mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2025-12-24 13:28:13 +08:00
99258e19c8
* benchmark工具支持受限解码场景指定response_format * Update backend_request_func.py output.success判断兼容思考内容超长截断时回复内容为空的情况 * Update benchmark_serving.py 更新benchmark_metrics * 支持Completions接口 * 支持Completions接口 * 支持Completions接口 * [Benchmark]支持Completions接口 * [Benchmark]支持Completions接口 --------- Co-authored-by: YuBaoku <49938469+EmmonsCurse@users.noreply.github.com>
106 lines
3.4 KiB
Python
106 lines
3.4 KiB
Python
#!/bin/env python3
|
||
# -*- coding: utf-8 -*-
|
||
# @author DDDivano
|
||
# encoding=utf-8 vi:ts=4:sw=4:expandtab:ft=python
|
||
|
||
from core import TEMPLATE, URL, build_request_payload, send_request
|
||
|
||
|
||
def test_incremental_image_reasoning_consistency():
|
||
"""
|
||
多模态增量推理一致性检查：
|
||
第一次请求携带图片和文本，并打开 return_token_ids。
|
||
第二次请求拼入 prompt_token_ids，要求输出一致，否则校验失败。
|
||
"""
|
||
|
||
# 第一次请求数据
|
||
data_1st = {
|
||
"stream": False,
|
||
"messages": [
|
||
{
|
||
"role": "user",
|
||
"content": [
|
||
{
|
||
"type": "image_url",
|
||
"image_url": {
|
||
"url": "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_images/example2.jpg"
|
||
},
|
||
},
|
||
{"type": "text", "text": "图中的文物属于哪个年代？"},
|
||
],
|
||
}
|
||
],
|
||
"return_token_ids": True,
|
||
}
|
||
|
||
print("==== 开始第一次请求 ====")
|
||
payload_1st = build_request_payload(TEMPLATE, data_1st)
|
||
res1 = send_request(URL, payload_1st)
|
||
print(f"第一次请求返回码: {res1.status_code}")
|
||
|
||
if res1.status_code != 200:
|
||
raise AssertionError(f"首次请求失败: {res1.text}")
|
||
|
||
result1 = res1.json()
|
||
choice1 = result1["choices"][0]
|
||
msg1 = choice1.get("message", {})
|
||
|
||
content1 = msg1.get("content", "")
|
||
reasoning1 = msg1.get("reasoning_content", "")
|
||
tokens1 = msg1.get("prompt_token_ids", [])
|
||
print(f"第一次请求结果: {tokens1}")
|
||
|
||
if not tokens1:
|
||
raise AssertionError("首次请求未返回 prompt_token_ids！")
|
||
|
||
print(f"第一次 content = {content1}")
|
||
print(f"第一次 reasoning_content = {reasoning1}")
|
||
print(f"第一次 prompt_token_ids 长度 = {len(tokens1)}")
|
||
|
||
# 构造第二次请求
|
||
print("\n==== 开始第二次请求（携带 prompt_token_ids）====")
|
||
data_2nd = {
|
||
"stream": False,
|
||
"messages": [
|
||
{
|
||
"role": "user",
|
||
"content": [
|
||
{
|
||
"type": "image_url",
|
||
"image_url": {
|
||
"url": "https://paddlenlp.bj.bcebos.com/datasets/paddlemix/demo_images/example2.jpg"
|
||
},
|
||
},
|
||
],
|
||
}
|
||
],
|
||
"prompt_token_ids": tokens1,
|
||
}
|
||
|
||
payload_2nd = build_request_payload(TEMPLATE, data_2nd)
|
||
res2 = send_request(URL, payload_2nd)
|
||
print(f"第二次请求返回码: {res2.status_code}")
|
||
|
||
if res2.status_code != 200:
|
||
raise AssertionError(f"二次请求失败: {res2.text}")
|
||
|
||
result2 = res2.json()
|
||
choice2 = result2["choices"][0]
|
||
msg2 = choice2.get("message", {})
|
||
|
||
content2 = msg2.get("content", "")
|
||
reasoning2 = msg2.get("reasoning_content", "")
|
||
|
||
print(f"第二次 content = {content2}")
|
||
print(f"第二次 reasoning_content = {reasoning2}")
|
||
|
||
# 一致性校验
|
||
assert content1 == content2, "content 不一致，增量推理一致性校验失败!"
|
||
assert reasoning1 == reasoning2, "reasoning_content 不一致，增量推理一致性校验失败!"
|
||
|
||
print("\n 一致性校验通过！增量推理行为正常。")
|
||
|
||
|
||
if __name__ == "__main__":
|
||
test_incremental_image_reasoning_consistency()
|