playground/Abbie-b200/trainer_utils/tivila_transform.py

import base64
import io
import json
import logging
import os
import traceback
import warnings
from copy import deepcopy
from functools import cached_property
from typing import List, Optional

import numpy as np
import torch
from arpeggio import Chord, TransformBase, register_transform
from arpeggio.utils.conversation_utils import chatml_input_ids_to_labels
from arpeggio.utils.qwen_vl_utils import get_mrope_index, get_mrope_index_qwen3_vl
from PIL import Image
from qwen_vl_utils import fetch_image, process_vision_info
from transformers import AutoConfig, AutoProcessor
from transformers.configuration_utils import PretrainedConfig
from transformers.processing_utils import ProcessorMixin
from transformers.tokenization_utils_base import PreTrainedTokenizerBase


DEFAULT_PROMPT_KEY = os.getenv("DEFAULT_PROMPT_KEY", "messages")

SYSTEM_PROMPT = '''
You are an assistant that must first decide whether reasoning is necessary before answering.

You must always start with a <choose>...</choose> tag.

If the question requires reasoning, complex analysis, multi-step deduction, calculation, or careful verification, output:

<choose>I need to think.</choose>
<think>
your reasoning process
</think>
answer

If the question is simple, factual, direct, or does not require reasoning, output:

<choose>I don't need to think.</choose>
answer

Rules:
- Never skip the <choose> tag.
- Only output <think> when you selected "I need to think."
'''


def decode_image_base64(image_base64: str) -> Image.Image:
    image_buf = base64.b64decode(image_base64)
    with io.BytesIO(image_buf) as bio:
        return Image.open(bio).convert("RGB")


class TiViLATransform(TransformBase):
    def __init__(
        self,
        processor: ProcessorMixin,
        model_config: PretrainedConfig,
        prompt_key: str = DEFAULT_PROMPT_KEY,
        add_raw_prompt: bool = False,
        allow_skip: bool = True,
        image_min_pixels: Optional[int] = None,
        image_max_pixels: Optional[int] = None,
        video_min_pixels: Optional[int] = None,
        video_max_pixels: Optional[int] = 307200,
        video_min_frames: Optional[int] = None,
        video_max_frames: Optional[int] = 16,
        max_seq_len: Optional[int] = None,
        **unused_kwargs,
    ):
        self.processor = processor
        self.model_config = model_config
        self.prompt_key = prompt_key
        self.add_raw_prompt = add_raw_prompt
        self.allow_skip = allow_skip
        self.max_seq_len = max_seq_len

        self.video_max_frames = video_max_frames
        self._image_ele_kwargs = {
            "min_pixels": image_min_pixels,
            "max_pixels": image_max_pixels,
        }
        self._video_ele_kwargs = {
            "min_pixels": video_min_pixels,
            "max_pixels": video_max_pixels,
            "min_frames": video_min_frames,
            "max_frames": video_max_frames,
        }

        self._image_ele_kwargs = {k: v for k, v in self._image_ele_kwargs.items() if v is not None}
        self._video_ele_kwargs = {k: v for k, v in self._video_ele_kwargs.items() if v is not None}

    @property
    def tokenizer(self) -> PreTrainedTokenizerBase:
        return self.processor.tokenizer

    @cached_property
    def _patch_size(self) -> int:
        return self.model_config.vision_config.patch_size

    @cached_property
    def _vision_start_id(self) -> int:
        return self.tokenizer.encode("<|vision_start|>", add_special_tokens=False)[0]

    @cached_property
    def _vision_end_id(self) -> int:
        return self.tokenizer.encode("<|vision_end|>", add_special_tokens=False)[0]

    @cached_property
    def _assistant_token_id(self) -> int:
        return self.tokenizer.encode("assistant", add_special_tokens=False)[0]

    @cached_property
    def _bos_token_id(self) -> int:
        return self.tokenizer.encode("<|im_start|>", add_special_tokens=False)[0]

    def process_input_ids_to_labels(self, input_ids: List[int]) -> List[int]:
        # Qwen uses ChatML
        return chatml_input_ids_to_labels(
            input_ids=input_ids,
            assistant_token_id=self._assistant_token_id,
            bos_token_id=self._bos_token_id,
            eos_token_id=self.tokenizer.eos_token_id,
        )

    def _input_ids_to_labels_pretrain(self, input_ids: List[int]) -> List[int]:
        # For stage1&stage2 interleaved data
        vision_start_id = self._vision_start_id
        vision_end_id = self._vision_end_id

        labels = []
        in_vision_block = False

        for i, token_id in enumerate(input_ids):
            if token_id == vision_start_id:
                in_vision_block = True
                labels.append(-100)
            elif token_id == vision_end_id:
                labels.append(-100)
                in_vision_block = False
            elif in_vision_block:
                labels.append(-100)
            else:
                labels.append(token_id)

        if labels and labels[0] != -100:
            labels[0] = -100

        return labels

    def make_model_inputs(
        self,
        text: str,
        images: Optional[List[Image.Image]] = None,
        videos: Optional[List[torch.Tensor]] = None,
        video_kwargs: Optional[dict] = None,
        video_metadatas: Optional[dict] = None,
        item: Optional[dict] = None,
        pt_label_func=None,
    ) -> Chord:
        if images is not None and len(images) == 0:
            images = None
        if videos is not None and len(videos) == 0:
            videos = video_metadatas = None
            video_kwargs = {}
        if video_kwargs is None:
            video_kwargs = {}

        # Process with model processor
        model_inputs = self.processor(
            text=text,
            images=images,
            videos=videos,
            video_metadata=video_metadatas,
            do_resize=False,
            **video_kwargs,
        )

        # Form labels
        input_ids = model_inputs["input_ids"][0]

        if pt_label_func is not None:
            labels = pt_label_func(input_ids)
        else:
            labels = self.process_input_ids_to_labels(input_ids)

        model_inputs["labels"] = [labels]
        model_inputs = model_inputs.convert_to_tensors("pt")

        # A bit out of place but seq_len will be used later
        seq_len = len(input_ids)

        if "qwen3" in self.model_config.model_type:
            model_inputs["position_ids"] = get_mrope_index_qwen3_vl(
                config=self.model_config,
                input_ids=model_inputs["input_ids"][0],
                image_grid_thw=model_inputs.get("image_grid_thw"),
                video_grid_thw=model_inputs.get("video_grid_thw"),
            )[:, None]  # [3, 1, seq_len]

        else:
            # Otherwise assume qwen2.5-vl
            model_inputs["position_ids"] = get_mrope_index(
                config=self.model_config,
                input_ids=model_inputs["input_ids"][0],
                image_grid_thw=model_inputs.get("image_grid_thw"),
                video_grid_thw=model_inputs.get("video_grid_thw"),
                second_per_grid_ts=model_inputs.get("second_per_grid_ts"),
            )[:, None]  # [3, 1, seq_len]

        # Fill stuff into extra_info
        extra_info = item.pop("extra_info", {})
        assert isinstance(extra_info, dict), "extra_info should be a dictionary"

        extra_info["seq_len"] = seq_len
        model_inputs["extra_info"] = [extra_info]

        # Emplace other keys into item
        for key, value in item.items():
            assert key not in model_inputs, (
                f"`{key}` conflicts with `model_inputs`. Please rename this field in your dataset."
            )
            model_inputs[key] = [value]

        return model_inputs

    def process_conversation(self, item: dict) -> Chord:
        messages = item.pop(self.prompt_key)
        if isinstance(messages, str):
            messages = json.loads(messages)

        for message in messages:
            content = message["content"]
            if isinstance(content, str):
                content = [{"type": "text", "text": content}]

            for i, part in enumerate(content):
                part_type = part["type"]

                if part_type == "text":
                    text = part["text"]

                    # add choose logic
                    if message.get("role") == "assistant":
                        if "</think>" in text:
                            text = "<choose>I need to think.</choose>\n" + text
                        else:
                            text = "<choose>I don't need to think.</choose>\n" + text

                    content[i] = {"type": "text", "text": text}

                elif part_type == "image":
                    # TODO: Decode this to PIL Image
                    img = decode_image_base64(part["image"])

                    content[i] = {
                        "type": "image",
                        "image": img,
                        **self._image_ele_kwargs,
                    }

                elif part_type == "video":
                    frames = [decode_image_base64(b) for b in part["video"]]

                    # Down sample frames if necessary
                    sample_fps = 1.0
                    if len(frames) > self.video_max_frames:
                        sample_idxs = np.linspace(
                            0,
                            len(frames) - 1,
                            self.video_max_frames,
                            dtype=int,
                        ).tolist()
                        sample_fps = len(sample_idxs) / len(frames)
                        frames = [frames[i] for i in sample_idxs]

                    content[i] = {
                        "type": "video",
                        "video": frames,
                        "sample_fps": sample_fps,
                        **self._video_ele_kwargs,
                    }

                else:
                    raise NotImplementedError(f"invalid part_type={part_type}")

        # messages = [{"role": "system", "content": SYSTEM_PROMPT}] + messages
        
        text = self.processor.apply_chat_template(messages, tokenize=False)

        images, videos, video_kwargs = process_vision_info(
            messages,
            image_patch_size=self._patch_size,
            return_video_kwargs=True,
            return_video_metadata=True,
        )

        if videos:
            videos, video_metadatas = zip(*videos)
            videos, video_metadatas = list(videos), list(video_metadatas)
        else:
            videos = video_metadatas = None

        # In some cases, extra_info is stored as a list
        extra_info = item["extra_info"]
        if isinstance(extra_info, list):
            item["extra_info"] = extra_info[0]

        return self.make_model_inputs(
            text=text,
            images=images,
            videos=videos,
            video_kwargs=video_kwargs,
            video_metadatas=video_metadatas,
            item=item,
        )

    def process_interleaved(self, item: dict) -> Chord:
        text_list = item.pop("texts")
        image_list = item.pop("images")
        assert len(text_list) == len(image_list), "text and image lists are not the same length"

        text_segments = []
        images = []

        for text, image_bytes in zip(text_list, image_list):
            if text is not None:
                text_segments.append(text)

            if image_bytes is not None and image_bytes != b"\x00":
                img = Image.open(io.BytesIO(image_bytes))
                width, height = img.size
                ratio = max(width / height, height / width)
                if ratio > 6:
                    continue

                img = img.convert("RGB")
                image_ele = {"type": "image", "image": img, **self._image_ele_kwargs}
                img = fetch_image(image_ele, image_patch_size=self._patch_size)

                images.append(img)

                ("data:image/jpeg;base64,{BASE64_IMAGE}",)

                text_segments.append("<|vision_start|><|image_pad|><|vision_end|>")

        text = "".join(text_segments)
        text = "<|im_start|>" + text + "<|im_end|>"

        return self.make_model_inputs(
            text=text,
            images=images,
            item=item,
            pt_label_func=self._input_ids_to_labels_pretrain,
        )

    def make_dummy_sample(self) -> Chord:
        # This is mainly used in map-style datasets to patch attention backward
        return {
            "input_ids": torch.zeros(1, 1, dtype=torch.long),
            "position_ids": torch.zeros(3, 1, 1, dtype=torch.long),  # Assume MRoPE
            "attention_mask": torch.ones(1, 1, dtype=torch.long),
            "labels": torch.zeros(1, 1, dtype=torch.long) - 100,
            "extra_info": [
                {
                    "file_name": "skip/skip",
                    "dataset_name": "skip",
                    "seq_len": 1,
                }
            ],
        }

    def _preprocess(self, item: dict) -> Chord:
        item = deepcopy(item)
        if self.prompt_key in item:
            sample = self.process_conversation(item)
        else:
            sample = self.process_interleaved(item)

        if self.max_seq_len is not None:
            seq_len = sample["input_ids"].size(-1)
            if seq_len > self.max_seq_len:
                extra_info = sample["extra_info"][0]
                msg = f"Found sample of length {seq_len} extra_info={extra_info}"
                warnings.warn(msg, UserWarning)
                sample = self.make_dummy_sample()

        return sample

    def preprocess(self, item: dict) -> Chord:
        try:
            return self._preprocess(item)
        except Exception as e:
            if not self.allow_skip:
                raise e

            err_trace = traceback.format_exc()
            logging.warning(f"Skipped sample due to {e}: {err_trace}")
            return {
                "input_ids": torch.zeros(1, 0, dtype=torch.long),
                "position_ids": torch.zeros(3, 1, 0, dtype=torch.long),
                "attention_mask": torch.zeros(1, 0, dtype=torch.long),
                "labels": torch.zeros(1, 0, dtype=torch.long),
                "extra_info": [{}],
            }

    @classmethod
    def from_pretrained(cls, pretrained_path, **kwargs):
        processor = AutoProcessor.from_pretrained(pretrained_path)
        config = AutoConfig.from_pretrained(pretrained_path)
        return cls(processor=processor, model_config=config, **kwargs)

    def save_pretrained(self, pretrained_path: str):
        self.processor.save_pretrained(pretrained_path)
        self.model_config.save_pretrained(pretrained_path)


register_transform("tivila", TiViLATransform)