Better HF grammar implementation (#4953)

modules/grammar/logits_process.py

@@ -0,0 +1,104 @@
+'''
+This file has been 100% copied from this PR to the Transformers library:
+https://github.com/huggingface/transformers/pull/27557
+
+Author: Saibo-creator
+Author GitHub: https://github.com/Saibo-creator
+
+All credits go to the author.
+'''
+
+import math
+
+import torch
+from transformers.generation.logits_process import LogitsProcessor
+from transformers.utils import add_start_docstrings
+
+LOGITS_PROCESSOR_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. [What are input IDs?](../glossary#input-ids)
+        scores (`torch.FloatTensor` of shape `(batch_size, config.vocab_size)`):
+            Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam
+            search or log softmax for each vocabulary token when using beam search
+
+    Return:
+        `torch.FloatTensor` of shape `(batch_size, config.vocab_size)`: The processed prediction scores.
+
+"""
+
+
+class GrammarConstrainedLogitsProcessor(LogitsProcessor):
+    def __init__(self, grammar_constraint):
+        self.last_size = None
+        self.grammar_constraint = grammar_constraint
+        self.batch_stacks = None
+
+    def filter_logits(self, logits, device):
+        # resolve each stack to a tensor of True/False for each token
+        # indicating acceptance
+        # acceptance = self.grammar_acceptor.filter_vocab(self.stacks, device)
+        acceptance = self.grammar_constraint.batch_filter_vocab(self.batch_stacks, device)
+        # logger.debug(acceptance)
+        # Logits to -inf where False
+        logits[~acceptance] = -math.inf
+
+    # TODO: batching
+    def process_logits(self, input_ids, scores, parse_start_index=None):
+        """
+        :param input_ids:
+        :param scores:
+        :param parse_start_index: default None, which means generate from scratch. Set to 0 to parse all input_ids
+        :return:
+        """
+        # we dynamically create stacks at the first call, so that we know the batch size and beam size
+        if self.batch_stacks is None:
+            self.batch_stacks = [self.grammar_constraint.init_stacks() for _ in range(len(input_ids))]
+
+        # if self.last_size is not set (which would be the case when processing the first token).
+        # In this case, do nothing.
+        if self.last_size is None:
+            prefix_to_parse = [
+                single_input_ids[parse_start_index:] if parse_start_index is not None else []
+                for single_input_ids in input_ids
+            ]
+            # self.grammar_acceptor.accept_token_ids(prefix_to_parse, self.stacks)
+            self.batch_stacks = [
+                self.grammar_constraint.accept_token_ids(prefix, stack)
+                for prefix, stack in zip(prefix_to_parse, self.batch_stacks)
+            ]
+        #  if the length of the current input IDs (input_ids[0]) is exactly one more than self.last_size.
+        #  This is expected in a scenario where inputs are processed incrementally, one token at a time.
+        elif len(input_ids[0]) == self.last_size + 1:
+            # self.stacks = self.grammar_acceptor.accept_token_id(input_ids[0][-1], self.stacks)
+            self.batch_stacks = [
+                self.grammar_constraint.accept_token_id(single_input_ids[-1], stack)
+                for single_input_ids, stack in zip(input_ids, self.batch_stacks)
+            ]
+        #  ensure that the input size is consistent with the expected incremental processing
+        #  (i.e., one token at a time).
+        else:
+            # here we check if the input_ids are one token longer than the last time we processed
+            # but we don't check if input_ids are actually valid.
+            # Imagine a scenario where we generate 10 tokens, then we replace the 10 generated tokens with 10 new tokens.
+            # In this case, the input_ids will be consistent with the last_size, but the input_ids are not valid.
+            # However, should we really check if the input_ids are valid here?
+            # If we do, then we need to reparse the whole input_ids at each call, which is not efficient.
+            # Maybe we should just trust the user to provide valid input_ids?
+            # The conclusion is that, we assume the input_ids are valid, and our generation will be correct.
+            # If the input_ids are not valid, then the generation result will be wrong and we don't take responsibility for that.
+            raise RuntimeError(
+                "Input ID's length is inconsistent with the current state of "
+                "the GrammarConstrainedLogitsProcessor. If you want to process "
+                "another input sequence, please instantiate a new "
+                "GrammarConstrainedLogitsProcessor."
+            )
+
+        self.filter_logits(scores, scores.device)
+
+        self.last_size = len(input_ids[0])
+        return scores
+
+    @add_start_docstrings(LOGITS_PROCESSOR_INPUTS_DOCSTRING)
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        return self.process_logits(input_ids, scores)