* Add draft model param to llama class, implement basic prompt lookup decoding draft model
* Use samplingcontext for sampling
* Use 1d array
* Use draft model for sampling
* Fix dumb mistake
* Allow for later extensions to the LlamaDraftModel api
* Cleanup
* Adaptive candidate prediction
* Update implementation to match hf transformers
* Tuning
* Fix bug where last token was not used for ngram prediction
* Remove heuristic for num_pred_tokens (no benefit)
* fix: n_candidates bug.
* Add draft_model_num_pred_tokens server setting
* Cleanup
* Update README
* Added mistral instruct chat format as "mistral"
* Fix stop sequence (merge issue)
* Update chat format name to `mistral-instruct`
---------
Co-authored-by: Andrei <abetlen@gmail.com>
* feat: Add support for jinja templating
Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com>
* fix: Refactor chat formatter and update interface for jinja templates
- Simplify the `llama2_template` in `llama_jinja_format.py` by removing unnecessary line breaks for readability without affecting functionality.
- Update `ChatFormatterInterface` constructor to accept a more generic `Optional[object]` type for the template parameter, enhancing flexibility.
- Introduce a `template` property to `ChatFormatterInterface` for standardized access to the template string.
- Replace `MetaSingleton` metaclass with `Singleton` for the `ChatFormatterFactory` to streamline the singleton implementation.
These changes enhance code readability, maintain usability, and ensure consistency in the chat formatter's design pattern usage.
* Add outline for Jinja2 templating integration documentation
Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com>
* Add jinja2 as a dependency with version range for Hugging Face transformers compatibility
Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com>
* Update jinja2 version constraint for mkdocs-material compatibility
Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com>
* Fix attribute name in AutoChatFormatter
- Changed attribute name from `self._renderer` to `self._environment`
---------
Signed-off-by: teleprint-me <77757836+teleprint-me@users.noreply.github.com>
The existing code often causes "LookupError: unknown encoding: ascii" when open() called in a destructor. Saving open in self.open is not enough to avoid this. Instead, we can avoid reopening /dev/null every time by doing it once when the module is loaded.
* Support defaulting to infinity or -1 for chat completions
* Check if completion_tokens is none in error handler.
* fix: max_tokens in create completion should match openai spec
* Fix __call__
---------
Co-authored-by: Andrei Betlen <abetlen@gmail.com>
F16_KV appears to have been removed here: af99c6fbfc
This addresses two issues:
- #995 which just requests to add the KV cache offloading param
- #1006 a NULL ptr exception when using the embeddings (introduced by
leaving f16_kv in the fields struct)
If the n_ctx is set to 0 the code should use the maximum context length of the selected model, but it didn't work. There was a problem with the initialization of this parameter and a related problem with 'n_batch'.
See #990. This change makes the logits_to_logprobs function equivalent to the version in the llama.cpp repository. It uses numpy so it's much faster than the previous version.
* bug fixing
* llava from readme got this error: UnsupportedOperation: fileno quick fix by checking hasattr
* multi modal params fix: add logits = True -> to make llava work
* multi modal params fix: add logits = True -> to make llava work
---------
Co-authored-by: Andrei <abetlen@gmail.com>
