Onnxruntime-directml 1.18.0 broken multithreading inference session

[Djdefrag]

Describe the issue

With the new version 1.18 it seems that trying to use different InferenceSession using the same DirectML device, all threads remain stalled without giving any exception or error

To reproduce

Thread 1

 AI_model_loaded = onnx_load(AI_model_path)

AI_model = onnxruntime_inferenceSession(
    path_or_bytes = AI_model_loaded.SerializeToString(), 
    providers =  [('DmlExecutionProvider', {"device_id": "0"})]
)    

onnx_input  = {AI_model.get_inputs()[0].name: image}
onnx_output = AI_model.run(None, onnx_input)[0]

Thread n (where n can be any number)

AI_model_loaded = onnx_load(AI_model_path)

AI_model = onnxruntime_inferenceSession(
    path_or_bytes = AI_model_loaded.SerializeToString(), 
    providers =  [('DmlExecutionProvider', {"device_id": "0"})]
)    

onnx_input  = {AI_model.get_inputs()[0].name: image}
onnx_output = AI_model.run(None, onnx_input)[0]

Urgency

No response

Platform

Windows

OS Version

10

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.18.0

ONNX Runtime API

Python

Architecture

X64

Execution Provider

DirectML

Execution Provider Library Version

1.18.0

[sophies927]

Tagging @PatriceVignola @smk2007 @fdwr for visibility.

[saulthu]

Same here on Windows, versions 1.16.0 to 1.17.3 work fine over multiple threads, however 1.18.0 gives Windows fatal exception: access violation with the following stack trace produced by my own Windows SEH handler:

-----------
Caught unhandled exception...
-----------

Terminating from thread id 10152

Non-C++ exception:
  Error: EXCEPTION_ACCESS_VIOLATION
  Type: Read
  Addr: 0x0

Trace:
 40:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 39:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 38:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 37:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 36:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 35:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 34:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 33:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 32:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 31:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 30:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 29:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 28:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 27:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 26:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 25:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 24:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 23:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 22:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 21:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 20:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 19:  ?: PyInit_onnxruntime_pybind11_state  (onnxruntime_pybind11_state.pyd)
 18:  ?: pybind11::error_already_set::discard_as_unraisable  (onnxruntime_pybind11_state.pyd)
 17:  ?: PyObject_MakeTpCall  (python311.dll)
 16:  ?: PyObject_Vectorcall  (python311.dll)
 15:  ?: PyEval_EvalFrameDefault  (python311.dll)
 14:  ?: PyFunction_Vectorcall  (python311.dll)
 13:  ?: PyFunction_Vectorcall  (python311.dll)
 12:  ?: PyObject_CallObject  (python311.dll)
 11:  ?: PyEval_EvalFrameDefault  (python311.dll)
 10:  ?: PyFunction_Vectorcall  (python311.dll)
  9:  ?: PyObject_CallObject  (python311.dll)
  8:  ?: PyEval_EvalFrameDefault  (python311.dll)
  7:  ?: PyFunction_Vectorcall  (python311.dll)
  6:  ?: PyFunction_Vectorcall  (python311.dll)
  5:  ?: PyObject_Call  (python311.dll)
  4:  ?: PyInterpreterState_Delete  (python311.dll)
  3:  ?: PyInterpreterState_Delete  (python311.dll)
  2:  ?: recalloc  (ucrtbase.dll)
  1:  ?: BaseThreadInitThunk  (KERNEL32.DLL)
  0:  ?: RtlUserThreadStart  (ntdll.dll)

[liuyunms]

We’ve noted the issue with GPU resource contention due to multiple threads. This usage pattern is not recommended as it makes multiple threads request all of the GPU resources, and can cause contention. Also, the allocator in python API (both CUDA and DML) is explicitly not thread safe because it initializes the allocator as a global singleton due it living outside of the session.

We’re investigating the recent failure and will address it. Meanwhile, please avoid this pattern to prevent GPU contention.

[Djdefrag]

Hi @liuyunms

Sorry to bother, I'm currently using an InferenceSession per tread, but you say it shouldn't be used this way.

4 threds -> 4 inference session with same gpu

Do you mean to use the same InferenceSession in multiple threads? Is it possible?

4 threds -> 1 inference session with same gpu

[Djdefrag]

@PatriceVignola @smk2007 @fdwr

Hi, sorry to bother, there are some news for this problem? Actually testing 1.18.1 and the problem is still present :(

Thank you

[zhangxiang1993]

#21566
This PR might be related. @Djdefrag Could you help verify if the problem is fixed?

[saulthu]

@zhangxiang1993 It still crashes using multiple threads in my application. I just tried the nightly build of 1.19 from here (I used the Python 3.11 build for Windows). I've reverted back to 1.17.3 which still works.

[Djdefrag]

@zhangxiang1993

Hi, I can confirm that the problem is also present on 1.19 nightly (python 3.11)

• ORT 1.19 [NOT working]
• ORT 1.18.1 [NOT working]
• ORT 1.18 [NOT working]
• ORT 1.17.3 [working]

[henryruhs]

Not sure if this helps, but I have this method to work around it.

import threading
from contextlib import nullcontext
from typing import ContextManager, Union

THREAD_SEMAPHORE : threading.Semaphore = threading.Semaphore()
NULL_CONTEXT : ContextManager[None] = nullcontext()

def conditional_thread_semaphore() -> Union[threading.Semaphore, ContextManager[None]]:
	if has_execution_provider('directml') or has_execution_provider('rocm'):
		return THREAD_SEMAPHORE
	return NULL_CONTEXT

with conditional_thread_semaphore():
	onnxruntime.run()

Sorry, but implement has_execution_provider yourself :)

[saulthu]

@henryruhs

with conditional_thread_semaphore():
	onnxruntime.run()

This works, however it defeats the purpose of running in multiple threads. In older versions that do not crash I can have the GPU running at 100%, but this workaround causes a very large performance hit.

[henryruhs]

yeah, the performance hit is something I am aware of.

[Djdefrag]

Hi @henryruhs

Thank you. I tried the solution using Semaphore and it works, but the performance is in line with using only one Thread.

Hopefully they will fix the problem with the next release.

[linyu0219]

This problem is significant, so most of us will remain on version 1.17.3. Please fix it.