[Build] Fix CI build (Llama2 Chat Templates) (#948)

A couple of issues in the CI build were not caught because the PR was from a fork. There are problems with the Llama2 chat templates, which are addressed here. There are some fixes to the templates themselves, but the handling of the `system` role is so different from how the Guidance context managers work that the test is XFAILed.

Author: riedgar-ms

guidance/chat.py

@@ -130,19 +130,19 @@ def get_role_start(self, role_name):
         if role_name == "system":
             return "[INST] <<SYS>>\n"
         elif role_name == "user":
-            return "<s>[INST]"
+            return "<s>[INST] "
         elif role_name == "assistant":
             return " "
         else:
             raise UnsupportedRoleException(role_name, self)
 
     def get_role_end(self, role_name=None):
         if role_name == "system":
-            return "\n<</SYS>"
+            return "\n<</SYS>>"
         elif role_name == "user":
             return " [/INST]"
         elif role_name == "assistant":
-            return "</s>"
+            return " </s>"
         else:
             raise UnsupportedRoleException(role_name, self)
 

tests/need_credentials/test_chat_templates.py

@@ -81,7 +81,13 @@ def test_chat_format_smoke(model_id: str):
     [
         "microsoft/Phi-3-mini-4k-instruct",
         "meta-llama/Meta-Llama-3-8B-Instruct",
-        "meta-llama/Llama-2-7b-chat-hf",
+        pytest.param(
+            "meta-llama/Llama-2-7b-chat-hf",
+            marks=pytest.mark.xfail(
+                reason="Handling of system prompt highly constrained; does not work well with context blocks",
+                raises=AssertionError,
+            ),
+        ),
     ],
 )
 def test_chat_format_smoke_with_system(model_id: str):