Model: Add TokenizerConfig stub and add_eos_token fallback

This stub fetches the add_eos_token field from the HF tokenizer config.
Ideally, this should be in the backend rather than tabby.

Signed-off-by: kingbri <8082010+kingbri1@users.noreply.github.com>

Author: kingbri1

backends/exllamav2/model.py

@@ -50,7 +50,7 @@
 from common.multimodal import MultimodalEmbeddingWrapper
 from common.sampling import BaseSamplerRequest
 from common.templating import PromptTemplate, find_prompt_template
-from common.transformers_utils import GenerationConfig
+from common.transformers_utils import GenerationConfig, TokenizerConfig
 from common.utils import calculate_rope_alpha, coalesce, unwrap
 from endpoints.core.types.model import ModelCard, ModelCardParameters
 
@@ -80,6 +80,7 @@ class ExllamaV2Container(BaseModelContainer):
     draft_cache_mode: str = "FP16"
     max_batch_size: Optional[int] = None
     generation_config: Optional[GenerationConfig] = None
+    tokenizer_config: Optional[TokenizerConfig] = None
 
     # GPU split vars
     gpu_split: List[float] = []
@@ -130,14 +131,27 @@ async def create(cls, model_directory: pathlib.Path, quiet=False, **kwargs):
         if generation_config_path.exists():
             try:
                 self.generation_config = await GenerationConfig.from_file(
-                    generation_config_path.parent
+                    model_directory
                 )
             except Exception:
                 logger.error(traceback.format_exc())
                 logger.warning(
                     "Skipping generation config load because of an unexpected error."
                 )
 
+        # Load tokenizer config overrides
+        tokenizer_config_path = model_directory / "tokenizer_config.json"
+        if tokenizer_config_path.exists():
+            try:
+                self.tokenizer_config = await TokenizerConfig.from_file(
+                    model_directory
+                )
+            except Exception:
+                logger.error(traceback.format_exc())
+                logger.warning(
+                    "Skipping tokenizer config load because of an unexpected error."
+                )
+
         # Set vision state and error if vision isn't supported on the current model
         self.use_vision = unwrap(kwargs.get("vision"), False)
         if self.use_vision and not self.config.vision_model_type:
@@ -1240,9 +1254,17 @@ async def generate_gen(
         ) and gen_settings.token_repetition_range == -1
 
         stop_conditions = params.stop
-        add_bos_token = unwrap(params.add_bos_token, True)
         ban_eos_token = params.ban_eos_token
 
+
+        print(self.tokenizer_config.add_bos_token)
+        # Set add_bos_token for generation
+        add_bos_token = coalesce(
+            params.add_bos_token, self.tokenizer_config.add_bos_token, True
+        )
+
+        print(add_bos_token)
+
         # Fetch EOS tokens from generation_config if they exist
         eos_tokens = (
             self.generation_config.eos_tokens()

common/templating.py

@@ -239,6 +239,7 @@ async def find_prompt_template(template_name, model_dir: pathlib.Path):
         ]
 
     # Add lookup from prompt template name if provided
+    # TODO: Possibly link to the TokenizerConfig class
     if template_name:
         find_template_functions[:0] = [
             lambda: PromptTemplate.from_file(pathlib.Path("templates") / template_name),

common/transformers_utils.py

@@ -53,3 +53,23 @@ async def from_file(cls, model_directory: pathlib.Path):
             contents = await hf_config_json.read()
             hf_config_dict = json.loads(contents)
             return cls.model_validate(hf_config_dict)
+
+
+class TokenizerConfig(BaseModel):
+    """
+    An abridged version of HuggingFace's tokenizer config.
+    """
+
+    add_bos_token: Optional[bool] = None
+
+    @classmethod
+    async def from_file(cls, model_directory: pathlib.Path):
+        """Create an instance from a tokenizer config file."""
+
+        tokenizer_config_path = model_directory / "tokenizer_config.json"
+        async with aiofiles.open(
+            tokenizer_config_path, "r", encoding="utf8"
+        ) as tokenizer_config_json:
+            contents = await tokenizer_config_json.read()
+            tokenizer_config_dict = json.loads(contents)
+            return cls.model_validate(tokenizer_config_dict)