collator.py

# Copyright 2025 OpenAccess AI Collective and the LlamaFactory team.
#
# This code is inspired by the OpenAccess AI Collective's axolotl library.
# https://github.com/OpenAccess-AI-Collective/axolotl/blob/main/src/axolotl/monkeypatch/utils.py
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, Literal, Optional

import numpy as np
import torch
import torch.nn.functional as F
from transformers import DataCollatorForSeq2Seq

from ..extras.constants import AUDIO_PLACEHOLDER, IGNORE_INDEX, IMAGE_PLACEHOLDER
from ..extras.misc import get_current_device
from ..extras.packages import is_pillow_available


if is_pillow_available():
    from PIL import Image


if TYPE_CHECKING:
    from transformers import ProcessorMixin

    from .template import Template


def prepare_4d_attention_mask(attention_mask_with_indices: "torch.Tensor", dtype: "torch.dtype") -> "torch.Tensor":
    r"""Expand 2d attention mask to 4d attention mask.

    Expand the attention mask with indices from (batch_size, seq_len) to (batch_size, 1, seq_len, seq_len),
    handle packed sequences and transforms the mask to lower triangular form to prevent future peeking.

    e.g.
    ```python
    # input
    [[1, 1, 2, 2, 2, 0]]
    # output
    [
        [
            [
                [o, x, x, x, x, x],
                [o, o, x, x, x, x],
                [x, x, o, x, x, x],
                [x, x, o, o, x, x],
                [x, x, o, o, o, x],
                [x, x, x, x, x, x],
            ]
        ]
    ]
    ```
    where `o` equals to `0.0`, `x` equals to `min_dtype`.
    """
    _, seq_len = attention_mask_with_indices.size()

    # Move to compute device if the source is CPU.
    source_device = attention_mask_with_indices.device
    compute_device = get_current_device() if source_device.type == "cpu" else source_device
    if compute_device != source_device:
        attention_mask_with_indices = attention_mask_with_indices.to(compute_device)

    min_dtype = torch.finfo(dtype).min
    zero_tensor = torch.tensor(0, dtype=dtype, device=compute_device)

    # Create a non-padding mask.
    non_padding = (attention_mask_with_indices != 0).unsqueeze(1).unsqueeze(2)
    # Create indices for comparison.
    indices = attention_mask_with_indices.unsqueeze(1).unsqueeze(2)  # [bsz, 1, 1, seq_len]
    indices_t = attention_mask_with_indices.unsqueeze(1).unsqueeze(3)  # [bsz, 1, seq_len, 1]
    # Create a lower triangular mask.
    tril_mask = torch.tril(torch.ones((seq_len, seq_len), dtype=torch.bool, device=compute_device))
    attention_mask_4d = (indices == indices_t) & non_padding & tril_mask
    # Invert the attention mask.
    attention_mask_4d = torch.where(attention_mask_4d, zero_tensor, min_dtype)

    # Move back to original device if needed.
    if compute_device != source_device:
        attention_mask_4d = attention_mask_4d.to(source_device)
    return attention_mask_4d


@dataclass
class MultiModalDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
    r"""Data collator that supports VLMs.

    Features should contain input_ids, attention_mask, labels, and optionally contain images, videos and audios.
    """

    template: Optional["Template"] = None
    processor: Optional["ProcessorMixin"] = None

    def __post_init__(self):
        if self.template is None:
            raise ValueError("Template is required for MultiModalDataCollator.")

    def __call__(self, features: list[dict[str, Any]]) -> dict[str, "torch.Tensor"]:
        batch_images, batch_videos, batch_audios = [], [], []
        batch_imglens, batch_vidlens, batch_audlens, batch_input_ids = [], [], [], []
        for feature in features:
            images = feature.pop("images", None) or []
            videos = feature.pop("videos", None) or []
            audios = feature.pop("audios", None) or []
            batch_images.extend(images)
            batch_videos.extend(videos)
            batch_audios.extend(audios)
            batch_imglens.append(len(images))
            batch_vidlens.append(len(videos))
            batch_audlens.append(len(audios))
            batch_input_ids.append(feature["input_ids"])

        fake_input_ids = []
        if (
            self.template.mm_plugin.image_token is not None and sum(batch_imglens) == 0 and sum(batch_vidlens) == 0
        ):  # avoid process hanging in zero3/fsdp case
            fake_messages = [{"role": "user", "content": IMAGE_PLACEHOLDER}]
            fake_images = [Image.new("RGB", (64, 64), (255, 255, 255))]
            fake_messages = self.template.mm_plugin.process_messages(
                fake_messages, fake_images, [], [], self.processor
            )
            _fake_input_ids = self.tokenizer.encode(fake_messages[0]["content"], add_special_tokens=False)
            _fake_input_ids, _ = self.template.mm_plugin.process_token_ids(
                _fake_input_ids, None, fake_images, [], [], self.tokenizer, self.processor
            )
            fake_input_ids.extend(_fake_input_ids)
            batch_images = fake_images
            batch_imglens[0] = 1

        if (
            self.template.mm_plugin.audio_token is not None and sum(batch_audlens) == 0
        ):  # avoid process hanging in zero3/fsdp case
            fake_messages = [{"role": "user", "content": AUDIO_PLACEHOLDER}]
            fake_audios = [np.zeros(1600)]
            fake_messages = self.template.mm_plugin.process_messages(
                fake_messages, [], [], fake_audios, self.processor
            )
            _fake_input_ids = self.tokenizer.encode(fake_messages[0]["content"], add_special_tokens=False)
            _fake_input_ids, _ = self.template.mm_plugin.process_token_ids(
                _fake_input_ids, None, [], [], fake_audios, self.tokenizer, self.processor
            )
            fake_input_ids.extend(_fake_input_ids)
            batch_audios = fake_audios
            batch_audlens[0] = 1

        if len(fake_input_ids) != 0:
            if self.tokenizer.padding_side == "right":
                features[0]["input_ids"] = features[0]["input_ids"] + fake_input_ids
                features[0]["attention_mask"] = features[0]["attention_mask"] + [0] * len(fake_input_ids)
                features[0]["labels"] = features[0]["labels"] + [IGNORE_INDEX] * len(fake_input_ids)
            else:
                features[0]["input_ids"] = fake_input_ids + features[0]["input_ids"]
                features[0]["attention_mask"] = [0] * len(fake_input_ids) + features[0]["attention_mask"]
                features[0]["labels"] = [IGNORE_INDEX] * len(fake_input_ids) + features[0]["labels"]

            batch_input_ids[0] = features[0]["input_ids"]

        mm_inputs = self.template.mm_plugin.get_mm_inputs(
            batch_images,
            batch_videos,
            batch_audios,
            batch_imglens,
            batch_vidlens,
            batch_audlens,
            batch_input_ids,
            self.processor,
        )
        if "token_type_ids" in mm_inputs:
            token_type_ids = mm_inputs.pop("token_type_ids")
            for i, feature in enumerate(features):
                feature["token_type_ids"] = token_type_ids[i]

        features: dict[str, torch.Tensor] = super().__call__(features)

        if self.model is not None and hasattr(self.model, "get_rope_index"):  # for qwen2vl mrope
            rope_index_kwargs = {
                "input_ids": features["input_ids"],
                "image_grid_thw": mm_inputs.get("image_grid_thw"),
                "video_grid_thw": mm_inputs.get("video_grid_thw"),
                "attention_mask": features["attention_mask"],
            }
            if "second_per_grid_ts" in mm_inputs:  # for qwen2vl
                rope_index_kwargs["second_per_grid_ts"] = mm_inputs.get("second_per_grid_ts")
            if "video_second_per_grid" in mm_inputs:  # for qwen2omni
                rope_index_kwargs["second_per_grids"] = mm_inputs.get("video_second_per_grid")

            if getattr(self.model.config, "model_type", None) == "qwen2_5_omni_thinker":  # for qwen2omni
                feature_attention_mask = mm_inputs.get("feature_attention_mask", None)
                if feature_attention_mask is not None:
                    audio_feature_lengths = torch.sum(
                        feature_attention_mask, dim=1
                    )  # FIXME need to get video image lengths
                    rope_index_kwargs["audio_seqlens"] = audio_feature_lengths  # prepare for input

                delta0 = (1 - rope_index_kwargs["attention_mask"]).sum(dim=-1).unsqueeze(1)
                # avoid conflict
                new_position_ids, rope_deltas = self.model.get_rope_index(**rope_index_kwargs)
                features["position_ids"], features["rope_deltas"] = (
                    new_position_ids.clone(),
                    rope_deltas - delta0,
                )  # avoid inplace operation FIXME
            else:  # for qwen2vl
                features["position_ids"], features["rope_deltas"] = self.model.get_rope_index(**rope_index_kwargs)

        if "cross_attention_mask" in mm_inputs:  # for mllama inputs when pad_to_multiple_of is enabled
            cross_attention_mask = mm_inputs.pop("cross_attention_mask")
            seq_len = features["input_ids"].size(1)
            orig_len = cross_attention_mask.size(1)
            mm_inputs["cross_attention_mask"] = F.pad(cross_attention_mask, (0, 0, 0, 0, 0, seq_len - orig_len))

        features.update(mm_inputs)

        if "image_bound" in features:  # for minicpmv inputs
            bsz, seq_length = features["input_ids"].shape
            features["position_ids"] = torch.arange(seq_length).long().repeat(bsz, 1)
            return {"data": features, "input_ids": features["input_ids"], "labels": features["labels"]}

        return features


@dataclass
class SFTDataCollatorWith4DAttentionMask(MultiModalDataCollatorForSeq2Seq):
    r"""Data collator for 4d attention mask."""

    block_diag_attn: bool = False
    attn_implementation: Literal["eager", "sdpa", "flash_attention_2"] = "eager"
    compute_dtype: "torch.dtype" = torch.float32

    def __call__(self, features: list[dict[str, Any]]) -> dict[str, "torch.Tensor"]:
        features = super().__call__(features)
        if self.block_diag_attn and self.attn_implementation != "flash_attention_2":
            features["attention_mask"] = prepare_4d_attention_mask(features["attention_mask"], self.compute_dtype)

        for key, value in features.items():  # cast data dtype for paligemma
            if torch.is_tensor(value) and torch.is_floating_point(value):
                features[key] = value.to(self.compute_dtype)

        return features


@dataclass
class PairwiseDataCollatorWithPadding(MultiModalDataCollatorForSeq2Seq):
    r"""Data collator for pairwise data."""

    def __call__(self, features: list[dict[str, Any]]) -> dict[str, "torch.Tensor"]:
        r"""Pad batched data to the longest sequence in the batch.

        We generate 2 * n examples where the first n examples represent chosen examples and
        the last n examples represent rejected examples.
        """
        concatenated_features = []
        for key in ("chosen", "rejected"):
            for feature in features:
                target_feature = {
                    "input_ids": feature[f"{key}_input_ids"],
                    "attention_mask": feature[f"{key}_attention_mask"],
                    "labels": feature[f"{key}_labels"],
                    "images": feature["images"],
                    "videos": feature["videos"],
                    "audios": feature["audios"],
                }
                concatenated_features.append(target_feature)

        return super().__call__(concatenated_features)


@dataclass
class KTODataCollatorWithPadding(MultiModalDataCollatorForSeq2Seq):
    r"""Data collator for KTO data."""

    def __call__(self, features: list[dict[str, Any]]) -> dict[str, "torch.Tensor"]:
        target_features = []
        kl_features = []
        kto_tags = []
        for feature in features:
            target_feature = {
                "input_ids": feature["input_ids"],
                "attention_mask": feature["attention_mask"],
                "labels": feature["labels"],
                "images": feature["images"],
                "videos": feature["videos"],
                "audios": feature["audios"],
            }
            kl_feature = {
                "input_ids": feature["kl_input_ids"],
                "attention_mask": feature["kl_attention_mask"],
                "labels": feature["kl_labels"],
                "images": feature["images"],
                "videos": feature["videos"],
                "audios": feature["audios"],
            }
            target_features.append(target_feature)
            kl_features.append(kl_feature)
            kto_tags.append(feature["kto_tags"])

        batch = super().__call__(target_features)
        kl_batch = super().__call__(kl_features)
        batch["kl_input_ids"] = kl_batch["input_ids"]
        batch["kl_attention_mask"] = kl_batch["attention_mask"]
        batch["kl_labels"] = kl_batch["labels"]
        if "cross_attention_mask" in kl_batch:  # for mllama inputs.
            batch["kl_cross_attention_mask"] = kl_batch["cross_attention_mask"]
        if "token_type_ids" in kl_batch:
            batch["kl_token_type_ids"] = kl_batch["token_type_ids"]

        batch["kto_tags"] = torch.tensor(kto_tags)
        return batch