diff --git a/simulator/README.md b/simulator/README.md
index afaf2983..55589af5 100644
--- a/simulator/README.md
+++ b/simulator/README.md
@@ -150,9 +150,11 @@ python -m simulator sweep \
| `length_mix` | `{tokens,weight}[]` | 持续模式的离散长度分布；`weight` 只需为正数，不要求预归一化。 |
| `csv_path` | `string|null` | CLI 读取的 CSV 路径；与 `csv_text` 互斥。页面上传会使用 `csv_text`。 |
| `csv_text` | `string|null` | CSV 原始内容；与 `csv_path` 互斥。 |
-| `csv_sampling` | `cycle|sample` / `"cycle"` | 无时间戳 CSV 在持续模式下的选取方式：按行循环或有放回随机采样。 |
+| `csv_sampling` | `cycle|sample` / `"cycle"` | CSV 在合成到达过程下的长度选取方式：按行循环或有放回随机采样。 |
-优先级：CSV > `fixed_lengths`/`length_mix`。固定模式下 CSV 每行回放一次；持续模式下无时间戳 CSV 提供经验长度分布。
+优先级：CSV > `fixed_lengths`/`length_mix`。固定模式下 CSV 每行回放一次；
+持续模式下，`constant`/`poisson` 把 CSV 当作经验长度分布，即使文件带
+`arrival_time_ms` 也会忽略该列；只有 `trace`/`scaled_trace` 会消费时间戳。
WebUI 还提供 MoonConv V4 Flash 的 `formal_0/1/2`（各 512 条）和
`screening`（128 条）内置 trace。它们只保留输入长度与相对到达时间；选择后
@@ -177,6 +179,10 @@ WebUI 还提供 MoonConv V4 Flash 的 `formal_0/1/2`（各 512 条）和
配置值。QPS sweep 可以使用 `scaled_trace`，每个扫描点重新缩放同一 arrival
形状。
+WebUI 上传带 `arrival_time_ms` 的 CSV 时默认切换到持续模式和
+`scaled_trace`。仍可手动切换为 `trace` 精确回放，或切换为
+`constant`/`poisson` 仅使用其长度分布。
+
### 2.3 `scheduler`
| 字段 | 类型 / 默认值 | 说明 |
diff --git a/simulator/tests/test_server.py b/simulator/tests/test_server.py
index 17ce5206..3adaccf2 100644
--- a/simulator/tests/test_server.py
+++ b/simulator/tests/test_server.py
@@ -187,6 +187,8 @@ def test_defaults_page_and_simulation_api(self) -> None:
self.assertIn("data.length_datasets.forEach", page)
self.assertIn("/api/length-datasets/", page)
self.assertIn('value="scaled_trace"', page)
+ self.assertIn("function useTimestampedCsvDefaults", page)
+ self.assertIn("columns.includes('arrival_time_ms')", page)
self.assertIn("item.num_devices===afd.num_devices", page)
self.assertIn('value="prefill_token_greedy"', page)
self.assertIn('value="prefill_token_square_greedy"', page)
diff --git a/simulator/tests/test_workload.py b/simulator/tests/test_workload.py
index fc7a7143..f736094a 100644
--- a/simulator/tests/test_workload.py
+++ b/simulator/tests/test_workload.py
@@ -104,6 +104,33 @@ def test_scaled_trace_keeps_zero_gap_arrivals_simultaneous(self) -> None:
self.assertEqual([request.arrival_ms for request in workload], [0.0, 0.0])
+ def test_synthetic_arrival_uses_timestamped_csv_as_length_pool(self) -> None:
+ config = SimulationConfig.from_mapping(
+ {
+ "mode": "continuous",
+ "csv_text": (
+ "arrival_time_ms,input_length\n"
+ "100,128\n"
+ "125,256\n"
+ "160,512\n"
+ ),
+ "arrival": {
+ "kind": "constant",
+ "qps": 2,
+ "duration_s": 1,
+ "warmup_s": 0,
+ },
+ }
+ )
+
+ workload = generate_workload(config)
+
+ self.assertEqual([request.arrival_ms for request in workload], [0.0, 500.0])
+ self.assertEqual(
+ [request.input_tokens for request in workload],
+ [128, 256],
+ )
+
def test_prefix_cache_sampling_is_deterministic_and_block_aligned(self) -> None:
raw = {
"mode": "fixed",
diff --git a/simulator/web/index.html b/simulator/web/index.html
index bd722fe4..f1327abe 100644
--- a/simulator/web/index.html
+++ b/simulator/web/index.html
@@ -334,11 +334,21 @@ 仿真参数
$('profile').textContent = `${metadata.model || 'DSV4-Flash'} · ${metadata.device || 'device unknown'} · ${afd.num_devices} die · ${afd.layer_count} 层`;
}
+ function useTimestampedCsvDefaults(text) {
+ const header = text.replace(/^\ufeff/, '').split(/\r?\n/, 1)[0];
+ const columns = header.split(',').map(column => column.trim());
+ if (!columns.includes('arrival_time_ms')) return;
+ $('mode').value = 'continuous';
+ $('arrival-kind').value = 'scaled_trace';
+ }
+
$('csv-file').addEventListener('change', async event => {
const file = event.target.files[0];
csvText = file ? await file.text() : null;
$('length-dataset').value = '';
$('csv-status').textContent = file ? `${file.name} · ${(file.size / 1024).toFixed(1)} KiB` : '未加载 CSV';
+ if (csvText) useTimestampedCsvDefaults(csvText);
+ updateControlState();
});
$('length-dataset').addEventListener('change', async event => {
@@ -356,8 +366,7 @@ 仿真参数
const response = await fetch(`/api/length-datasets/${encodeURIComponent(datasetId)}`);
if (!response.ok) throw new Error(`加载长度数据集失败：HTTP ${response.status}`);
csvText = await response.text();
- $('mode').value = 'continuous';
- $('arrival-kind').value = 'scaled_trace';
+ useTimestampedCsvDefaults(csvText);
$('csv-status').textContent = `${dataset.label} · ${dataset.request_count} 条 · ${(csvText.length / 1024).toFixed(1)} KiB`;
updateControlState();
} catch (error) {
diff --git a/simulator/workload.py b/simulator/workload.py
index 28efd7d9..123b17e0 100644
--- a/simulator/workload.py
+++ b/simulator/workload.py
@@ -129,7 +129,11 @@ def generate_workload(config: SimulationConfig) -> tuple[RequestSpec, ...]:
]
return _materialize_requests(raw_requests, arrivals, config.prefix_cache)
- if csv_requests and all(item.arrival_time_ms is not None for item in csv_requests):
+ if (
+ csv_requests
+ and config.arrival.kind in {"trace", "scaled_trace"}
+ and all(item.arrival_time_ms is not None for item in csv_requests)
+ ):
if config.arrival.kind == "scaled_trace":
raw_requests, arrivals = _scale_and_repeat_trace(
csv_requests,
@@ -141,11 +145,6 @@ def generate_workload(config: SimulationConfig) -> tuple[RequestSpec, ...]:
arrivals,
config.prefix_cache,
)
- if config.arrival.kind != "trace":
- raise ValueError(
- "CSV arrival_time_ms requires arrival.kind=trace for exact replay "
- "or scaled_trace for QPS-scaled replay"
- )
first_arrival = float(csv_requests[0].arrival_time_ms or 0.0)
arrivals = [
float(item.arrival_time_ms or 0.0) - first_arrival for item in csv_requests