modeling_internvl_chat.py

# --------------------------------------------------------
# InternVL
# Copyright (c) 2024 OpenGVLab
# Licensed under The MIT License [see LICENSE for details]
# --------------------------------------------------------

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

import torch
import torch.utils.checkpoint
import transformers
from torch import nn
from torch.nn import CrossEntropyLoss
from transformers import GenerationConfig
from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import logging
from transformers import LlamaForCausalLM, Qwen2ForCausalLM, Qwen3ForCausalLM, Qwen3MoeForCausalLM
from .configuration_internvl_chat import InternVLChatConfig
from .conversation import get_conv_template
from .modeling_intern_vit import InternVisionModel, has_flash_attn
from .speech_encoder import DualWrappedEncoder
from .speech_projector import EncoderProjectorConcat
from beats_model import BEATsConfig, BEATs


logger = logging.get_logger(__name__)

# Speech related constants
IGNORE_INDEX = -100
SPEECH_TOKEN_INDEX = -200
DEFAULT_SPEECH_TOKEN = "<speech>"

# Image related constants  
IMAGE_TOKEN_INDEX = -201


def tokenizer_speech_token(prompt, tokenizer, speech_token_index=SPEECH_TOKEN_INDEX, return_tensors=None):
    """Tokenize prompt with speech tokens, similar to OLA's implementation"""
    prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<speech>')]

    def insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]

    input_ids = []
    offset = 0
    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    for x in insert_separator(prompt_chunks, [speech_token_index] * (offset + 1)):
        input_ids.extend(x[offset:])

    if return_tensors is not None:
        if return_tensors == 'pt':
            return torch.tensor(input_ids, dtype=torch.long)
        raise ValueError(f'Unsupported tensor type: {return_tensors}')
    return input_ids


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))


# SpeechProjector is now imported from speech_projector.py as EncoderProjectorConcat


def build_speech_encoder(audio_config):
    if audio_config.speech_encoder_type is None:
        return None
    return DualWrappedEncoder(audio_config)


def build_speech_projector(audio_config, llm_hidden_size):
    # Check if fallback speech projector path is specified in config
    fallback_path = getattr(audio_config, 'speech_projector', None)
    
    if fallback_path and os.path.exists(fallback_path):
        print(f"Loading speech projector from fallback path: {fallback_path}")
        
        # Load the pretrained weights first to determine the expected dimensions
        import torch
        try:
            state_dict = torch.load(fallback_path, map_location='cpu')
            
            # Handle the 'model.speech_projector.' prefix in state_dict keys
            speech_projector_state_dict = {}
            for key, value in state_dict.items():
                if key.startswith('model.speech_projector.'):
                    # Remove the 'model.speech_projector.' prefix
                    new_key = key.replace('model.speech_projector.', '')
                    speech_projector_state_dict[new_key] = value
                else:
                    # If no prefix, use the key as-is
                    speech_projector_state_dict[key] = value
            
            # Determine the expected input dimensions from the loaded weights
            linear1_weight_shape = speech_projector_state_dict.get('linear1.weight', None)
            if linear1_weight_shape is not None:
                expected_input_dim = linear1_weight_shape.shape[1]  # [out_features, in_features]
                print(f"Detected expected input dimension from weights: {expected_input_dim}")
                
                # Calculate the encoder hidden size and ds_rate that match this
                # expected_input_dim = encoder_hidden_size * ds_rate
                # We know current encoder outputs 2048 dim (combined Whisper+BEATs)
                current_encoder_dim = 2048
                required_ds_rate = expected_input_dim // current_encoder_dim
                
                if expected_input_dim == current_encoder_dim * required_ds_rate:
                    print(f"Using ds_rate={required_ds_rate} to match loaded weights")
                    ds_rate = required_ds_rate
                    encoder_hidden_size = current_encoder_dim
                else:
                    print(f"Warning: Cannot perfectly match dimensions. Expected {expected_input_dim}, current encoder {current_encoder_dim}")
                    print(f"Will use closest match: ds_rate={required_ds_rate}")
                    ds_rate = max(1, required_ds_rate)  # Ensure at least 1
                    encoder_hidden_size = current_encoder_dim
            else:
                print("Warning: Could not determine input dimensions from weights, using defaults")
                ds_rate = 5
                encoder_hidden_size = 2048
                
        except Exception as e:
            print(f"Warning: Failed to analyze speech projector weights: {e}")
            print("Using default dimensions")
            ds_rate = 5
            encoder_hidden_size = 2048
        
        # Create a config with the determined dimensions
        class ConfigWrapper:
            def __init__(self, llm_hidden_size, ds_rate, encoder_hidden_size):
                self.speech_encoder_ds_rate = ds_rate
                self.speech_encoder_hidden_size = encoder_hidden_size
                self.hidden_size = llm_hidden_size
        
        wrapper_config = ConfigWrapper(llm_hidden_size, ds_rate, encoder_hidden_size)
        projector = EncoderProjectorConcat(wrapper_config)
        
        # Load the weights
        try:
            projector.load_state_dict(speech_projector_state_dict, strict=False)
            print(f"Successfully loaded speech projector weights from {fallback_path}")
        except Exception as e:
            print(f"Warning: Failed to load speech projector weights: {e}")
            print("Using randomly initialized speech projector")
        
        return projector
    
    # If no fallback path or normal speech encoder is configured
    if audio_config.speech_encoder_type is None:
        return None
    
    class ConfigWrapper:
        def __init__(self, audio_config, llm_hidden_size):
            self.speech_encoder_ds_rate = audio_config.speech_encoder_ds_rate
            self.speech_encoder_hidden_size = audio_config.speech_encoder_hidden_size
            self.hidden_size = llm_hidden_size  # Note: EncoderProjectorConcat uses hidden_size
    
    wrapper_config = ConfigWrapper(audio_config, llm_hidden_size)
    return EncoderProjectorConcat(wrapper_config)


class InternVLChatModel(PreTrainedModel):
    config_class = InternVLChatConfig
    main_input_name = 'pixel_values'
    base_model_prefix = 'language_model'
    _supports_flash_attn_2 = True
    supports_gradient_checkpointing = True
    _no_split_modules = [
        "InternVisionModel",
        "Qwen3DecoderLayer",
    ]

    # support transformers 4.51.+
    _tp_plan = ''

    def __init__(self, config: InternVLChatConfig, vision_model=None, language_model=None, use_flash_attn=True):
        super().__init__(config)

        assert version_cmp(transformers.__version__, '4.37.0', '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.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
        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'

        logger.info(f'num_image_token: {self.num_image_token}')
        logger.info(f'ps_version: {self.ps_version}')
        if vision_model is not None:
            self.vision_model = vision_model
        else:
            self.vision_model = InternVisionModel(config.vision_config)
        if language_model is not None:
            self.language_model = language_model
        else:
            architecture: str = config.llm_config.architectures[0]
            if architecture == 'LlamaForCausalLM':
                self.language_model = LlamaForCausalLM(config.llm_config)
            elif architecture == 'Qwen2ForCausalLM':
                self.language_model = Qwen2ForCausalLM(config.llm_config)
            elif architecture == 'Qwen3MoeForCausalLM':
                self.language_model = Qwen3MoeForCausalLM(config.llm_config)
            elif architecture == 'Qwen3ForCausalLM':
                self.language_model = Qwen3ForCausalLM(config.llm_config)
            else:
                raise NotImplementedError(f'{architecture} is not implemented.')

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

        self.mlp1 = nn.Sequential(
            nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2),
            nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size),
            nn.GELU(),
            nn.Linear(llm_hidden_size, llm_hidden_size)
        )

        # Initialize speech encoder and projector
        self.speech_encoder = build_speech_encoder(config.audio_config)
        self.speech_projector = build_speech_projector(config.audio_config, llm_hidden_size)
        
        # Add a dimension adjustment layer if needed
        self.speech_dim_adapter = None
        if self.speech_projector is not None and hasattr(self.speech_projector, 'encoder_dim'):
            expected_encoder_dim = self.speech_projector.encoder_dim
            # Our current encoder outputs 2048 (combined Whisper+BEATs)
            actual_encoder_dim = 2048
            if expected_encoder_dim != actual_encoder_dim:
                print(f"Adding dimension adapter: {actual_encoder_dim} -> {expected_encoder_dim}")
                self.speech_dim_adapter = nn.Linear(actual_encoder_dim, expected_encoder_dim)
        
        self.img_context_token_id = None
        self.speech_context_token_id = None
        self.conv_template = get_conv_template(self.template)
        self.system_message = self.conv_template.system_message

    def get_speech_encoder(self):
        return self.speech_encoder
    
    def get_speech_projector(self):
        return self.speech_projector

    def encode_speech(self, speech, speech_lengths, speech_wav):
        """Encode speech similar to Ola's implementation"""
        speech_encoder = self.get_speech_encoder()
        if speech_encoder is None:
            return None
            
        # Process raw_wav to handle BEATs input format requirements
        processed_raw_wav = speech_wav
        if speech_wav is not None and speech_wav.dim() == 2:
            processed_raw_wav = speech_wav
        elif speech_wav is not None and isinstance(speech_wav, list):
            processed_raw_wav = torch.stack(speech_wav, dim=0)
        
        # Call speech encoder with processed raw_wav parameter
        try:
            encoder_outs = speech_encoder(speech.permute(0, 2, 1), raw_wav=processed_raw_wav)
        except Exception as e:
            print(f"⚠️  BEATs processing failed: {e}")
            print("🔄 Falling back to Whisper-only processing")
            encoder_outs = speech_encoder(speech.permute(0, 2, 1), raw_wav=None)
        
        speech_lengths = (speech_lengths + 1) // 2
        
        # Apply dimension adapter if needed
        if self.speech_dim_adapter is not None:
            encoder_outs = self.speech_dim_adapter(encoder_outs)
        
        # Apply speech projector using config-based approach like your version
        speech_projector_type = getattr(self.config.audio_config, 'speech_projector_type', 'linear')
        speech_projector = self.get_speech_projector()
        if speech_projector_type == "linear" and speech_projector is not None:
            encoder_outs = speech_projector(encoder_outs)
            # Note: speech_projector.k is the downsampling rate
            if hasattr(speech_projector, 'k'):
                speech_lengths = speech_lengths // speech_projector.k
        elif speech_projector_type != "linear":
            raise ValueError(f'Unknown speech projector: {speech_projector_type}')
        
        return encoder_outs

    def prepare_inputs_labels_for_speech_vision_text(
        self, input_ids, position_ids, attention_mask, past_key_values, labels,
        speech, speech_lengths, speech_chunks, speech_wav, pixel_values, modalities, image_sizes=None, image_flags=None
    ):
        """Prepare inputs similar to Ola's implementation"""
        speech_encoder = self.speech_encoder
        
        if speech_encoder is None or input_ids.shape[1] == 1:
            return input_ids, position_ids, attention_mask, past_key_values, None, labels

        # Encode speech
        if speech is not None:
            if not isinstance(speech, list):
                if speech_chunks is not None:
                    speech = torch.split(speech, speech_chunks.tolist(), dim=0)
                    speech_lengths = torch.split(speech_lengths, speech_chunks.tolist(), dim=0)
                    speech_wav = torch.split(speech_wav, speech_chunks.tolist(), dim=0)
                else:
                    speech = [speech]
                    speech_lengths = [speech_lengths]
                    speech_wav = [speech_wav]
                    
            speech_features = []
            for idx in range(len(speech)):
                speech_feat = self.encode_speech(speech[idx], speech_lengths[idx], speech_wav[idx])
                if speech_feat is not None:
                    speech_features.append(speech_feat)
        else:
            speech_features = []

        # Encode vision
        if isinstance(modalities, str):
            modalities = [modalities]

        image_features = []
        if pixel_values is not None:
            if image_flags is not None:
                image_flags = image_flags.squeeze(-1)
                vit_embeds = self.extract_feature(pixel_values)
                vit_embeds = vit_embeds[image_flags == 1]
                # Apply vision projector
                for idx in range(len(modalities)):
                    img_feat = self.mlp1(vit_embeds[idx:idx+1])
                    image_features.append(img_feat.flatten(0, 1))
            else:
                vit_embeds = self.extract_feature(pixel_values)
                for idx in range(vit_embeds.shape[0]):
                    img_feat = vit_embeds[idx:idx+1]
                    image_features.append(img_feat.flatten(0, 1))

        # Save original values
        _labels = labels
        _position_ids = position_ids
        _attention_mask = attention_mask
        
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
        else:
            attention_mask = attention_mask.bool()
        if position_ids is None:
            position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
        if labels is None:
            labels = torch.full_like(input_ids, IGNORE_INDEX)

        # Remove padding using attention_mask
        input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
        labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]

        new_input_embeds = []
        new_labels = []
        cur_speech_idx = 0
        cur_image_idx = 0
        
        for batch_idx, cur_input_ids in enumerate(input_ids):
            num_speech = (cur_input_ids == SPEECH_TOKEN_INDEX).sum()
            num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
            num_speech_images = num_images + num_speech
            
            if num_speech_images == 0:
                # No speech or image tokens
                cur_input_embeds_1 = self.language_model.get_input_embeddings()(cur_input_ids)
                if len(speech_features) > cur_speech_idx:
                    cur_speech_features = speech_features[cur_speech_idx]
                    cur_input_embeds = torch.cat([cur_input_embeds_1, cur_speech_features[0:0]], dim=0)
                else:
                    cur_input_embeds = cur_input_embeds_1
                if len(image_features) > cur_image_idx:
                    cur_images_features = image_features[cur_image_idx]
                    cur_input_embeds = torch.cat([cur_input_embeds, cur_images_features[0:0]], dim=0)
                new_input_embeds.append(cur_input_embeds)
                new_labels.append(labels[batch_idx])
                cur_speech_idx += 1
                cur_image_idx += 1
                continue

            # Handle speech and image tokens
            speech_image_token_indices = [-1] + torch.where((cur_input_ids == SPEECH_TOKEN_INDEX) | (cur_input_ids == IMAGE_TOKEN_INDEX))[0].tolist() + [cur_input_ids.shape[0]]

            cur_input_ids_nospeech_image = []
            cur_labels = labels[batch_idx]
            cur_labels_nospeech_image = []
            
            for i in range(len(speech_image_token_indices) - 1):
                cur_input_ids_nospeech_image.append(cur_input_ids[speech_image_token_indices[i]+1:speech_image_token_indices[i+1]])
                cur_labels_nospeech_image.append(cur_labels[speech_image_token_indices[i]+1:speech_image_token_indices[i+1]])
            
            split_sizes = [x.shape[0] for x in cur_labels_nospeech_image]
            cur_input_embeds = self.language_model.get_input_embeddings()(torch.cat(cur_input_ids_nospeech_image))
            cur_input_embeds_no_speech_image = torch.split(cur_input_embeds, split_sizes, dim=0)
            cur_new_input_embeds = []
            cur_new_labels = []
            
            # Process tokens in order, similar to OLA's approach
            speech_idx_in_sequence = 0
            image_idx_in_sequence = 0
            
            for i in range(num_speech_images + 1):
                cur_new_input_embeds.append(cur_input_embeds_no_speech_image[i])
                cur_new_labels.append(cur_labels_nospeech_image[i])
                
                if i < num_speech_images:
                    # Determine which token type comes next based on position
                    if i < len(speech_image_token_indices) - 1:
                        token_pos = speech_image_token_indices[i + 1]
                        token_type = cur_input_ids[token_pos].item()
                        
                        if token_type == SPEECH_TOKEN_INDEX and len(speech_features) > cur_speech_idx:
                            cur_speech_features = speech_features[cur_speech_idx]
                            cur_speech_idx += 1
                            cur_new_input_embeds.append(cur_speech_features)
                            cur_new_labels.append(torch.full((cur_speech_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
                        elif token_type == IMAGE_TOKEN_INDEX and len(image_features) > cur_image_idx:
                            cur_images_features = image_features[cur_image_idx]
                            cur_image_idx += 1
                            cur_new_input_embeds.append(cur_images_features)
                            cur_new_labels.append(torch.full((cur_images_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))

            cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds]
            cur_new_input_embeds = torch.cat(cur_new_input_embeds)
            cur_new_labels = torch.cat(cur_new_labels)

            # Handle missing modalities
            if num_images == 0 and len(image_features) > cur_image_idx:
                cur_new_input_embeds = torch.cat([cur_new_input_embeds, image_features[cur_image_idx][0:0]], dim=0)
                cur_image_idx += 1

            if num_speech == 0 and len(speech_features) > cur_speech_idx:
                cur_new_input_embeds = torch.cat([cur_new_input_embeds, speech_features[cur_speech_idx][0:0]], dim=0)
                cur_speech_idx += 1

            new_input_embeds.append(cur_new_input_embeds)
            new_labels.append(cur_new_labels)

        # Truncate sequences to max length
        tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
        if tokenizer_model_max_length is not None:
            new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
            new_labels = [x[:tokenizer_model_max_length] for x in new_labels]

        # Combine and pad
        max_len = max(x.shape[0] for x in new_input_embeds) if new_input_embeds else 0
        batch_size = len(new_input_embeds)

        if max_len > 0:
            new_input_embeds_padded = []
            new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
            attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
            position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)

            for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
                cur_len = cur_new_embed.shape[0]
                if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
                    new_input_embeds_padded.append(torch.cat((
                        torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
                        cur_new_embed
                    ), dim=0))
                    if cur_len > 0:
                        new_labels_padded[i, -cur_len:] = cur_new_labels
                        attention_mask[i, -cur_len:] = True
                        position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
                else:
                    new_input_embeds_padded.append(torch.cat((
                        cur_new_embed,
                        torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
                    ), dim=0))
                    if cur_len > 0:
                        new_labels_padded[i, :cur_len] = cur_new_labels
                        attention_mask[i, :cur_len] = True
                        position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)

            new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)

            if _labels is None:
                new_labels = None
            else:
                new_labels = new_labels_padded

            if _attention_mask is None:
                attention_mask = None
            else:
                attention_mask = attention_mask.to(dtype=_attention_mask.dtype)

            if _position_ids is None:
                position_ids = None

            return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
        else:
            return input_ids, position_ids, attention_mask, past_key_values, None, labels

    def forward(
            self,
            pixel_values: torch.FloatTensor = None,
            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,
            # Speech related parameters
            speech: Optional[torch.FloatTensor] = None,
            speech_lengths: Optional[torch.LongTensor] = None,
            speech_chunks: Optional[torch.LongTensor] = None,
            speech_wav: Optional[torch.FloatTensor] = None,
            modalities: Optional[List[str]] = ["image"],
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        # Use Ola-style multimodal processing if speech or complex multimodal input is provided
        if speech is not None or (pixel_values is not None and speech_chunks is not None):
            (
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                inputs_embeds,
                labels
            ) = self.prepare_inputs_labels_for_speech_vision_text(
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                labels,
                speech,
                speech_lengths,
                speech_chunks,
                speech_wav,
                pixel_values,
                modalities,
                image_sizes=None,
                image_flags=image_flags
            )
            
            if inputs_embeds is not None:
                input_embeds = inputs_embeds
            else:
                # Fallback to simple processing
                input_embeds = self.language_model.get_input_embeddings()(input_ids).clone()
                B, N, C = input_embeds.shape
                input_embeds = input_embeds.reshape(B * N, C)
                input_ids_flat = input_ids.reshape(B * N)

                # Process speech input if provided
                if speech is not None and hasattr(self, 'speech_context_token_id') and self.speech_context_token_id is not None:
                    speech_features = self.encode_speech(speech, speech_lengths, speech_wav)
                    if speech_features is not None:
                        speech_selected = (input_ids_flat == self.speech_context_token_id)
                        if speech_selected.sum() > 0:
                            try:
                                input_embeds[speech_selected] = input_embeds[speech_selected] * 0.0 + speech_features.reshape(-1, C)[:speech_selected.sum()]
                            except Exception as e:
                                print(f'warning: {e}, speech processing fallback')
                                n_token = min(speech_selected.sum(), speech_features.size(0))
                                input_embeds[speech_selected][:n_token] = input_embeds[speech_selected][:n_token] * 0.0 + speech_features.reshape(-1, C)[:n_token]

                # Process vision input if provided
                if pixel_values is not None:
                    image_flags = image_flags.squeeze(-1)
                    vit_embeds = self.extract_feature(pixel_values)
                    vit_embeds = vit_embeds[image_flags == 1]

                    selected = (input_ids_flat == self.img_context_token_id)
                    try:
                        input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, C)
                    except Exception as e:
                        vit_embeds = vit_embeds.reshape(-1, C)
                        print(f'warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, '
                              f'vit_embeds.shape={vit_embeds.shape}')
                        n_token = min(selected.sum(), vit_embeds.size(0))
                        input_embeds[selected][:n_token] = input_embeds[selected][:n_token] * 0.0 + vit_embeds[:n_token]

                input_embeds = input_embeds.reshape(B, N, C)
        else:
            # Original simple processing for vision-only inputs
            input_embeds = self.language_model.get_input_embeddings()(input_ids).clone()
            B, N, C = input_embeds.shape
            input_embeds = input_embeds.reshape(B * N, C)
            input_ids_flat = input_ids.reshape(B * N)

            # Process speech input if provided
            if speech is not None and hasattr(self, 'speech_context_token_id') and self.speech_context_token_id is not None:
                speech_features = self.encode_speech(speech, speech_lengths, speech_wav)
                if speech_features is not None:
                    speech_selected = (input_ids_flat == self.speech_context_token_id)
                    if speech_selected.sum() > 0:
                        try:
                            input_embeds[speech_selected] = input_embeds[speech_selected] * 0.0 + speech_features.reshape(-1, C)[:speech_selected.sum()]
                        except Exception as e:
                            print(f'warning: {e}, speech processing fallback')
                            n_token = min(speech_selected.sum(), speech_features.size(0))
                            input_embeds[speech_selected][:n_token] = input_embeds[speech_selected][:n_token] * 0.0 + speech_features.reshape(-1, C)[:n_token]

            # Process vision input if provided
            if pixel_values is not None:
                image_flags = image_flags.squeeze(-1)
                vit_embeds = self.extract_feature(pixel_values)
                vit_embeds = vit_embeds[image_flags == 1]

                selected = (input_ids_flat == self.img_context_token_id)
                try:
                    input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, C)
                except Exception as e:
                    vit_embeds = vit_embeds.reshape(-1, C)
                    print(f'warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, '
                          f'vit_embeds.shape={vit_embeds.shape}')
                    n_token = min(selected.sum(), vit_embeds.size(0))
                    input_embeds[selected][:n_token] = input_embeds[selected][:n_token] * 0.0 + vit_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,
        )

    def pixel_shuffle(self, x, scale_factor=0.5):
        n, w, h, c = x.size()
        # N, W, H, C --> N, W, H * scale, C // scale
        x = x.view(n, w, int(h * scale_factor), int(c / scale_factor))
        # N, W, H * scale, C // scale --> N, H * scale, W, C // scale
        x = x.permute(0, 2, 1, 3).contiguous()
        # N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2)
        x = x.view(n, int(h * scale_factor), int(w * scale_factor),
                   int(c / (scale_factor * scale_factor)))
        if self.ps_version == 'v1':
            warnings.warn("In ps_version 'v1', the height and width have not been swapped back, "
                          'which results in a transposed image.')
        else:
            x = x.permute(0, 2, 1, 3).contiguous()
        return x

    def extract_feature(self, pixel_values):
        if self.select_layer == -1:
            vit_embeds = self.vision_model(
                pixel_values=pixel_values,
                output_hidden_states=False,
                return_dict=True).last_hidden_state
        else:
            vit_embeds = self.vision_model(
                pixel_values=pixel_values,
                output_hidden_states=True,
                return_dict=True).hidden_states[self.select_layer]
        vit_embeds = vit_embeds[:, 1:, :]

        h = w = int(vit_embeds.shape[1] ** 0.5)
        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
        vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio)
        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1])
        vit_embeds = self.mlp1(vit_embeds)
        return vit_embeds

    def batch_chat(self, tokenizer, pixel_values, questions, generation_config, num_patches_list=None,
                   history=None, return_history=False, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>',
                   IMG_CONTEXT_TOKEN='<IMG_CONTEXT>', verbose=False, image_counts=None,
                   speech=None, speech_lengths=None, speech_wav=None, SPEECH_CONTEXT_TOKEN='<SPEECH_CONTEXT>'):
        if history is not None or return_history:
            print('Now multi-turn chat is not supported in batch_chat.')
            raise NotImplementedError

        if image_counts is not None:
            num_patches_list = image_counts
            print('Warning: `image_counts` is deprecated. Please use `num_patches_list` instead.')

        img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
        self.img_context_token_id = img_context_token_id

        # Set up speech context token
        if speech is not None:
            speech_context_token_id = tokenizer.convert_tokens_to_ids(SPEECH_CONTEXT_TOKEN)
            self.speech_context_token_id = speech_context_token_id

        if verbose and pixel_values is not None:
            image_bs = pixel_values.shape[0]
            print(f'dynamic ViT batch size: {image_bs}')
        
        if verbose and speech is not None:
            speech_bs = speech.shape[0]
            print(f'speech batch size: {speech_bs}')

        queries = []
        for idx, num_patches in enumerate(num_patches_list):
            question = questions[idx]
            if pixel_values is not None and '<image>' not in question:
                question = '<image>\n' + question
            if speech is not None and '<speech>' not in question:
                question = '<speech>\n' + question
            
            template = get_conv_template(self.template)
            template.system_message = self.system_message
            template.append_message(template.roles[0], question)
            template.append_message(template.roles[1], None)
            query = template.get_prompt()

            # Replace image tokens
            if pixel_values is not None:
                image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
                query = query.replace('<image>', image_tokens, 1)
            
            queries.append(query)

        tokenizer.padding_side = 'left'
        # Use OLA-style tokenization for speech inputs
        if speech is not None:
            input_ids = []
            for idx, query in enumerate(queries):
                if '<speech>' in query:
                    # Use OLA-style tokenization directly
                    tokens = tokenizer_speech_token(query, tokenizer, return_tensors='pt')
                else:
                    # Replace speech tokens with context tokens for non-speech queries
                    speech_len = speech_lengths[idx] if speech_lengths is not None else speech.shape[1]
                    num_downsampled_frames = speech_len // self.config.audio_config.speech_encoder_ds_rate
                    num_speech_tokens = num_downsampled_frames + 3
                    speech_tokens = SPEECH_CONTEXT_TOKEN * num_speech_tokens
                    processed_query = query.replace('<speech>', speech_tokens, 1)
                    tokens = tokenizer(processed_query, return_tensors='pt').input_ids.squeeze(0)
                input_ids.append(tokens)
            
            # Pad sequences
            max_len = max(len(ids) for ids in input_ids)
            padded_input_ids = []
            attention_mask = []
            
            for ids in input_ids:
                pad_len = max_len - len(ids)
                if pad_len > 0:
                    padded_ids = torch.cat([torch.full((pad_len,), tokenizer.pad_token_id, dtype=ids.dtype), ids])
                    mask = torch.cat([torch.zeros(pad_len, dtype=torch.bool), torch.ones(len(ids), dtype=torch.bool)])
                else:
                    padded_ids = ids
                    mask = torch.ones(len(ids), dtype=torch.bool)
                
                padded_input_ids.append(padded_ids)
                attention_mask.append(mask)
            
            input_ids = torch.stack(padded_input_ids).to(self.device)
            attention_mask = torch.stack(attention_mask).to(self.device)
        else:
            # Replace speech tokens with context tokens for non-OLA processing
            processed_queries = []
            for idx, query in enumerate(queries):
                if speech is not None and '<speech>' in query:
                    speech_len = speech_lengths[idx] if speech_lengths is not None else speech.shape[1]
                    num_downsampled_frames = speech_len // self.config.audio_config.speech_encoder_ds_rate
                    num_speech_tokens = num_downsampled_frames + 3
                    speech_tokens = SPEECH_CONTEXT_TOKEN * num_speech_tokens
                    query = query.replace('<speech>', speech_tokens, 1)
                processed_queries.append(query)
            
            model_inputs = tokenizer(processed_queries, return_tensors='pt', padding=True)
            input_ids = model_inputs['input_ids'].to(self.device)
            attention_mask = model_inputs['attention_mask'].to(self.device)
        eos_token_id = tokenizer.convert_tokens_to_ids(template.sep.strip())
        generation_config['eos_token_id'] = eos_token_id
        generation_output = self.generate(
            pixel_values=pixel_values,
            input_ids=input_ids,
            attention_mask=attention_mask,
            speech=speech,
            speech_lengths=speech_lengths,
            speech_chunks=None,
            speech_wav=speech_wav if speech_wav is not None else speech,  # Use speech_wav if provided, otherwise fallback to speech
            modalities=["image"],
            **generation_config
        )
        responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True)
        responses = [response.split(template.sep.strip())[0].strip() for response in responses]
        return responses

    def chat(self, tokenizer, pixel_values, question, generation_config, history=None, return_history=False,
             num_patches_list=None, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>', IMG_CONTEXT_TOKEN='<IMG_CONTEXT>',
             verbose=False, speech=None, speech_lengths=None, speech_wav=None, SPEECH_CONTEXT_TOKEN='<SPEECH_CONTEXT>'):

        if history is None and pixel_values is not None and '<image>' not in question:
            question = '<image>\n' + question
        if history is None and speech is not None and '<speech>' not in question:
            question = '<speech>\n' + question

        if num_patches_list is None:
            num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else []
        assert pixel_values is None or len(pixel_values) == sum(num_patches_list)

        img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
        self.img_context_token_id = img_context_token_id
        
        # Set up speech context token
        if speech is not None:
            speech_context_token_id = tokenizer.convert_tokens_to_ids(SPEECH_CONTEXT_TOKEN)
            self.speech_context_token_id = speech_context_token_id

        template = get_conv_template(self.template)
        template.system_message = self.system_message
        eos_token_id = tokenizer.convert_tokens_to_ids(template.sep.strip())

        history = [] if history is None else history
        for (old_question, old_answer) in history:
            template.append_message(template.roles[0], old_question)
            template.append_message(template.roles[1], old_answer)
        template.append_message(template.roles[0], question)
        template.append_message(template.roles[1], None)
        query = template.get_prompt()

        if verbose and pixel_values is not None:
            image_bs = pixel_values.shape[0]
            print(f'dynamic ViT batch size: {image_bs}')
        
        if verbose and speech is not None:
            speech_bs = speech.shape[0]
            print(f'speech batch size: {speech_bs}')

        # Replace image tokens
        for num_patches in num_patches_list:
            image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
            query = query.replace('<image>', image_tokens, 1)
        
        # Use OLA-style tokenization for speech inputs
        if speech is not None and '<speech>' in query:
            # Use OLA-style tokenization directly with <speech> tokens

            input_ids = tokenizer_speech_token(query, tokenizer, return_tensors='pt').unsqueeze(0).to(self.device)
            attention_mask = torch.ones_like(input_ids, dtype=torch.bool).to(self.device)
        else:
            # Replace speech tokens with context tokens for non-OLA processing
            if speech is not None:
                speech_len = speech_lengths[0] if speech_lengths is not None else speech.shape[1]
                # Account for downsampling and special tokens (begin, end, newline)
                num_downsampled_frames = speech_len // self.config.audio_config.speech_encoder_ds_rate
                # Add 3 for begin, end, and newline tokens
                num_speech_tokens = num_downsampled_frames + 3
                speech_tokens = SPEECH_CONTEXT_TOKEN * num_speech_tokens
                query = query.replace('<speech>', speech_tokens, 1)
            
            model_inputs = tokenizer(query, return_tensors='pt')
            input_ids = model_inputs['input_ids'].to(self.device)
            attention_mask = model_inputs['attention_mask'].to(self.device)
        generation_config['eos_token_id'] = eos_token_id
        generation_output = self.generate(
            pixel_values=pixel_values,
            input_ids=input_ids,
            attention_mask=attention_mask,
            speech=speech,
            speech_lengths=speech_lengths,
            speech_chunks=None,
            speech_wav=speech_wav if speech_wav is not None else speech,  # Use speech_wav if provided, otherwise fallback to speech
            modalities=["image"],
            **generation_config
        )
        response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0]
        response = response.split(template.sep.strip())[0].strip()
        history.append((question, response))
        if return_history:
            return response, history
        else:
            query_to_print = query.replace(IMG_CONTEXT_TOKEN, '')
            query_to_print = query_to_print.replace(f'{IMG_START_TOKEN}{IMG_END_TOKEN}', '<image>')
            if verbose:
                print(query_to_print, response)
            return response

    @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,
            speech: Optional[torch.FloatTensor] = None,
            speech_lengths: Optional[torch.LongTensor] = None,
            speech_chunks: Optional[torch.LongTensor] = None,
            speech_wav: Optional[torch.FloatTensor] = None,
            modalities: Optional[List[str]] = ["image"],
            **generate_kwargs,
    ) -> torch.LongTensor:

        # Use Ola-style multimodal processing if speech or complex multimodal input is provided
        if speech is not None or (pixel_values is not None and speech_chunks is not None):
            (
                input_ids,
                position_ids,
                attention_mask,
                past_key_values,
                inputs_embeds,
                labels
            ) = self.prepare_inputs_labels_for_speech_vision_text(
                input_ids,
                None,  # position_ids
                attention_mask,
                None,  # past_key_values
                None,  # labels
                speech,
                speech_lengths,
                speech_chunks,
                speech_wav,
                pixel_values,
                modalities,
                image_sizes=None,
                image_flags=None
            )
            
            if inputs_embeds is not None:
                input_embeds = inputs_embeds
            else:
                # Fallback to simple processing
                input_embeds = self.language_model.get_input_embeddings()(input_ids)
                B, N, C = input_embeds.shape
                input_embeds = input_embeds.reshape(B * N, C)
                input_ids_flat = input_ids.reshape(B * N)

                # Process speech input if provided
                if speech is not None and hasattr(self, 'speech_context_token_id') and self.speech_context_token_id is not None:
                    speech_features = self.encode_speech(speech, speech_lengths, speech_wav)
                    if speech_features is not None:
                        speech_selected = (input_ids_flat == self.speech_context_token_id)
                        if speech_selected.sum() > 0:
                            input_embeds[speech_selected] = speech_features.reshape(-1, C)[:speech_selected.sum()].to(input_embeds.device)

                # Process vision input if provided
                if pixel_values is not None:
                    assert self.img_context_token_id is not None
                    if visual_features is not None:
                        vit_embeds = visual_features
                    else:
                        vit_embeds = self.extract_feature(pixel_values)
                    
                    selected = (input_ids_flat == self.img_context_token_id)
                    assert selected.sum() != 0
                    input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device)

                input_embeds = input_embeds.reshape(B, N, C)
        else:
            # Original simple processing for vision-only inputs
            input_embeds = self.language_model.get_input_embeddings()(input_ids)
            B, N, C = input_embeds.shape
            input_embeds = input_embeds.reshape(B * N, C)
            input_ids_flat = input_ids.reshape(B * N)

            # Process speech input if provided
            if speech is not None and hasattr(self, 'speech_context_token_id') and self.speech_context_token_id is not None:
                speech_features = self.encode_speech(speech, speech_lengths, speech_wav)
                if speech_features is not None:
                    speech_selected = (input_ids_flat == self.speech_context_token_id)
                    if speech_selected.sum() > 0:
                        input_embeds[speech_selected] = speech_features.reshape(-1, C)[:speech_selected.sum()].to(input_embeds.device)

            # Process vision input if provided
            if pixel_values is not None:
                assert self.img_context_token_id is not None
                if visual_features is not None:
                    vit_embeds = visual_features
                else:
                    vit_embeds = self.extract_feature(pixel_values)
                
                selected = (input_ids_flat == self.img_context_token_id)
                assert selected.sum() != 0
                input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device)

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

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

        return outputs

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

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

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

    def set_input_embeddings(self, value):
        return self.language_model.set_input_embeddings(value)

    def set_output_embeddings(self, value):
        return self.language_model.set_output_embeddings(value)