Model: Add support for HuggingFace config and bad_words_ids

This is necessary for Kobold's API. Current models use bad_words_ids
in generation_config.json, but for some reason, they're also present
in the model's config.json.

Signed-off-by: kingbri <bdashore3@proton.me>

backends/exllamav2/model.py

@@ -47,7 +47,7 @@ from common.templating import (
     TemplateLoadError,
     find_template_from_model,
 )
-from common.transformers_utils import GenerationConfig
+from common.transformers_utils import GenerationConfig, HuggingFaceConfig
 from common.utils import coalesce, unwrap
 
 
@@ -72,6 +72,7 @@ class ExllamaV2Container:
     draft_cache_mode: str = "FP16"
     max_batch_size: int = 20
     generation_config: Optional[GenerationConfig] = None
+    hf_config: Optional[HuggingFaceConfig] = None
 
     # GPU split vars
     gpu_split: Optional[list] = None
@@ -186,6 +187,9 @@ class ExllamaV2Container:
         except AttributeError:
             pass
 
+        # Create the hf_config
+        self.hf_config = HuggingFaceConfig.from_file(model_directory)
+
         # Then override the base_seq_len if present
         override_base_seq_len = kwargs.get("override_base_seq_len")
         if override_base_seq_len:
@@ -268,15 +272,8 @@ class ExllamaV2Container:
         else:
             self.cache_size = self.config.max_seq_len
 
-        # Try to set prompt template
-        self.prompt_template = self.find_prompt_template(
-            kwargs.get("prompt_template"), model_directory
-        )
-
         # Load generation config overrides
-        generation_config_path = (
-            pathlib.Path(self.config.model_dir) / "generation_config.json"
-        )
+        generation_config_path = model_directory / "generation_config.json"
         if generation_config_path.exists():
             try:
                 self.generation_config = GenerationConfig.from_file(
@@ -288,6 +285,11 @@ class ExllamaV2Container:
                     "Skipping generation config load because of an unexpected error."
                 )
 
+        # Try to set prompt template
+        self.prompt_template = self.find_prompt_template(
+            kwargs.get("prompt_template"), model_directory
+        )
+
         # Catch all for template lookup errors
         if self.prompt_template:
             logger.info(

common/transformers_utils.py

@@ -1,6 +1,7 @@
 import json
 import pathlib
 from typing import List, Optional, Union
+from loguru import logger
 from pydantic import BaseModel
 
 
@@ -11,6 +12,7 @@ class GenerationConfig(BaseModel):
     """
 
     eos_token_id: Optional[Union[int, List[int]]] = None
+    bad_words_ids: Optional[List[List[int]]] = None
 
     @classmethod
     def from_file(self, model_directory: pathlib.Path):
@@ -30,3 +32,40 @@ class GenerationConfig(BaseModel):
             return [self.eos_token_id]
         else:
             return self.eos_token_id
+
+
+class HuggingFaceConfig(BaseModel):
+    """
+    An abridged version of HuggingFace's model config.
+    Will be expanded as needed.
+    """
+
+    badwordsids: Optional[str] = None
+
+    @classmethod
+    def from_file(self, model_directory: pathlib.Path):
+        """Create an instance from a generation config file."""
+
+        hf_config_path = model_directory / "config.json"
+        with open(
+            hf_config_path, "r", encoding="utf8"
+        ) as hf_config_json:
+            hf_config_dict = json.load(hf_config_json)
+            return self.model_validate(hf_config_dict)
+
+    def get_badwordsids(self):
+        """Wrapper method to fetch badwordsids."""
+
+        if self.badwordsids:
+            try:
+                bad_words_list = json.loads(self.badwordsids)
+                return bad_words_list
+            except json.JSONDecodeError:
+                logger.warning(
+                    "Skipping badwordsids from config.json "
+                    "since it's not a valid array."
+                )
+
+                return []
+        else:
+            return []

common/utils.py

@@ -18,3 +18,9 @@ def prune_dict(input_dict):
     """Trim out instances of None from a dictionary."""
 
     return {k: v for k, v in input_dict.items() if v is not None}
+
+
+def flat_map(input_list):
+    """Flattens a list of lists into a single list."""
+
+    return [item for sublist in input_list for item in sublist]

endpoints/Kobold/types/generation.py

@@ -1,7 +1,9 @@
 from typing import List, Optional
 
 from pydantic import BaseModel, Field
+from common import model
 from common.sampling import BaseSamplerRequest
+from common.utils import flat_map, unwrap
 
 
 class GenerateRequest(BaseSamplerRequest):
@@ -14,6 +16,16 @@ class GenerateRequest(BaseSamplerRequest):
         if self.penalty_range == 0:
             self.penalty_range = -1
 
+        # Move badwordsids into banned tokens for generation
+        if self.use_default_badwordsids:
+            bad_words_ids = unwrap(
+                model.container.generation_config.bad_words_ids,
+                model.container.hf_config.get_badwordsids()
+            )
+
+            if bad_words_ids:
+                self.banned_tokens += flat_map(bad_words_ids)
+
         return super().to_gen_params(**kwargs)