251421102d
cleanup; these mallocs are no longer needed
Erik Schultheis
2024-07-18 20:00:51 +02:00
bfce9bdbcf
bugfixes
Erik Schultheis
2024-07-18 19:35:37 +02:00
0bcb3e5840
allocate_weights utility function reverted some earlier changes + more error checking
Erik Schultheis
2024-07-18 19:29:49 +02:00
cde79ffb49
allocate_state utility function
Erik Schultheis
2024-07-18 19:19:06 +02:00
5c89416adf
also allocate gradient eagerly moved helper buffer allocation to forward, so backward now no longer does any memory allocation
Erik Schultheis
2024-07-18 18:54:51 +02:00
72698a5ba3
update gpt3 125M script to be better. should exactly reproduce gpt-3 125M, except we don't use banded attention and we use fineweb
Andrej Karpathy
2024-07-18 16:43:14 +00:00
b20def92d4
consolidate model init
Erik Schultheis
2024-07-18 17:23:06 +02:00
cbe04c3f92
moved dtype utilities to cuda_utils
Erik Schultheis
2024-07-18 17:13:38 +02:00
85d17f4aa2
Merge pull request #668 from abhilash1910/tpc
Andrej
2024-07-15 18:22:15 -07:00
835bb2e6d7
delete segmentation fault when B < 4 atm
Andrej Karpathy
2024-07-16 01:20:16 +00:00
88514df231
Merge pull request #689 from karpathy/refactor/code-to-zerocuh
Andrej
2024-07-15 12:51:01 -07:00
a6dc64533a
Merge pull request #688 from karpathy/feature/gpt3v1
Andrej
2024-07-15 11:53:05 -07:00
220125f39c
ok let's allow T < max_seq_len because our tests use it. make warning very prominent
feature/gpt3v1
Andrej Karpathy
2024-07-15 18:21:34 +00:00
faf8d159a3
use int instead of bool
Andrej Karpathy
2024-07-15 18:01:01 +00:00
528b1b0e31
fix fat fingers
Andrej Karpathy
2024-07-15 17:56:41 +00:00
039904ee9f
revert legacy behavior in .sh scripts for -e option
Andrej Karpathy
2024-07-15 17:55:26 +00:00
64db3ded85
refactor gpt2 gpt3 descriptor to keep legacy behavior of -e, be strict everywhere, remove interpolation for now
Andrej Karpathy
2024-07-15 17:51:27 +00:00
1ae3243185
Add check versions of socket close
Aleksa Gordic
2024-06-25 14:20:27 +00:00
c6efd90a86
upload good 1558M sh script
Andrej Karpathy
2024-07-02 16:49:09 +00:00
39270cc69a
Merge pull request #665 from ngc92/zero-grad
Andrej
2024-07-01 16:08:07 -07:00
486f98e3cc
zero-grad is async and part of backward call
Erik Schultheis
2024-07-02 00:47:41 +02:00
79223b950b
Merge branch 'ademeure-optimize_attproj_fcproj'
Andrej Karpathy
2024-07-01 22:07:24 +00:00
cee396ab5a
resolve merge conflict
Andrej Karpathy
2024-07-01 22:07:14 +00:00
48fee049cb
add edu fineweb sh download script, change paths to my HF repo, add hellaswag to starter pack
Andrej Karpathy
2024-07-01 21:41:15 +00:00
a876282eb8
Merge pull request #653 from ademeure/cublaslt_refactor
Andrej
2024-07-01 11:53:02 -07:00
00247dca9b
Disable GELU fusion by deffault & fix potential argv[] out of bounds access
ademeure
2024-07-01 18:40:21 +00:00
2e6a397d5f
Merge pull request #663 from karpathy/fix/reshuffle_zero_grad
Andrej
2024-07-01 10:39:23 -07:00
7b92930021
move things around a bit, a few comments, and also update python script to correspond to C more
fix/reshuffle_zero_grad
Andrej Karpathy
2024-07-01 17:20:46 +00:00
942fed59fe
Merge pull request #662 from karpathy/overfit-batch-actually
Andrej
2024-07-01 10:12:20 -07:00
7fd7717816
fix the overfit single batch behavior to actually overfit batch, not microbatch, we do this more cleanly by simply resetting the dataloader every step
overfit-batch-actually
Andrej Karpathy
2024-07-01 17:00:40 +00:00