Add support for chat templates (#408)

* Add basic support for chat templates

* Cleanup

* JSDoc improvements

* Support conversion of user-defined functions

* Cleanup

* Fix function creation

* Add unit tests for templates

* Cleanup

* Improve JSDoc

* Add missing return types

* Add chat templates docs to table of contents

* Add support for logical negation

* Fix nested logical negation

* Add unit tests for logical operators

* Add loop variables

* Add support for `RuntimeValue` built-in functions

* Add unit tests for string instance methods

* Fix conversion of normal function to `FunctionValue`

* Update object method unit tests

* Save chat template to tokenizer_config.json during conversion

* Fix `raise_exception` error

* Add `!=` operator for booleans

* Remember to increment loop index

* Cleanup for loop evaluator

* Use `is` helper function

* Add support for text nodes

i.e., non Jinja statements/expressions

* Add auto-generated templating tests

* Update unit tests

* Remove unused function

* Add default chat templates

* Use repo with up-to-date tokenizer config

* Temporarily disable zephyr test

* Delete templates.test.js

* Move Jinja functionality to `@huggingface/jinja`

* Fix template cache type

* Update chat template unit tests

* Update `@huggingface/jinja` version

* Fix default llama2 system prompt usage

* Add unit test for llama2 w/o chat template set

* Update jinja version

* Update jinja version

* Add unit test for user-defined chat templates

Example from https://discuss.huggingface.co/t/issue-with-llama-2-chat-template-and-out-of-date-documentation/61645/3

* Add `AddedToken` for improved tokenization

* Add example usage for chat templates

* Add 'first' Metaspace pretokenizer prepend scheme

* Formatting

* Update wav2vec2 converter special tokens whitespace split

* Fix Metaspace pretokenizer split criteria

* Update inputs of `PreTokenizerSequence`

* Improve Metaspace pretokenizer

* Update llama tokenizer tests

* Improve handling of legacy llama tokenizer

* Re-enable SPM tests

* Add static tokenizer test cases

* Add llama2 static tests

* Allow user to override legacy tokenizer behaviour in `.from_pretrained`

* Add legacy tokenizer unit tests

* Bump jinja version to 0.1.0

Author: xenova

package-lock.json

@@ -44,6 +44,9 @@
   "optionalDependencies": {
     "onnxruntime-node": "1.14.0"
   },
+  "peerDependencies": {
+    "@huggingface/jinja": "^0.1.0"
+  },
   "devDependencies": {
     "@types/jest": "^29.5.1",
     "catharsis": "github:xenova/catharsis",

package.json

@@ -283,6 +283,13 @@ def main():
         # Load tokenizer
         tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
 
+        # To avoid inserting all chat templates into tokenizers.js, we save the chat template
+        # to the tokenizer_config.json file, and load it when the tokenizer is loaded.
+        if getattr(tokenizer, 'chat_template', None) is None and \
+            getattr(tokenizer, 'use_default_system_prompt', False):
+            # No chat template specified, and we use the default
+            setattr(tokenizer, 'chat_template', tokenizer.default_chat_template)
+
     except KeyError:
         pass  # No Tokenizer
 

scripts/convert.py

@@ -20,8 +20,8 @@ def generate_tokenizer_json(tokenizer):
                 "id": v,
                 "content": k,
                 "single_word": False,
-                "lstrip": False,
-                "rstrip": False,
+                "lstrip": True,
+                "rstrip": True,
                 "normalized": False,
                 "special": True
             }