Summary
Compiling a regular ONNX model with WinMLCG emits a warning containing EPContext fallback, has_attrs=0, and is_epctx=0, even though this is the initial creation of an EP context from a source ONNX model. Compilation and subsequent GPU execution succeed.
The wording and warning severity make a successful first-time compilation look like a failed cache load or execution fallback. This report concerns the WinMLCG diagnostic surfaced by winml compile; it may need routing to the Windows ML / WinMLCG EP owners.
Environment
- Windows 11 x64; Python 3.11.14
winml-cli: 0.4.0windowsml: 2.7.52a0onnxruntime-windowsml: 1.30.0.202609102321- GPU: NVIDIA GeForce RTX 5080
- Model:
microsoft/resnet-50
Steps to reproduce
Run in Command Prompt with the environment activated:
winml export -m microsoft/resnet-50 -o temp\bugbash-cgc\resnet50.onnx --overwrite
winml compile -m temp\bugbash-cgc\resnet50.onnx --ep winmlcg --device gpu -o temp\bugbash-cgc\resnet50-winmlcg-ctx.onnx --overwrite
winml perf -m temp\bugbash-cgc\resnet50-winmlcg-ctx.onnx --ep winmlcg --device gpu --runtime winml-ort --iterations 100 --output temp\bugbash-cgc\perf-compiled-ep.json --overwrite
On a machine with multiple GPUs, use --device-luid from winml sys to select the intended adapter for the performance command. The observed run explicitly selected the RTX 5080.
Actual behavior
During the compile command, stderr includes the following warning (the generated fused-node suffix is abbreviated):
[W:onnxruntime:, winmlcg_ep.cpp:3185 winmlcg_ep::WinMLCGExecutionProvider::CompileEPContextSubgraph] EPContext fallback: fused_op=WinMLCGExecutionProvider_<generated-id>_0 fused_domain=WinMLCGExecutionProvider has_attrs=0 is_epctx=0
The command then reports Success! Model compiled and exits 0. The subsequent performance run completes 100/100 iterations and exits 0.
Validation
- The input to
compile is the original source ONNX, not an already compiled EP context.
- The output contains one
com.microsoft:EPContext node with winmlcg_compiled=1 and an existing external MLIR sidecar.
- A separate session reloaded this existing context with
session.disable_cpu_ep_fallback=1 and profiling enabled. Three inferences succeeded with finite [1, 1000] outputs.
- The profile contained three
WinMLCGExecutionProvider kernel events and zero CPUExecutionProvider kernel events. These observations do not support interpreting this warning as a CPU fallback or an invalid compiled result.
Expected behavior
An ordinary first-time ONNX-to-EP-context compilation should not emit a warning that implies cache failure or execution fallback solely because its source graph has no EPContext attributes. Use an informational/debug message for normal path selection, and reserve warning-level diagnostics for an unexpected condition requiring user attention. If a real fallback occurs, identify its source and destination explicitly.
Impact
Users cannot tell from this warning whether compilation is following the expected path or whether the requested provider/cache failed. This is a diagnostic-quality issue; no inference failure was observed in the tested workflow.
Summary
Compiling a regular ONNX model with WinMLCG emits a warning containing
EPContext fallback,has_attrs=0, andis_epctx=0, even though this is the initial creation of an EP context from a source ONNX model. Compilation and subsequent GPU execution succeed.The wording and warning severity make a successful first-time compilation look like a failed cache load or execution fallback. This report concerns the WinMLCG diagnostic surfaced by
winml compile; it may need routing to the Windows ML / WinMLCG EP owners.Environment
winml-cli: 0.4.0windowsml: 2.7.52a0onnxruntime-windowsml: 1.30.0.202609102321microsoft/resnet-50Steps to reproduce
Run in Command Prompt with the environment activated:
winml export -m microsoft/resnet-50 -o temp\bugbash-cgc\resnet50.onnx --overwrite winml compile -m temp\bugbash-cgc\resnet50.onnx --ep winmlcg --device gpu -o temp\bugbash-cgc\resnet50-winmlcg-ctx.onnx --overwrite winml perf -m temp\bugbash-cgc\resnet50-winmlcg-ctx.onnx --ep winmlcg --device gpu --runtime winml-ort --iterations 100 --output temp\bugbash-cgc\perf-compiled-ep.json --overwriteOn a machine with multiple GPUs, use
--device-luidfromwinml systo select the intended adapter for the performance command. The observed run explicitly selected the RTX 5080.Actual behavior
During the compile command, stderr includes the following warning (the generated fused-node suffix is abbreviated):
The command then reports
Success! Model compiledand exits 0. The subsequent performance run completes 100/100 iterations and exits 0.Validation
compileis the original source ONNX, not an already compiled EP context.com.microsoft:EPContextnode withwinmlcg_compiled=1and an existing external MLIR sidecar.session.disable_cpu_ep_fallback=1and profiling enabled. Three inferences succeeded with finite[1, 1000]outputs.WinMLCGExecutionProviderkernel events and zeroCPUExecutionProviderkernel events. These observations do not support interpreting this warning as a CPU fallback or an invalid compiled result.Expected behavior
An ordinary first-time ONNX-to-EP-context compilation should not emit a warning that implies cache failure or execution fallback solely because its source graph has no EPContext attributes. Use an informational/debug message for normal path selection, and reserve warning-level diagnostics for an unexpected condition requiring user attention. If a real fallback occurs, identify its source and destination explicitly.
Impact
Users cannot tell from this warning whether compilation is following the expected path or whether the requested provider/cache failed. This is a diagnostic-quality issue; no inference failure was observed in the tested workflow.