问题描述 当从位于美国硅谷的基础设施向 Vertex AI API ( aiplatform.googleapis.com ) 模型: gemini-3-pro-preview 发起流式预测调用时,我们观察到响应流中首个 Token 的延迟异常偏高。首 Token 延迟( TTFT )持续超过 17 秒,而通常情况应低于 2 秒。
server address: 142.250.191.42
1 、Basic Ping Tests (Connectivity & Baseline Latency) Run these commands from the affected server/client in Silicon Valley. ping(base) [root@usa-gg-test01 ~]# ping aiplatform.googleapis.com PING aiplatform.googleapis.com (142.250.191.42) 56(84) bytes of data. 64 bytes from nuq04s42-in-f10.1e100.net (142.250.191.42): icmp_seq=1 ttl=118 time=2.67 ms 64 bytes from nuq04s42-in-f10.1e100.net (142.250.191.42): icmp_seq=2 ttl=118 time=2.62 ms 64 bytes from nuq04s42-in-f10.1e100.net (142.250.191.42): icmp_seq=3 ttl=118 time=2.64 ms
2 、python code test Using the model:gemini-3-pro-preview
import requests import json import time
def stream_gemini_content(): api_key='xxx' url = "https://aiplatform.googleapis.com/v1/publishers/google/models/gemini-3-pro-preview:streamGenerateContent?alt=sse"
headers = {
"x-goog-api-key": api_key,
"Content-Type": "application/json"
}
data = {
"contents": [{
"role": "user",
"parts": [{
"text": "请讲一个 200 字的故事,不要用推理,直接回答。"
}]
}],
"generationConfig": {
"thinkingConfig": {
"includeThoughts": False
}
}
}
print(f"begin requests: {url} ...")
start_time = time.time()
first_token_time = None
last_chunk_time = None
try:
with requests.post(url, headers=headers, json=data, stream=True) as response:
if response.status_code != 200:
print(f"status: {response.status_code}")
print(response.text)
return
print("-" * 50)
for line in response.iter_lines():
if not line:
continue
decoded_line = line.decode('utf-8').strip()
if not decoded_line.startswith("data: "):
continue
json_str = decoded_line[6:]
if json_str == "[DONE]":
break
try:
now = time.time()
if first_token_time is None:
first_token_time = now
print(f"\n[total] frist token TTFT: {(now - start_time) * 1000:.2f} ms")
print("-" * 50)
last_chunk_time = now
chunk_data = json.loads(json_str)
candidates = chunk_data.get("candidates", [])
total_elapsed = (now - start_time) * 1000
chunk_gap = (now - last_chunk_time) * 1000 if last_chunk_time else 0
last_chunk_time = now
if candidates:
content = candidates[0].get("content", {})
parts = content.get("parts", [])
if parts:
text_chunk = parts[0].get("text", "")
print(text_chunk, end="", flush=True)
except Exception as e:
pass
except Exception as e:
pass
end_time = time.time()
print("\n\n" + "-" * 50)
print(f"total time: {(end_time - start_time) * 1000:.2f} ms")
if name == "main": stream_gemini_content()
代码测试非常慢,200 个字故事就超过 17s 了
1
heqing 1 天前
第一、第二次调用首个 Token 输出延迟是否有显著差异?更换其他模型是否出现相同的现象?
|
2
xuliang12187 OP 用了 gemini-2.0-flash 模型首个 token 输出 300ms 200 字的故事,3-4s 就返回了全部内容了 gemini-2.5-flash 首 token 超过 3s 很慢,总时间长度超过 5s ,gemini-3-pro-preview 首个 token 超过 12s ,我们用的 google cloud 企业服务 vertex AI apI 接口。
|
3
chenluo0429 1 天前 via Android
你也没指定不思考啊,gemeni3 默认思考级别是高,这不是得先思考再给你回答吗
|
4
xuliang12187 OP @chenluo0429 调过一样,很慢都超过 17s
|
5
fov6363 1 天前
+1 ,不加 thinking 太弱智了,加了就是得 10s+,即使是简单的 QA 也不行。问了 chatGPT 说是 vertex 要开那个 endpoint 独占的实例概念,不了会有冷启动,first chunk 只有几百 ms ,但是等到第一次返回就得 10s+
|
6
xuliang12187 OP @fov6363 vertex 先阶段 没有 endpoint 独立实例概念,现在只有 global 全球的。说是有不同付费级别。那个是针对业务并发量高。并不能解决 接口延迟问题
|
7
GXD 1 天前
gemini3 得用`thinking_level`参数来指定推理深度吧,默认是 high
|
8
fov6363 19 小时 33 分钟前
|