Execution providers belong in the record
Record the provider configuration used to execute a model, not only the model filename.
Research note · November 14, 2026Research question
Can the same model artifact run through different execution paths?
Method
- Use a model supported by a controlled ONNX Runtime installation. Record the requested providers, their order, runtime version and provider-specific options before creating the session.
- Run the same input under two explicitly selected configurations. Capture supported evidence of node assignment or profiling where available, rather than assuming that a requested provider executed every operation.
- Preserve failed session creation and fallback behavior as outcomes. Compare outputs and timing only after documenting the actual execution conditions.
Records to retain
- Runtime and provider versions
- Provider order and options
- Available assignment or profiling evidence
- Inputs, outputs and failures
Source basis
ONNX Runtime uses execution-provider capability information to allocate nodes or subgraphs. Provider selection is therefore part of the execution context, and a provider’s presence does not alone establish that the entire graph ran there.
Interpretation & limits
This protocol does not extend Contour’s historical CPU-only experiment to other providers. Output agreement across two runs is not a general numerical-equivalence guarantee.
Next experiments
- Choose one small graph with known provider support.
- Record an unsupported-operation case without silently excluding it.
References
Technical reference
Implementation details supporting this note, drawn directly from the maintained documentation. These are technical references, not additional papers or new experimental results. Expand a topic to inspect its method, fields and limits.