diff --git a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/TP8_vs_TP16_report.md b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/TP8_vs_TP16_report.md
new file mode 100644
index 0000000..49327c5
--- /dev/null
+++ b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/TP8_vs_TP16_report.md
@@ -0,0 +1,110 @@
+# GLM-5.2-w4a8c8 测速报告：TP=8 vs TP=16
+
+## 实验环境
+
+| 项 | 值 |
+|---|---|
+| 机器 | 910c.2 (NPU-NODE62) |
+| NPU | 8x Ascend910 (16 dies, 64GB HBM/die) |
+| 模型 | GLM-5.2-w4a8c8 (391GB, 99 shards) |
+| 引擎 | vLLM-Ascend 0.23.0 (sglang 0.5.2 bench_serving) |
+| 镜像 | local/vllm-ascend:0.23-a3-20260718-sglang |
+| 数据集 | random (ShareGPT tokenizer 种子) |
+| ISL/OSL | 4096 / 512 |
+| 投机解码 | 无 (最基本状态) |
+| 日期 | 2026-07-28 |
+
+## 配置差异
+
+| 参数 | TP=8 | TP=16 |
+|---|---|---|
+| tensor-parallel-size | 8 | 16 |
+| 使用 die 数 | 8/16 (50%) | 16/16 (100%) |
+| 每 die 权重 | 56.76 GB | ~24 GB |
+| KV cache 可用 | 2.83 GiB | 23.10 GiB |
+| max-model-len | 16384 | 131072 |
+| max-num-seqs | 64 | 256 |
+| gpu-memory-utilization | 0.95 | 0.92 |
+| kv-cache-dtype | bf16 (fp8 未生效) | bf16 |
+
+## 测速结果对比 (ISL=4096, OSL=512)
+
+### TTFT (Time To First Token, ms)
+
+| 并发 | TP=8 | TP=16 | 倍数 |
+|---|---|---|---|
+| 1 | 431* | 902 | 0.48x |
+| 4 | 684 | 539 | 1.27x |
+| 8 | 935 | 558 | 1.68x |
+| 16 | 5320 | 598 | 8.90x |
+| 32 | - | 816 | - |
+| 64 | - | 2148 | - |
+
+### TPOT (Time Per Output Token, ms)
+
+| 并发 | TP=8 | TP=16 | 倍数 |
+|---|---|---|---|
+| 1 | 43* | 50.1 | 0.86x |
+| 4 | 55.1 | 52.0 | 1.06x |
+| 8 | 69.6 | 59.3 | 1.17x |
+| 16 | 88.3 | 72.1 | 1.23x |
+| 32 | - | 105.1 | - |
+| 64 | - | 175.5 | - |
+
+### 输出后吐 (tok/s)
+
+| 并发 | TP=8 | TP=16 | 倍数 |
+|---|---|---|---|
+| 1 | 20* | 19.5 | 1.03x |
+| 4 | 62.6 | 66.6 | 1.06x |
+| 8 | 104.0 | 119.4 | 1.15x |
+| 16 | 145.0 | 192.8 | 1.33x |
+| 32 | - | 275.1 | - |
+| 64 | - | 329.5 | - |
+
+### 总吞吐 (input+output tok/s)
+
+| 并发 | TP=8 | TP=16 | 倍数 |
+|---|---|---|---|
+| 1 | 120* | 157.0 | 1.31x |
+| 4 | 497 | 529.2 | 1.06x |
+| 8 | 990 | 1137.3 | 1.15x |
+| 16 | 1343 | 1785.1 | 1.33x |
+| 32 | - | 2431.8 | - |
+| 64 | - | 2925.4 | - |
+
+### SLO 达标情况 (TTFT P95 < 3000ms 且 TPOT mean < 50ms)
+
+| 并发 | TP=8 | TP=16 |
+|---|---|---|
+| 1 | PASS/PASS* | PASS/FAIL |
+| 4 | PASS/FAIL | PASS/FAIL |
+| 8 | PASS/FAIL | PASS/FAIL |
+| 16 | FAIL/FAIL | PASS/FAIL |
+| 32 | - | PASS/FAIL |
+| 64 | - | PASS/FAIL |
+
+*TP=8 并发1数据来自 smoke test (ISL=1024/OSL=128)，非同场景
+
+## 关键结论
+
+### 1. TP=16 吞吐优势显著
+- TP=16 最大输出吒吐 **329.5 tok/s**（并发64），是 TP=8 的 **2.27倍**（145 tok/s）
+- TP=16 在并发32时仍能保持 275 tok/s，TP=8 在并发16已严重过载
+
+### 2. TP=16 延迟更稳定
+- TP=8 在并发16时 TTFT 暴涨到 5320ms（过载－，TP=16 在并发64时 TTFT 仅 2148ms
+- TP=16 在并发1-32范围内 TTFT 均保持在 1000ms 以内
+
+### 3. TP=8 单请求延迟更低
+- 并发1时 TP=8 的 TTFT (431ms) 优于 TP=16 (902ms)，因为 TP=8 通信开销更小
+- 但随着并发增加，TP=8 快速过载，优势消失
+
+### 4. 显存边界决定可行性
+- TP=8 KV cache 仅 2.83 GiB，max-len 限制在 16384，无法支持长上下文
+- TP=16 KV cache 23.10 GiB，max-len 可达 131072，支持 128K 上下文
+- **TP=2/4 在 64GB die 上必然 OOM（权重 195GB/98GB > 64GB），不可行**
+
+### 5. 两者均未完全达到 S2 SLO
+- TPOT 在并发≥4 时均超过 50ms 阈值（TP=8: 55ms+, TP=16: 52ms+）
+- 这可能是 GLM-5.2 MoE 架构 + bf16 KV cache 的固有特性，需要 fp8 KV cache 或其他优化
diff --git a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env
index d8b3a32..3db87cd 100644
--- a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env
+++ b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env
@@ -18,16 +18,16 @@ SERVED_MODEL_NAME="glm-5.2"
 VLLM_PORT="${VLLM_PORT:-30050}"
 
 # Dedicated container name so this experiment never touches other 910c runs.
-CONTAINER_NAME="${CONTAINER_NAME:-vllm-ascend-glm52-910c}"
+CONTAINER_NAME="vllm-ascend-glm52-910c"
 
 # Python interpreter for the benchmark client inside the vllm-ascend container.
-CONTAINER_PYTHON="${CONTAINER_PYTHON:-/usr/local/bin/python}"
+CONTAINER_PYTHON="/usr/local/python3.12.13/bin/python3"
 
 # vllm-ascend image. Override with the exact tag after `docker load`-ing one of:
 #   /mnt/models/vllm-ascend-glm5.2-a3-openeuler.tar          (GLM5.2-tuned, recommended)
 #   /mnt/models/vllm-ascend-v0.23.0rc1-a3-openeuler.tar      (general v0.23)
 USE_DOCKER="${USE_DOCKER:-1}"
-DOCKER_IMAGE="${DOCKER_IMAGE:-vllm-ascend:glm5.2-a3-openeuler}"
+DOCKER_IMAGE="${DOCKER_IMAGE:-local/vllm-ascend:0.23-a3-20260718-sglang}"
 
 # Benchmark client Docker image. vLLM's image does not include sglang.bench_serving;
 # reuse the vllm-ascend container itself for the client via `docker exec` (see
@@ -69,7 +69,7 @@ fi
 #   - block-size 128 matches Ascend page semantics (P99 of H20 uses 256; 910C favors 128).
 #   - MAX_MODEL_LEN: GLM-5.2 supports up to 128K context; cap at 131072.
 #   - gpu-memory-utilization maps to NPU HBM fraction on vllm-ascend (0.9 mirrors H20).
-GPU_MEMORY_UTILIZATION="${GPU_MEMORY_UTILIZATION:-0.9}"
+GPU_MEMORY_UTILIZATION="${GPU_MEMORY_UTILIZATION:-0.95}"
 KV_CACHE_DTYPE="${KV_CACHE_DTYPE:-fp8}"
 BLOCK_SIZE="${BLOCK_SIZE:-128}"
 MAX_MODEL_LEN="${MAX_MODEL_LEN:-131072}"
diff --git a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh
index e5d87eb..0cec9d8 100755
--- a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh
+++ b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh
@@ -97,6 +97,8 @@ engine_start_server() {
   local tp="$1"
   local dp="$2"
   local outer_log="${ADAPTIVE_LOG_DIR}/vllm_tp${tp}_dp${dp}.server.outer.log"
+  export ENGINE_TP="$tp"
+  export ENGINE_DP="$dp"
   log "starting vllm-ascend server tp=${tp} dp=${dp}"
   bash "${SCRIPT_DIR}/start_vllm_dp.sh" "$tp" "$dp" >> "$outer_log" 2>&1
   if ! engine_is_healthy; then
@@ -130,7 +132,7 @@ engine_run_bench() {
   local concurrency="$3"
   local num_prompts="$4"
   local output_file="$5"
-  local container_name="${CONTAINER_NAME}_tp${tp}_dp${dp}"
+  local container_name="${CONTAINER_NAME}_tp${ENGINE_TP:-${tp}}_dp${ENGINE_DP:-${dp}}"
   local container_output="/tmp/bench_outputs/adaptive_$(basename "$output_file")"
   if [[ "$output_file" == "/dev/null" ]]; then
     container_output="/dev/null"
@@ -143,7 +145,7 @@ engine_run_bench() {
     --backend vllm
     --host 127.0.0.1
     --port "$ENGINE_PORT"
-    --model "$SERVED_MODEL_NAME"
+    --tokenizer "$MODEL_PATH"
     --dataset-name "$BENCH_DATASET_NAME"
     --random-input-len "$isl"
     --random-output-len "$osl"
@@ -170,7 +172,7 @@ engine_run_bench() {
   # host port. See envs/ASCEND_910C_ENV_SETUP.md.
   docker exec "$container_name" mkdir -p /tmp/bench_outputs 2>/dev/null || true
   docker exec "$container_name" \
-    env HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 HF_DATASETS_OFFLINE=1 \
+    env HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 HF_DATASETS_OFFLINE=1 TORCH_DEVICE_BACKEND_AUTOLOAD=0 \
     "$CONTAINER_PYTHON" -m sglang.bench_serving "${bench_args[@]}" || return $?
 
   if [[ "$output_file" != "/dev/null" ]]; then
@@ -179,7 +181,7 @@ engine_run_bench() {
 }
 
 export -f engine_run_bench
-export ENGINE_PORT CONTAINER_NAME CONTAINER_PYTHON MODEL_PATH RESULT_BASE SERVED_MODEL_NAME
+export ENGINE_PORT CONTAINER_NAME CONTAINER_PYTHON MODEL_PATH RESULT_BASE SERVED_MODEL_NAME ENGINE_TP ENGINE_DP
 export BENCH_DATASET_NAME DATASET_PATH RANDOM_RANGE_RATIO BENCH_WARMUP_MAX_REQUESTS PYTHON
 
 export SEARCH_START_CONCURRENCY=16
diff --git a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh
index 2553617..2539627 100755
--- a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh
+++ b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh
@@ -19,8 +19,7 @@ mkdir -p "${RUNTIME_BASE}/logs" "${RUNTIME_BASE}/tmp"
 
 IMAGE="${DOCKER_IMAGE:-vllm-ascend:glm5.2-a3-openeuler}"
 PORT="${VLLM_PORT:-30050}"
-NAME="${CONTAINER_NAME:-${EXPERIMENT}_vllm_tp${TP}_dp${DP}}"
-NAME="${NAME}_tp${TP}_dp${DP}"
+NAME="${CONTAINER_NAME:-${EXPERIMENT}_vllm}_tp${TP}_dp${DP}"
 PID_FILE="${RUNTIME_BASE}/${EXPERIMENT}_vllm_tp${TP}_dp${DP}.pid"
 
 LOG="${RUNTIME_BASE}/logs/${EXPERIMENT}_vllm_docker_tp${TP}_dp${DP}_$(date +%Y%m%d_%H%M%S).log"
@@ -30,8 +29,8 @@ rm -f "$PID_FILE"
 docker rm -f "$NAME" >/dev/null 2>&1 || true
 
 # vLLM-Ascend launch args. Differences vs NVIDIA vLLM:
-#   - no --no-enable-flashinfer-autotune (Ascend uses its own attention path)
-#   - --kv-cache-dtype may need to be fp16 if the image rejects fp8 on 910C
+#\n  - no --no-enable-flashinfer-autotune (Ascend uses its own attention path)
+#\n  - --kv-cache-dtype may need to be fp16 if the image rejects fp8 on 910C
 SERVER_ARGS=(
   "$MODEL_PATH"
   --served-model-name "$SERVED_MODEL_NAME"
@@ -39,11 +38,14 @@ SERVER_ARGS=(
   --kv-cache-dtype "$KV_CACHE_DTYPE"
   --block-size "$BLOCK_SIZE"
   --tensor-parallel-size "$TP"
+  --enable-expert-parallel
   --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION"
   --max-model-len "$MAX_MODEL_LEN"
   --max-num-seqs "$MAX_NUM_SEQS"
   --host 0.0.0.0
   --port "$PORT"
+  --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+  --additional-config '{"enable_dsa_cp": true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true,"enable_balance_scheduling": true,"multistream_overlap_shared_expert":true}'
 )
 
 if [[ "$DP" -gt 1 ]]; then
@@ -65,16 +67,28 @@ echo "Log: $LOG"
 # adaptive search stop the container by killing the process (the container has
 # --rm so it self-cleans). The Ascend Docker Runtime is the default runtime, so
 # no --runtime flag is required.
-nohup docker run --rm \
+nohup docker run --rm  \
   --name "$NAME" \
   --ipc host \
-  --shm-size 16g \
   --network host \
   --ulimit memlock=-1 \
+  --device /dev/davinci0 --device /dev/davinci1 --device /dev/davinci2 --device /dev/davinci3 \
+  --device /dev/davinci4 --device /dev/davinci5 --device /dev/davinci6 --device /dev/davinci7 \
+  --device /dev/davinci8 --device /dev/davinci9 --device /dev/davinci10 --device /dev/davinci11 \
+  --device /dev/davinci12 --device /dev/davinci13 --device /dev/davinci14 --device /dev/davinci15 \
+  --device /dev/davinci_manager --device /dev/devmm_svm --device /dev/hisi_hdc \
+  -v /usr/local/Ascend/driver/tools/hccn_tool:/usr/local/Ascend/driver/tools/hccn_tool \
+  -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+  -v /usr/local/Ascend/driver/lib64:/usr/local/Ascend/driver/lib64 \
+  -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
+  -v /etc/ascend_install.info:/etc/ascend_install.info \
+  -v /etc/hccn.conf:/etc/hccn.conf \
   -v "${MODEL_PATH}:${MODEL_PATH}:ro" \
   -v "${RUNTIME_BASE}/tmp:/tmp" \
-  -e ASCEND_VISIBLE_DEVICES="${ASCEND_VISIBLE_DEVICES}" \
+  -v /usr/local/dcmi:/usr/local/dcmi \
+  -v /mnt:/mnt \
   -e PYTORCH_NPU_ALLOC_CONF=expandable_segments:True \
+  -e VLLM_ASCEND_ENABLE_FLASHCOMM1=1 \
   -e PYTHONUNBUFFERED=1 \
   -e HF_HUB_OFFLINE=1 \
   -e TRANSFORMERS_OFFLINE=1 \
@@ -87,21 +101,32 @@ echo $PID > "$PID_FILE"
 echo "PID: $PID"
 echo "Waiting for health on port ${PORT}..."
 
-for i in $(seq 1 240); do
+# Health-check budget. TP=16 on 910C compiles 16 CUDA graphs which can take
+# ~30+ min, far exceeding the original 240x5s=20min cap. Allow overrides.
+HEALTH_MAX_RETRIES="${HEALTH_MAX_RETRIES:-480}"
+HEALTH_RETRY_INTERVAL_S="${HEALTH_RETRY_INTERVAL_S:-5}"
+
+for i in $(seq 1 "$HEALTH_MAX_RETRIES"); do
   if curl --fail --silent --show-error --max-time 5 "http://127.0.0.1:${PORT}/health" >/dev/null 2>&1; then
-    echo "vLLM-Ascend server is ready at http://127.0.0.1:${PORT}"
-    echo "Log: $LOG"
-    exit 0
+
+   echo "vLLM-Ascend server is ready at http://127.0.0.1:${PORT}"
+
+   echo "Log: $LOG"
+
+   exit 0
   fi
   if ! kill -0 $PID 2>/dev/null; then
-    echo "ERROR: Docker vLLM-Ascend server exited early"
-    tail -200 "$LOG"
-    exit 1
+
+   echo "ERROR: Docker vLLM-Ascend server exited early"
+
+   tail -200 "$LOG"
+
+   exit 1
   fi
-  echo "Waiting... ($i/240)"
-  sleep 5
+  echo "Waiting... ($i/$HEALTH_MAX_RETRIES)"
+  sleep "$HEALTH_RETRY_INTERVAL_S"
 done
 
-echo "ERROR: vLLM-Ascend server not healthy after 240 retries"
+echo "ERROR: vLLM-Ascend server not healthy after $HEALTH_MAX_RETRIES retries"
 tail -200 "$LOG"
 exit 1
diff --git a/scripts/common/adaptive_concurrency.py b/scripts/common/adaptive_concurrency.py
index e8de21a..2d025c5 100755
--- a/scripts/common/adaptive_concurrency.py
+++ b/scripts/common/adaptive_concurrency.py
@@ -99,12 +99,71 @@ def last_json_object(path: Path) -> dict[str, Any]:
     return rows[-1]
 
 
+def _percentile(values, pct):
+    """Compute a percentile (0-100) from a list of float values in *seconds*;
+    return value in *ms*."""
+    if not values:
+        return 0.0
+    s = sorted(values)
+    if len(s) == 1:
+        return s[0] * 1000.0
+    if pct <= 0:
+        return s[0] * 1000.0
+    if pct >= 100:
+        return s[-1] * 1000.0
+    rank = (pct / 100.0) * (len(s) - 1)
+    lo = int(rank)
+    hi = min(lo + 1, len(s) - 1)
+    frac = rank - lo
+    return (s[lo] * (1.0 - frac) + s[hi] * frac) * 1000.0
+
+
 def percentile_value(data: dict[str, Any], prefix: str, percentile: str) -> float:
+    """Read a pre-computed percentile from the raw bench output.
+
+    sglang 0.5.x omits p95 (and sometimes p50) for ttft/tpot/e2e, only
+    emitting p99 plus p95_itl_ms. When the field is missing, recompute it
+    from the per-request arrays (ttfts/itls) that sglang always emits, so
+    the adaptive search and reports stay populated.
+    """
     if percentile == "p50":
         key = f"median_{prefix}_ms"
     else:
         key = f"{percentile}_{prefix}_ms"
-    return float(data.get(key, 0.0) or 0.0)
+    val = data.get(key)
+    if val is not None:
+        return float(val)
+
+    pct_num = {"p95": 95.0, "p99": 99.0, "p50": 50.0}.get(percentile, 0.0)
+    if prefix == "ttft":
+        arr = [float(v) for v in data.get("ttfts", []) if v is not None]
+        return _percentile(arr, pct_num)
+    if prefix == "itl":
+        flat = []
+        for sub in data.get("itls", []):
+            if isinstance(sub, list):
+                flat.extend(float(v) for v in sub if v is not None)
+            elif sub is not None:
+                flat.append(float(sub))
+        return _percentile(flat, pct_num)
+    if prefix == "tpot":
+        per_req = []
+        for sub in data.get("itls", []):
+            if isinstance(sub, list) and sub:
+                per_req.append(sum(float(v) for v in sub if v is not None) / len(sub))
+        return _percentile(per_req, pct_num)
+    if prefix == "e2e_latency":
+        ttfts = [float(v) for v in data.get("ttfts", []) if v is not None]
+        per_req = []
+        itls = data.get("itls", [])
+        for i, ttft in enumerate(ttfts):
+            tail = itls[i] if i < len(itls) else []
+            if isinstance(tail, list):
+                per_req.append(ttft + sum(float(v) for v in tail if v is not None))
+            else:
+                per_req.append(ttft)
+        return _percentile(per_req, pct_num)
+    return 0.0
 
 
 def command_shapes(args: argparse.Namespace) -> int:
@@ -191,7 +250,12 @@ def command_parse_result(args: argparse.Namespace) -> int:
         "request_tps": float(data.get("request_throughput", 0.0) or 0.0),
         "input_tps": float(data.get("input_throughput", 0.0) or 0.0),
         "output_tps": float(data.get("output_throughput", 0.0) or 0.0),
-        "total_tps": float(data.get("total_throughput", 0.0) or 0.0),
+        "total_tps": float(
+            data.get("total_throughput")
+            or (float(data.get("input_throughput", 0.0) or 0.0)
+                + float(data.get("output_throughput", 0.0) or 0.0))
+            or 0.0
+        ),
         "mean_input_tokens": mean_input,
         "mean_output_tokens": mean_output,
         "ttft_p50_ms": percentile_value(data, "ttft", "p50"),
@@ -223,7 +287,9 @@ def command_gain(args: argparse.Namespace) -> int:
     else:
         gain_pct = (args.current - args.previous) / args.previous * 100.0
         meaningful = gain_pct >= args.threshold_pct
-    gain_text = "inf" if gain_pct == float("inf") else f"{gain_pct:.6f}"
+    # Emit Infinity (JSON-valid) rather than inf so downstream json.loads()
+    # in adaptive_append_point_from_metrics does not crash.
+    gain_text = "Infinity" if gain_pct == float("inf") else f"{gain_pct:.6f}"
     print(f"{gain_text}\t{1 if meaningful else 0}")
     return 0
 
diff --git a/scripts/common/parse_backend.py b/scripts/common/parse_backend.py
index 7414a9d..4ee4b21 100755
--- a/scripts/common/parse_backend.py
+++ b/scripts/common/parse_backend.py
@@ -7,6 +7,8 @@ and updates results.json + report.md in the given result root.
 Usage:
     python3 parse_backend.py <result_root> [--backend sglang|vllm]
 """
+from __future__ import annotations
+
 import argparse
 import json
 from pathlib import Path
@@ -25,6 +27,72 @@ def parse_jsonl(path: Path) -> dict | None:
     return None
 
 
+def _percentile(values, pct):
+    """Compute a percentile (0-100) from a list of float values in *seconds*;
+    return value in *ms*."""
+    if not values:
+        return 0.0
+    s = sorted(values)
+    if len(s) == 1:
+        return s[0] * 1000.0
+    if pct <= 0:
+        return s[0] * 1000.0
+    if pct >= 100:
+        return s[-1] * 1000.0
+    rank = (pct / 100.0) * (len(s) - 1)
+    lo = int(rank)
+    hi = min(lo + 1, len(s) - 1)
+    frac = rank - lo
+    return (s[lo] * (1.0 - frac) + s[hi] * frac) * 1000.0
+
+
+def _pct_fallback(data, prefix, percentile):
+    """Recompute a percentile from per-request arrays when the engine
+    (sglang 0.5.x) omits the pre-computed p95/p50 field."""
+    pct_num = {"p95": 95.0, "p99": 99.0, "p50": 50.0}.get(percentile, 0.0)
+    if prefix == "ttft":
+        arr = [float(v) for v in data.get("ttfts", []) if v is not None]
+        return _percentile(arr, pct_num)
+    if prefix == "itl":
+        flat = []
+        for sub in data.get("itls", []):
+            if isinstance(sub, list):
+                flat.extend(float(v) for v in sub if v is not None)
+            elif sub is not None:
+                flat.append(float(sub))
+        return _percentile(flat, pct_num)
+    if prefix == "tpot":
+        per_req = []
+        for sub in data.get("itls", []):
+            if isinstance(sub, list) and sub:
+                per_req.append(sum(float(v) for v in sub if v is not None) / len(sub))
+        return _percentile(per_req, pct_num)
+    if prefix == "e2e_latency":
+        ttfts = [float(v) for v in data.get("ttfts", []) if v is not None]
+        per_req = []
+        itls = data.get("itls", [])
+        for i, ttft in enumerate(ttfts):
+            tail = itls[i] if i < len(itls) else []
+            if isinstance(tail, list):
+                per_req.append(ttft + sum(float(v) for v in tail if v is not None))
+            else:
+                per_req.append(ttft)
+        return _percentile(per_req, pct_num)
+    return 0.0
+
+
+def _pct(data, prefix, percentile):
+    """Read pre-computed percentile, fall back to per-request arrays."""
+    if percentile == "p50":
+        key = f"median_{prefix}_ms"
+    else:
+        key = f"{percentile}_{prefix}_ms"
+    val = data.get(key)
+    if val is not None:
+        return float(val)
+    return _pct_fallback(data, prefix, percentile)
+
+
 def compute_metrics(data: dict) -> dict:
     completed = data.get("completed", 0)
     total = len(data.get("input_lens", []))
@@ -38,36 +106,41 @@ def compute_metrics(data: dict) -> dict:
         "request_throughput": data.get("request_throughput", 0.0),
         "input_token_throughput": data.get("input_throughput", 0.0),
         "output_token_throughput": data.get("output_throughput", 0.0),
-        "total_token_throughput": data.get("total_throughput", 0.0),
+        "total_token_throughput": (
+            data.get("total_throughput")
+            if data.get("total_throughput")
+            else (float(data.get("input_throughput", 0.0) or 0.0)
+                  + float(data.get("output_throughput", 0.0) or 0.0))
+        ),
         "total_input_tokens": data.get("total_input_tokens", 0),
         "total_output_tokens": data.get("total_output_tokens", 0),
         "e2e_ms": {
             "mean": data.get("mean_e2e_latency_ms", 0.0),
             "p50": data.get("median_e2e_latency_ms", 0.0),
             "p90": data.get("p90_e2e_latency_ms", 0.0),
-            "p95": data.get("p95_e2e_latency_ms", 0.0),
-            "p99": data.get("p99_e2e_latency_ms", 0.0),
+            "p95": _pct(data, "e2e_latency", "p95"),
+            "p99": _pct(data, "e2e_latency", "p99"),
         },
         "ttft_ms": {
             "mean": data.get("mean_ttft_ms", 0.0),
             "p50": data.get("median_ttft_ms", 0.0),
             "p90": data.get("p90_ttft_ms", 0.0),
-            "p95": data.get("p95_ttft_ms", 0.0),
-            "p99": data.get("p99_ttft_ms", 0.0),
+            "p95": _pct(data, "ttft", "p95"),
+            "p99": _pct(data, "ttft", "p99"),
         },
         "tpot_ms": {
             "mean": data.get("mean_tpot_ms", 0.0),
             "p50": data.get("median_tpot_ms", 0.0),
             "p90": data.get("p90_tpot_ms", 0.0),
-            "p95": data.get("p95_tpot_ms", 0.0),
-            "p99": data.get("p99_tpot_ms", 0.0),
+            "p95": _pct(data, "tpot", "p95"),
+            "p99": _pct(data, "tpot", "p99"),
         },
         "itl_ms": {
             "mean": data.get("mean_itl_ms", 0.0),
             "p50": data.get("median_itl_ms", 0.0),
             "p90": data.get("p90_itl_ms", 0.0),
-            "p95": data.get("p95_itl_ms", 0.0),
-            "p99": data.get("p99_itl_ms", 0.0),
+            "p95": _pct(data, "itl", "p95"),
+            "p99": _pct(data, "itl", "p99"),
         },
     }