Ryan Nett

That's more a TF-core thing, I don't know if there's support for it, although it seems like a common enough use case. If you don't find a way, you may...

Yeah, that's what #375 is for. Some of the ops weren't generated when they were added, and I forgot to commit the new generation for the init scope PR.

Mostly what you generate locally, but apparently there's a few ops that are misclassified.

Check your CUDA and cudnn versions, 0.4.0 is on tensorflow 2.5.0, 3.1.0 was on 2.4.1, and I think the required CUDA version changed.

Here's the test: ```kotlin import org.tensorflow.DeviceSpec import org.tensorflow.op.Ops fun main() { val tf = Ops.create().withDevice(DeviceSpec.newBuilder().deviceType(DeviceSpec.DeviceType.GPU).build()) println(tf.constant(10).asTensor().getInt()) } ``` here's the log: ``` Debug: Loading class org.bytedeco.javacpp.presets.javacpp Debug: Loading class org.bytedeco.javacpp.Loader...

Comparing out `.bazelrc` to tensorflow's, it seems like the cuda options changed, i.e. there is no `using_cuda` any more. Additionally, our `.bazelrc` refers to cuda `10.1` while tensorflow's uses `11.2`....

We really should try to get some GPU CI tests.

I should be able to test it locally, I'm running native builds for the custom gradient PR anyways. I'm also seeing a new op `QuantizedConcatV2` when I build locally, that...

Also, when I run generation locally I get a bunch of new op classes like `CacheDatasetV2` that are in `ops.pb` but don't seem to be generated in `master`, can anyone...

I also added a `.gitattributes` to set the line endings of most files to `LF`, since I was running into issues with it w/ the formatter when switching between Windows...