Otherwise, on many machines, the size of the OpenMP thread pool will
change between MKL and our OpenMP enabled functions. The constant thread
creation and destruction results in worse performance and leaks memory
on GCC 5.4
* A pile of misc doc fixes.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
* Handle @apaszke review comments.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
* Initial csrc documentation.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Because of this Variables can no longer appear in the graph.
Every usage of a leaf Variable will leave an AccumulateGrad
function that has no outputs, but modifies var.grad as a side
effect.
Fixes#1267
This fixes a number of issues when PyTorch was compiled with CUDA
support but run on a machine without any GPUs. Now, we treat all errors
from cudaGetDeviceCount() as if the machine has no devices.
The core autograd Variable, Function, and Engine no longer depend on the
Python API. This let's us implement functions in C++. In the future, we
can also multithread engine and release the GIL for most of the
non-Python backwards.
- Non differentiable outputs could prevent a gradient computation (see
test_dep_nograd)
- Crash in backward on variable which doesn't requires_grad (issue
#438)
- Stochastic functions could be backproped through multiple times
Only references to their data and version counters are stored.
Also, it is now possible to have None arguments in save_for_backward
and return too many values from backward (as long as the excessive
results are None).
