Merge branch 'main' of https://github.com/gittb/tabbyAPI-function

Author: gittb

backends/exllamav2/model.py

@@ -585,6 +585,7 @@ def progress(loaded_modules: int, total_modules: int)
                 cache_class=draft_cache_class,
                 autosplit=True,
                 use_tp=False,
+                model=self.draft_model,
             )
 
             for value in self.draft_model.load_autosplit_gen(
@@ -635,6 +636,7 @@ def progress(loaded_modules: int, total_modules: int)
             cache_class=cache_class,
             autosplit=self.gpu_split_auto,
             use_tp=self.use_tp,
+            model=self.model,
         )
 
         # Load model with autosplit (without TP)
@@ -669,20 +671,24 @@ def get_cache_class(self, cache_mode: str):
                 return ExLlamaV2Cache
 
     def create_cache(
-        self, cache_class: ExLlamaV2CacheBase, autosplit: bool, use_tp: bool
+        self,
+        cache_class: ExLlamaV2CacheBase,
+        autosplit: bool,
+        use_tp: bool,
+        model: ExLlamaV2,
     ):
         """Utility function to create a model cache."""
 
         if has_tp and use_tp:
             return ExLlamaV2Cache_TP(
-                self.model,
+                model,
                 base=cache_class,
                 max_seq_len=self.cache_size,
                 batch_size=1,
             )
         else:
             return cache_class(
-                self.model,
+                model,
                 max_seq_len=self.cache_size,
                 lazy=autosplit,
                 batch_size=1,
@@ -865,7 +871,7 @@ def get_special_tokens(
     def get_logprobs(self, token_ids: torch.Tensor, token_probs: torch.Tensor):
         top_tokens = [
             self.tokenizer.extended_id_to_piece.get(
-                index, self.tokenizer.id_to_piece[index]
+                index, self.tokenizer.get_id_to_piece_list(True)[index]
             )
             for index in token_ids.flatten().tolist()
         ]
@@ -1140,7 +1146,7 @@ async def generate_gen(
 
             # Map logits to the tensor with their biases
             for token_id, bias in logit_bias.items():
-                if 0 <= token_id < len(self.tokenizer.id_to_piece):
+                if 0 <= token_id < len(self.tokenizer.get_id_to_piece_list(True)):
                     gen_settings.token_bias[token_id] = bias
                 else:
                     logger.warning(