Generate: add min_p sampling (huggingface#30639)

* min_p

* more relaxed test to avoid numerical issues

* Update src/transformers/generation/logits_process.py

Co-authored-by: menhguin <minh1228@gmail.com>

* Update src/transformers/generation/configuration_utils.py

Co-authored-by: menhguin <minh1228@gmail.com>

* docstring clarifications

* PR comments

* Update tests/generation/test_logits_process.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* make fixup

---------

Co-authored-by: menhguin <minh1228@gmail.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

docs/source/en/internal/generation_utils.md

@@ -167,6 +167,9 @@ generation.
 [[autodoc]] MinNewTokensLengthLogitsProcessor
     - __call__
 
+[[autodoc]] MinPLogitsWarper
+    - __call__
+
 [[autodoc]] NoBadWordsLogitsProcessor
     - __call__
 

src/transformers/__init__.py

@@ -1222,6 +1222,7 @@
             "MaxTimeCriteria",
             "MinLengthLogitsProcessor",
             "MinNewTokensLengthLogitsProcessor",
+            "MinPLogitsWarper",
             "NoBadWordsLogitsProcessor",
             "NoRepeatNGramLogitsProcessor",
             "PhrasalConstraint",
@@ -5777,6 +5778,7 @@
             MaxTimeCriteria,
             MinLengthLogitsProcessor,
             MinNewTokensLengthLogitsProcessor,
+            MinPLogitsWarper,
             NoBadWordsLogitsProcessor,
             NoRepeatNGramLogitsProcessor,
             PhrasalConstraint,

src/transformers/generation/__init__.py

@@ -64,6 +64,7 @@
         "LogitsWarper",
         "MinLengthLogitsProcessor",
         "MinNewTokensLengthLogitsProcessor",
+        "MinPLogitsWarper",
         "NoBadWordsLogitsProcessor",
         "NoRepeatNGramLogitsProcessor",
         "PrefixConstrainedLogitsProcessor",
@@ -204,6 +205,7 @@
             LogitsWarper,
             MinLengthLogitsProcessor,
             MinNewTokensLengthLogitsProcessor,
+            MinPLogitsWarper,
             NoBadWordsLogitsProcessor,
             NoRepeatNGramLogitsProcessor,
             PrefixConstrainedLogitsProcessor,

src/transformers/generation/configuration_utils.py

@@ -139,6 +139,10 @@ class GenerationConfig(PushToHubMixin):
         top_p (`float`, *optional*, defaults to 1.0):
             If set to float < 1, only the smallest set of most probable tokens with probabilities that add up to
             `top_p` or higher are kept for generation.
+        min_p (`float`, *optional*):
+            Minimum token probability, which will be scaled by the probability of the most likely token. It must be a
+            value between 0 and 1. Typical values are in the 0.01-0.2 range, comparably selective as setting `top_p` in
+            the 0.99-0.8 range (use the opposite of normal `top_p` values).
         typical_p (`float`, *optional*, defaults to 1.0):
             Local typicality measures how similar the conditional probability of predicting a target token next is to
             the expected conditional probability of predicting a random token next, given the partial text already
@@ -324,6 +328,7 @@ def __init__(self, **kwargs):
         self.temperature = kwargs.pop("temperature", 1.0)
         self.top_k = kwargs.pop("top_k", 50)
         self.top_p = kwargs.pop("top_p", 1.0)
+        self.min_p = kwargs.pop("min_p", None)
         self.typical_p = kwargs.pop("typical_p", 1.0)
         self.epsilon_cutoff = kwargs.pop("epsilon_cutoff", 0.0)
         self.eta_cutoff = kwargs.pop("eta_cutoff", 0.0)

src/transformers/generation/logits_process.py

@@ -520,6 +520,83 @@ def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> to
         return scores_processed
 
 
+class MinPLogitsWarper(LogitsWarper):
+    """
+    [`LogitsWarper`] that performs min-p, i.e. keeps all tokens that are above a minimum probability, scaled by the
+    probability of the most likely token. As a result, the filter becomes more agressive in the presence of
+    high-probability tokens, which is a sign of a confident output that we shouldn't deviate from.
+
+    Often used together with [`TemperatureLogitsWarper`]. Used as an alternative to [`TopPLogitsWarper`] and
+    [`TopKLogitsWarper`].
+
+    Created by @menhguin and @kalomaze (github handles). Code adapted from [this external PR](https://github.com/oobabooga/text-generation-webui/pull/4449/files)
+
+    Args:
+        min_p (`float`):
+            Minimum token probability, which will be scaled by the probability of the most likely token. It must be a
+            value between 0 and 1. Typical values are in the 0.01-0.2 range, comparably selective as setting `top_p` in
+            the 0.99-0.8 range (use the opposite of normal `top_p` values).
+        filter_value (`float`, *optional*, defaults to -inf):
+            All filtered values will be set to this float value.
+        min_tokens_to_keep (`int`, *optional*, defaults to 1):
+            Minimum number of tokens that cannot be filtered.
+
+    Examples:
+
+    ```python
+    >>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed
+
+    >>> set_seed(1)
+    >>> model = AutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
+    >>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilgpt2")
+
+    >>> inputs = tokenizer("A sequence: 1, 2", return_tensors="pt")
+
+    >>> # With sampling, the output is unexpected -- sometimes too unexpected.
+    >>> outputs = model.generate(**inputs, do_sample=True)
+    >>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
+    A sequence: 1, 2, 3 | < 4 (left-hand pointer) ;
+    <BLANKLINE>
+    <BLANKLINE>
+
+    >>> # With `min_p` sampling, the output gets restricted to high-probability tokens.
+    >>> # Pro tip: In practice, LLMs use `min_p` in the 0.01-0.2 range.
+    >>> outputs = model.generate(**inputs, do_sample=True, min_p=0.1)
+    >>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
+    A sequence: 1, 2, 3, 4, 5, 6, 7, 8, 9
+    ```
+    """
+
+    def __init__(self, min_p: float, filter_value: float = -float("Inf"), min_tokens_to_keep: int = 1):
+        if not (0 <= min_p <= 1.0):
+            raise ValueError(f"`min_p` has to be a float in the [0, 1] interval, but is {min_p}")
+        if not isinstance(min_tokens_to_keep, int) or (min_tokens_to_keep < 1):
+            raise ValueError(f"`min_tokens_to_keep` has to be a positive integer, but is {min_tokens_to_keep}")
+
+        self.min_p = min_p
+        self.filter_value = filter_value
+        self.min_tokens_to_keep = min_tokens_to_keep
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        # Convert logits to probabilities
+        probs = torch.softmax(scores, dim=-1)
+        # Get the probability of the top token for each sequence in the batch
+        top_probs, _ = probs.max(dim=-1, keepdim=True)
+        # Calculate the actual min_p threshold by scaling min_p with the top token's probability
+        scaled_min_p = self.min_p * top_probs
+        # Create a mask for tokens that have a probability less than the scaled min_p
+        tokens_to_remove = probs < scaled_min_p
+
+        sorted_indices = torch.argsort(scores, descending=True, dim=-1)
+        sorted_indices_to_remove = torch.gather(tokens_to_remove, dim=-1, index=sorted_indices)
+        # Keep at least min_tokens_to_keep
+        sorted_indices_to_remove[..., : self.min_tokens_to_keep] = False
+
+        indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+        scores_processed = scores.masked_fill(indices_to_remove, self.filter_value)
+        return scores_processed
+
+
 class TypicalLogitsWarper(LogitsWarper):
     r"""
     [`LogitsWarper`] that performs typical decoding. Inspired on how humans use language, it prioritizes tokens whose

src/transformers/generation/utils.py

@@ -61,6 +61,7 @@
     LogitsProcessorList,
     MinLengthLogitsProcessor,
     MinNewTokensLengthLogitsProcessor,
+    MinPLogitsWarper,
     NoBadWordsLogitsProcessor,
     NoRepeatNGramLogitsProcessor,
     PrefixConstrainedLogitsProcessor,
@@ -739,6 +740,9 @@ def _get_logits_warper(
             warpers.append(TopKLogitsWarper(top_k=generation_config.top_k, min_tokens_to_keep=min_tokens_to_keep))
         if generation_config.top_p is not None and generation_config.top_p < 1.0:
             warpers.append(TopPLogitsWarper(top_p=generation_config.top_p, min_tokens_to_keep=min_tokens_to_keep))
+        if generation_config.min_p is not None:
+            # Applied after temperature scaling (see https://github.com/ggerganov/llama.cpp/pull/3841#issuecomment-2073826084)
+            warpers.append(MinPLogitsWarper(min_p=generation_config.min_p, min_tokens_to_keep=min_tokens_to_keep))
         if generation_config.typical_p is not None and generation_config.typical_p < 1.0:
             warpers.append(
                 TypicalLogitsWarper(mass=generation_config.typical_p, min_tokens_to_keep=min_tokens_to_keep)

src/transformers/utils/dummy_pt_objects.py

@@ -317,6 +317,13 @@ def __init__(self, *args, **kwargs):
         requires_backends(self, ["torch"])
 
 
+class MinPLogitsWarper(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class NoBadWordsLogitsProcessor(metaclass=DummyObject):
     _backends = ["torch"]
 

tests/generation/test_logits_process.py

@@ -42,6 +42,7 @@
         LogitsProcessorList,
         MinLengthLogitsProcessor,
         MinNewTokensLengthLogitsProcessor,
+        MinPLogitsWarper,
         NoBadWordsLogitsProcessor,
         NoRepeatNGramLogitsProcessor,
         PrefixConstrainedLogitsProcessor,
@@ -304,6 +305,52 @@ def test_top_p_dist_warper(self):
         # first batch should keep three tokens, second batch would keep only 1, but due to `min_tokens_to_keep=2` keeps 2.
         self.assertListEqual((filtered_dist != 0.0).to(torch.long).sum(dim=-1).tolist(), [3, 2])
 
+    def test_min_p_dist_warper(self):
+        input_ids = None
+        vocab_size = 10
+        batch_size = 2
+
+        # create distribution and take log (inverse to Softmax as taken in MinPLogitsWarper)
+        dist = torch.log(
+            torch.tensor(
+                [
+                    [0.9, 0.0274, 0.047, 0.0274],  # two tokens should be kept (0.047 > 0.9*0.05=0.045)
+                    [0.15, 0.3, 0.3, 0.25],  # all should be kept -- no high-probability token
+                    [0.97, 0.01, 0.01, 0.01],  # only the first token should be kept
+                ],
+                device=torch_device,
+                dtype=torch.float,
+            )
+        )
+
+        min_p_warp = MinPLogitsWarper(0.05)
+        filtered_dist = torch.exp(min_p_warp(input_ids, dist))
+
+        # exp (-inf) => 0
+        EXPECTED_FILTERED_DIST = torch.tensor(
+            [[0.9, 0.0, 0.047, 0.0], [0.15, 0.3, 0.3, 0.25], [0.97, 0.0, 0.0, 0.0]],
+            device=torch_device,
+            dtype=torch.float,
+        )
+        self.assertTrue(torch.allclose(filtered_dist, EXPECTED_FILTERED_DIST, atol=1e-3))
+
+        # processor should not change logits in-place
+        self.assertFalse(torch.all(min_p_warp(input_ids, dist) == dist))
+
+        # check edge cases with negative and extreme logits
+        ramp_logits = torch.arange(vocab_size, device=torch_device, dtype=torch.float) - (vocab_size // 2)
+        ramp_logits = ramp_logits.unsqueeze(0).repeat(batch_size, 1)
+
+        # make ramp_logits more extreme
+        ramp_logits[1] = ramp_logits[1] * 100.0
+
+        # make sure at least 2 tokens are kept
+        min_p_warp = MinPLogitsWarper(0.9, min_tokens_to_keep=2, filter_value=0.0)
+        filtered_dist = min_p_warp(input_ids, ramp_logits)
+
+        # first batch should keep two tokens, second batch would keep only 1, but due to `min_tokens_to_keep=2` keeps 2.
+        self.assertListEqual((filtered_dist != 0.0).to(torch.long).sum(dim=-1).tolist(), [2, 2])
+
     def test_typical_dist_warper(self):
         input_ids = None
         vocab_size = 10