Add bpb and n_bytes to metric logging #41
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
facebook-github-bot
added
the
CLA Signed
sriniiyer
reviewed
Feb 5, 2025
bytelatent/metrics.py
Outdated
| if self.fs is None: | ||
| self.jsonl_writer = open(self.outdir, "a") | ||
| else: | ||
| self.jsonl_writer = self.fs.open(self.outdir, "a") |
There was a problem hiding this comment.
What functionality does this add?
There was a problem hiding this comment.
Eventually (once we fix torch.distributed.checkpointing.save to work with fsspec), all of dump_dir should be compatible with writing to nfs or s3/blob store.
sriniiyer
reviewed
Feb 5, 2025
bytelatent/train.py
Outdated
| @@ -403,6 +419,24 @@ def train(args: TrainArgs): | |||
| batch_patch_lengths = torch.from_numpy(batch.patch_lengths).cuda() | |||
| mask = None if batch.mask is None else torch.from_numpy(batch.mask).cuda() | |||
|
|
|||
| if args.data.tokenizer_args.name in ["bytes", "blt"]: | |||
| if mask is None: | |||
| n_bytes += batch_y.numel() | |||
There was a problem hiding this comment.
n_bytes += mask.sum() if mask else batch_y.numel()
sriniiyer
reviewed
Feb 5, 2025
bytelatent/train.py
Outdated
| f" grad: {grad_norm:.2e}" | ||
| f" flops: {FLOPS:.2e}" | ||
| f" wps: {wps:.2e}" | ||
| f" iter: {curr_iter_time:>7}" | ||
| f" data: {data_load_time:>5}" | ||
| f" lr: {curr_lr:.2e}" | ||
| f" n_bytes={total_n_bytes}" |
There was a problem hiding this comment.
this is confusing, we need a way to indicate per_gpu vs. across all gpus
sriniiyer
reviewed
Feb 5, 2025
bytelatent/train.py
Outdated
| logger.info( | ||
| f"step: {train_state.step}" | ||
| f" acc: {train_state.acc_step}" | ||
| f" loss: {round(loss.item(),4):>7}" | ||
| f" loss: step={round(loss.item(),4):>7} avg={avg_loss}" | ||
| f" bpb: {avg_bpb:3f}" |
There was a problem hiding this comment.
This is inconsistent, we normally report bpb per gpu
This was referenced Feb 6, 2025
Draft
EntilZha
force-pushed
the
pr41
branch
2 times, most recently
from
4e2ed0a to
b6396eb
Compare
sriniiyer
reviewed
Feb 7, 2025
bytelatent/train.py
Outdated
| logger.info( | ||
| f"step: {train_state.step}" | ||
| f" acc: {train_state.acc_step}" | ||
| f" loss: {round(loss.item(),4):>7}" | ||
| f" loss: [step_local={round(step_loss_per_gpu, 4):>7} interval_local={round(interval_loss_per_gpu, 4):>7} step_global={round(step_loss_across_gpus, 4):>7} interval_global={round(interval_loss_across_gpus, 4):>7}]" |
There was a problem hiding this comment.
this is too many things - I think interval local and interval global is good enough
Summary: Test Plan:
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
Test Plan: