Core v1 Conformance

Core v1 defines compatibility expectations for runtimes targeting the MIND IR. This page explains the CPU and GPU profiles and shows how to run the official conformance suite.

Core v1 CPU baseline

  • Guarantees deterministic execution for the full Core v1 op set on CPU, including shape inference, gradients, and canonical rewriting behavior.
  • Conformance expectations are defined in the Core spec under conformance.md.
  • Run the official suite locally with: mindc conformance --profile cpu

Core v1 GPU profile

  • Optional profile for implementations that expose GPU or accelerator devices.
  • Supports the full Core v1 op surface (17 operations) with GPU-specific constraints.
  • Reference implementation: MockGpuBackend executes via CPU delegation for conformance testing.

GPU tensor constraints:

  • DeviceKind::Gpu device targeting
  • dtype: f32 only (f64 not supported on GPU profile)
  • numel % 4 == 0 alignment requirement
  • fill reads value from inputs[0].data[0] (first element)

Running the conformance suite

From the star-ga/mind repo, run the Rust test harness:

cargo test --test conformance

From any environment with mindc installed:

mindc conformance --profile cpu
mindc conformance --profile gpu

The full operator catalogue the suite is checked against is printable from the same binary: mindc ops --core-v1 lists every operator with its arity, admissible dtypes, and whether it is differentiable. --profile gpu exercises the GPU profile— the open-source build ships no GPU backend, so it runs against the CPU-delegating reference implementation; requesting the GPU target from the compiler returns a structured no backend available for target gpu error rather than falling back silently.

For other implementations

The conformance corpus is intended to be reused—port the tests to your runtime or run them through a compatibility shim. Opening issues or PRs against mind-spec is the best way to clarify expectations.