A tracing JIT for PyTorch

Перейти к файлу

Nuno Lopes 304b2470d2 add mobilenet inference bench		2021-05-31 18:44:02 +01:00
autogen	special case argmax return value	2021-05-27 18:20:48 +01:00
backends	record TLS data and replay it during interpretation	2021-05-27 16:59:57 +01:00
benchmarks/torchvision	add mobilenet inference bench	2021-05-31 18:44:02 +01:00
scripts	dont override the key set after materializing a tensor	2021-05-20 18:52:14 +01:00
tests	fix deep copy of TensorList	2021-05-18 19:45:56 +01:00
torchy	stubs	2021-03-25 20:18:06 +00:00
.gitattributes	add unit test framework	2021-04-29 17:20:42 +01:00
.gitignore	fix crashes	2021-05-06 14:32:20 +01:00
LICENSE	Initial commit	2021-03-25 20:15:04 +00:00
README.md	make the example work (almost) :D	2021-04-16 18:14:12 +01:00
config.h	stash stats	2021-05-28 14:30:09 +01:00
dispatch.h	add autograd fallback handler	2021-04-27 18:27:03 +01:00
gen.py	special case argmax return value	2021-05-27 18:20:48 +01:00
init.cpp	add tape	2021-03-29 19:22:44 +01:00
ops.cpp	stash	2021-04-26 17:06:06 +01:00
ops.h	sync with pytorch latest git	2021-04-30 13:17:39 +01:00
pytorch.patch	sync with latest pytorch	2021-05-27 14:18:43 +01:00
run-tests.sh	add unit test framework	2021-04-29 17:20:42 +01:00
setup.py	go back to -O3 as it produces smaller binaries and compile time is similar	2021-05-18 18:29:47 +01:00
special_fns.py	special case argmax return value	2021-05-27 18:20:48 +01:00
stats.cpp	more stats	2021-05-31 18:13:43 +01:00
stats.h	more stats	2021-05-31 18:13:43 +01:00
tensor.cpp	fix build	2021-05-31 13:08:50 +01:00
tensor.h	fix bugs with tensor copying	2021-05-11 17:53:42 +01:00
trace.cpp	more stats	2021-05-31 18:13:43 +01:00
trace.h	more stats	2021-05-31 18:13:43 +01:00
utils.h	fix crash	2021-05-27 14:32:52 +01:00

README.md

Torchy

A tracing JIT for PyTorch. WIP; don't use.

Install

$ python setup.py install

Run

Torchy shouldn't require any change beyond adding a call to torchy.enable(). Example:

import torch
import torchy

torchy.enable()

x = torch.tensor(((3.,2.), (4.,5.)))
y = torch.tensor(((5.,6.), (7.,1.)))

w = x.add(x)
z = x.add(y)
w = None  # w not computed
print(z)