sandbox setting details

[monster476]

作者您好，感谢您杰出的工作，想询问sandbox环境的设置细节，在internal_sandbox.py文件中的https://seed-sandbox.byteintl.net/faas/sandbox/在正常训练时是否需要替换为http://127.0.0.1:12345/faas/sandbox/，同时在运行之前是否需要执行uvicorn sandbox_api:app --host 127.0.0.1 --port 12345 --workers 4启动；能否提供更加详细的运行说明，感谢。
当前我按照如下步骤运行报错：
step 1: 替换https://seed-sandbox.byteintl.net/faas/sandbox/为http://127.0.0.1:12345/faas/sandbox/；
step 2: 执行uvicorn sandbox_api:app --host 127.0.0.1 --port 12345 --workers 4；
step 3: 执行训练脚本；
报错信息：Traceback (most recent call last):
File "xxx/SimpleTIR/recipe/simpletir/main_simpletir.py", line 59, in main
run_simpletir(config)
File "xxx/SimpleTIR/recipe/simpletir/main_simpletir.py", line 81, in run_simpletir
ray.get(runner.run.remote(config))
File "/usr/local/conda/lib/python3.10/site-packages/ray/_private/auto_init_hook.py", line 21, in auto_init_wrapper
return fn(*args, **kwargs)
File "/usr/local/conda/lib/python3.10/site-packages/ray/_private/client_mode_hook.py", line 103, in wrapper
return func(*args, **kwargs)
File "/usr/local/conda/lib/python3.10/site-packages/ray/_private/worker.py", line 2782, in get
values, debugger_breakpoint = worker.get_objects(object_refs, timeout=timeout)
File "/usr/local/conda/lib/python3.10/site-packages/ray/_private/worker.py", line 929, in get_objects
raise value.as_instanceof_cause()
ray.exceptions.RayTaskError: ray::TaskRunner.run() (pid=100716, ip=10.164.126.42, actor_id=cbc06dbbad9d239b86903e8201000000, repr=<main_simpletir.TaskRunner object at 0x7f58bc319300>)
File "xxx/SimpleTIR/recipe/simpletir/main_simpletir.py", line 213, in run
trainer.fit()
File "xxx/SimpleTIR/recipe/simpletir/simpletir_ray_trainer.py", line 1074, in fit
gen_batch_output = generation_manager.run_llm_loop(
File "xxx/SimpleTIR/recipe/simpletir/agent_utils.py", line 435, in run_llm_loop
next_obs, dones, is_void_turn, code_info = self.execute_predictions(
File "xxx/SimpleTIR/recipe/simpletir/agent_utils.py", line 606, in execute_predictions
sandbox_success, sandbox_stdout, sandbox_stderr = asyncio.run(
File "/usr/local/conda/lib/python3.10/asyncio/runners.py", line 44, in run
return loop.run_until_complete(main)
File "uvloop/loop.pyx", line 1518, in uvloop.loop.Loop.run_until_complete
File "xxx/SimpleTIR/sandbox/internal_sandbox.py", line 35, in parallel_sandbox
results = await asyncio.gather(*tasks_async, return_exceptions=False)
File "xxx/SimpleTIR/sandbox/internal_sandbox.py", line 21, in single_sandbox
response = await run_code_async(request, client_timeout=30.0, max_attempts=2)
File "xxx/.local/lib/python3.10/site-packages/sandbox_fusion/async_client.py", line 48, in run_code
return await _run_code(request)
File "xxx/.local/lib/python3.10/site-packages/tenacity/asyncio/init.py", line 189, in async_wrapped
return await copy(fn, *args, **kwargs)
File "xxx/.local/lib/python3.10/site-packages/tenacity/asyncio/init.py", line 111, in call
do = await self.iter(retry_state=retry_state)
File "xxx/.local/lib/python3.10/site-packages/tenacity/asyncio/init.py", line 153, in iter
result = await action(retry_state)
File "xxx/.local/lib/python3.10/site-packages/tenacity/_utils.py", line 99, in inner
return call(*args, **kwargs)
File "xxx/.local/lib/python3.10/site-packages/sandbox_fusion/client.py", line 49, in on_retry_error
raise e
File "xxx/.local/lib/python3.10/site-packages/tenacity/asyncio/init.py", line 114, in call
result = await fn(*args, **kwargs)
File "xxx/.local/lib/python3.10/site-packages/sandbox_fusion/client.py", line 70, in async_wrapper
return await func(*args, **kwargs)
File "xxx/.local/lib/python3.10/site-packages/sandbox_fusion/async_client.py", line 42, in _run_code
raise Exception(f'Faas api responded with code {result.status}: {await result.text()}')
Exception: Faas api responded with code 404: {"detail":"Not Found"}

[ltzheng]

使用本地sandbox设置好环境变量即可, 请参考sandbox目录下的README和local_sandbox.py

[monster476]

作者您好，感谢您的回复，上面的问题已解决。想进一步询问训练时间，由于权限问题无法安装firejail，当前采用readme中的设置（相同的数据集，verl版本，batch_size等）训练Qwen2.5-7B模型，所使用设备为8*A100-80G，采用默认的sandbox

超过半小时没有跑出第一步，请问是否是正常现象，firejail的加速效果如何？当前运行结果如图所示

[seTalent]

看来是超时等问题，我尝试了sandbox-fusion这个库，速度还行，但是也会有超时等问题，即便我的cpu和内存是充足的，非常好奇如何解决这个问题。

{
	'case_index': 0,
	'input': None,
	'expected_output': None,
	'api_request_error': None,
	'api_response': {
		'status': 'Failed',
		'message': '',
		'compile_result': None,
		'run_result': {
			'status': 'TimeLimitExceeded',
			'execution_time': 120.00120234489441,
			'return_code': None,
			'stdout': '',
			'stderr': ''
		},
		'executor_pod_name': None,
		'files': {}
	},
	'status': 'timeout',
	'stdout': '',
	'stderr': '',
	'exit_code': None,
	'duration': 120.00120234489441,
	'compile_duration': None,
	'compile_stderr': None,
	'api_status': 'Failed',
	'compile_status': None,
	'run_status': 'TimeLimitExceeded'
}，

[monster476]

我把batch size改成32，采用python3作为沙盒而不用firejail时好像在200step左右能开始收敛，曲线如下：

[seTalent]

本地sandbox如何使用的，非常好奇，我使用了sandbox-fusion在该代码中，并发速度是ok的，但是非常容易出现请求超时导致训练被终止。

[monster476]

我修改了sandbox_api.py中的cmd代码为cmd = ["python3", src.name]

[seTalent]

速度怎么样啊，大概多久一个step呢？
我使用sanbox一个step大概7-8分钟，2个step左右就会因为超时挂掉。

[monster476]

200步大概16h，设备为8*A100-80G

[seTalent]

非常感谢，我看到您提到缩小了训练的batch_size，这是为什么，并发的限制么？
此外，除了这个，您还有修改其他的内容吗，我想尝试你的复现方法。

[monster476]

改小batch_size只是为了加快出一个step的时间，别的内容我没有修改了，只替换了python3的部分

[seTalent]

(TaskRunner pid=206133) WARNING:2025-09-16 11:15:22,213:single_sandbox attempt 1/2 failed: 500, message='Internal Server Error', url='http://127.0.0.1:12345/faas/sandbox'
(TaskRunner pid=206133) WARNING:2025-09-16 11:15:22,216:single_sandbox attempt 1/2 failed: 500, message='Internal Server Error', url='http://127.0.0.1:12345/faas/sandbox'
(TaskRunner pid=206133) WARNING:2025-09-16 11:15:22,217:single_sandbox attempt 1/2 failed: 500, message='Internal Server Error', url='http://127.0.0.1:12345/faas/sandbox'
(TaskRunner pid=206133) WARNING:2025-09-16 11:15:22,218:single_sandbox attempt 1/2 failed: 500, message='Internal Server Error', url='http://127.0.0.1:12345/faas/sandbox'
(TaskRunner pid=206133) WARNING:2025-09-16 11:15:22,219:single_sandbox attempt 1/2 failed: 500, message='Internal Server Error', url='http://127.0.0.1:12345/faas/sandbox'
(TaskRunner pid=206133) WARNING:2025-09-16 11:15:22,219:single_sandbox attempt 1/2 failed: 500, message='Internal Server Error', url='http://127.0.0.1:12345/faas/sandbox'
(TaskRunner pid=206133) WARNING:2025-09-16 11:15:22,220:single_sandbox attempt 1/2 failed: 500, message='Internal Server Error', url='http://127.0.0.1:12345/faas/sandbox'
(TaskRunner pid=206133) [debug] void turn number: 0 out of 1216 samples

我遇到了上面提到的问题，请问这个对实际训练的有影响么？
我后面会尝试一下增大unicorn的workers个数。

[monster476]

我的时常也会请求失败，但大多数情况能请求成功，因为代码执行失败可能一定程度降低评估精度。

[vengdeng]

作者您好，感谢您的回复，上面的问题已解决。想进一步询问训练时间，由于权限问题无法安装firejail，当前采用readme中的设置（相同的数据集，verl版本，batch_size等）训练Qwen2.5-7B模型，所使用设备为8*A100-80G，采用默认的sandbox

超过半小时没有跑出第一步，请问是否是正常现象，firejail的加速效果如何？当前运行结果如图所示

你好，请问一下404的问题是怎么解决的啊，感谢！