Generate: PT's `top_p` enforces `min_tokens_to_keep` when it is `1` (#24111)

src/transformers/generation/flax_logits_process.py

@@ -129,6 +129,8 @@ class FlaxTopPLogitsWarper(FlaxLogitsWarper):
     def __init__(self, top_p: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
         if not isinstance(top_p, float) or (top_p < 0 or top_p > 1.0):
             raise ValueError(f"`top_p` has to be a float > 0 and < 1, but is {top_p}")
+        if not isinstance(min_tokens_to_keep, int) or (min_tokens_to_keep < 0):
+            raise ValueError(f"`min_tokens_to_keep` has to be a non-negative integer, but is {min_tokens_to_keep}")
 
         self.top_p = top_p
         self.filter_value = filter_value

src/transformers/generation/logits_process.py

@@ -255,6 +255,8 @@ class TopPLogitsWarper(LogitsWarper):
         top_p = float(top_p)
         if top_p < 0 or top_p > 1.0:
             raise ValueError(f"`top_p` has to be a float > 0 and < 1, but is {top_p}")
+        if not isinstance(min_tokens_to_keep, int) or (min_tokens_to_keep < 0):
+            raise ValueError(f"`min_tokens_to_keep` has to be a non-negative integer, but is {min_tokens_to_keep}")
 
         self.top_p = top_p
         self.filter_value = filter_value
@@ -266,9 +268,8 @@ class TopPLogitsWarper(LogitsWarper):
 
         # Remove tokens with cumulative top_p above the threshold (token with 0 are kept)
         sorted_indices_to_remove = cumulative_probs <= (1 - self.top_p)
-        if self.min_tokens_to_keep > 1:
-            # Keep at least min_tokens_to_keep
-            sorted_indices_to_remove[..., -self.min_tokens_to_keep :] = 0
+        # Keep at least min_tokens_to_keep
+        sorted_indices_to_remove[..., -self.min_tokens_to_keep :] = 0
 
         # scatter sorted tensors to original indexing
         indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)

src/transformers/generation/tf_logits_process.py

@@ -160,6 +160,8 @@ class TFTopPLogitsWarper(TFLogitsWarper):
     def __init__(self, top_p: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
         if not isinstance(top_p, float) or (top_p < 0 or top_p > 1.0):
             raise ValueError(f"`top_p` has to be a float > 0 and < 1, but is {top_p}")
+        if not isinstance(min_tokens_to_keep, int) or (min_tokens_to_keep < 0):
+            raise ValueError(f"`min_tokens_to_keep` has to be a non-negative integer, but is {min_tokens_to_keep}")
 
         self.top_p = top_p
         self.filter_value = filter_value