Add dlpack support

[SunDoge]

DLPack is an open in-memory tensor structure for sharing tensors among frameworks. DLPack enables

• Easier sharing of operators between deep learning frameworks.
• Easier wrapping of vendor level operator implementations, allowing collaboration when introducing new devices/ops.
• Quick swapping of backend implementations, like different version of BLAS
• For final users, this could bring more operators, and possibility of mixing usage between frameworks.

Supporting dlpack will make it easier to share tensors with dfdx, tch-rs and even numpy, torch in Python (with pyo3 feature).

[nilgoyette]

I don't know much about dlpack, but I'm quite sure this won't be merged if

• it's not hidden behind a feature gate
• it doesn't use a real crates.io version (not git)

[SunDoge]

This is still a draft. The api of dlpark is still evolving and I'll release v0.3.0 this week. I want to utilize the ownership system to make sure dlpack can only be consumed once, otherwise there will be double free error. And thanks for your advice, I will add the feature gate.

[bluss]

this function takes a raw pointer (wrapped in NonNull) and it must be an unsafe function, otherwise we can trivially violate memory safety unfortunately.

The only way to remove this requirement - the requirement of using unsafe - would be if you have a "magical" function that can take an arbitrary pointer and say whether it's a valid, live, non-mutably aliased pointer to a tensor.

Here's how to create a dangling bad pointer: NonNull::new(1 as *mut u8 as *mut dlpark::ffi::DLManagedTensor) does this code crash if we run with this pointer? I think it would..

[SunDoge]

I agree with you. from_dlpack should be unsafe, and users should use it at their own risk.

[bluss]

I would say, we normally don't commit to public dependencies that are not stable (yes, not a very fair policy since ndarray itself is not so stable.), and dlpark is a public dependency here because it becomes part of our API. It could mean it takes a long time between version bumps.

[SunDoge]

Maybe we don't need to include dlpark as a dependency. We can create an ArrayView using ArrayView::from_shape_ptr and ManagedTensor. I can implement ToTensor for ArrayD in dlpark with a new feature ndarray. I'll do some quick experiments.

[bluss]

Later work, check compatibility of dlpack and ndarray strides, how they work, their domains etc.