16 Коммитов

Автор	SHA1	Сообщение	Дата
Nuno Lopes	09c486d625	add basic support for autograd	2021-10-28 13:52:56 +01:00
Nuno Lopes	81fbcdd8ee	fix bug with missing event of 'tensor.data = tensor2' ... this calls TensorImpl::shallow_copy_from which may change an input tensor since we don't freeze input tensors (as we expect to catch all events), nor have shallow copy events added to the trace, we need to flush straight away This makes TorchVision's models work with Cuda	2021-10-19 15:14:38 +01:00
Nuno Lopes	bb0ccea023	fix some assertions	2021-10-14 11:53:35 +01:00
Nuno Lopes	427936015c	stats: fix top frequent traces	2021-09-28 10:28:44 +01:00
Nuno Lopes	5db8fa456d	add caching mechanism for compiled traces	2021-09-24 18:02:01 +01:00
Nuno Lopes	e258320f75	generate dispatch wrappers for unsupported ops ... This is needed otherwise the dispatcher may not go to the device as fallback more info: https://github.com/nunoplopes/torchy/issues/7	2021-09-16 17:24:08 +01:00
Nuno Lopes	a419897bf4	enable torchscript backend by default and fallback to interpreter on failure	2021-09-07 11:58:06 +01:00
Nuno Lopes	c5f085234e	use PyTorch fallback mechanism to handle our pass-through functions ... PyTorch has a more efficient way of handling pass throughs	2021-06-16 14:17:20 +01:00
Nuno Lopes	5842fdfcc2	fix dot output for ...	2021-06-08 17:30:24 +01:00
Nuno Lopes	470834041d	stats: add dot printer	2021-06-02 16:33:21 +01:00
Nuno Lopes	ca22e15166	stats: fix top trace printing	2021-06-01 18:27:46 +01:00
Nuno Lopes	0a9ffd32c5	add some hugging face benchmarks	2021-06-01 17:37:47 +01:00
Nuno Lopes	1c160ef1e5	stats: collect run-time per trace	2021-06-01 12:20:04 +01:00
Nuno Lopes	21a1f9309a	more stats	2021-05-31 18:13:43 +01:00
Nuno Lopes	3d60d8f86f	fix build	2021-05-31 13:08:50 +01:00
Nuno Lopes	dc4ca34e0d	stash stats	2021-05-28 14:30:09 +01:00