PyPI 1.16.0 release requires specifying the execution provider during InferenceSession creation
cbourjau opened this issue · 5 comments
Describe the issue
The latest 1.16.0 CPU-release (aka onnxruntime) on PyPI appears to have AzureExecutionProvider as a default provider (introduced by #17025 ?). This means that the EP is now ambiguous with respect to the second default CPUExecutionProvider causing an exception:
ValueError: This ORT build has ['AzureExecutionProvider', 'CPUExecutionProvider'] enabled. Since ORT 1.9, you are
required to explicitly set the providers parameter when instantiating InferenceSession. For example,
onnxruntime.InferenceSession(..., providers=['AzureExecutionProvider', 'CPUExecutionProvider'], ...)
The text in the exception is misleading. I just checked and as of 1.15.1 it was not necessary to specify an explicit provider through the Python interface. No warning is given in 1.15.1 about this change either.
This is a breaking change in the Python interface. Furthermore, the choice of making the AzureExecutionProvider EP one of the defaults seems strange given that it appears to be undocumented.
To reproduce
Create an inference session via the Python API (using a random model from the onnx/onnx repository):
import onnxruntime as ort
ort.InferenceSession("onnx/onnx/backend/test/data/node/test_abs/model.onnx")Urgency
This is a breaking change with no prior warnings on possibly the most common way to initialize a session.
Platform
Mac
OS Version
12.3.1
ONNX Runtime Installation
Released Package
ONNX Runtime Version or Commit ID
1.16.0
ONNX Runtime API
Python
Architecture
X64
Execution Provider
Default CPU
Execution Provider Library Version
No response
Edit: Clarified that this is about the CPU-release
The providers parameter is required for onnxruntime-gpu 1.15.1 in linux/windows.
Maybe this issue is for cpu-only package.
Yes, this is for the cpu package, or at least that is what I think. I installed using:
$ pip install onnxruntime==1.16
Yes, this looks like a valid issue. The workaround is to explicitly supply the CPUExecutionProvider in the list. @RandySheriff - can you take a look?
+1 also having this issue