Open
Conversation
On top of the missing links highlighted in #1622, there were a bunch of broken links in the README. This PR makes sure all links lead to actually existing page. In some cases, the links do not lead to the exact section it should, but that's a problem to fix in the documentation in a separate PR. In the meantime, links at least lead to a related page from which the user should be able to find their way.
Per the XDG Base Directory Specification, $XDG_CACHE_HOME already defines the base cache directory (typically ~/.cache). The previous code joined it with ".cache" again, producing a redundant path like ~/.cache/.cache/outlines instead of ~/.cache/outlines. Signed-off-by: JiangNan <[email protected]>
Bug 1 – Attention mask incorrectly masks EOS tokens: LlamaCppTokenizer.encode() used pad_token_id (== eos_token_id) to generate the attention mask. Any EOS token appearing in the middle of a real prompt was masked as padding (attention=0). Since batch encoding is not supported, there is never any real padding, so all tokens should always be attended. Bug 2 – Fallback vocab extraction truncates long token pieces: The fallback path (no HF tokenizer) used a fixed 32-byte ctypes buffer. If llama_token_to_piece reports a piece longer than 32 bytes, the piece was silently truncated. Two tokens that share the first 32 bytes but differ afterwards would collide in the vocabulary dict, losing an entry. Now retry with a correctly-sized buffer when truncation is detected. Also hoists the llama_model_get_vocab call out of the per-token loop to avoid redundant work. Fixes #1819
Cover the retry branch when llama_token_to_piece returns n > buffer size (32 bytes) and verify attention mask is all-ones even when the prompt contains EOS tokens. Signed-off-by: Giulio Leone <[email protected]>
…piece Addresses review feedback from @kudos07: llama_token_to_piece can return negative values as error codes. These are now skipped during vocabulary construction instead of producing invalid token entries. Added test_negative_n_skips_invalid_token to cover this path.
The compound 'or all(...)' clause only differed from the first check when an empty-string key mapped to the error token id, which would itself be a bug. Replaced with the simpler direct check.
Add _ensure_json_quoted() helper that wraps bare String terms in double-quote delimiters when they appear inside container types (List, Tuple, Dict). This ensures Literal/Enum string values produce valid JSON regex patterns. Includes 16 unit tests covering: - String/Alternatives/passthrough quoting - List, Tuple (fixed and variadic), Dict containers - Non-string types unchanged (int, bool) - Nested alternatives, special characters, single-variant literals
Add 14 comprehensive tests that verify the full pipeline (python_types_to_terms → to_regex → re.fullmatch) for: - List/Tuple/Dict with Literal and Enum string values - Mixed string+int Literals (only strings quoted) - Empty string literals - Special characters (spaces, hyphens) - Standalone Literal unaffected (no quotes) - Negative assertions (bare words correctly rejected)
Apply reviewer suggestion to avoid creating a Sequence for a simple
quoted string — String(f'"{{term.value}}"') is cleaner and produces
the same regex output.
All tests now check for String('"...") instead of Sequence wrapping,
matching the simplified implementation suggested by @RobinPicard.
`SPIECE_UNDERLINE` is imported from `transformers.file_utils`, which is an internal module that may be removed or relocated in future versions. Add try/except fallback that inlines the constant (U+2581 "▁") when the import fails, making outlines resilient to transformers refactors. Fixes #1829
Cover the except-ImportError branch in both llamacpp.py and transformers.py convert_token_to_string() by mocking the import to fail.
Co-authored-by: Copilot <[email protected]>
…tring In `create_outlines_core_vocabulary`, the assignment `formatted_vocab[token_as_str] = [token_id]` overwrites previous entries when multiple BPE tokens decode to the same string via `token_to_str`. For Mistral-7B, this silently drops 17.2% of token IDs. Replace with `formatted_vocab.setdefault(token_as_str, []).append(token_id)` to accumulate all token IDs for each decoded string. Fixes #1830 Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]>
- Rename client import from Dottxt to DotTxt - Replace model_name + model_revision with single model identifier - Replace client.json() with client.generate(input=, response_format=) - Add AsyncDottxt class wrapping AsyncDotTxt client - Single from_dottxt dispatching on client type (DotTxt vs AsyncDotTxt) - Add explicit ValueError when model identifier is missing - Export AsyncDottxt, add to AsyncBlackBoxModel
- Update client import to DotTxt/AsyncDotTxt - Replace list_models() with models.list() and model.id - Replace model_name + model_revision fixtures with single model_name - Add async model fixtures and async test coverage - Add missing model validation tests - Fix test for wrong inference parameters to use model with identifier set
- Rewrite dottxt.md to reflect new SDK: single model param, async support, inference arguments, and API key request link - Add Dottxt to API Support row in README - Add API key request link at the top of the README - Add Dottxt API and schema audit call-to-action sections to docs homepage
Removed duplicate early access request link from README.
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
20 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.