token healing is added

generator/generate.py

@@ -1,5 +1,3 @@
-from dataclasses import dataclass
-
 from comfy.model_management import get_torch_device
 from generator.model import get_model_tokenizer
 
@@ -7,10 +5,12 @@ from generator.utility import (
     get_accelerator_type,
     get_variable_dictionary,
     str_to_quant_type,
+    check_transformers_version,
+    ModelType,
 )
 from generator.preprocess import preprocess
 
-from .utility import ModelType
+from dataclasses import dataclass
 
 
 @dataclass
@@ -37,6 +37,7 @@ class Generator:
     model = None
     tokenizer = None
     dev = None
+    extra_params = {}
 
     def __init__(
         self, model_path: str, is_accelerate: bool, model_quant_type: str
@@ -54,6 +55,15 @@ class Generator:
                 model_path, accelerator_type, quantize_type
             )
 
+        self.tokenizer.padding_side = "left"
+
+        self.extra_params["renormalize_logits"] = True
+        self.extra_params["pad_token_id"] = self.tokenizer.eos_token_id
+
+        if check_transformers_version(4, 6):
+            self.extra_params["token_healing"] = True
+            self.extra_params["tokenizer"] = self.tokenizer
+
         self.dev = get_torch_device()
 
     # generate single output
@@ -69,12 +79,13 @@ class Generator:
 
         args.num_return_sequences = 1
 
-        inputs = self.tokenizer(input, padding=True, return_tensors="pt").to(self.dev)
+        inputs = self.tokenizer(
+            input.strip(), padding=True, truncation=True, return_tensors="pt"
+        ).to(self.dev)
         generated_ids = self.model.generate(
             **inputs,
             **get_variable_dictionary(args),
-            pad_token_id=self.tokenizer.eos_token_id,
-            renormalize_logits=True,
+            **self.extra_params,
         )
 
         output = self.tokenizer.decode(
@@ -96,15 +107,16 @@ class Generator:
 
         args.num_return_sequences = 5
 
-        inputs = self.tokenizer(input, padding=True, return_tensors="pt").to(self.dev)
+        inputs = self.tokenizer(
+            input.strip(), padding=True, truncation=True, return_tensors="pt"
+        ).to(self.dev)
         generated_ids = self.model.generate(
             **inputs,
             **get_variable_dictionary(args),
-            pad_token_id=self.tokenizer.eos_token_id,
-            renormalize_logits=True,
+            **self.extra_params,
         )
         outputs = self.tokenizer.batch_decode(
-            generated_ids, skip_special_tokens=True, cleanup_tokenization_spaces=True
+            generated_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True
         )
 
         return outputs