• Draft state
• Closes Run tests with miri #897
• Closes Reduce test sizes #898

• Temporary patch on crossbeam.
  • Note that crossbeam is an indirect dependency and they are still to release a version that passes on this miri check. So to temporarily get this out of the way, you can patch crossbeam on the workspace dfdx/Cargo.toml.
  • Wait for crossbeam to make a new release and patch to that.
  • Then try to update dependencies so the patch is no longer needed.
• Check all test pass by some feature combinations.
  • std,fast-alloc,cpu,nightly,safetensors
  • std,cpu,nightly,safetensors
  • no-std,fast-alloc,nightly
  • no-std,nightly
• Reduce test sizes so a miri pass is practical

To run a particular test, add it's rust path as a filter at the end of the miri run command:

# cargo clean (run once when `cargo miri` is not used)
cargo clean

# run all tests (note: very slow and should take hours, assuming there are no errors)
MIRIFLAGS='-Zmiri-strict-provenance -Zmiri-symbolic-alignment-check -Zmiri-retag-fields -Zmiri-disable-isolation -Zmiri-tree-borrows' \
cargo miri nextest run --test-threads 1 --no-default-features --features std,fast-alloc,cpu,nightly,safetensors

Miri UB errors

• None found, both under dfdx package (74 tests) and under dfdx-core(293 tests).

Memory/Thread leaks

• nn::layers::add_into::tests::longer_network
  • Thread-leak error, disabling rayon on gemm for matmul seems to fix it. Commit: c6e0642
  • Investigate why the leak happened, try to re-enable rayon usage.
• tensor::cpu::tests::test_reuse_allocation_on_clone_tensor
  • Memory-leak error, deallocating the cached allocations on TensorCache drop seems to fix it. Commit: ae777d4
  • Avoid panics during drops? Instead of panicking, can emit error message warning about possible memory leaks.
• None other found, both under dfdx package (74 tests) and under dfdx-core(293 tests).

Smaller tensors

Some tests have been reduced in tensor size so miri runs faster. Those changes are specific to the following commits:

• 9deca1b
• 86c9cbb

The tensor reduction is quite optional, they only makes it simpler to run miri "on everything". But if this change ends up persisting, this wiki page should be updated. Finally, Idk how mha/transformer work, so I can't tell if the reduction shouldn't happen. That is, each test reduction should be reviewed to ack their reduction doesn't lose anything that should be tested.

Note that Miri depends on the test suite to find something, so having a bigger test suite may help. This is to say that a Miri check could be re-run in the future.
I haven't verified the cpu test coverage over the cpu source code, but a good way to go would be to have a high coverage (every relevant line of code getting "tested"). If more assurance would be desired, there could be more tests aiming at different characteristics of the source code, such as those related to branch coverage etc.