Add `min_new_tokens` argument in generate() (implementation based on...

Add `min_new_tokens` argument in generate() (implementation based on `MinNewTokensLengthLogitsProcessor`) (#21044)

add a new parameter min_new_tokens for generate()

src/transformers/generation/configuration_utils.py

@@ -75,7 +75,11 @@ class GenerationConfig(PushToHubMixin):
         max_new_tokens (`int`, *optional*):
             The maximum numbers of tokens to generate, ignoring the number of tokens in the prompt.
         min_length (`int`, *optional*, defaults to 0):
-            The minimum length of the sequence to be generated.
+            The minimum length of the sequence to be generated. Corresponds to the length of the input prompt +
+            `min_new_tokens`. In general, prefer the use of `min_new_tokens`, which ignores the number of tokens in the
+            prompt.
+        min_new_tokens (`int`, *optional*):
+            The minimum numbers of tokens to generate, ignoring the number of tokens in the prompt.
         early_stopping (`bool`, *optional*, defaults to `False`):
             Whether to stop the beam search when at least `num_beams` sentences are finished per batch or not.
         max_time(`float`, *optional*):
@@ -207,6 +211,7 @@ class GenerationConfig(PushToHubMixin):
         self.max_length = kwargs.pop("max_length", 20)
         self.max_new_tokens = kwargs.pop("max_new_tokens", None)
         self.min_length = kwargs.pop("min_length", 0)
+        self.min_new_tokens = kwargs.pop("min_new_tokens", None)
         self.early_stopping = kwargs.pop("early_stopping", False)
         self.max_time = kwargs.pop("max_time", None)
 

src/transformers/generation/utils.py

@@ -48,6 +48,7 @@ from .logits_process import (
     LogitNormalization,
     LogitsProcessorList,
     MinLengthLogitsProcessor,
+    MinNewTokensLengthLogitsProcessor,
     NoBadWordsLogitsProcessor,
     NoRepeatNGramLogitsProcessor,
     PrefixConstrainedLogitsProcessor,
@@ -822,6 +823,16 @@ class GenerationMixin:
             and generation_config.min_length > 0
         ):
             processors.append(MinLengthLogitsProcessor(generation_config.min_length, generation_config.eos_token_id))
+        if (
+            generation_config.min_new_tokens is not None
+            and generation_config.eos_token_id is not None
+            and generation_config.min_new_tokens > 0
+        ):
+            processors.append(
+                MinNewTokensLengthLogitsProcessor(
+                    input_ids_seq_length, generation_config.min_new_tokens, generation_config.eos_token_id
+                )
+            )
         if prefix_allowed_tokens_fn is not None:
             processors.append(
                 PrefixConstrainedLogitsProcessor(