Fixing the update to outlines.

.gitignore

@@ -22,3 +22,4 @@ server/fbgemmm
 
 .direnv/
 .venv/
+out/

integration-tests/models/__snapshots__/test_grammar_response_format_llama/test_grammar_response_format_llama_json.json

@@ -1,23 +1,23 @@
 {
   "choices": [
     {
-      "finish_reason": "eos_token",
+      "finish_reason": "stop",
       "index": 0,
       "logprobs": null,
       "message": {
-        "content": "{\n  \"temperature\": [\n    35,\n    34,\n    36\n  ],\n  \"unit\": \"°c\"\n}",
+        "content": "{ \"temperature\": [ 26, 30, 33, 29 ] ,\"unit\": \"Fahrenheit\" }",
         "role": "assistant"
       }
     }
   ],
-  "created": 1718044128,
+  "created": 1726319532,
   "id": "",
   "model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
-  "object": "text_completion",
-  "system_fingerprint": "2.0.5-dev0-native",
+  "object": "chat.completion",
+  "system_fingerprint": "2.2.1-dev0-native",
   "usage": {
-    "completion_tokens": 39,
+    "completion_tokens": 36,
     "prompt_tokens": 136,
-    "total_tokens": 175
+    "total_tokens": 172
   }
 }

integration-tests/models/test_grammar_response_format_llama.py

@@ -56,10 +56,7 @@ async def test_grammar_response_format_llama_json(llama_grammar, response_snapsh
     called = chat_completion["choices"][0]["message"]["content"]
 
     assert response.status_code == 200
-    assert (
-        called
-        == '{\n  "temperature": [\n    35,\n    34,\n    36\n  ],\n  "unit": "°c"\n}'
-    )
+    assert called == '{ "temperature": [ 26, 30, 33, 29 ] ,"unit": "Fahrenheit" }'
     assert chat_completion == response_snapshot