George Hotz
c143c92828
adam threestep
2022-06-06 09:38:28 -07:00
George Hotz
c26ed64cef
slightly cleaner graph
2022-06-06 09:31:48 -07:00
George Hotz
d302049e53
don't use div
2022-06-06 09:25:31 -07:00
George Hotz
7229d65235
shortcut nops
2022-06-06 09:15:41 -07:00
George Hotz
b482e1db85
minor cleanups, one less buffer
2022-06-06 08:54:24 -07:00
George Hotz
1ec786960f
fix ci with f
2022-06-06 08:34:05 -07:00
George Hotz
f177cd550f
make binary ops respect needs_input_grad
2022-06-06 08:29:08 -07:00
George Hotz
a1dff4061b
minor cleanups
2022-06-06 08:14:52 -07:00
George Hotz
613f0ca6e5
remove new_buffer
2022-06-06 07:57:39 -07:00
George Hotz
30f55eaaba
rename llops to opencl
2022-06-06 07:49:36 -07:00
George Hotz
233c71a7ba
support requires_grad
2022-06-06 07:47:31 -07:00
George Hotz
9f9cf076c0
comments
2022-06-05 21:51:31 -07:00
George Hotz
a7ea00dae7
remove useless
2022-06-05 21:37:12 -07:00
George Hotz
5526f36f44
all buffer alloc in top level
2022-06-05 21:34:34 -07:00
George Hotz
85e130a0c7
line count
2022-06-05 21:12:02 -07:00
George Hotz
d9e57954d8
factor out unary and binary op allocations
2022-06-05 20:48:36 -07:00
George Hotz
bb7e7b48f9
output edges aren't intermediates either
2022-06-05 17:27:32 -07:00
George Hotz
927a238dc8
better printing for intermediate tensors
2022-06-05 17:19:47 -07:00
George Hotz
3dac8fa728
this fix the gc
2022-06-05 17:16:40 -07:00
George Hotz
0ee21ba115
add ViT test and car
2022-06-05 17:12:43 -07:00
George Hotz
ebd7290f56
simpler batchnorm
2022-06-05 16:51:51 -07:00
George Hotz
30025f2bc7
bn: don't track_running_stats if you aren't training
2022-06-05 16:41:20 -07:00
George Hotz
1de75b67d5
fix bug in graph with use of id
2022-06-05 16:31:20 -07:00
George Hotz
c193088c97
fix bug in purple edges
2022-06-05 16:04:51 -07:00
George Hotz
81c9438ea1
keepdim avoids reshapes
2022-06-05 15:56:42 -07:00
George Hotz
60b8689ea2
code2 was unneeded for reduce op
2022-06-05 15:15:04 -07:00
George Hotz
46aaac5aaf
clean up loops
2022-06-05 15:00:21 -07:00
George Hotz
44168c4c40
remove all ctxes
2022-06-05 14:55:39 -07:00
George Hotz
f01bad36c2
factor out convs
2022-06-05 14:48:42 -07:00
George Hotz
b49bfb6e02
fix conv args. fix spacing
2022-06-05 14:35:31 -07:00
George Hotz
365e62a609
refactor out matmul
2022-06-05 14:33:06 -07:00
George Hotz
11d0cfec77
more readable and faster
2022-06-05 14:13:08 -07:00
George Hotz
d1b6f9822c
UnaryOp class should make linecount
2022-06-05 14:07:02 -07:00
George Hotz
24ed95e667
gpu refactors
2022-06-05 13:55:49 -07:00
George Hotz
7a3fe34db1
GPU llops
2022-06-05 13:49:39 -07:00
George Hotz
f0fe37bd34
simpler graph demo
2022-06-05 12:40:12 -07:00
George Hotz
89acf6742d
more graph docs
2022-06-05 12:16:50 -07:00
George Hotz
88de42fb6e
document graph mode
2022-06-05 12:13:05 -07:00
George Hotz
2097d814f6
Sum doesn't need to save the tensor
2022-06-05 12:04:51 -07:00
George Hotz
c8b569a8c7
cleaner comments
2022-05-14 21:28:39 -07:00
Adrian Bakke
8a7bbcc240
Accel/cuda ( #319 )
...
* start work
* right a wrong
* mimimal reduce example
* save
* get tests to pass
* fix stupid bug
* tests passing
* out of bounds error fixed, all tests passing
* revert changes to other files
* clean
* getting better, will drink a glass of wine and finish over the weekend
* clean
2022-05-14 21:25:30 -07:00
Drew Hintz
7ad60eb8b2
remove redundant list comprehensions ( #317 )
...
any() accepts a generator. This saves a few characters and can also return early without evaluating each item.
2022-04-05 10:30:50 -07:00
namliz
8ce8664024
Golfed -1 ( #315 )
...
* Golfed -1
To please the orange website
* Typo
2022-04-05 10:16:27 -07:00
Edward Betts
a6203b5ecd
Correct spelling mistakes ( #314 )
2022-04-05 05:22:18 -07:00
Andrei Augustin
050636bcb1
here's two extra lines of precious code ( #307 )
2022-02-21 11:45:19 -08:00
Ildar Kashaev
d065040df0
Fix yolo param name ( #306 )
2022-02-10 01:40:24 -07:00
George Hotz
cfb7a4c41a
back to 1000 lines
2022-01-17 10:44:39 -08:00
George Hotz
a482d56229
don't save for backward pass if we don't need to. add purple edges to graph
2022-01-16 12:18:41 -08:00
George Hotz
8dfdc4e11f
GRAPH and DEBUG can be independent. under 1k lines
2022-01-16 12:04:10 -08:00
George Hotz
2f531e35be
add ctx saved tensors to graph
2022-01-16 11:18:36 -08:00