Improve normalization of losses and metrics

fast_llm/data/document/abstract.py

@@ -9,6 +9,7 @@
 if typing.TYPE_CHECKING:
     import torch
 
+    from fast_llm.engine.distributed.distributed import Distributed
     from fast_llm.tensor import TensorMeta
 
 
@@ -59,6 +60,15 @@ def to_kwargs(self) -> dict[str, typing.Any]:
             AttentionKwargs.presents: self.presents,
         }
 
+    @classmethod
+    def share_batch_data(cls, model_inputs: "list[ModelInput]", distributed: "Distributed"):
+        """
+        Gather values depending on the entire data-parallel batch, ex. the total number of labels or documents.
+        Should be called in the main process because distributed operations are not available during preprocessing.
+        Implemented as a class method so quantities shared by all models inputs are only computed once.
+        TODO: ====== Use as entry point for batch broadcasting? ======
+        """
+
 
 @dataclasses.dataclass(kw_only=True)
 class Batch(Document):

fast_llm/data/document/config.py

@@ -29,6 +29,11 @@ class LengthPreprocessingConfig(BatchPreprocessingConfig):
     return_position_index: bool = Field(default=False)
 
 
+@config_class()
+class TokenPreprocessingConfig(LengthPreprocessingConfig):
+    return_document_count: bool = Field(default=False)
+
+
 @config_class()
 class ImageNormalizationConfig(Config):
     scale: float = Field(default=255.0)
@@ -62,7 +67,7 @@ def get_batch_meta(self, size: int = 1) -> "PatchBatch":
 
 
 @config_class()
-class LanguageModelBatchPreprocessingConfig(LengthPreprocessingConfig):
+class LanguageModelBatchPreprocessingConfig(TokenPreprocessingConfig):
     _abstract = False
     phase: PhaseType = Field(default=PhaseType.training)
     micro_batch_splits: int = Field(default=1)

fast_llm/data/document/language_model.py

@@ -4,12 +4,14 @@
 
 import torch
 
+from fast_llm.core.distributed import allreduce_scalar
 from fast_llm.data.document.abstract import ModelInput
 from fast_llm.data.document.config import LanguageModelBatchPreprocessingConfig
 from fast_llm.data.document.patch import PatchBatch, PatchDocument, PatchModelInput
 from fast_llm.data.document.range import RangeBatch, RangeDocument
 from fast_llm.data.document.token import TokenBatch, TokenDocument, TokenModelInput
 from fast_llm.data.document.token_data import TokenDataBatch, TokenDataDocument
+from fast_llm.engine.distributed.distributed import Distributed
 from fast_llm.layers.language_model.config import LanguageModelKwargs
 from fast_llm.utils import div
 
@@ -33,13 +35,36 @@ class LanguageModelTargetInput(ModelInput):
     advantages: torch.Tensor | None = None
     old_log_probabilities: torch.Tensor | None = None
     label_counts: torch.Tensor | None = None
+    num_labels: int | None = None
+    num_labels_in_batch: int | None = None
+
+    @classmethod
+    def share_batch_data(cls, model_inputs: "list[LanguageModelTargetInput]", distributed: "Distributed"):
+        if model_inputs[0].num_labels is not None and model_inputs[0].num_labels_in_batch is None:
+            # We sum over sequences but not within a sequence.
+            num_labels_in_batch = allreduce_scalar(
+                sum(model_input.num_labels for model_input in model_inputs),
+                dtype=torch.int32,
+                group=distributed.batch_data_group,
+            )
+            for model_input in model_inputs:
+                model_input.num_labels_in_batch = num_labels_in_batch
 
 
 @dataclasses.dataclass(kw_only=True)
 class LanguageModelInput(TokenModelInput):
     targets: list[LanguageModelTargetInput] = dataclasses.field(default_factory=list)
     image_patches: PatchModelInput | None = None
 
+    @classmethod
+    def share_batch_data(cls, model_inputs: "list[LanguageModelInput]", distributed: "Distributed"):
+        super().share_batch_data(model_inputs, distributed)
+        for targets in zip(*(model_input.targets for model_input in model_inputs), strict=True):
+            targets[0].share_batch_data(targets, distributed)
+        model_inputs[0].image_patches.share_batch_data(
+            [model_input.image_patches for model_input in model_inputs], distributed
+        )
+
     def set_children_attributes(self) -> None:
         if self.image_patches is not None:
             self.image_patches.set_parent_attributes(self)
@@ -58,6 +83,7 @@ def to_kwargs(self) -> dict[str, typing.Any]:
             LanguageModelKwargs.advantages: [target.advantages for target in self.targets],
             LanguageModelKwargs.old_log_probabilities: [target.old_log_probabilities for target in self.targets],
             LanguageModelKwargs.label_counts: [target.label_counts for target in self.targets],
+            LanguageModelKwargs.num_labels_in_batch: [target.num_labels_in_batch for target in self.targets],
         }
         if self.image_patches is not None:
             out.update(self.image_patches.to_kwargs())
@@ -113,6 +139,12 @@ def get_model_inputs(self, config: LanguageModelBatchPreprocessingConfig) -> lis
             )
         ):
             model_input = self._get_model_input(sequence_k_past, sequence_k_past + local_input_length, config)
+            model_input.phase = config.phase
+
+            if config.use_image_patches:
+                model_input.image_patches = self.image_patches.get_model_input(
+                    sequence_k_past, sequence_k_past + local_input_length, config.vision_encoder
+                )
 
             model_input.pasts = presents
             presents = None if micro_sequence_index == config.micro_batch_splits - 1 else []
@@ -121,73 +153,66 @@ def get_model_inputs(self, config: LanguageModelBatchPreprocessingConfig) -> lis
 
             model_inputs.append(model_input)
 
+        self._set_target_inputs(model_inputs, config)
+
         return model_inputs
 
-    def _get_model_input(
-        self, begin: int, end: int, config: LanguageModelBatchPreprocessingConfig
-    ) -> LanguageModelInput:
-        model_input = super()._get_model_input(begin, end, config)
-        model_input.phase = config.phase
+    def _set_target_inputs(
+        self, model_inputs: list[LanguageModelInput], config: LanguageModelBatchPreprocessingConfig
+    ):
+        labels = self.tokens.clone()
 
-        if config.use_image_patches:
-            model_input.image_patches = self.image_patches.get_model_input(begin, end, config.vision_encoder)
+        # Apply loss masking spans.
+        if config.use_loss_masking_spans and self.loss_masking_spans is not None:
+            for span_begin, span_end in self.loss_masking_spans.ranges:
+                labels[span_begin:span_end] = -100
 
         for prediction_distance in range(1, config.num_labels + 1):
-            label_begin = begin + prediction_distance
-            label_end = end + prediction_distance
-            # Keep complete documents to simplify preprocessing.
-            _, first_document_begin, last_document_end = self._get_cropped_lengths(begin, label_end)
-            cropped_lengths, _, _ = self._get_cropped_lengths(first_document_begin, last_document_end)
-            labels = self.tokens[first_document_begin:last_document_end].clone()
-            labels_in_range = labels[label_begin - first_document_begin : label_end - first_document_begin]
-
-            # Apply loss masking spans.
-            if config.use_loss_masking_spans and self.loss_masking_spans is not None:
-                for span_begin, span_end in self.loss_masking_spans.get_cropped_ranges(
-                    first_document_begin, last_document_end
-                ):
-                    labels[span_begin:span_end] = -100
-
             # Mask cross-document predictions.
             document_begin = 0
-            for length in cropped_lengths:
-                labels[document_begin : document_begin + prediction_distance] = -100
+            for length in self.lengths:
+                labels[document_begin + prediction_distance - 1] = -100
                 document_begin += length
 
-            if config.return_label_counts:
-                # Count the number of non-masked labels in each document through cumulative sums.
-                mask = labels >= 0
-                mask_cumsum = torch.cat([mask.new_zeros(1), mask.cumsum(0)])
-                length_cumsum = torch.tensor([0] + cropped_lengths, device=self.device).cumsum(0)
-                label_count_cumsum = mask_cumsum[length_cumsum]
-                labels_per_document = label_count_cumsum[1:] - label_count_cumsum[:-1]
-                # Expand to one entry per token: find each token's document index via the sorted
-                # length cumsum, then look up that document's label count.
-                # TODO: Document index already computed in `LengthModelInputPreprocessor`.
-                document_index = torch.searchsorted(
-                    length_cumsum[1:], torch.arange(len(mask), device=self.device), side="right"
-                )
-                label_counts = labels_per_document[document_index][
-                    label_begin - first_document_begin : label_end - first_document_begin
-                ]
-                mask = (
-                    mask[label_begin - first_document_begin : label_end - first_document_begin]
-                    if config.return_prediction_mask
-                    else None
-                )
-            else:
-                label_counts = None
-                mask = labels_in_range >= 0 if config.return_prediction_mask else None
-
-            # Labels contain all four sources of masking: padding, user-defined spans, image placeholders, cross-document predictions.
-            target_input = LanguageModelTargetInput(tokens=labels_in_range, mask=mask, label_counts=label_counts)
-
-            if config.use_grpo_data and not model_input.is_meta:
-                target_input.advantages = self.advantages.get_cropped_data(label_begin, label_end)
-                target_input.old_log_probabilities = self.old_log_probabilities.get_cropped_data(
-                    label_begin, label_end
+            prediction_labels = labels[
+                prediction_distance : len(self.tokens) - config.num_labels + prediction_distance
+            ].clone()
+            mask = prediction_labels >= 0
+            label_counts = self._get_label_counts(mask) if config.return_label_counts else None
+
+            for input_index, model_input in enumerate(model_inputs):
+                begin = model_input.sequence_k_dim.size
+                end = begin + model_input.token_dim.size
+
+                # Labels contain all four sources of masking: padding, user-defined spans, image placeholders, cross-document predictions.
+                target_input = LanguageModelTargetInput(
+                    tokens=labels[begin:end],
+                    mask=mask[begin:end] if config.return_prediction_mask else None,
+                    label_counts=label_counts[begin:end] if config.return_label_counts else None,
+                    # Set value for the first input only so `share_batch_data` generated the correct sum.
+                    # TODO: ====== Make optional?
+                    num_labels=mask.sum(dtype=torch.int32).item() if input_index == 0 else 0,
                 )
-
-            model_input.targets.append(target_input)
-
-        return model_input
+                if config.use_grpo_data and not model_input.is_meta:
+                    target_input.advantages = self.advantages.get_cropped_data(
+                        begin + prediction_distance, end + prediction_distance
+                    )
+                    target_input.old_log_probabilities = self.old_log_probabilities.get_cropped_data(
+                        begin + prediction_distance, end + prediction_distance
+                    )
+
+                model_input.targets.append(target_input)
+
+    def _get_label_counts(self, mask: torch.Tensor):
+        # Count the number of non-masked labels in each document through cumulative sums.
+        mask_cumsum = torch.cat([mask.new_zeros(1), mask.cumsum(0)])
+        length_cumsum = torch.tensor([0] + self.lengths, device=self.device).cumsum(0)
+        label_count_cumsum = mask_cumsum[length_cumsum]
+        labels_per_document = label_count_cumsum[1:] - label_count_cumsum[:-1]
+        # Expand to one entry per token: find each token's document index via the sorted
+        # length cumsum, then look up that document's label count.
+        # TODO: Document index already computed in `LengthModelInputPreprocessor`.
+        document_index = torch.searchsorted(
+            length_cumsum[1:], torch.arange(len(mask), device=self.device), side="right"
+        )
+        return labels_per_document[document_index]

fast_llm/data/document/range.py

@@ -32,6 +32,6 @@ def from_documents(
             document_begin += size
         return cls(ranges=ranges) if ranges else None
 
-    def get_cropped_ranges(self, begin: int, end: int) -> list[tuple[int, int]]:
-        cropped_ranges = ((max(begin_ - begin, 0), min(end_ - begin, end - begin)) for begin_, end_ in self.ranges)
-        return [(begin_, end_) for begin_, end_ in cropped_ranges if end_ > begin_]
+    # def get_cropped_ranges(self, begin: int, end: int) -> list[tuple[int, int]]:
+    #    cropped_ranges = ((max(begin_ - begin, 0), min(end_ - begin, end - begin)) for begin_, end_ in self.ranges)
+    #    return [(begin_, end_) for begin_, end_ in cropped_ranges if end_ > begin_]

fast_llm/data/document/token.py

@@ -1,12 +1,14 @@
 import dataclasses
-import functools
 import typing
 
 import torch
 
+from fast_llm.core.distributed import allreduce_scalar
 from fast_llm.data.document.abstract import Batch, Document
 from fast_llm.data.document.block import BlockModelInput, LengthModelInputPreprocessor
-from fast_llm.data.document.config import LengthPreprocessingConfig
+from fast_llm.data.document.config import TokenPreprocessingConfig
+from fast_llm.engine.distributed.distributed import Distributed
+from fast_llm.layers.language_model.config import LanguageModelKwargs
 from fast_llm.tensor import TensorMeta
 from fast_llm.utils import Assert
 
@@ -22,12 +24,34 @@ def __len__(self) -> int:
     def device(self) -> torch.device:
         return self.tokens.device
 
+    @property
+    def is_meta(self) -> bool:
+        return self.device.type == "meta"
+
 
 @dataclasses.dataclass(kw_only=True)
 class TokenModelInput(BlockModelInput, TokenDocument):
-    @functools.cached_property
-    def is_meta(self) -> bool:
-        return isinstance(self.tokens, TensorMeta)
+    num_documents: int | None = None
+    num_documents_in_batch: int | None = None
+
+    @classmethod
+    def share_batch_data(cls, model_inputs: "list[TokenModelInput]", distributed: "Distributed"):
+        if model_inputs[0].num_documents is not None and model_inputs[0].num_documents_in_batch is None:
+            # We sum over sequences but not within a sequence.
+            num_documents_in_batch = allreduce_scalar(
+                sum(model_input.num_documents for model_input in model_inputs),
+                dtype=torch.int32,
+                group=distributed.batch_data_group,
+            )
+            for model_input in model_inputs:
+                model_input.num_documents_in_batch = num_documents_in_batch
+
+    def to_kwargs(self) -> dict[str, typing.Any]:
+        # TODO: Avoid conversion, use `LanguageModelMicroBatch` directly instead.
+        return {
+            **super().to_kwargs(),
+            LanguageModelKwargs.num_documents_in_batch: self.num_documents_in_batch,
+        }
 
 
 @dataclasses.dataclass(kw_only=True)
@@ -74,10 +98,13 @@ def _get_cropped_lengths(self, begin: int, end: int) -> tuple[list[int], int, in
 
         return lengths, first_document_begin, document_end
 
-    def _get_model_input(self, begin: int, end: int, config: LengthPreprocessingConfig):
+    def _get_model_input(self, begin: int, end: int, config: TokenPreprocessingConfig):
         model_input = self._model_input_class(tokens=self.tokens[begin:end])
         lengths, first_document_begin, last_document_end = self._get_cropped_lengths(begin, end)
 
+        if config.return_document_count:
+            model_input.num_documents = len(self.lengths) if begin == 0 else 0
+
         LengthModelInputPreprocessor(
             lengths=lengths,
             sequence_k_past=begin,
@@ -89,7 +116,7 @@ def _get_model_input(self, begin: int, end: int, config: LengthPreprocessingConf
         ).preprocess(model_input, config)
 
         Assert.eq(model_input.token_dim.size, end - begin)
-        if self.tokens.device.type == "meta":
+        if self.is_meta:
             model_input.tokens = TensorMeta.from_dims(
                 (model_input.token_dim,), tensor_name=f"tokens_{begin}_to_{end}", dtype=torch.int64
             )

fast_llm/engine/base_model/base_model.py

@@ -178,14 +178,13 @@ def __init__(
     @abc.abstractmethod
     def preprocess_batch(
         self,
-        model_inputs: list[ModelInput],
+        model_input: ModelInput,
         *,
         phase: PhaseType,
         iteration: int,
         metrics: dict | None = None,
         extra_kwargs: dict[str, typing.Any] | None = None,
-        device: torch.device | None,
-    ) -> list[tuple[torch.Tensor, dict]]:
+    ) -> tuple[torch.Tensor, dict]:
         # TODO Move batch splitting elsewhere, align interface with LayerBase
         pass
 

fast_llm/engine/inference/runner.py

@@ -1,6 +1,7 @@
 import abc
 import typing
 
+from fast_llm.data.document.abstract import ModelInput
 from fast_llm.engine.distributed.config import PhaseType
 from fast_llm.engine.multi_stage.fast_llm_model import FastLLMModel
 from fast_llm.engine.schedule.config import ScheduleConfig
@@ -57,15 +58,14 @@ def setup(self):
             Assert.is_(self._runner._distributed, self._fast_llm_model.distributed)
 
     def forward(
-        self, input_, kwargs: dict, *, iteration: int = 1, return_metrics: bool = False
+        self, model_input: ModelInput, *, iteration: int = 1, return_metrics: bool = False
     ) -> tuple[dict[str, float | int], dict[str, typing.Any] | None]:
         # TODO: Return an actual model output.
         reduced_losses, update_successful, metrics = self._runner.run_step(
-            iter((((input_, kwargs),),)),
+            iter(((model_input,),)),
             self._schedule,
             iteration=iteration,
             return_metrics=return_metrics,
-            preprocessed=True,
         )
         assert update_successful
         return reduced_losses, metrics