Unintuitive out of memory errors, and odd OOMs

I'm experimenting with TensorScope, and thus using large amounts of memory for tensors. However, when I try this, I get odd oom-related errors.

Example (kotlin):

repeat(1000) {
    TFloat32.tensorOf(Shape.of(10000, 10000))
}

gives

2021-01-11 20:36:45.599992: W external/org_tensorflow/tensorflow/core/framework/cpu_allocator_impl.cc:81] Allocation of 400000000 exceeds 10% of free system memory.
2021-01-11 20:36:45.625897: W external/org_tensorflow/tensorflow/core/framework/cpu_allocator_impl.cc:81] Allocation of 400000000 exceeds 10% of free system memory.
2021-01-11 20:36:45.626092: W external/org_tensorflow/tensorflow/core/framework/cpu_allocator_impl.cc:81] Allocation of 400000000 exceeds 10% of free system memory.
2021-01-11 20:36:45.626180: W external/org_tensorflow/tensorflow/core/framework/cpu_allocator_impl.cc:81] Allocation of 400000000 exceeds 10% of free system memory.
2021-01-11 20:36:45.626262: W external/org_tensorflow/tensorflow/core/framework/cpu_allocator_impl.cc:81] Allocation of 400000000 exceeds 10% of free system memory.
Exception in thread "main" java.lang.NullPointerException
	at org.tensorflow.internal.buffer.TensorBuffers.tensorMemory(TensorBuffers.java:180)
	at org.tensorflow.internal.buffer.TensorBuffers.toFloats(TensorBuffers.java:102)
	at org.tensorflow.internal.types.TFloat32Mapper.mapDense(TFloat32Mapper.java:34)
	at org.tensorflow.internal.types.TFloat32Mapper.mapDense(TFloat32Mapper.java:30)
	at org.tensorflow.RawTensor.asTypedTensor(RawTensor.java:215)
	at org.tensorflow.Tensor.of(Tensor.java:90)
	at org.tensorflow.Tensor.of(Tensor.java:63)
	at org.tensorflow.types.TFloat32.tensorOf(TFloat32.java:79)
	at tester.TensorScopeKt.main(TensorScope.kt:8)
	at tester.TensorScopeKt.main(TensorScope.kt)

I'd think this is a fairly straight forward OOM, just without nice error messages, but the used memory in Task Manager (I'm on Windows) barely moves. I do start using heap memory, but it caps 100MB, even with -Xmx6g. Is there some limit on tensor allocations I'm not aware of?

Commencez par reproduire l’exemple Kotlin via TFloat32.tensorOf et retracez l’échec signalé à org/tensorflow/internal/buffer/TensorBuffers.java:180. Comparez les avertissements d’allocation native avec l’exception côté Java et définissez done comme le fait de signaler les échecs d’allocation avec une erreur précise et exploitable au lieu de la NullPointerException observée.