
-USE_SGLANG_ROUTER=false
-SGLANG_BACKEND_PORT="$PORT"
-ROUTER_LOG="$RESULT_DIR/router.log"
-if [ "$DP_ATTENTION" = "true" ]; then
-    USE_SGLANG_ROUTER=true
-    export AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true
-    SGLANG_BACKEND_PORT=$((PORT + 1))
-    SGLANG_ROUTER_METRICS_PORT=$((PORT + 10000))
-    SGLANG_ROUTER_CMD=(python3 -m sglang_router.launch_router)
-fi
-
 PARALLEL_ARGS=(
     --tp "$TP"
     --dp 1
@@ -93,20 +82,6 @@ PARALLEL_ARGS=(
 )
 TOKENIZER_WORKERS=6
 STREAM_INTERVAL=50
-if [ "$DP_ATTENTION" = "true" ]; then
-    PARALLEL_ARGS=(
-        --tp "$TP"
-        --dp "$TP"
-        --ep-size "$EP_SIZE"
-        --enable-dp-attention
-        --enable-dp-attention-local-control-broadcast
-        --enable-dp-lm-head
-        --dist-init-addr "127.0.0.1:$((PORT + 2000))"
-        --incremental-streaming-output
-    )
-    TOKENIZER_WORKERS="$TP"
-    STREAM_INTERVAL=20
-fi
 
 # AgentX concurrency counts live session trees rather than individual HTTP
 # requests. Leave room for subagent fan-out and avoid spending HBM on graphs
@@ -132,7 +107,7 @@ SGLANG_CMD=(
     --model-path "$MODEL_PATH"
     --served-model-name "$MODEL"
     --host 0.0.0.0
-    --port "$SGLANG_BACKEND_PORT"
+    --port "$PORT"
     --trust-remote-code
     "${PARALLEL_ARGS[@]}"
     --enable-symm-mem
@@ -170,32 +145,14 @@ SERVER_PID=$!
 capture_cache_metrics() {
     {
         echo "=== SGLang cache metrics snapshot $(date --iso-8601=seconds) ==="
-        curl -fsS "http://localhost:$SGLANG_BACKEND_PORT/metrics" 2>/dev/null \
+        curl -fsS "http://localhost:$PORT/metrics" 2>/dev/null \
             | grep -E '^(sglang:(cache_hit_rate|cached_tokens_total|prompt_tokens_total|hicache_host_used_tokens|hicache_host_total_tokens|token_usage|num_requests_running|num_requests_waiting))' \
             || true
         echo "============================================================"
     } >> "$SERVER_LOG"
 }
 
-wait_for_server_ready --port "$SGLANG_BACKEND_PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
-
-if [ "$USE_SGLANG_ROUTER" = "true" ]; then
-    "${SGLANG_ROUTER_CMD[@]}" \
-        --worker-urls "http://localhost:$SGLANG_BACKEND_PORT" \
-        --policy consistent_hashing \
-        --request-id-headers x-correlation-id \
-        --dp-aware \
-        --host 0.0.0.0 \
-        --port "$PORT" \
-        --prometheus-host 127.0.0.1 \
-        --prometheus-port "$SGLANG_ROUTER_METRICS_PORT" \
-        --connect-timeout-secs 900 \
-        --request-timeout-secs 14400 \
-        --disable-health-check \
-        --disable-retries > "$ROUTER_LOG" 2>&1 &
-    ROUTER_PID=$!
-    wait_for_server_ready --port "$PORT" --server-log "$ROUTER_LOG" --server-pid "$ROUTER_PID"
-fi
+wait_for_server_ready --port "$PORT" --server-log "$SERVER_LOG" --server-pid "$SERVER_PID"
 
 capture_cache_metrics
 trap capture_cache_metrics EXIT
@@ -204,6 +161,6 @@ if [ "${EVAL_ONLY:-false}" = "true" ]; then
     run_eval --port "$PORT"
 else
     build_replay_cmd "$RESULT_DIR"
-    REPLAY_CMD+=" --server-metrics http://localhost:$SGLANG_BACKEND_PORT/metrics"
+    REPLAY_CMD+=" --server-metrics http://localhost:$PORT/metrics"
     run_agentic_replay_and_write_outputs "$RESULT_DIR"
 fi
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 38704e96e7..6e9fb846fd 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7183,7 +7183,7 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # hits by c96, placing its useful no-offload/offload crossover below c64.
 # TP4 remains healthy at c64; HiCache sustains 87% prefix hits and 234k input
 # tok/s at c80, so resolve the crossover densely and carry offload past it.
-# DEP8 probes the throughput end with session-sticky attention-DP routing.
+# DEP8 was rejected after fragmenting the trace across eight radix caches.
 qwen3.5-fp4-b300-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
   model: nvidia/Qwen3.5-397B-A17B-NVFP4
@@ -7200,7 +7200,6 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 72, 80, 88, 96, 104, 112] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72] }
-      - { tp: 8, ep: 8, dp-attn: true, spec-decoding: mtp, kv-offloading: none, conc-list: [64, 128, 256], router: { name: sglang-router, version: "0.3.2" } }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:
   image: vllm/vllm-openai:v0.22.0

From e15c32d701f87a605988084f9c00e2a22d56b8d6 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:19:50 -0500
Subject: [PATCH 13/30] perf: extend B300 TP2 HiCache tail
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Probe three post-c72 offload points to locate saturation of the measured TP2 efficiency frontier.\n\n中文：在 c72 之后增加三个卸载并发点，用于定位实测 B300 TP2 效率前沿的饱和位置。
---
 configs/nvidia-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 6e9fb846fd..2e1cb7c20a 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7199,7 +7199,7 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72, 80, 88, 96] }
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 72, 80, 88, 96, 104, 112] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72, 76, 80, 88] }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:
   image: vllm/vllm-openai:v0.22.0

From 3bb4e4d33275a38f3cfeb3e8ccfd502d0abf43ef Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:22:07 -0500
Subject: [PATCH 14/30] perf: prune dominated B300 TP4 tail
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Remove the no-offload c96 point after it was dominated by HiCache c80 in both throughput and latency, and refresh the measured crossover note.\n\n中文：无卸载 c96 在吞吐与延迟上均被 HiCache c80 支配，因此移除该点并更新实测切换区间说明。
---
 configs/nvidia-master.yaml | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 2e1cb7c20a..89e5bbe6bc 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7181,7 +7181,7 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
 # p75 ISL, TP2 is healthy through c32 but reaches 97% GPU KV with zero prefix
 # hits by c96, placing its useful no-offload/offload crossover below c64.
-# TP4 remains healthy at c64; HiCache sustains 87% prefix hits and 234k input
+# TP4 remains healthy at c64; HiCache sustains 90% prefix hits and >260k input
 # tok/s at c80, so resolve the crossover densely and carry offload past it.
 # DEP8 was rejected after fragmenting the trace across eight radix caches.
 qwen3.5-fp4-b300-sglang-agentic-mtp:
@@ -7196,7 +7196,7 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72, 80, 88, 96] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72, 80, 88] }
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 72, 80, 88, 96, 104, 112] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72, 76, 80, 88] }

From c5b8203b6d6b5445e491b53097d765e77a13a926 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:24:00 -0500
Subject: [PATCH 15/30] perf: cap B300 HiCache at valid working sets
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Remove offload points that exceed the measured host-cache validity boundary and densify the final safe interval below c80.\n\n中文：移除超过实测主机缓存有效边界的卸载配置，并在 c80 以下的安全区间细化最终并发点。
---
 configs/nvidia-master.yaml | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 89e5bbe6bc..72c64d8ff9 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7181,8 +7181,8 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
 # p75 ISL, TP2 is healthy through c32 but reaches 97% GPU KV with zero prefix
 # hits by c96, placing its useful no-offload/offload crossover below c64.
-# TP4 remains healthy at c64; HiCache sustains 90% prefix hits and >260k input
-# tok/s at c80, so resolve the crossover densely and carry offload past it.
+# TP4 remains healthy at c64, but HiCache c80 becomes invalid once its host
+# pool is full, so resolve c64-76 without carrying that offload tail forward.
 # DEP8 was rejected after fragmenting the trace across eight radix caches.
 qwen3.5-fp4-b300-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
@@ -7197,9 +7197,9 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
     - dram-utilization: 0.80
       search-space:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72, 80, 88] }
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 72, 80, 88, 96, 104, 112] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72, 76, 80, 88] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72, 76, 80] }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:
   image: vllm/vllm-openai:v0.22.0

From a3dac8d0b4a1ca2701e053f5cb5d4f2ec622e56b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:25:06 -0500
Subject: [PATCH 16/30] perf: cap B300 TP2 before the cache cliff
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Remove no-offload c64 after it stalled under the long-context working set; retain the matched HiCache points beyond c56.\n\n中文：无卸载 c64 在长上下文工作集下停滞，因此移除该点，并保留 c56 之后对应的 HiCache 配置。
---
 configs/nvidia-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 72c64d8ff9..f17968164a 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7198,7 +7198,7 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
       search-space:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72, 80, 88] }
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72, 76, 80] }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:

From 79dd0c71bef7765314211d64100be35ec003fed7 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:36:15 -0500
Subject: [PATCH 17/30] perf: probe B300 TP4 HiCache at the cliff
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：在缓存工作集拐点测试 B300 TP4 HiCache。
---
 configs/nvidia-master.yaml | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index f17968164a..93e862662a 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7181,8 +7181,9 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
 # p75 ISL, TP2 is healthy through c32 but reaches 97% GPU KV with zero prefix
 # hits by c96, placing its useful no-offload/offload crossover below c64.
-# TP4 remains healthy at c64, but HiCache c80 becomes invalid once its host
-# pool is full, so resolve c64-76 without carrying that offload tail forward.
+# TP4 remains healthy through c88, while c96 loses roughly half its throughput
+# and raises p50 ITL from 15 ms to 44 ms as prefix hits fall. Resolve the
+# c88-96 cliff and compare HiCache directly across that boundary.
 # DEP8 was rejected after fragmenting the trace across eight radix caches.
 qwen3.5-fp4-b300-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
@@ -7197,7 +7198,7 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
     - dram-utilization: 0.80
       search-space:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72, 80, 88] }
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76, 80, 88, 96] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72, 76, 80] }
 

From 020a1aa8a698a2f8eb5698db272c5bb083dd9136 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:40:09 -0500
Subject: [PATCH 18/30] perf: prune the B300 TP2 no-offload cliff
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：裁剪 B300 TP2 无卸载配置的缓存拐点尾部。
---
 configs/nvidia-master.yaml | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 93e862662a..725993c033 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7179,8 +7179,8 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # Qwen3.5 target FP8 KV is 15,360 bytes/token before the native MTP layer.
 # TP4 replicates the model's two KV heads and measured 30.5 KB/token node-wide;
 # TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
-# p75 ISL, TP2 is healthy through c32 but reaches 97% GPU KV with zero prefix
-# hits by c96, placing its useful no-offload/offload crossover below c64.
+# p75 ISL, TP2 is healthy through c32 but falls to zero prefix hits by c56,
+# placing its useful no-offload/offload crossover below that point.
 # TP4 remains healthy through c88, while c96 loses roughly half its throughput
 # and raises p50 ITL from 15 ms to 44 ms as prefix hits fall. Resolve the
 # c88-96 cliff and compare HiCache directly across that boundary.
@@ -7199,7 +7199,7 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
       search-space:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72, 80, 88] }
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76, 80, 88, 96] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72, 76, 80] }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:

From d2224b2e2e31632d2bbaf25a794bf8bb4aad31a2 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:52:51 -0500
Subject: [PATCH 19/30] fix(agentx): refine B300 cache frontier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：优化 B300 缓存悬崖附近的并发点，并使用 SGLang 默认单 tokenizer 启动路径以避免 HiCache 初始化期间的多 tokenizer 共享内存竞态。
---
 .../agentic/qwen3.5_fp4_b300_sglang_mtp.sh      |  2 --
 configs/nvidia-master.yaml                      | 17 +++++++++--------
 2 files changed, 9 insertions(+), 10 deletions(-)

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
index b0aa51657e..23bf3c6d27 100755
--- a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
@@ -80,7 +80,6 @@ PARALLEL_ARGS=(
     --dp 1
     --ep-size "$EP_SIZE"
 )
-TOKENIZER_WORKERS=6
 STREAM_INTERVAL=50
 
 # AgentX concurrency counts live session trees rather than individual HTTP
@@ -124,7 +123,6 @@ SGLANG_CMD=(
     --mem-fraction-static 0.80
     --stream-interval "$STREAM_INTERVAL"
     --scheduler-recv-interval "$SCHEDULER_RECV_INTERVAL"
-    --tokenizer-worker-num "$TOKENIZER_WORKERS"
     --tokenizer-path "$MODEL"
     --reasoning-parser qwen3
     --tool-call-parser qwen3_coder
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 725993c033..234014734a 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7179,11 +7179,12 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # Qwen3.5 target FP8 KV is 15,360 bytes/token before the native MTP layer.
 # TP4 replicates the model's two KV heads and measured 30.5 KB/token node-wide;
 # TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
-# p75 ISL, TP2 is healthy through c32 but falls to zero prefix hits by c56,
-# placing its useful no-offload/offload crossover below that point.
-# TP4 remains healthy through c88, while c96 loses roughly half its throughput
-# and raises p50 ITL from 15 ms to 44 ms as prefix hits fall. Resolve the
-# c88-96 cliff and compare HiCache directly across that boundary.
+# p75 ISL, TP2 no-offload peaks at c32; c40/c48 plateau at the same input
+# throughput with worse latency, and c56 loses all prefix hits. HiCache takes
+# over at c40. TP4 no-offload is healthy through c72, while c96 loses roughly
+# half its throughput as prefix hits fall. HiCache c88 is clean and dominates
+# no-offload c88. Both c80 trials produced shortened outputs, so that invalid
+# point is excluded while the clean c72-96 cache-cliff neighborhood stays dense.
 # DEP8 was rejected after fragmenting the trace across eight radix caches.
 qwen3.5-fp4-b300-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
@@ -7197,9 +7198,9 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72, 80, 88] }
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76, 80, 88, 96] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76, 88, 96] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72, 76, 80] }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:

From 89cd5926cfc4b3eee36640baacf4536d456e412b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 02:55:39 -0500
Subject: [PATCH 20/30] perf(agentx): densify B300 HiCache cliff
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：在 B300 HiCache 缓存悬崖附近增加 c84、c92 和 c94 测试点，以更细粒度覆盖有效的帕累托前沿。
---
 configs/nvidia-master.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 234014734a..2b0dbad9c2 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7199,7 +7199,7 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
     - dram-utilization: 0.80
       search-space:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72] }
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76, 88, 96] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76, 84, 88, 92, 94, 96] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72, 76, 80] }
 

From c43766e2d03a0f948796795f4e0441432bb298a0 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 03:08:07 -0500
Subject: [PATCH 21/30] perf(agentx): densify B300 TP2 cache crossover
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：在 B300 TP2 的 c32 至 c56 区间增加 HiCache 重叠测试点，以覆盖一小时 AgentX 预热下可能提前出现的缓存悬崖。
---
 configs/nvidia-master.yaml | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 2b0dbad9c2..49a5f3fbd2 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7180,8 +7180,9 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # TP4 replicates the model's two KV heads and measured 30.5 KB/token node-wide;
 # TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
 # p75 ISL, TP2 no-offload peaks at c32; c40/c48 plateau at the same input
-# throughput with worse latency, and c56 loses all prefix hits. HiCache takes
-# over at c40. TP4 no-offload is healthy through c72, while c96 loses roughly
+# throughput with worse latency, and c56 loses all prefix hits. HiCache overlaps
+# from c32 and densely resolves that crossover for the one-hour sweep. TP4
+# no-offload is healthy through c72, while c96 loses roughly
 # half its throughput as prefix hits fall. HiCache c88 is clean and dominates
 # no-offload c88. Both c80 trials produced shortened outputs, so that invalid
 # point is excluded while the clean c72-96 cache-cliff neighborhood stays dense.
@@ -7201,7 +7202,7 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72] }
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76, 84, 88, 92, 94, 96] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [40, 48, 56, 60, 64, 68, 72, 76, 80] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [32, 36, 40, 44, 48, 52, 56, 60, 64, 68, 72, 76, 80] }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:
   image: vllm/vllm-openai:v0.22.0

From e0b8e280681dcc83f72113c21d6a0640ab3b71fd Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 03:19:34 -0500
Subject: [PATCH 22/30] fix(agentx): scale Qwen tokenization by topology
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Use six tokenizer workers for TP4 long-context replay while retaining SGLang default single-worker startup for TP2 HiCache.

中文：TP4 长上下文回放使用 6 个 tokenizer worker；TP2 HiCache 保持 SGLang 默认单 worker 启动路径，避免共享内存初始化竞态。
---
 .../single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh   | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
index 23bf3c6d27..05aafe16c2 100755
--- a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
@@ -82,6 +82,14 @@ PARALLEL_ARGS=(
 )
 STREAM_INTERVAL=50
 
+# TP4 needs parallel tokenization to keep 256k AgentX warmups below the client
+# request timeout. Keep TP2 on SGLang's single-worker default: multi-tokenizer
+# startup races with the TP2 HiCache shared-memory initialization path.
+TOKENIZER_ARGS=()
+if [ "$TP" -ge 4 ]; then
+    TOKENIZER_ARGS=(--tokenizer-worker-num 6)
+fi
+
 # AgentX concurrency counts live session trees rather than individual HTTP
 # requests. Leave room for subagent fan-out and avoid spending HBM on graphs
 # above the batch sizes that remain useful for this long-context workload.
@@ -123,6 +131,7 @@ SGLANG_CMD=(
     --mem-fraction-static 0.80
     --stream-interval "$STREAM_INTERVAL"
     --scheduler-recv-interval "$SCHEDULER_RECV_INTERVAL"
+    "${TOKENIZER_ARGS[@]}"
     --tokenizer-path "$MODEL"
     --reasoning-parser qwen3
     --tool-call-parser qwen3_coder

From f8b5e3c3d9df431f0b218aa640100866ee99d4af Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 03:30:25 -0500
Subject: [PATCH 23/30] perf(agentx): densify B300 TP2 cache boundary
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Add HiCache concurrency 34 and 38 around the measured c36-to-c40 transition.

中文：在实测 c36 到 c40 的转折区间补充 HiCache 并发点 34 和 38。
---
 configs/nvidia-master.yaml | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 49a5f3fbd2..dcd0350850 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7181,7 +7181,8 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
 # p75 ISL, TP2 no-offload peaks at c32; c40/c48 plateau at the same input
 # throughput with worse latency, and c56 loses all prefix hits. HiCache overlaps
-# from c32 and densely resolves that crossover for the one-hour sweep. TP4
+# from c32; warm-cache probes put a sharp transition between c36 and c40, so
+# c34/c38 densely resolve that crossover for the one-hour sweep. TP4
 # no-offload is healthy through c72, while c96 loses roughly
 # half its throughput as prefix hits fall. HiCache c88 is clean and dominates
 # no-offload c88. Both c80 trials produced shortened outputs, so that invalid
@@ -7202,7 +7203,7 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72] }
       - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76, 84, 88, 92, 94, 96] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [32, 36, 40, 44, 48, 52, 56, 60, 64, 68, 72, 76, 80] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [32, 34, 36, 38, 40, 44, 48, 52, 56, 60, 64, 68, 72, 76, 80] }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:
   image: vllm/vllm-openai:v0.22.0

From 023f5438d7f24653bd777ac1866f5deada4dfcd8 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 04:12:41 -0500
Subject: [PATCH 24/30] perf(agentx): refine B300 cache frontier
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Densify the TP4 no-offload cliff and remove TP4 HiCache points that are dominated by the TP4 no-offload and TP2 HiCache arms. Keep TP2 HiCache dense through the measured host-KV saturation region.\n\n中文：细化 B300 缓存帕累托前沿；加密 TP4 无卸载缓存临界区间，移除被 TP4 无卸载与 TP2 HiCache 方案支配的 TP4 HiCache 测点，并保留覆盖主机 KV 饱和区间的密集 TP2 HiCache 测点。
---
 configs/nvidia-master.yaml | 22 +++++++++++-----------
 1 file changed, 11 insertions(+), 11 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index dcd0350850..b2fc4a0166 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -7179,14 +7179,15 @@ qwen3.5-fp8-b300-sglang-agentic-hicache:
 # Qwen3.5 target FP8 KV is 15,360 bytes/token before the native MTP layer.
 # TP4 replicates the model's two KV heads and measured 30.5 KB/token node-wide;
 # TP2 measured 15.3 KB/token. Against the 256k trace's 88.8k median and 148.8k
-# p75 ISL, TP2 no-offload peaks at c32; c40/c48 plateau at the same input
-# throughput with worse latency, and c56 loses all prefix hits. HiCache overlaps
-# from c32; warm-cache probes put a sharp transition between c36 and c40, so
-# c34/c38 densely resolve that crossover for the one-hour sweep. TP4
-# no-offload is healthy through c72, while c96 loses roughly
-# half its throughput as prefix hits fall. HiCache c88 is clean and dominates
-# no-offload c88. Both c80 trials produced shortened outputs, so that invalid
-# point is excluded while the clean c72-96 cache-cliff neighborhood stays dense.
+# p75 ISL, TP2 no-offload peaks at c32. HiCache overlaps from c32; probes put
+# the throughput transition between c34 and c40, while host KV reaches 98% at
+# c52 and c56 is the last useful stress point. c64+ warmups only deepen the
+# queue and are excluded. TP4 no-offload remains healthy into the c56-72
+# neighborhood, which is sampled every four requests around the cache cliff.
+# TP4 HiCache is dominated there: c68 sustained 197k tok/s versus 224k for
+# no-offload c64, while c76 fell to 158k with host KV full. Clean HiCache c94
+# was also dominated by no-offload c56 (210k vs 234k tok/s, with 44 ms vs 7 ms
+# median ITL), so the TP4 HiCache arm is excluded.
 # DEP8 was rejected after fragmenting the trace across eight radix caches.
 qwen3.5-fp4-b300-sglang-agentic-mtp:
   image: lmsysorg/sglang:v0.5.16-cu130
@@ -7200,10 +7201,9 @@ qwen3.5-fp4-b300-sglang-agentic-mtp:
     agentic-coding:
     - dram-utilization: 0.80
       search-space:
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 64, 72] }
-      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [64, 68, 72, 76, 84, 88, 92, 94, 96] }
+      - { tp: 4, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32, 40, 48, 56, 60, 64, 68, 72] }
       - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 4, 8, 12, 16, 20, 24, 28, 32] }
-      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [32, 34, 36, 38, 40, 44, 48, 52, 56, 60, 64, 68, 72, 76, 80] }
+      - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [32, 34, 36, 38, 40, 44, 48, 52, 56] }
 
 kimik2.5-fp4-b200-vllm-agentic-lmcache:
   image: vllm/vllm-openai:v0.22.0

From 012e78fc41110dc11de182cb57362e9fe07b905b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 12:05:45 -0500
Subject: [PATCH 25/30] fix(agentx): cap Qwen trace idle gaps

Pin the AIPerf trace idle-gap branch and set the Qwen B300 replay cap to 300 seconds.
---
 benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh | 1 +
 perf-changelog.yaml                                           | 1 +
 utils/aiperf                                                  | 2 +-
 3 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
index 05aafe16c2..e0781a3d21 100755
--- a/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
+++ b/benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh
@@ -168,6 +168,7 @@ if [ "${EVAL_ONLY:-false}" = "true" ]; then
     run_eval --port "$PORT"
 else
     build_replay_cmd "$RESULT_DIR"
+    REPLAY_CMD+=" --trace-idle-gap-cap-seconds 300"
     REPLAY_CMD+=" --server-metrics http://localhost:$PORT/metrics"
     run_agentic_replay_and_write_outputs "$RESULT_DIR"
 fi
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 4ab147d2a4..2553ab2bcf 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5345,5 +5345,6 @@
   description:
     - "Add Qwen3.5-397B-A17B NVFP4 AgentX benchmark on B300 with SGLang native NEXTN MTP"
     - "Use the 256k trace dataset and golden synthetic acceptance length 3.39"
+    - "Pin AIPerf trace idle-gap support and cap per-trace idle gaps at 300 seconds"
     - "Image: lmsysorg/sglang:v0.5.16-cu130"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2421
diff --git a/utils/aiperf b/utils/aiperf
index be758d6218..c1f4ea78f1 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit be758d6218268171e2957fbec9d4f557275bca2d
+Subproject commit c1f4ea78f1226cddc00ec45f553569d3b1b7c3f6

From c9187d787fc50ccff52c153628f276261743cd48 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 12:14:31 -0500
Subject: [PATCH 26/30] chore(aiperf): pin merged trace idle cap branch

---
 utils/aiperf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/utils/aiperf b/utils/aiperf
index c1f4ea78f1..63d8be31b6 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit c1f4ea78f1226cddc00ec45f553569d3b1b7c3f6
+Subproject commit 63d8be31b6d03f0f30a76efa8db35fa077ffefc8

From 6144fe059d2c468b2371f0623b3fb11aab9cbc5b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 13:16:19 -0500
Subject: [PATCH 27/30] chore(agentx): bump AIPerf trace idle cap fix

Signed-off-by: Cam Quilici <cjquilici@gmail.com>
---
 utils/aiperf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/utils/aiperf b/utils/aiperf
index 63d8be31b6..1e5015f2c6 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit 63d8be31b6d03f0f30a76efa8db35fa077ffefc8
+Subproject commit 1e5015f2c6f5ea220f2025fe8c4c6c1534e87f77

From 9c65228f9d6e95b1b1db5051a7c5788810ac328b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 13:35:04 -0500
Subject: [PATCH 28/30] fix(agentx): update AIPerf trace-cap cleanup

Signed-off-by: Cam Quilici <cjquilici@gmail.com>
---
 utils/aiperf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/utils/aiperf b/utils/aiperf
index 1e5015f2c6..040766e5e0 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit 1e5015f2c6f5ea220f2025fe8c4c6c1534e87f77
+Subproject commit 040766e5e0ea1656283b6a76c870d1bd35bda092

From 94353501dbf6bede75e26b700a641c296998b74e Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 14:58:13 -0500
Subject: [PATCH 29/30] chore(agentx): pin reconstruction-only idle cap

Update AIPerf after removing runtime trace idle enforcement.

Signed-off-by: Cam Quilici <cjquilici@gmail.com>
---
 utils/aiperf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/utils/aiperf b/utils/aiperf
index 040766e5e0..61a29cda34 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit 040766e5e0ea1656283b6a76c870d1bd35bda092
+Subproject commit 61a29cda34b828c4624278a1ae70910d6487156d

From 2a7e8613152b381af7633370863197ab0ad6c584 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Thu, 30 Jul 2026 16:39:59 -0500
Subject: [PATCH 30/30] fix(agentx): update AIPerf warmup handoff

Signed-off-by: Cam Quilici <cjquilici@gmail.com>
---
 utils/aiperf | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/utils/aiperf b/utils/aiperf
index 61a29cda34..b6e4ecf4ba 160000
--- a/utils/aiperf
+++ b/utils/aiperf
@@ -1 +1 @@
-Subproject commit 61a29cda34b828c4624278a1ae70910d6487156d
+Subproject commit b6e4ecf4ba87ed1ab085380828c464ea601343d2