feat: add Jinja2 loop controls to chat templates

CHANGELOG.md

@@ -7,6 +7,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+- feat: add Jinja2 loop controls to chat templates by @handshape in #2018
 - fix: avoid cleanup errors for partially initialized `LlamaModel` objects by @usernames122 in #2173
 - fix: suppress stdout and stderr in Jupyter notebooks by @Anai-Guo in #2181
 - feat: enable arm64 musl builds by @acon96 in #2221

llama_cpp/llama_chat_format.py

@@ -223,7 +223,12 @@ def __init__(
             loader=jinja2.BaseLoader(),
             trim_blocks=True,
             lstrip_blocks=True,
-            extensions=[Jinja2ChatFormatter.IgnoreGenerationTags],
+            # Keep this aligned with Transformers' chat-template Jinja extensions.
+            # https://github.com/huggingface/transformers/blob/39603d0e5cdb6f00e8d473d7fcbb01032d709181/src/transformers/utils/chat_template_utils.py#L489-L490
+            extensions=[
+                Jinja2ChatFormatter.IgnoreGenerationTags,
+                jinja2.ext.loopcontrols,
+            ],
         ).from_string(self.template)
 
     @staticmethod