Andrej Karpathy
f92efce169
add negative result about not allowing attention across BOS tokens. A lot more code complexity for basically no gain in performance
2026-01-13 21:33:54 +00:00
Andrej Karpathy
43c29dd9d5
Big DataLoader refactor: BOS-aligned dataloaders with epoch tracking for pre/mid-training
...
The new DataLoader ensures that every token sequence in train/val batches has a BOS token
at the beginning. Therefore, no token streams start abruptly in the middle of a document,
which could be confusing for the model. Note that this changes the loss scale because there
are fewer confusing tokens in the train/val batches. The main downside is that we now waste
about 35% of tokens due to cropping. This is ok because we have a lot of data. See dev/LOG.md
entry for this change for a lot more information.
2026-01-13 20:05:47 +00:00
Andrej Karpathy
64b48d0e5c
validated that \p{N}{1,2} is the correct number of digits to group up to in the regex pattern of the GPT-4 tokenizer (2 down from 3), leading to the best val_bpb for 32K vocabs
2026-01-13 17:45:06 +00:00
Andrej Karpathy
238353c998
document my struggle with fp8 integration yesterday, it's not working like i thought it would and i suffered. one day i will return to continue the fight.
2026-01-13 17:14:29 +00:00
Andrej Karpathy
4610a838a1
record negative result on MTP
2026-01-12 05:23:47 +00:00
Andrej Karpathy
fbc1484e8c
add alternating window size patterns for the GPT layers, following GPT-3. Experimented a bit and found the pattern SSSL to work well - 3 short, 1 long alternating. This is now the new default and the plots look quite a bit better on flops vs. bpb
2026-01-11 21:49:54 +00:00
Andrej Karpathy
2ff7d51252
integrate Flash Attention 3. +9% tok_per_sec for d12 with ctx even as low as 2048 out of the box nice. also, ready to tune windows huge
2026-01-11 20:33:19 +00:00
Andrej Karpathy
aa530cdad5
Add learnable lambdas that gate the residual connection and a skip connection to the input embeddings, solid bump to val_bpb
2026-01-11 18:47:35 +00:00
Andrej Karpathy
2c4473dd1b
Big Muon optimizer changes inspired by latest of modded-nanogpt. Added Polar Express, Adafactor-style variance reduction, cautious weight decay, schedule weight decay linearly to ramp down to zero. Tuned optimum weight decay for multiple model sizes d8, d12, d16, d20 and found a scaling law with optimum wd \propto 1/channels^2, including it as default into code. --weight_decay of base_train is now default on and configured optimally according to all of these experiments. Solid bump to val_bpb observed as a result of these changes.
2026-01-11 16:56:59 +00:00
Andrej Karpathy
061f83c152
delete grad_clip. appears to not be necessary at all. not only was it buggy because the clipping happened per gpu before grad synchronization, but it costs ~2% MFU, and it also doesn't even help. I tried deleting it a while ago and back then it did help. So I'm guessing that some hyperparameter tuning obviated the reason for it since then
2026-01-08 02:16:50 +00:00
