Move from torch.cuda.amp to torch.autocast; Add tests for amp #838
Conversation
…IR-Chem/fairchem into move_to_autocast_general_and_amp_test
| if node_energy.device.type == "cuda": | ||
| energy.index_add_(0, data.batch, node_energy.view(-1)) | ||
| else: | ||
| energy.index_add_(0, data.batch, node_energy.float().view(-1)) |
There was a problem hiding this comment.
do we loose a lot of performance or use up too much memory when casting? if not then should we cast regardless?
…en-Catalyst-Project/ocp into move_to_autocast_general_and_amp_test
…en-Catalyst-Project/ocp into move_to_autocast_general_and_amp_test
…en-Catalyst-Project/ocp into move_to_autocast_general_and_amp_test
…IR-Chem/fairchem into move_to_autocast_general_and_amp_test
| @@ -36,6 +36,8 @@ def __init__( | |||
| precon=None, | |||
| cpu=False, | |||
| batch_size=4, | |||
| seed=0, # set a seed for reproducibility | |||
| amp=None, | |||
There was a problem hiding this comment.
why not just set amp=True as the default. From the lines bellow it looks like thats the case
| @@ -110,11 +112,13 @@ def __init__( | |||
| local_rank=config.get("local_rank", 0), | |||
| is_debug=config.get("is_debug", True), | |||
| cpu=cpu, | |||
| amp=True, | |||
| amp=(amp==None or amp), # AMP on by default | |||
There was a problem hiding this comment.
nit, its more pythonic to check None with amp is None (its a "singleton")
| if self.lmax > 0: | ||
| num_m_components = (self.lmax + 1) ** 2 | ||
| feature = node_input.narrow(1, 1, num_m_components - 1) | ||
| with torch.autocast(device_type=node_input.device.type, enabled=False): |
There was a problem hiding this comment.
Is it possible to double up on autocast decorators like the forward method above?
I am just worried about potential incorrect indentation bugs in the future.
| out = self._forward(batch) | ||
| out = {k: v.float() for k, v in out.items()} |
There was a problem hiding this comment.
Did we decide to always cast to float32 on predictions? Thats sounds ok with me, just making sure because this is different than before.
This PR updates torch.cuda.amp.autocast(args...) and torch.cuda.amp.GradScaler(args...) (deprecated) and also adds CPU AMP tests. ( https://pytorch.org/docs/stable/amp.html )
Recently there was an eSCN model that did not run on GPU AMP and the current set of tests did not catch it. This PR adds those equivalent tests on CPU AMP, which will test this going forward.
Additional fixes haven been made to eSCN / SCN / Gemnet OC in this PR.