dataloader checkpoint callback - #60
Merged
Merged
Conversation
taylormjs
reviewed
Apr 11, 2025
| if self._is_s3_uri: | ||
| with tempfile.NamedTemporaryFile() as tmp_file: | ||
| temp_path = tmp_file.name | ||
| torch.save(dataloader.state_dict(), temp_path) |
Collaborator
There was a problem hiding this comment.
Just curious, about how large are the dataloader state dicts?
Collaborator
Author
There was a problem hiding this comment.
I don't have an example on hand but I think they are pretty small since it just stores the item indices and some metadata about input_dir etc
Collaborator
Author
There was a problem hiding this comment.
see here: https://github.com/Lightning-AI/litData/blob/06353697d17bd71f1bd1af3e47c5c3d5760a2e93/src/litdata/streaming/dataloader.py#L680
and here:
https://github.com/Lightning-AI/litData/blob/ee033833589ce82119892a1640fa0e3864061678/src/litdata/streaming/dataset.py#L438
taylormjs
reviewed
Apr 11, 2025
|
|
||
| # Instantiate the model | ||
| if ckpt_path is not None: | ||
| self.model = FlexBERT.load_from_checkpoint(ckpt_path) |
Collaborator
There was a problem hiding this comment.
Good catch. So by default, always use Ume.load_from_checkpoint rather than specify a ckpt_path in model instantiation?
Collaborator
Author
There was a problem hiding this comment.
yes exactly, ckpt_path in the model parameters is only needed because of this line: https://github.com/prescient-design/lobster/blob/main/src/lobster/cmdline/_train.py#L54
taylormjs
pushed a commit
that referenced
this pull request
Apr 29, 2025
* dataloader callback * utils * ume * gitignore dev * tests
karinazad
added a commit
that referenced
this pull request
May 14, 2025
* peer fixes, add evaluate method * dataloader checkpoint callback (#60) * dataloader callback * utils * ume * gitignore dev * tests * update flash attention wheels (#61) * lock * torch 2.5 * torch 2.5 * part * .env * unpin flash attn (#62) * fix scheduler params (#64) * scheduler * fix scheduler * fix scheduler * Add AtomicaDataset (#63) Processed Atomica interactions dataset * Ume conversion/interaction tokenizer + fix SMILES and nucleotide tokenizers (#65) add two special tokens: <convert> and <interact> for later stages of Ume training: will be used as this: (or something like that) [CLS] PROT_SEQ [SEP] <convert> PROT_STRUCT(masked) [SEP] [CLS] PROT_SEQ [SEP] <interact> SMILES(masked) [SEP] extend functionality of UmeTokenizerTransform to handle dual modalities change the name of Ume embedding method and allow embedding from existing input_ids fix existing tokenizers: add lowercase normalized to nucleotide tokenizer (OG2 dataset contains a mix of upper and lowercase letters) BPE handled SMILES tokenization incorrectly, switch to WordLevel * Ume SMILES tokenizer fix (#66) * tokenizer * fix tests * lowercase normalizer for nt * tests * remove mod conv dataset * embed * Test * merge 2mod into UmeTokenizerTransform * fix tests * all * type hints * docstrings * tests * fix SMILES tokenizer * switch all tokenizer to BPE * Revert "switch all tokenizer to BPE" This reverts commit 367e77d. * tok * fix SMILES tokenizer * remove print statement * Ume perplexity logging (#67) * pplx * tests * src * ignore torchmetrics warnings * docstrings * docstrings * Update README.md (#69) * Ume fix perplexity device (#68) * pplx as attr * pplx as attr * pplx * comments * on step * comment * update tests, fix ruff * ruff * ruff ruff * Add <cls_modality> to Ume tokenizers (#71) * add <cls_modality> tokens * add <cls_modality> tokens * docstring * RNS metric implementation (#73) * add <cls_modality> tokens * add <cls_modality> tokens * modality embeddings * module dict * embeddings * tests * modality and device * rank zero only * rank zero * fix back modality mask * sync dist * RNS implementation * restore from main * restore * docstrings * docstrings * review * test * Ume modality-specific embeddings (#72) * add <cls_modality> tokens * add <cls_modality> tokens * modality embeddings * module dict * embeddings * tests * modality and device * rank zero only * rank zero * fix back modality mask * sync dist * add conversion transforms (#74) * add initial smiles to peptide and peptide to smiles transforms * remove smiles -> * transforms and touch up conversion functions * rename * add option to randomize smiles and caps --------- Co-authored-by: Colin Grambow <grambowc@gene.com> * fix def pad token, replace process_and_embed w/ ume.embed * update tests w -100 pad token --------- Co-authored-by: Taylor Joren <joren.taylor@gene.com> Co-authored-by: Karina Zadorozhny <karina.zadorozhny@gmail.com> Co-authored-by: Nathan Frey <ncfrey@users.noreply.github.com> Co-authored-by: Colin Grambow <17198155+cgrambow@users.noreply.github.com> Co-authored-by: Colin Grambow <grambowc@gene.com>
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
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.
No description provided.