Expand S2S range and make pad trimming controllable (#235)

examples/llm/src/data/denoising.py

@@ -42,7 +42,7 @@ def ul2_prefix_function(
     """
     if mean_length is None:
         # This is the case for "sequence to sequence"
-        prefix = '[S2S]'
+        prefix = '[S2S]' if mask_ratio < 1.0 else '[CLM]'
     elif mean_length >= 12 or mask_ratio >= 0.3:
         # UL2 tags this corruption rate "extreme"
         prefix = '[NLG]'
@@ -101,13 +101,20 @@ class MixtureOfDenoisersCollator:
             Default: ``None`` does not add any span corruption tasks.
         sequence_mask_ratios (optional): A float or list of floats, one for each
             sequence corruption denoising task to add to the task mixture. Each
-            sequence mask ratio must be greater than 0.0 and less than 0.5.
+            sequence mask ratio must be greater than 0.0 and at most 1.0.
             This type of task is a special instance of span corruption, with
             exactly one masked span take from the end of the sequence. The
-            length of the span is sampled uniformly from
-            [1, 2*mask_ratio*n_tokens], where n_tokens is the length of the
-            unmasked token sequence.
+            length of the span is sampled uniformly such that the average
+            portion of masked tokens equals sequence_mask_ratio.
+            Note: A value of 1.0 essentially yields causal LM examples.
             Default: ``None` does not add any sequence corruption tasks.
+        allow_pad_trimming (bool, optional): Whether to allow the collator to
+            trim away sequence regions that are entirely padding (i.e. padding
+            for each example in the batch). If ``True``, shorter sequences may
+            improve throughput but at a potentially higher memory cost
+            owing to variable sequence lengths from batch to batch.
+            Default: ``False`` yields batches that are always padded to
+            max_seq_length.
         prefix_function (callable, optional): A function that maps denoising
             task parameters (e.g. mean_length=3, mask_ratio=0.15) to a prefix
             that will be added to sequences when the associated "noiser" is
@@ -124,13 +131,18 @@ def __init__(
         decoder_only_format: bool = False,
         span_mean_lengths_and_ratios: Optional[List] = None,
         sequence_mask_ratios: Optional[Union[List[float], float]] = None,
+        allow_pad_trimming: bool = False,
         prefix_function: Optional[PREFIX_FUNCTION] = ul2_prefix_function,
     ):
         self.tokenizer = tokenizer
         self.max_seq_length = max_seq_length
         self.decoder_only_format = decoder_only_format
         self._sentinel_token_ids = np.array(self.tokenizer.sentinel_token_ids)
 
+        # Trimming will always be skipped on at least the first __call__
+        self._allow_pad_trimming = allow_pad_trimming
+        self._seen_first_batch = False
+
         # Prepare the tokenizer for denoising tasks
         utils.adapt_tokenizer_for_denoising(self.tokenizer)
 
@@ -165,9 +177,9 @@ def __init__(
         else:
             # In this case, there is one or more sequence corruption tasks
             for ratio in sequence_mask_ratios:
-                if not (0 < ratio < 0.5):
+                if not (0 < ratio <= 1.0):
                     raise ValueError('`sequence_mask_ratios` must be a float (or list '+\
-                                    'of floats) between 0.0 and 0.5, or None. '+\
+                                    'of floats) that are each >0.0 and <=1.0, or None. '+\
                                     f'Got {sequence_mask_ratios}.')
             self.sequence_mask_ratios = sequence_mask_ratios
 
@@ -274,6 +286,12 @@ def __call__(self, examples: List[Dict[str,
                     decoder_only_format=self.decoder_only_format))
         batch = self.tokenizer.pad(processed_examples)
 
+        # This logic prevents trimming on at least the first batch
+        if not (self._allow_pad_trimming and self._seen_first_batch):
+            self._seen_first_batch = True
+            return batch
+        self._seen_first_batch = True
+
         # Truncate portions of the inputs that are purely padding
         # (up to a multiple of 8)
         multiple_of = 8
@@ -349,15 +367,18 @@ def build_text_denoising_dataloader(cfg: DictConfig,
             cfg.mixture_of_denoisers.decoder_only_format (bool): Whether the
                 batches should use the format required for training a decoder-only
                 model (if ``True``) or an encoder-decoder model (if ``False``).
-            cfg.mixture_of_denoisers.span_mean_lengths_and_ratios (optiona): The
+            cfg.mixture_of_denoisers.span_mean_lengths_and_ratios (optional): The
                 parameters for any span corruption denoising tasks to include in
                 the task mixture.
                 See :class:`MixtureOfDenoisersCollator` docstring for details.
-            cfg.mixture_of_denoisers.sequence_mask_ratios (optiona): The
+            cfg.mixture_of_denoisers.sequence_mask_ratios (optional): The
                 parameters for any sequence denoising tasks to include in the
                 task mixture.
                 See :class:`MixtureOfDenoisersCollator` docstring for details.
-            cfg.mixture_of_denoisers.prefix_function (optiona): Set to ``None``
+            cfg.mixture_of_denoisers.allow_pad_trimming (optional): Whether to
+                allow the collator to trim padding when possible (if ``True``).
+                Defaults to ``False``.
+            cfg.mixture_of_denoisers.prefix_function (optional): Set to ``None``
                 to disable the UL2-style prefixes that will be automatically
                 added by default.
             ---
@@ -377,6 +398,8 @@ def build_text_denoising_dataloader(cfg: DictConfig,
             'span_mean_lengths_and_ratios'),
         sequence_mask_ratios=cfg.mixture_of_denoisers.get(
             'sequence_mask_ratios'),
+        allow_pad_trimming=cfg.mixture_of_denoisers.get('allow_pad_trimming',
+                                                        False),
         prefix_function=cfg.mixture_of_denoisers.get('prefix_function',
                                                      ul2_prefix_function))
 
@@ -463,20 +486,21 @@ def noise_token_sequence(
 
     prefix_tokens = prefix_tokens or []
 
+    if length < 1:
+        raise ValueError('Example cannot be empty but token length <1.')
+
     # mean_span_length==None is a special case for "sequential" denoising
     # (where a single span at the end of the sequence is masked)
     if mean_span_length is None:
         # This ensures that exactly 1 span will be produced and that
         # trimming to max_seq_length won't cut off any <EOS> token.
         # In the decoder-only case, this won't insert new tokens.
-        min_span_length = np.maximum(
-            1, length + len(prefix_tokens) - max_seq_length)
-        max_span_length = np.maximum(
-            min_span_length, np.minimum(length - 1, 2 * mask_ratio * length))
-        mean_span_length = float(
-            np.floor(
-                np.random.uniform(low=min_span_length, high=max_span_length)))
-        mask_ratio = mean_span_length / length
+        if mask_ratio <= 0.5:
+            u = np.random.uniform(low=0.0, high=mask_ratio * 2)
+        else:
+            u = np.random.uniform(low=(mask_ratio * 2) - 1, high=1.0)
+        mean_span_length = float(np.round(1 + u * (length - 1)))
+        mask_ratio = mean_span_length / length  # type: ignore
         use_sentinels = False
     else:
         use_sentinels = True
@@ -714,7 +738,7 @@ def _format_tokens_for_decoder_only(
     n_input = len(tokens_inputs)
     n_label = len(tokens_labels)
     n_concat = n_input + n_label
-    assert n_concat <= max_seq_length
+    assert n_concat <= max_seq_length, f'{n_concat=}, {n_input=}, {n_label=}'
 
     tokens_concat = torch.concat([tokens_inputs, tokens_labels], dim=0)