modeling_sa2va_qwen.py

import warnings
from typing import Any, List, Optional, Tuple, Union

import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode

import torch.utils.checkpoint
import transformers

from .modeling_internlm2 import InternLM2ForCausalLM
from .modeling_phi3 import Phi3ForCausalLM
from peft import LoraConfig, get_peft_model
from torch import nn
from torch.nn import CrossEntropyLoss
from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM,
                          LlamaTokenizer, Qwen2ForCausalLM)
from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import ModelOutput, logging
from transformers import StoppingCriteriaList, StoppingCriteria, AutoProcessor
from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import Qwen2_5_VisionTransformerPretrainedModel

from .configuration_sa2va_chat import Sa2VAChatConfig
from .modeling_intern_vit import InternVisionModel, has_flash_attn

from .sam2 import SAM2
from .templates import PROMPT_TEMPLATE

import numpy as np
from torchvision.transforms.functional import resize, to_pil_image

from types import MethodType
import torch.nn.functional as F

try:
    from .flash_attention import FlashAttention
    has_flash_attn = True
except:
    print('FlashAttention is not installed.')
    has_flash_attn = False

logger = logging.get_logger(__name__)

def version_cmp(v1, v2, op='eq'):
    import operator

    from packaging import version
    op_func = getattr(operator, op)
    return op_func(version.parse(v1), version.parse(v2))

class StopWordStoppingCriteria(StoppingCriteria):
    """StopWord stopping criteria."""

    def __init__(self, tokenizer, stop_word):
        self.tokenizer = tokenizer
        self.stop_word = stop_word
        self.length = len(self.stop_word)

    def __call__(self, input_ids, *args, **kwargs) -> bool:
        cur_text = self.tokenizer.decode(input_ids[0])
        cur_text = cur_text.replace('\r', '').replace('\n', '')
        return cur_text[-self.length:] == self.stop_word

def get_stop_criteria(
    tokenizer,
    stop_words=[],
):
    stop_criteria = StoppingCriteriaList()
    for word in stop_words:
        stop_criteria.append(StopWordStoppingCriteria(tokenizer, word))
    return stop_criteria

class DirectResize:
    def __init__(self, target_length: int) -> None:
        self.target_length = target_length

    def apply_image(self, image: np.ndarray) -> np.ndarray:
        """
        Expects a numpy array with shape HxWxC in uint8 format.
        """
        img = to_pil_image(image, mode='RGB')
        return np.array(img.resize((self.target_length, self.target_length)))

class Sa2VAChatModelQwen(PreTrainedModel):
    config_class = Sa2VAChatConfig
    main_input_name = 'pixel_values'
    base_model_prefix = 'language_model'
    _no_split_modules = ['Qwen2_5_VisionTransformerPretrainedModel', 'Qwen2_5_VLDecoderLayer', 'SAM2']
    _supports_flash_attn_2 = True
    supports_gradient_checkpointing = True

    def __init__(self, config: Sa2VAChatConfig, visual=None, language_model=None, use_flash_attn=True):
        super().__init__(config)

        assert version_cmp(transformers.__version__, '4.49.0', 'ge')
        ##image_size = config.force_image_size or config.vision_config.image_size
        ##patch_size = config.vision_config.patch_size
        #self.patch_size = patch_size
        self.select_layer = config.select_layer
        self.template = config.template
        self.template = self.template.replace('-', '_')
        #self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2))
        #self.downsample_ratio = config.downsample_ratio
        #self.ps_version = config.ps_version
        self.llm_arch_name = config.llm_config.architectures[0]

        self.min_pixels = 512*28*28
        self.max_pixels = 2048*28*28

        use_flash_attn = use_flash_attn if has_flash_attn else False
        config.vision_config.use_flash_attn = True if use_flash_attn else False
        config.llm_config._attn_implementation = 'flash_attention_2' if use_flash_attn else 'eager'

        print(config.vision_config.keys())
        print('-'*33)
        print(config.keys())
        config.vision_config.spatial_merge_size=2

        self.merge_length = config.vision_config.spatial_merge_size ** 2

        #logger.info(f'num_image_token: {self.num_image_token}')
        #logger.info(f'ps_version: {self.ps_version}')
        if visual is not None:
            self.visual = visual
        else:
            self.visual = Qwen2_5_VisionTransformerPretrainedModel._from_config(config.vision_config)
        if language_model is not None:
            self.language_model = language_model
        else:
            if config.llm_config.architectures[0] == 'Qwen2ForCausalLM':
                self.language_model = Qwen2ForCausalLM(config.llm_config)

            elif config.llm_config.architectures[0] == 'Qwen3ForCausalLM':
                assert transformers.__version__ >= '4.56.0', 'Please upgrade transformers to >=4.56.0 for Qwen3 support.'
                from transformers import Qwen3ForCausalLM
                self.language_model = Qwen3ForCausalLM(config.llm_config)
            else:
                raise NotImplementedError(f'{config.llm_config.architectures[0]} is not implemented.')

        #vit_hidden_size = config.vision_config.hidden_size
        llm_hidden_size = config.llm_config.hidden_size

        self.lm_head = nn.Linear(llm_hidden_size, config.llm_config.vocab_size, bias=False)


        self.img_context_token_id = None
        self.conv_template = PROMPT_TEMPLATE[self.template]
        self.template = self.conv_template
        if hasattr(config, 'system_message'):
            self.system_message = config.system_message
        self.num_samples = 0

        if config.use_backbone_lora:
            self.wrap_backbone_lora(r=config.use_backbone_lora, lora_alpha=2 * config.use_backbone_lora)

        if config.use_llm_lora:
            self.wrap_llm_lora(r=config.use_llm_lora, lora_alpha=2 * config.use_llm_lora)

        self.grounding_encoder = SAM2()
        out_dim = self.grounding_encoder.hidden_dim
        in_dim = llm_hidden_size
        self.text_hidden_fcs = nn.Sequential(
            nn.Linear(in_dim, in_dim), nn.ReLU(inplace=True),
            nn.Linear(in_dim, out_dim), nn.Dropout(0.0)
        )

        self.init_prediction_config = False

    def wrap_backbone_lora(self, r=128, lora_alpha=256, lora_dropout=0.05):
        lora_config = LoraConfig(
            r=r,
            target_modules=['attn.qkv', 'attn.proj', 'mlp.fc1', 'mlp.fc2'],
            lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
        )
        self.visual = get_peft_model(self.visual, lora_config)
        self.visual.print_trainable_parameters()

    def wrap_llm_lora(self, r=128, lora_alpha=256, lora_dropout=0.05):
        # Determine the target modules based on the architecture of the language model
        if self.llm_arch_name == 'InternLM2ForCausalLM':
            target_modules = ['attention.wqkv', 'attention.wo', 'feed_forward.w1', 'feed_forward.w2', 'feed_forward.w3']
        elif self.llm_arch_name == 'Phi3ForCausalLM':
            target_modules = ['mlp.down_proj', 'mlp.gate_up_proj', 'self_attn.o_proj', 'self_attn.qkv_proj']
        elif self.llm_arch_name in ['Qwen2ForCausalLM', 'LlamaForCausalLM']:
            target_modules = ['self_attn.q_proj', 'self_attn.k_proj', 'self_attn.v_proj', 'self_attn.o_proj',
                              'mlp.gate_proj', 'mlp.down_proj', 'mlp.up_proj']
        else:
            raise NotImplemented
        lora_config = LoraConfig(
            r=r,
            target_modules=target_modules,
            lora_alpha=lora_alpha,
            lora_dropout=lora_dropout,
            task_type='CAUSAL_LM'
        )
        self.language_model = get_peft_model(self.language_model, lora_config)
        self.language_model.enable_input_require_grads()
        self.language_model.print_trainable_parameters()

    def extract_feature(self, pixel_values, image_grid_thw):
        if self.select_layer == -1:
            vit_embeds = self.visual(
                pixel_values=pixel_values,
                grid_thw=image_grid_thw)
        # else:
        #     vit_embeds = self.visual(
        #         pixel_values=pixel_values,
        #         output_hidden_states=True,
        #         return_dict=True).hidden_states[self.select_layer]
        # vit_embeds = vit_embeds[:, 1:, :]


        split_sizes = (image_grid_thw.prod(-1) // self.visual.spatial_merge_size**2).tolist()
        vit_embeds = torch.split(vit_embeds, split_sizes)

        return vit_embeds

    @property
    def lm_head(self):
        return self.language_model.get_output_embeddings()

    def get_input_embeddings(self):
        return self.language_model.get_input_embeddings()

    def get_output_embeddings(self):
        return self.language_model.get_output_embeddings()

    def forward(self, data, data_samples=None, mode='loss'):
        pixel_values = data['pixel_values']
        image_grid_thw = data['image_grid_thw']

        if type(pixel_values) is list or pixel_values.ndim == 5:
            if type(pixel_values) is list:
                pixel_values = [
                    x.unsqueeze(0) if x.ndim == 3 else x for x in pixel_values
                ]
            # b*n, c, h, w
            concat_images = torch.cat(
                [image.to(self.visual.dtype) for image in pixel_values], dim=0)
        else:
            raise NotImplementedError()

        input_ids = data['input_ids']
        position_ids = data['position_ids']
        attention_mask = data['attention_mask']
        # sum is 0 are text
        image_flags = torch.sum(concat_images, dim=(1, 2, 3)) != 0
        image_flags = image_flags.long()

        labels = data['labels']
        use_cache = False

        if 'vp_overall_mask' not in data.keys():
            vp_overall_mask = None
        else:
            vp_overall_mask = data['vp_overall_mask']

        if 'prompt_masks' in data.keys():
            prompt_masks = data['prompt_masks']
        else:
            prompt_masks = None

        outputs = self._llm_forward(
            input_ids=input_ids,
            position_ids=position_ids,
            attention_mask=attention_mask,
            image_flags=image_flags,
            pixel_values=concat_images,
            labels=labels,
            use_cache=use_cache,
            output_hidden_states=True,
            vp_overall_mask=vp_overall_mask,
            prompt_masks=prompt_masks,
            image_grid_thw=image_grid_thw,
        )

        return outputs

    def _llm_forward(
            self,
            pixel_values: torch.FloatTensor,
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            image_flags: Optional[torch.LongTensor] = None,
            past_key_values: Optional[List[torch.FloatTensor]] = None,
            labels: Optional[torch.LongTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            vp_overall_mask=None,
            prompt_masks=None,
            image_grid_thw=None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        return_dict = return_dict if return_dict is not None \
            else self.config.use_return_dict

        image_flags = image_flags.squeeze(-1)
        # We only added the clone code here to avoid the error.
        input_embeds = self.language_model.get_input_embeddings()(
            input_ids).clone()

        vit_embeds = self.extract_feature(pixel_values, image_grid_thw)
        vit_embeds = vit_embeds.to(input_embeds.dtype)  # FIXME: why vit_embeds is float16?
        fast_vit_embeds = None

        vit_embeds = vit_embeds[image_flags == 1]
        vit_batch_size = pixel_values.shape[0]

        B, N, C = input_embeds.shape
        input_embeds = input_embeds.reshape(B * N, C)

        self._count += 1

        if vp_overall_mask is not None and prompt_masks is not None:
            vp_embeds = []
            vp_overall_mask = vp_overall_mask.to(vit_embeds.device).bool()
            prompt_masks = [item.to(vit_embeds.device).bool() for item in prompt_masks]

            vp_overall_mask = vp_overall_mask[image_flags == 1]
            overall_tile_vit_embeds = vit_embeds[vp_overall_mask]  # (n_img, hw, c)

            i_vp_img = 0
            for i_img in range(len(vit_embeds)):
                vp_embeds.append(vit_embeds[i_img].reshape(-1, C))
                if vp_overall_mask[i_img]:
                    tile_vit_embeds = overall_tile_vit_embeds[i_vp_img].reshape(-1, C)  # (hw, C)
                    objects_prompt_masks = prompt_masks[i_vp_img]
                    n_obj = len(objects_prompt_masks)
                    tile_vit_embeds = tile_vit_embeds.unsqueeze(0).repeat(n_obj, 1, 1)
                    objects_prompt_masks = objects_prompt_masks.reshape(n_obj, -1)
                    vp_embeds.append(tile_vit_embeds[objects_prompt_masks])
                    i_vp_img += 1
            vp_embeds = torch.cat(vp_embeds, dim=0)
        else:
            vp_embeds = None

        input_ids = input_ids.reshape(B * N)
        selected = (input_ids == self.img_context_token_id)

        if vp_embeds is None:
            try:
                input_embeds[selected] = vit_embeds.reshape(-1, C)
            except Exception as e:
                vit_embeds = vit_embeds.reshape(-1, C)
                print(f'warning: {e}, input_embeds[selected].shape='
                      f'{input_embeds[selected].shape}, '
                      f'vit_embeds.shape={vit_embeds.shape}')
                n_token = selected.sum()
                if n_token > len(vit_embeds):
                    print(f"Wrong !!! {n_token} image tokens in text but only {len(vit_embeds)} vit embeds !!!")
                    expand_ratio = n_token // len(vit_embeds) + 1
                    vit_embeds = torch.cat([vit_embeds] * expand_ratio, dim=0)

                input_embeds[selected] = vit_embeds[:n_token]
        else:
            try:
                input_embeds[selected] = vp_embeds.reshape(-1, C)
            except Exception as e:
                vp_embeds = vp_embeds.reshape(-1, C)
                print(f'warning: {e}, input_embeds[selected].shape='
                      f'{input_embeds[selected].shape}, '
                      f'vp_embeds.shape={vp_embeds.shape}')
                n_token = selected.sum()
                if n_token > len(vp_embeds):
                    print(f"Wrong !!! {n_token} image tokens in text but only {len(vp_embeds)} vit embeds !!!")
                    expand_ratio = n_token // len(vp_embeds) + 1
                    vp_embeds = torch.cat([vp_embeds] * expand_ratio, dim=0)

                input_embeds[selected] = vp_embeds[:n_token]

        input_embeds = input_embeds.reshape(B, N, C)

        outputs = self.language_model(
            inputs_embeds=input_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        logits = outputs.logits

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(
                -1, self.language_model.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    @torch.no_grad()
    def generate(
            self,
            pixel_values: Optional[torch.FloatTensor] = None,
            input_ids: Optional[torch.FloatTensor] = None,
            attention_mask: Optional[torch.LongTensor] = None,
            visual_features: Optional[torch.FloatTensor] = None,
            generation_config: Optional[GenerationConfig] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            prompt_masks=None,
            vp_overall_mask=None,
            **generate_kwargs,
    ) -> torch.LongTensor:
        device = self.device
        assert self.img_context_token_id is not None

        if pixel_values is not None:
            if visual_features is not None:
                vit_embeds = visual_features
            else:
                if type(pixel_values) is list or pixel_values.ndim == 5:
                    if type(pixel_values) is list:
                        pixel_values = [
                            x.unsqueeze(0) if x.ndim == 3 else x for x in pixel_values
                        ]
                    # b*n, c, h, w
                    pixel_values = torch.cat(
                        [image.to(self.visual.dtype) for image in pixel_values], dim=0)

                vit_embeds = self.extract_feature(pixel_values.to(device))
            image_flags = torch.sum(pixel_values, dim=(1, 2, 3)) != 0
            image_flags = image_flags.long()
            vit_embeds = vit_embeds[image_flags == 1]

            input_embeds = self.language_model.get_input_embeddings()(input_ids.to(device))
            B, N, C = input_embeds.shape
            input_embeds = input_embeds.reshape(B * N, C)

            if vp_overall_mask is not None and prompt_masks is not None:
                vp_embeds = []
                vp_overall_mask = vp_overall_mask.to(vit_embeds.device).bool()
                prompt_masks = [item.to(vit_embeds.device).bool() for item in prompt_masks]

                vp_overall_mask = vp_overall_mask[image_flags == 1]
                overall_tile_vit_embeds = vit_embeds[vp_overall_mask]  # (n_img, hw, c)

                i_vp_img = 0
                for i_img in range(len(vit_embeds)):
                    vp_embeds.append(vit_embeds[i_img].reshape(-1, C))
                    if vp_overall_mask[i_img]:
                        tile_vit_embeds = overall_tile_vit_embeds[i_vp_img].reshape(-1, C)  # (hw, C)
                        objects_prompt_masks = prompt_masks[i_vp_img]
                        n_obj = len(objects_prompt_masks)
                        tile_vit_embeds = tile_vit_embeds.unsqueeze(0).repeat(n_obj, 1, 1)
                        objects_prompt_masks = objects_prompt_masks.reshape(n_obj, -1)
                        vp_embeds.append(tile_vit_embeds[objects_prompt_masks])
                        i_vp_img += 1

                vp_embeds = torch.cat(vp_embeds, dim=0)
            else:
                vp_embeds = None

            input_ids = input_ids.reshape(B * N)
            selected = (input_ids == self.img_context_token_id)
            assert selected.sum() != 0
            if vp_embeds is None:
                input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device)
            else:
                if len(input_embeds[selected]) != len(vp_embeds.reshape(-1, C)):
                    print("Shape mismatch, selected is {}, vp embeds is {} !!!" \
                          .format(len(input_embeds[selected]), len(vp_embeds.reshape(-1, C))))
                    min_tokens = min(len(input_embeds[selected]), len(vp_embeds.reshape(-1, C)))
                    input_embeds[selected][:min_tokens] = vp_embeds.reshape(-1, C)[:min_tokens].to(input_embeds.device)
                else:
                    input_embeds[selected] = vp_embeds.reshape(-1, C).to(input_embeds.device)

            input_embeds = input_embeds.reshape(B, N, C)
        else:
            input_embeds = self.language_model.get_input_embeddings()(input_ids)

        outputs = self.language_model.generate(
            inputs_embeds=input_embeds,
            attention_mask=attention_mask.to(device),
            generation_config=generation_config,
            output_hidden_states=output_hidden_states,
            # return_dict=return_dict,
            use_cache=True,
            **generate_kwargs,
        )

        return outputs

    def preparing_for_generation(self, tokenizer, max_new_tokens=2048, torch_dtype=torch.bfloat16):
        # set stop criteria and generation configs for model
        if not hasattr(self, 'tokenizer'):
            self.tokenizer = tokenizer
        self.bot_name = 'BOT'
        stop_words = []
        stop_words += self.template.get('STOP_WORDS', [])
        stop_criteria = get_stop_criteria(
            tokenizer=self.tokenizer, stop_words=stop_words)
        self.stop_criteria = stop_criteria

        default_generation_kwargs = dict(
            max_new_tokens=max_new_tokens,
            do_sample=False,
            eos_token_id=self.tokenizer.eos_token_id,
            pad_token_id=(
                self.tokenizer.pad_token_id
                if self.tokenizer.pad_token_id is not None
                else self.tokenizer.eos_token_id
            ),
        )

        self.gen_config = GenerationConfig(**default_generation_kwargs)
        self.gen_config.transformers_version = '4.56.1'
        self.init_prediction_config = True
        self.torch_dtype = torch_dtype
        # self.to(torch_dtype)
        self.extra_image_processor = DirectResize(target_length=1024, )
        # for multi image process

        #self.min_dynamic_patch = 1
        #self.max_dynamic_patch = 12
        #self.downsample_ratio = 0.5
        #self.image_size = 448
        self.use_thumbnail = True
        #patch_size = 14
        #self.patch_size = patch_size

        #self.patch_token = int((self.image_size // patch_size) ** 2 * (self.downsample_ratio ** 2))
        self.IMAGENET_MEAN = (0.485, 0.456, 0.406)
        self.IMAGENET_STD = (0.229, 0.224, 0.225)
        self.IMG_PAD_TOKEN = '<|image_pad|>'
        self.IMG_START_TOKEN = '<|vision_start|>'
        self.IMG_END_TOKEN = '<|vision_end|>'

        # self.transformer = T.Compose([
        #     T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        #     T.Resize((self.image_size, self.image_size), interpolation=InterpolationMode.BICUBIC),
        #     T.ToTensor(),
        #     T.Normalize(mean=self.IMAGENET_MEAN, std=self.IMAGENET_STD)
        # ])
        self.VP_START_TOKEN = '<vp>'
        self.VP_END_TOKEN = '</vp>'

        
        img_context_token_id = tokenizer.convert_tokens_to_ids('<|image_pad|>')
        self.img_context_token_id = img_context_token_id
        self.seg_token_idx = tokenizer.convert_tokens_to_ids('[SEG]')
        
        # Get special token IDs
        self.image_token_id = self.config.image_token_id
        self.video_token_id = self.config.video_token_id
        self.vision_start_token_id = self.config.vision_start_token_id
        
        # Vision config parameters
        self.spatial_merge_size = self.config.vision_config.spatial_merge_size
        self.temporal_patch_size = self.config.vision_config.temporal_patch_size
        self.patch_size = self.config.vision_config.patch_size

        return

    def predict_forward(
            self,
            image=None,
            video=None,
            text=None,
            past_text='',
            mask_prompts=None,
            tokenizer=None,
            processor=None,
    ):
        if not self.init_prediction_config:
            assert tokenizer
            self.preparing_for_generation(tokenizer=tokenizer)
            self.processor = processor

        if image is None and video is None and '<image>' not in past_text:
            text = text.replace('<image>', "")
            input_text = ''
            input_text += self.template['INSTRUCTION'].format(
                input=text, round=1, bot_name=self.bot_name)
            input_text = past_text + input_text
            ids = self.tokenizer.encode(input_text)
            ids = torch.tensor(ids).cuda().unsqueeze(0)

            attention_mask = torch.ones_like(ids, dtype=torch.bool)

            mm_inputs = {
                'pixel_values': None,
                'input_ids': ids,
                'attention_mask': attention_mask,
                'position_ids': None,
                'past_key_values': None,
                'labels': None,
                'prompt_masks': None,
                'vp_overall_mask': None,
                'image_grid_thw': None,
            }
            ret_masks = []
        else:
            input_dict = {}
            if video is not None:
                pixel_values = []
                extra_pixel_values = []
                images = []
                ori_image_size = video[0].size
                for frame_idx, frame_image in enumerate(video):
                    # assert ori_image_size == frame_image.size
                    g_image = np.array(frame_image)  # for grounding
                    g_image = self.extra_image_processor.apply_image(g_image)
                    g_image = torch.from_numpy(g_image).permute(2, 0, 1).contiguous()
                    extra_pixel_values.append(g_image)
                    if frame_idx < 5:
                        #img = self.transformer(frame_image)
                        #pixel_values.append(img)
                        images.append(frame_image)

                _data_dict = self.processor(
                    images=images, min_pixels=self.min_pixels, max_pixels=self.max_pixels
                )
                num_image_tokens = int(_data_dict['image_grid_thw'][0].prod()) // self.merge_length
                pixel_values = _data_dict['pixel_values']

                #pixel_values = torch.stack(pixel_values, dim=0).to(self.torch_dtype)  # (n_f, 3, h, w)
                g_pixel_values = torch.stack([
                    self.grounding_encoder.preprocess_image(pixel) for pixel in extra_pixel_values
                ]).to(self.torch_dtype)
                #num_image_tokens = self.patch_token
                num_frames = len(pixel_values)

                input_dict['vp_overall_mask'] = None
            else:
                #ori_image_size = image.size

                # prepare grounding images
                g_image = np.array(image)  # for grounding
                g_image = self.extra_image_processor.apply_image(g_image)
                g_pixel_values = torch.from_numpy(g_image).permute(2, 0, 1).contiguous().to(self.torch_dtype)
                extra_pixel_values = [g_pixel_values]
                g_pixel_values = torch.stack([
                    self.grounding_encoder.preprocess_image(pixel) for pixel in extra_pixel_values
                ]).to(self.torch_dtype)

                # images = dynamic_preprocess(image, self.min_dynamic_patch,
                #                             self.max_dynamic_patch,
                #                             self.image_size, self.use_thumbnail)

                if mask_prompts is not None:
                    vp_overall_mask = torch.Tensor([False] * (len(images) - 1) + [True])
                    input_dict['vp_overall_mask'] = vp_overall_mask
                else:
                    input_dict['vp_overall_mask'] = None

                #pixel_values = [self.transformer(image) for image in images]
                #pixel_values = torch.stack(pixel_values).to(self.torch_dtype)

                _data_dict = self.processor(
                    images=[image], min_pixels=self.min_pixels, max_pixels=self.max_pixels
                )

                num_image_tokens = int(_data_dict['image_grid_thw'][0].prod()) // self.merge_length
                num_frames = 1
            input_dict['g_pixel_values'] = g_pixel_values
            input_dict['pixel_values'] = _data_dict['pixel_values'].to(self.torch_dtype)

            if mask_prompts is not None:
                # reshape mask prompts to feature size
                mask_prompts = [torch.Tensor(item).to(pixel_values.device) for item in mask_prompts]
                mask_prompts = [F.interpolate(
                    item.unsqueeze(0),
                    size=(int(self.image_size // self.patch_size * self.downsample_ratio),
                          int(self.image_size // self.patch_size * self.downsample_ratio)),
                    mode='nearest').squeeze(0) for item in mask_prompts]
                region_pixels = []
                for mask_prompt in mask_prompts[0]:
                    region_pixels.append(mask_prompt.bool().to(torch.int64).sum())

                vp_token_str = '\nThere are {} part regions in the picture: '.format(len(mask_prompts[0]))
                for i in range(len(mask_prompts[0])):
                    vp_token_str = vp_token_str + \
                                   f"region{i + 1}" + self.VP_START_TOKEN + \
                                   self.IMG_PAD_TOKEN * region_pixels[i] + \
                                   self.VP_END_TOKEN
                    if i == len(mask_prompts[0]) - 1:
                        vp_token_str = vp_token_str + '.\n'
                    else:
                        vp_token_str = vp_token_str + ', '
            else:
                vp_token_str = ''

            image_token_str = f'{self.IMG_START_TOKEN}' \
                              f'{self.IMG_PAD_TOKEN * num_image_tokens}' \
                              f'{self.IMG_END_TOKEN}'
            image_token_str = image_token_str + '\n'
            image_token_str = image_token_str * num_frames
            image_token_str = image_token_str.strip()

            ret_masks = []

            if '<image>' in text or mask_prompts is not None:
                assert past_text is None or len(past_text) == 0
            text = text.replace('<image>', image_token_str + vp_token_str)
            input_text = ''
            input_text += self.template['INSTRUCTION'].format(
                input=text, round=1, bot_name=self.bot_name)
            input_text = past_text + input_text
            ids = self.tokenizer.encode(input_text)
            ids = torch.tensor(ids).cuda().unsqueeze(0)

            attention_mask = torch.ones_like(ids, dtype=torch.bool)

            mm_inputs = {
                'pixel_values': input_dict['pixel_values'],
                'input_ids': ids,
                'attention_mask': attention_mask,
                'position_ids': None,
                'past_key_values': None,
                'labels': None,
                'prompt_masks': mask_prompts,
                'vp_overall_mask': input_dict['vp_overall_mask'],
                'image_grid_thw': _data_dict['image_grid_thw'],
            }

        generate_output = self.generate(
            **mm_inputs,
            generation_config=self.gen_config,
            streamer=None,
            bos_token_id=self.tokenizer.bos_token_id,
            stopping_criteria=self.stop_criteria,
            output_hidden_states=True,
            return_dict_in_generate=True
        )
        predict = self.tokenizer.decode(
            generate_output.sequences[0], skip_special_tokens=False).strip()

        if image is None and video is None and '<image>' not in past_text:
            return {'prediction': predict, 'prediction_masks': ret_masks, }

        # if have seg result, find the seg hidden states
        hidden_states = generate_output.hidden_states
        last_hidden_states = [item[-1][0] for item in hidden_states]
        last_hidden_states = torch.cat(last_hidden_states, dim=0)
        seg_hidden_states = get_seg_hidden_states(
            last_hidden_states, generate_output.sequences[0][:-1],
            seg_id=self.seg_token_idx
        )
        all_seg_hidden_states = self.text_hidden_fcs(seg_hidden_states)

        for seg_hidden_states in all_seg_hidden_states:
            seg_hidden_states = seg_hidden_states.unsqueeze(0)
            g_pixel_values = input_dict['g_pixel_values']
            sam_states = self.grounding_encoder.get_sam2_embeddings(g_pixel_values)
            pred_masks = self.grounding_encoder.language_embd_inference(sam_states, [seg_hidden_states] * num_frames)
            w, h = ori_image_size
            masks = F.interpolate(pred_masks, size=(h, w), mode='bilinear', align_corners=False)
            masks = masks[:, 0]
            masks = masks.sigmoid() > 0.5
            masks = masks.cpu().numpy()
            ret_masks.append(masks)

        return {'prediction': predict, 'prediction_masks': ret_masks,}

def get_seg_hidden_states(hidden_states, output_ids, seg_id):
    seg_mask = output_ids == seg_id
    n_out = len(seg_mask)
    if n_out == 0:
        return hidden_states[0:0]
    return hidden_states[-n_out:][seg_mask]