allow for empty before_ids and after_ids#8
Merged
Conversation
Collaborator
|
this is great, thanks you! |
justinwangx
pushed a commit
that referenced
this pull request
Feb 3, 2025
* chore: housekeeping * chore: linter/formatter * chore: more linter/formatter + default model * feat: probe sampling for nanoGCG - plumbing (#3) * feat: probe sampling - checkpointing * a bit more cleanups * feat: probe sampling for nanoGCG - parallelization (#4) Major thing is the parallelization. It's actually relatively straightforward, what took the most time was actually debugging the cuda device-side assertions. * feat: allow retrying (#5) During probe sampling, quite a noticeable bunch of iterations don't actually bring down the losses. Adding a retry functionality to hopefully tackle the issue. * fix: correct condition of buffer.size == 0 * docs: Update README.md * docs: wordsmithing on README * debug: try using another pad token * fix: correct `optim_ids` assignment (#6) * A trivial that unfortunately blocked me for a couple of days. * Symptom: When probe sampling is enabled, the losses don't seem to be optimized at all. * Debugging: Manually tweaked `R` such that probe sampling effectively looks at all `B` candidates, the issue still persists. * Approach: In the end a silly mistake was discovered during the `current_loss, optim_ids` assignment stage where in probe sampling, the indices in `sampled_ids` don't match the actual calculated optimal candidate. * Fixed by making probe sampling function return the optimal candidate. * chore: undo retry feature (#7) Thanks to psyclaudeZ#6, the retry mechanism is probably not helpful. Hence reverting. * chore: code cleanup (#8) * refactor: better capture draft mode with a dedicated config * docs: comments * perf: GPU memory cleanup (#9) * perf: actually might not need the cleanup * perf: no_grad for draft loss calculation (#10) * docs: revert changes to README * address feedback: excessive logger.debug, linter, simply.py * address feedback: pad tokens, deps * chore: lowerbound for transformers dep
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.
Currently if the chat template does not have prefix/suffix text to add, the run will fail because torch defaults to float32 dtype for empty tensors, meaning that the embedding lookup in gcg.py:223 will cause an error. This is useful for models that do not have prefixes/suffixes in their templates (e.g., a base model).