Pedro Rodriguez
45bfe94c1e
Broken train reproducing bf16 error
...
Summary:
Test Plan:
2025-02-06 01:27:23 +00:00
Pedro Rodriguez
8f1a9a858e
Minimal working eval
...
Summary:
Test Plan:
2025-02-05 22:47:01 +00:00
Pedro Rodriguez
48cf4dfee1
Allow ArrowIterator to read from json
...
Summary:
Test Plan:
2025-02-05 22:47:01 +00:00
Pedro Rodriguez
2f42633b07
Add bpb and n_bytes to metric logging
...
Summary:
Test Plan:
2025-02-05 22:27:15 +00:00
Pedro Rodriguez
1450464031
Update checkpointing to use fsspec
...
Lint with Black / lint (push) Waiting to run
Lint with isort / lint (push) Waiting to run
Summary:
- Make the data/checkpoint code fsspec compatible
Test Plan:
Run unit tests and the commands below
```
python -m bytelatent.train config=internal/configs/s3_debug.yaml eval=null checkpoint.dump.every=100
```
```
torchrun --nproc-per-node 8 -m bytelatent.train config=internal/configs/s3_debug.yaml eval=null checkpoint.dump.every=100
```
```
python -m bytelatent.train config=internal/configs/s3_debug.yaml eval=null checkpoint.dump.every=100 dump_dir=s3://blt/scratch/checkpoint-test/
```
```
torchrun --nproc-per-node 8 -m bytelatent.train config=internal/configs/s3_debug.yaml eval=null checkpoint.dump.every=100 dump_dir=s3://blt/scratch/checkpoint-test/
```
2025-02-05 19:09:14 +00:00
Pedro Rodriguez
c79b1fdbd0
Fix distributed all reduce grad norm ( #40 )
...
Lint with Black / lint (push) Waiting to run
Lint with isort / lint (push) Waiting to run
Summary:
With >1 GPU, but only 1 node, all reduces fail when inputs are not bf16. This uses a modified copy of torch's grad norm to avoid failures
Test Plan:
- Run unit tests:
- Run single gpu training: `python -m bytelatent.train config=internal/configs/s3_debug.yaml eval=null checkpoint.dump.every=100`
- Run 1 node, multi-gpu training `torchrun --nproc-per-node 8 -m bytelatent.train config=internal/configs/s3_debug.yaml eval=null checkpoint.dump.every=100`
2025-02-04 16:53:50 -08:00
Pedro Rodriguez
7044771a12
This includes fixes that make checkpointing and reloading work correctly. ( #35 )
...
Lint with Black / lint (push) Has been cancelled
Lint with isort / lint (push) Has been cancelled
It also batches in a first set of changes for fixing eval code
Summary:
Test Plan:
2025-01-27 16:56:42 -08:00
Pedro Rodriguez
7622d28b74
Initial codes and scripts for training entropy model ( #34 )
...
Lint with Black / lint (push) Waiting to run
Lint with isort / lint (push) Waiting to run
Summary:
Test Plan:
2025-01-27 09:46:44 -08:00
Pedro Rodriguez
a809259e71
Use load_async flag to not start MP iterator ( #33 )
...
Lint with Black / lint (push) Has been cancelled
Lint with isort / lint (push) Has been cancelled
Summary:
Test Plan:
2025-01-24 10:57:20 -08:00
Pedro Rodriguez
bc42cebd7d
Update file check script to check sizes ( #32 )
...
Lint with isort / lint (push) Has been cancelled
Lint with Black / lint (push) Has been cancelled
Summary:
Test Plan:
2025-01-22 13:06:46 -08:00
Ink
392117bff2
Fix realtime entropy patching ( #26 )
...
Lint with Black / lint (push) Waiting to run
Lint with isort / lint (push) Waiting to run
* allow loading of the entropy model directly
* remove unused argument
* remove spammy warning
* allow patch_batch_size to be adjusted in the forward() method
* revert to original patcher style, fix warning
* allow grads when calculating entropies
* fix grad flow
* return preds from calculate_entropies()
* remove legacy arg
* fix an error with monotonicity and small sequence lengths
* ensure patcher is serializable
* revert patcher to original
* remove unused import
2025-01-21 16:34:23 -08:00
Pedro Rodriguez
6ffeb66b53
Changes for training entropy model and correcting attention in local models ( #25 )
...
Lint with Black / lint (push) Has been cancelled
Lint with isort / lint (push) Has been cancelled
Summary:
- Refactor local model configs to be separate and clearer
- Add attention arguments and correct which attention is used in local models
- Preparation for being able to have an entropy train script
- Fix failing unit tests
Test Plan:
2025-01-17 14:23:01 -08:00
Ink
caec8d2621
allow flex-attention to be disabled ( #19 )
...
Lint with Black / lint (push) Has been cancelled
Lint with isort / lint (push) Has been cancelled
* allow flex-attention to silently fail
* allow flex-attn to be disabled via an env var
2025-01-14 09:32:07 -08:00
Pedro Rodriguez
1da3dd9315
Update preprocess_entropies script to blt inference + add fsspec support ( #23 )
...
Lint with Black / lint (push) Waiting to run
Lint with isort / lint (push) Waiting to run
Summary:
Test Plan:
2025-01-13 15:28:14 -08:00
Pedro Rodriguez
b0120da72f
Replace regular filesystem calls with fsspec + add s3 support ( #18 )
...
Lint with Black / lint (push) Has been cancelled
Lint with isort / lint (push) Has been cancelled
Summary:
For compatibility with either local/nfs or S3 datasets, swap to fsspec.
Add a tool to compare local and remote filesystems
Test Plan:
- Ran regular train script
- Ran with config with data in S3
2025-01-10 11:04:41 -08:00
Pedro Rodriguez
d4ddb95322
Add plotting code from paper ( #17 )
...
Lint with Black / lint (push) Waiting to run
Lint with isort / lint (push) Waiting to run
Summary:
Test Plan:
2025-01-09 12:11:50 -08:00
Ink
2fdc6f3cc9
Package bytelatent as a module ( #7 )
...
Lint with Black / lint (push) Has been cancelled
Lint with isort / lint (push) Has been cancelled
* make installable via pip
* fix missing xformers deps
* remove non-core dependencies
* fix linting
* fix isort
2025-01-06 16:44:50 -08:00
Ikko Eltociear Ashimine
9065bb1cce
docs: update README.md ( #1 )
...
Lint with Black / lint (push) Has been cancelled
Lint with isort / lint (push) Has been cancelled
folowing -> following
2025-01-03 12:08:00 -08:00
Daniele Sartiano
898671b66b
Update README.md ( #13 )
...
Fixed typo on Meta Lingua
2025-01-03 12:06:47 -08:00
Pedro Rodriguez
bcc039bb75
Initial commit
2024-12-12 15:32:30 -08:00
