modeling_internvl_chat.py

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

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

import torch.utils.checkpoint
import transformers
from torch import nn
from torch.nn import CrossEntropyLoss
from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM,
                          Qwen2ForCausalLM)
from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.modeling_utils import PreTrainedModel
from transformers.utils import ModelOutput, logging
from transformers import WhisperConfig, WhisperModel, WhisperProcessor
from .configuration_internvl_chat import InternVLChatConfig
from .conversation import get_conv_template
from .modeling_intern_vit import InternVisionModel, has_flash_attn

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 InternVLChatModel(PreTrainedModel):
    config_class = InternVLChatConfig
    main_input_name = 'pixel_values'
    base_model_prefix = 'language_model'
    _supports_flash_attn_2 = True
    _no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer', 'Qwen2DecoderLayer']

    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:
            if config.llm_config.architectures[0] == 'LlamaForCausalLM':
                self.language_model = LlamaForCausalLM(config.llm_config)
            elif config.llm_config.architectures[0] == 'Qwen2ForCausalLM':
                self.language_model = Qwen2ForCausalLM(config.llm_config)
            else:
                raise NotImplementedError(f'{config.llm_config.architectures[0]} is not implemented.')

        
        # whisper model
        whisper_config = WhisperConfig(**self.config.audio_config)
        self.audio_model = WhisperModel.from_pretrained(
            "/data/nvme5n1p1/vladimir_workspace/audio_internvl/models/whisper-large-v3-turbo",
            config=whisper_config,
            torch_dtype=torch.float16,
            low_cpu_mem_usage=True,
        )
        # Remove decoder since we only need the encoder
        del self.audio_model.decoder
        
        # Initialize audio processor
        self.audio_processor = WhisperProcessor.from_pretrained("/data/nvme5n1p1/vladimir_workspace/audio_internvl/models/whisper-large-v3-turbo")

        # Get hidden sizes
        vit_hidden_size = config.vision_config.hidden_size
        llm_hidden_size = config.llm_config.hidden_size
        whisper_hidden_size = self.audio_model.config.d_model

        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)
        )

        # Audio projection
        self.mlp2 = nn.Sequential(
            nn.LayerNorm(whisper_hidden_size),
            nn.Linear(whisper_hidden_size, llm_hidden_size),
            nn.GELU(),
            nn.Linear(llm_hidden_size, llm_hidden_size)
        )
        
        self.audio_context_token_id = None
        
        self.img_context_token_id = None
        self.conv_template = get_conv_template(self.template)
        self.system_message = self.conv_template.system_message

    def process_audio_feature(self, audio_values, audio_flags):
        print("\n=== Processing Audio Features ===")
        print(f"Input audio shape: {audio_values.shape}")
        print(f"Audio flags shape: {audio_flags.shape}")
        
        # Ensure float32 for audio input
        audio_values = audio_values.to(torch.float32)
        print(f"Audio values min/max: {audio_values.min():.3f}/{audio_values.max():.3f}")
        
        # Convert audio to features
        if len(audio_values.shape) == 2:
            audio_list = [arr.cpu().numpy() for arr in audio_values]
        else:
            audio_list = [audio_values.cpu().numpy()]
        
        processed_audio = self.audio_processor(
            audio_list, 
            sampling_rate=16000,
            return_tensors="pt"
        )
        audio_features = processed_audio["input_features"].to(self.device)
        print(f"Processed audio features shape: {audio_features.shape}")
        
        # Convert to float32 before encoder
        audio_features = audio_features.to(torch.float32)
        
        # Get encoder outputs
        with torch.cuda.amp.autocast(enabled=False):  # Disable mixed precision
            audio_outputs = self.audio_model.encoder(audio_features)
            audio_embeds = audio_outputs.last_hidden_state
        
        print(f"Whisper encoder output shape: {audio_embeds.shape}")
        audio_embeds = audio_embeds.to(torch.float32)  # Ensure float32
        print(f"Encoder output min/max: {audio_embeds.min():.3f}/{audio_embeds.max():.3f}")
        
        # Reshape to match the expected number of tokens (300)
        B, T, C = audio_embeds.shape
        target_length = 300
        
        # Use adaptive pooling to get the desired length
        adaptive_pool = torch.nn.AdaptiveAvgPool1d(target_length)
        audio_embeds = audio_embeds.transpose(1, 2)  # [B, C, T]
        audio_embeds = adaptive_pool(audio_embeds)  # [B, C, 300]
        audio_embeds = audio_embeds.transpose(1, 2)  # [B, 300, C]
        print(f"After pooling shape: {audio_embeds.shape}")
        
        # More robust normalization before MLP2
        audio_embeds = audio_embeds.float()
        
        # First normalize per-token with more stable computation
        mean = audio_embeds.mean(dim=-1, keepdim=True)
        std = audio_embeds.std(dim=-1, keepdim=True)
        # Add larger epsilon and clip std to avoid division by zero
        std = torch.clamp(std, min=1e-6)
        audio_embeds = (audio_embeds - mean) / std
        
        # Clip extreme values more conservatively
        audio_embeds = torch.clamp(audio_embeds, -2.0, 2.0)
        
        # Apply LayerNorm with larger eps
        layer_norm = nn.LayerNorm(audio_embeds.shape[-1], eps=1e-4).to(audio_embeds.device)
        audio_embeds = layer_norm(audio_embeds)
        
        print(f"Pre-MLP2 stats - min: {audio_embeds.min():.3f}, max: {audio_embeds.max():.3f}")
        
        # Project to LLM dimension with gradient scaling and additional checks
        with torch.cuda.amp.autocast(enabled=False):
            # Pre-normalize and scale more carefully
            mean = audio_embeds.mean(dim=-1, keepdim=True)
            std = audio_embeds.std(dim=-1, keepdim=True)
            std = torch.clamp(std, min=1e-6)
            audio_embeds = (audio_embeds - mean) / std
            
            # Scale down more conservatively before MLP2
            audio_embeds = audio_embeds * 0.05  # Reduced from 0.1
            
            # Apply MLP2 with gradient scaling
            audio_embeds = self.mlp2(audio_embeds)
            
            if torch.isnan(audio_embeds).any() or torch.isinf(audio_embeds).any():
                print("WARNING: NaN/Inf detected after MLP2! Using robust recovery...")
                audio_embeds = torch.nan_to_num(audio_embeds, nan=0.0, posinf=1.0, neginf=-1.0)
                
                # Normalize with small noise
                mean = audio_embeds.mean(dim=-1, keepdim=True)
                std = audio_embeds.std(dim=-1, keepdim=True)
                std = torch.clamp(std, min=1e-6)
                audio_embeds = (audio_embeds - mean) / std
                audio_embeds = audio_embeds + torch.randn_like(audio_embeds) * 0.0001
        
        # Final scaling to match LLM exactly
        llm_std = 0.009
        audio_embeds = audio_embeds * llm_std
        
        return audio_embeds

    def forward(
            self,
            pixel_values: torch.FloatTensor = None,
            audio_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,
            audio_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,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        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 = input_ids.reshape(B * N)

        # Process images if present
        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]
            vit_batch_size = pixel_values.shape[0]
            
            if torch.distributed.is_initialized() and torch.distributed.get_rank() == 0:
                print(f'dynamic ViT batch size: {vit_batch_size}, images per sample: {vit_batch_size / B}, dynamic token length: {N}')

            image_selected = (input_ids == self.img_context_token_id)
            try:
                input_embeds[image_selected] = input_embeds[image_selected] * 0.0 + vit_embeds.reshape(-1, C)
            except Exception as e:
                vit_embeds = vit_embeds.reshape(-1, C)
                n_token = image_selected.sum()
                input_embeds[image_selected] = input_embeds[image_selected] * 0.0 + vit_embeds[:n_token]

        # Process audio if present
        if audio_values is not None and audio_flags is not None:
            audio_flags = audio_flags.squeeze(-1)
            audio_embeds = self.process_audio_feature(audio_values, audio_flags)
            audio_batch_size = audio_values.shape[0] if len(audio_values.shape) > 1 else 1
            
            if torch.distributed.is_initialized() and torch.distributed.get_rank() == 0:
                print(f'dynamic Audio batch size: {audio_batch_size}, audio per sample: {audio_batch_size / B}, dynamic token length: {N}')

            audio_selected = (input_ids == self.audio_context_token_id)
            try:
                input_embeds[audio_selected] = input_embeds[audio_selected] * 0.0 + audio_embeds.reshape(-1, C)
            except Exception as e:
                audio_embeds = audio_embeds.reshape(-1, C)
                n_token = audio_selected.sum()
                input_embeds[audio_selected] = input_embeds[audio_selected] * 0.0 + audio_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_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            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):
        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

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

            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'
        model_inputs = tokenizer(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,
            **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=None, question=None, generation_config=None, 
            history=None, return_history=False, num_patches_list=None,
            IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>', IMG_CONTEXT_TOKEN='<IMG_CONTEXT>',
            AUDIO_START_TOKEN='<audio>', AUDIO_END_TOKEN='</audio>', AUDIO_CONTEXT_TOKEN='<AUDIO_CONTEXT>',
            verbose=False, **kwargs):  # Add **kwargs to catch extra arguments
        """Chat function that handles both text-only and multimodal inputs"""
        print("=== Starting Chat Process ===")
        print(f"Question: {question}")
        print(f"Input types - Pixel values: {type(pixel_values)}, Audio values: {type(kwargs.get('audio_values'))}")

        # Basic input validation
        if question is None:
            raise ValueError("Question cannot be None")
        if not isinstance(question, str):
            raise ValueError(f"Question must be string, got {type(question)}")

        audio_values = kwargs.get('audio_values', None)

        # Handle image prompt
        if history is None and pixel_values is not None and '<image>' not in question:
            question = '<image>\n' + question
            print("Added image token to question")

        # Handle audio prompt
        if history is None and audio_values is not None and '<audio>' not in question:
            question = '<audio>\n' + question
            print("Added audio token to question")

        # Process image patches
        if num_patches_list is None:
            num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else []
        if pixel_values is not None:
            assert len(pixel_values) == sum(num_patches_list)
            print(f"Image patches: {num_patches_list}")

        # Set context token IDs
        img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
        audio_context_token_id = tokenizer.convert_tokens_to_ids(AUDIO_CONTEXT_TOKEN)
        self.img_context_token_id = img_context_token_id
        self.audio_context_token_id = audio_context_token_id
        print(f"Token IDs - Image: {img_context_token_id}, Audio: {audio_context_token_id}")

        # Process template and history
        # Prepare conversation template
        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()
        print(f"Processed query: {query[:100]}...")  # Print first 100 chars

        if verbose:
            if pixel_values is not None:
                print(f'dynamic ViT batch size: {pixel_values.shape[0]}')
            if audio_values is not None:
                print(f'dynamic Audio batch size: {audio_values.shape[0]}')

        # Insert 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)

        # Insert audio tokens (assuming one audio per query)
        if audio_values is not None:
            audio_tokens = AUDIO_START_TOKEN + AUDIO_CONTEXT_TOKEN * 300 + AUDIO_END_TOKEN
            query = query.replace('<audio>', audio_tokens, 1)
            
            # Add debug prints
            print("\n=== Audio Token Debug ===")
            print(f"AUDIO_START_TOKEN: {AUDIO_START_TOKEN} (id: {tokenizer.convert_tokens_to_ids(AUDIO_START_TOKEN)})")
            print(f"AUDIO_CONTEXT_TOKEN: {AUDIO_CONTEXT_TOKEN} (id: {tokenizer.convert_tokens_to_ids(AUDIO_CONTEXT_TOKEN)})")
            print(f"AUDIO_END_TOKEN: {AUDIO_END_TOKEN} (id: {tokenizer.convert_tokens_to_ids(AUDIO_END_TOKEN)})")
            
            # Verify token presence
            test_tokens = tokenizer(query, return_tensors='pt')['input_ids'][0]
            context_token_count = (test_tokens == self.audio_context_token_id).sum()
            print(f"Number of AUDIO_CONTEXT tokens found: {context_token_count} (should be 300)")
            
            # Print full tokenization of a small segment
            audio_segment = query[query.find(AUDIO_START_TOKEN):query.find(AUDIO_END_TOKEN)+len(AUDIO_END_TOKEN)]
            print("\nTokenization of audio segment:")
            tokens = tokenizer.tokenize(audio_segment)
            print(f"First 10 tokens: {tokens[:10]}...")
            print(f"Last 10 tokens: {tokens[-10:]}")

        # Prepare model inputs
        model_inputs = tokenizer(query, return_tensors='pt')
        print("Model inputs:")
        print(f"input_ids shape: {model_inputs['input_ids'].shape}")  
        print(f"First few tokens: {tokenizer.convert_ids_to_tokens(model_inputs['input_ids'][0][:20])}")
        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

        # Ensure generation_config is a proper dictionary
        if generation_config is None or not isinstance(generation_config, dict):
            generation_config = {}
        
        # Set default generation parameters
        default_config = {
            "do_sample": True,
            "temperature": 0.7,
            "top_p": 0.9,
            "max_new_tokens": 256,
            "repetition_penalty": 1.2,
            "no_repeat_ngram_size": 3,
            "pad_token_id": tokenizer.pad_token_id,
            "eos_token_id": eos_token_id
        }
        
        # Update with user provided config
        for k, v in default_config.items():
            if k not in generation_config:
                generation_config[k] = v
        
        # Create GenerationConfig object
        generation_config = GenerationConfig(**generation_config)
        
        # Generate response
        generation_output = self.generate(
            pixel_values=pixel_values,
            audio_values=audio_values,
            input_ids=input_ids,
            attention_mask=attention_mask,
            generation_config=generation_config,
        )

        # Process response
        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:
            # Clean up query for printing
            query_to_print = query.replace(IMG_CONTEXT_TOKEN, '').replace(AUDIO_CONTEXT_TOKEN, '')
            query_to_print = query_to_print.replace(f'{IMG_START_TOKEN}{IMG_END_TOKEN}', '<image>')
            query_to_print = query_to_print.replace(f'{AUDIO_START_TOKEN}{AUDIO_END_TOKEN}', '<audio>')
            if verbose:
                print(query_to_print, response)
            return response

    @torch.no_grad()
    def generate(self, pixel_values=None, audio_values=None, input_ids=None, 
                attention_mask=None, visual_features=None, generation_config=None, 
                output_hidden_states=None, **generate_kwargs):
        
        print("\n=== Generate Method Debug ===")
        
        # Get initial embeddings and check statistics
        input_embeds = self.language_model.get_input_embeddings()(input_ids)
        initial_std = input_embeds.std().item()
        print(f"LLM embedding stats - mean: {input_embeds.mean():.3f}, std: {initial_std:.3f}")
        
        B, N, C = input_embeds.shape
        print(f"Initial embeddings shape: {input_embeds.shape}")
        
        input_embeds = input_embeds.reshape(B * N, C)
        input_ids = input_ids.reshape(B * N)
        
        if audio_values is not None:
            assert self.audio_context_token_id is not None
            print("\nAudio Processing:")
            print(f"Audio context token ID: {self.audio_context_token_id}")
            
            audio_embeds = self.process_audio_feature(
                audio_values, 
                torch.ones(audio_values.shape[0]).to(audio_values.device)
            )
            
            # Scale audio embeddings to match LLM embeddings
            audio_embeds = audio_embeds * (initial_std / audio_embeds.std().item())
            
            print(f"Processed audio embeds shape: {audio_embeds.shape}")
            print(f"Audio embedding stats after scaling - mean: {audio_embeds.mean():.3f}, std: {audio_embeds.std():.3f}")
            
            audio_selected = (input_ids == self.audio_context_token_id)
            num_audio_tokens = audio_selected.sum()
            print(f"Number of audio context tokens found: {num_audio_tokens}")
            
            try:
                input_embeds[audio_selected] = audio_embeds.reshape(-1, C).to(input_embeds.device)
                print("Successfully inserted audio embeddings")
            except Exception as e:
                print(f"Error inserting audio embeddings: {e}")
                raise
        
        # Reshape back
        input_embeds = input_embeds.reshape(B, N, C)
        
        # Final normalization of combined embeddings
        input_embeds = torch.nn.functional.layer_norm(
            input_embeds,
            input_embeds.shape[-1:],
            eps=1e-5
        ) * initial_std  # Scale back to original magnitude
        
        print("\nFinal combined embedding stats:")
        print(f"Mean: {input_embeds.mean():.3f}, Std: {input_embeds.std():.3f}")
        print(f"Min: {input_embeds.min():.3f}, Max: {input_embeds.max():.3f}")
        
        # Final safety check before generation
        if torch.isnan(input_embeds).any() or torch.isinf(input_embeds).any():
            raise ValueError("Critical: Found NaN/Inf values in embeddings before generation!")
        
        # Generate output with additional error handling
        try:
            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,
            )
        except RuntimeError as e:
            if "probability tensor contains either `inf`" in str(e):
                print("ERROR: Invalid probability distribution. Attempting recovery...")
                # Try with more conservative generation settings
                generate_kwargs["temperature"] = 1.0  # Reset temperature
                generate_kwargs["top_p"] = 1.0  # Disable top_p
                generate_kwargs["do_sample"] = False  # Fall back to greedy
                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,
                )
            else:
                raise

        return outputs