6bb92403d5
changes and optimizations to muon, making it more efficient and simpler/cleaner a bit
2026-01-15 03:20:48 +00:00
3142ca1a28
minor helpful message
2026-01-15 03:20:21 +00:00
3b50b77ed3
fix base_loss to report correct loss by switching the dataloader to the new default
2026-01-13 22:09:36 +00:00
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
23985413aa
adjust the comment on the regex pattern per recent experimnet see dev/LOG.md
2026-01-13 17:50:39 +00:00
21608ec51e
allow base_loss to report the loss of any arbitrary huggingface model similar to base_eval. had to change dataloader to be a lot better and just take tokenizer, not load the nanochat one. much better this way anyway
2026-01-12 03:10:13 +00:00
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
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
201d705957
recover the ability to load old checkpoints by patching the lambdas if they don't exist in checkpoints
2026-01-11 20:13:12 +00:00
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
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
4ddc803797
fix adamw slight bug. this chunk was copy pasted originally from modded-nanogpt, which still seems to have the bug
2026-01-08 18:18:42 +00:00
ccf4b7f9bf
nudge hyperparameters of the base script with the results of the sweeps and miniseries. vocab size down to 32K. D:N ratio from 20 to 8. add miniseries script
2026-01-07 22:11:59 +00:00
962b6bfba3
alright add transformers as a dep of the repo because it should be easy to evaluate the CORE score of HF models. Not super happy about it but i tried it and the uv.lock doesn't get bloated as much as i expected
2026-01-04 20:37:28 +00:00
eb7bbc1b66
delete the configurator in favor of argparse and clean up a lot of kwarg details to make them more consistent across all scripts
2026-01-04 19:14:23 +00:00
507d54224a
fix small bug where this would break if git stage has deleted files
2026-01-04 19:11:43 +00:00
be56d29b87
simplify redundant if/elif in bloat metrics
...
🤖 Generated with [Claude Code](https://claude.com/claude-code )
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com >
2026-01-04 01:40:42 +00:00
ee79f29fbd
replace files-to-prompt with git ls-files for bloat metrics
...
files-to-prompt was including untracked files (knowledge/, dev scripts, etc.) which inflated the bloat metrics. now we use git ls-files to only count tracked source files, which is more accurate and removes an external dependency.
🤖 Generated with [Claude Code](https://claude.com/claude-code )
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com >
2026-01-04 01:38:15 +00:00
48abd7d85f
simplify, clarify and slightly tune model initialization. should be very slightly better possibly, but certainly a lot clearer
2026-01-01 21:15:09 +00:00
10231dfb40
Fix conversation scroll to bottom on some browsers + remove duplicated padding ( #348 )
2025-12-31 13:03:22 -08:00
8f979a8bda
fix: sample first token independently for each row in multi-sample generation
...
Previously, when generating multiple samples (num_samples > 1), the first
token after prefill was sampled once and broadcast to all rows, causing
all samples to start identically. Now the prefill logits are expanded to
num_samples and sampled independently for each row.
Also simplified the generation loop by moving the forward pass to the end
of the loop, eliminating the first_iteration flag and if/else branching.
🤖 Generated with [Claude Code](https://claude.com/claude-code )
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com >
2025-12-28 04:52:13 +00:00
2f2d7ab80c
fix: safe DDP cleanup (check initialized PG, not just env) ( #256 )
2025-12-27 20:27:40 -08:00
e1770a3061
remove spurious cast, gets compiled away anyway but it's confusing people
2025-12-27 23:07:48 +00:00
49389ecaa8
fix tf32 warning for deprecated api use
2025-12-27 22:03:06 +00:00
d314e96aa2
formatting
2025-12-09 12:48:46 +01:00
bbc57da7d5
slightly nicer error message
2025-12-09 12:46:48 +01:00
f1bf69d562
feat: pad vocab size to 64 for DDP optimizers and efficiency
2025-12-09 12:38:18 +01:00
7931e0903a
rename checkpoint_dir to checkpoints_dir for consistency.
2025-12-08 18:32:12 -08:00
849d95ae1f
remove unnecessary check to make the logic in CausalSelfAttention.forward() clearer
2025-12-08 18:30:37 -08:00
1b2a675c88
Improve KV cache code readability
2025-12-08 18:19:05 -08:00
72a7cf2bc4
Fix distributed Parquet dataloader resume for multi-epoch training
2025-12-08 18:15:02 -08:00
bffdb2ef91
group common code to make things neater in gpt logit computation
2025-12-09 02:01:05 +00:00
cbf30c842c
apply float32 cast before logits softcapping so the tanh is in fp32. torch compile fuses this correctly with no extra memory costs.
2025-12-08 14:17:43 -08:00
90442de35f
fix bug where any rank has to be able to create checkpoint_dir if saving optim
2025-12-08 20:45:19 +00:00
01ea71be39
Fix distributed Parquet dataloader resume for multi-epoch training
2025-12-08 00:10:19 -06:00
06677c30e0
Refactor dimension validation for KV cache
2025-11-28 15:22:18 -05:00
a770dcef2e
Fix kv_cache indexing to explicitly include head dimension
2025-11-28 15:00:14 -05:00
16788eed3c
fix(model): apply float32 cast before logits softcapping
...
This change ensures that the logits softcapping operation (tanh) is performed in float32 precision rather than bfloat16. Previously, the code cast to float32 after the tanh operation, which meant the non-linearity was computed with bfloat16 precision
2025-11-23 20:12:09 +05:30
5c93a56be5
remove unnecessary check
2025-11-19 16:31:41 -05:00
a4a0959c73
renamed find_largest_model() argument checkpoint_dir to checkpoints_dir for clarity
2025-11-19 15:33:36 -05:00
11e68bf442
Fix comment: rotary embeddings final dimension size
2025-11-17 11:32:56 -05:00
bc1fca39f3
mqa -> gqa to reduce confusion
2025-11-15 15:43:37 +00:00
f66a780f68
Fix torch.dtype mismatching when running engine inline test.
2025-11-14 07:28:29 -08:00
4763ce612a
Small fixes to typos
2025-11-14 07:25:59 -08:00
c6f5bd67db
revert change of base to sft for quick inline test
2025-11-14 12:20:03 +01:00
a2fb3c83a6
fix typos
2025-11-14 11:20:25 +01:00
c6abcdfe3a
big change: add pretraining resumption logic so that checkpoints can now be approximately resumed and training can continue. this is useful for very long runs when you don't want the anxiety of your run crashing for some reason. alternatively, it's a way to recover training in the event of loss spikes. i mean, this should have been there in v0 but it's ok. the resumption is approximate to control complexity and bloat, but it's possible we want to change that in the future. to use, set --save_every to a step interval to write checkpoints with, and then use --resume_from_step to resume optimization from a given step. only base model training (pretraining) supports this atm, but it's ok because midtraining is comparably quite a bit faster.
2025-11-13 15:34:40 +00:00
91f09ccd0d
minor fix comment in engine
2025-11-13 15:28:18 +00:00
b399e43168
fix engine test bug
2025-11-06 08:56:45 +08:00
3a2ae631c4
Merge branch 'master' into master
2025-11-04 16:35:02 -08:00
Andrej Karpathy