Upload folder using huggingface_hub

README.md

@@ -1,348 +1,51 @@
 ---
-license_name: lfm1.0
-license_link: LICENSE
 language:
 - en
-base_model:
-- LiquidAI/LFM2.5-Audio-1.5B
-pipeline_tag: audio-to-audio
-library_name: mlx-audio
 tags:
-- audio-to-audio
-- speech
-- speech generation
-- voice isolation
-- sts
-- mlx
 - liquid
-- lfm2.5
-- edge
+- lfm2
+- audio
+- lfm2-audio
+- speech-to-speech
+- liquid-audio
+- mlx
+- speech-to-speech
+- speech
 - audio
+- speech enhancement
+- audio separation
+- sts
+- mlx-audio
+license: other
+license_name: lfm1.0
+license_link: LICENSE
+library_name: mlx-audio
+pipeline_tag: audio-to-audio
+base_model:
+- LiquidAI/LFM2-1.2B
 ---
-# mlx-community/LFM2.5-Audio-1.5B-bf16
-This model was converted to MLX format from [`LiquidAI/LFM2.5-Audio-1.5B`](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B) using mlx-audio version **0.3.0**.
-Refer to the [original model card](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B) for more details on the model.
-
-## Use with mlx
-```bash
-pip install -U mlx-audio
-```
-
-## Features
-
-- **Text-to-Speech (TTS)**: Generate natural speech from text
-- **Speech-to-Text (ASR)**: Transcribe audio to text
-- **Speech-to-Speech (STS)**: Voice conversations with audio input and output
-- **Interleaved Generation**: Mixed text and audio responses in a single turn
-- **Streaming**: Real-time token-by-token generation for low-latency applications
-
-## Quick Start
-
-### Text-to-Speech (TTS)
-
-```python
-import mlx.core as mx
-from mlx_audio.sts.models.lfm_audio import (
-    LFM2AudioModel,
-    LFM2AudioProcessor,
-    ChatState,
-    LFMModality,
-)
-
-# Load model and processor
-model = LFM2AudioModel.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-bf16")
-processor = LFM2AudioProcessor.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-bf16")
-
-# Create chat state
-chat = ChatState(processor)
-chat.new_turn("system")
-chat.add_text("Respond with audio.")
-chat.end_turn()
-chat.new_turn("user")
-chat.add_text("Say: Hello, welcome to MLX Audio!")
-chat.end_turn()
-chat.new_turn("assistant")
-
-# Generate with interleaved text and audio
-text_out, audio_out = [], []
-for token, modality in model.generate_interleaved(**dict(chat), max_new_tokens=2048):
-    mx.eval(token)
-    if modality == LFMModality.TEXT:
-        text_out.append(token)
-        print(processor.decode_text(token[None]), end="", flush=True)
-    else:
-        audio_out.append(token)
-
-# Decode audio - each token is (8,) for all codebooks
-if audio_out:
-    audio_codes = mx.stack(audio_out[:-1], axis=1)[None, :]  # (1, 8, T)
-    waveform = processor.decode_with_detokenizer(audio_codes)
-    # Or use Mimi codec: waveform = processor.decode_audio(audio_codes[0])
-
-    # Save audio (24kHz sample rate)
-    import soundfile as sf
-    sf.write("output.wav", waveform[0].tolist(), 24000)
-```
-
-### Speech-to-Text (ASR)
-
-```python
-import mlx.core as mx
-import numpy as np
-import soundfile as sf
-from mlx_audio.sts.models.lfm_audio import (
-    LFM2AudioModel,
-    LFM2AudioProcessor,
-    ChatState,
-    LFMModality,
-)
-
-# Load model and processor
-model = LFM2AudioModel.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-bf16")
-processor = LFM2AudioProcessor.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-bf16")
-
-# Load audio (must be 24kHz for audio input)
-audio, sr = sf.read("input.wav")
-audio = mx.array(audio.astype(np.float32))
-
-# Create chat state with audio input
-chat = ChatState(processor)
-chat.new_turn("user")
-chat.add_audio(audio, sample_rate=sr)
-chat.add_text("Transcribe the audio.")
-chat.end_turn()
-chat.new_turn("assistant")
-
-# Generate text response
-text_out = []
-for token, modality in model.generate_interleaved(**dict(chat), max_new_tokens=512):
-    mx.eval(token)
-    if modality == LFMModality.TEXT:
-        text_out.append(token)
-        print(processor.decode_text(token[None]), end="", flush=True)
-```
-
-### Speech-to-Speech (STS)
-
-```python
-import mlx.core as mx
-import numpy as np
-import soundfile as sf
-from mlx_audio.sts.models.lfm_audio import (
-    LFM2AudioModel,
-    LFM2AudioProcessor,
-    ChatState,
-    LFMModality,
-)
-
-# Load model and processor
-model = LFM2AudioModel.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-bf16")
-processor = LFM2AudioProcessor.from_pretrained("mlx-community/LFM2.5-Audio-1.5B-bf16")
-
-# Load input audio (24kHz)
-audio, sr = sf.read("input.wav")
-audio = mx.array(audio.astype(np.float32))
-
-# Create chat state with audio input
-chat = ChatState(processor)
-chat.new_turn("system")
-chat.add_text("Respond with interleaved text and audio.")
-chat.end_turn()
-chat.new_turn("user")
-chat.add_audio(audio, sample_rate=sr)
-chat.end_turn()
-chat.new_turn("assistant")
-
-# Generate response with both text and audio
-text_out, audio_out = [], []
-for token, modality in model.generate_interleaved(**dict(chat), max_new_tokens=2048):
-    mx.eval(token)
-    if modality == LFMModality.TEXT:
-        text_out.append(token)
-        print(processor.decode_text(token[None]), end="", flush=True)
-    else:
-        audio_out.append(token)
-
-# Decode audio response
-if audio_out:
-    audio_codes = mx.stack(audio_out[:-1], axis=1)[None, :]  # (1, 8, T)
-    waveform = processor.decode_with_detokenizer(audio_codes)
-    sf.write("response.wav", waveform[0].tolist(), 24000)
-```
-
-## Interleaved Text and Audio Generation
-
-LFM2.5-Audio uses `generate_interleaved` for mixed text and audio output. The model can respond with text, audio, or both interleaved together.
-
-Each audio token returned by `generate_interleaved` is a complete frame of shape `(8,)` containing all 8 codebook values:
-
-```python
-from mlx_audio.sts.models.lfm_audio import LFMModality
-
-text_out, audio_out = [], []
-for token, modality in model.generate_interleaved(**dict(chat), max_new_tokens=2048):
-    mx.eval(token)
-    if modality == LFMModality.TEXT:
-        text_out.append(token)
-        # Stream text output
-        print(processor.decode_text(token[None]), end="", flush=True)
-    else:  # LFMModality.AUDIO_OUT
-        audio_out.append(token)  # token shape: (8,)
-
-# Stack audio frames: list of (8,) -> (8, T)
-if audio_out:
-    audio_codes = mx.stack(audio_out[:-1], axis=1)[None, :]  # (1, 8, T)
-    waveform = processor.decode_with_detokenizer(audio_codes)
-```
-
-## Audio Decoding Options
-
-LFM2.5-Audio supports two methods for decoding audio codes to waveforms:
-
-### 1. Detokenizer (Recommended for TTS)
-
-The neural detokenizer reconstructs audio using ISTFT from predicted spectrograms:
-
-```python
-# Decode using detokenizer
-audio = processor.decode_with_detokenizer(codes[None])  # (1, T_audio)
-```
-
-### 2. Mimi Codec
-
-The Mimi neural codec provides an alternative decoding path:
-
-```python
-# Decode using Mimi codec
-audio = processor.decode_audio(codes, codec="mimi")  # (1, 1, T_audio)
-```
-
-## Generation Configuration
-
-```python
-from mlx_audio.sts.models.lfm_audio import GenerationConfig
-
-config = GenerationConfig(
-    max_new_tokens=2048,    # Maximum tokens to generate
-    temperature=0.9,        # Text sampling temperature
-    top_k=50,               # Text top-k sampling
-    top_p=1.0,              # Text nucleus sampling
-    audio_temperature=0.7,  # Audio sampling temperature
-    audio_top_k=30,         # Audio top-k sampling
-)
-```
-
-## Streaming Generation
-
-For real-time audio playback during generation:
-
-```python
-from mlx_audio.sts.models.lfm_audio import LFMModality
-
-FRAMES_PER_CHUNK = 10  # Decode every 10 audio frames
-
-audio_buffer = []
-for token, modality in model.generate_interleaved(**dict(chat), max_new_tokens=2048):
-    mx.eval(token)
-    if modality == LFMModality.AUDIO_OUT:
-        audio_buffer.append(token)
-
-        # Decode when we have enough frames
-        if len(audio_buffer) >= FRAMES_PER_CHUNK:
-            codes = mx.stack(audio_buffer, axis=1)[None, :]  # (1, 8, T)
-            chunk = processor.decode_with_detokenizer(codes)
-            # Play chunk with your audio library...
-            audio_buffer = []
-
-    elif modality == LFMModality.TEXT:
-        # Stream text output
-        print(processor.decode_text(token[None]), end="", flush=True)
-```
-
-## Model Architecture
-
-LFM2.5-Audio consists of:
-
-- **Audio Encoder**: Conformer-based encoder for processing input audio
-- **LFM Backbone**: 1.5B parameter Liquid Foundation Model for multimodal reasoning
-- **Audio Decoder**: Depthformer for generating audio codes
-- **Detokenizer**: ISTFT-based neural vocoder for waveform reconstruction
-
-## API Reference
-
-### LFM2AudioModel
-
-```python
-class LFM2AudioModel:
-    @classmethod
-    def from_pretrained(cls, model_name: str) -> "LFM2AudioModel":
-        """Load pretrained model from HuggingFace Hub."""
-
-    def generate_interleaved(
-        self,
-        text_tokens: mx.array,
-        audio_features: mx.array,
-        modalities: mx.array,
-        max_new_tokens: int = 512,
-        temperature: float = 0.9,
-        audio_temperature: float = 0.7,
-        audio_top_k: int = 30,
-    ) -> Generator[Tuple[mx.array, LFMModality], None, None]:
-        """Generate interleaved text and audio tokens.
-
-        Yields:
-            (token, modality) tuples where:
-            - For TEXT: token is scalar, modality is LFMModality.TEXT
-            - For AUDIO_OUT: token is (8,) array, modality is LFMModality.AUDIO_OUT
-        """
-```
-
-### LFM2AudioProcessor
-
-```python
-class LFM2AudioProcessor:
-    @classmethod
-    def from_pretrained(cls, model_name: str) -> "LFM2AudioProcessor":
-        """Load pretrained processor from HuggingFace Hub."""
-
-    def preprocess_audio(self, audio: mx.array, sample_rate: int) -> mx.array:
-        """Convert audio to mel spectrogram features."""
-
-    def tokenize_audio(self, audio: mx.array, sample_rate: int) -> mx.array:
-        """Tokenize audio using Mimi codec."""
-
-    def decode_audio(self, codes: mx.array, codec="detokenizer") -> mx.array:
-        """Decode audio codes using Detokenizer or Mimi codec."""
-
-    def tokenize_text(self, text: str) -> mx.array:
-        """Tokenize text."""
-
-    def decode_text(self, tokens: mx.array) -> str:
-        """Decode text tokens."""
-```
-
-### ChatState
-
-```python
-class ChatState:
-    def __init__(self, processor: LFM2AudioProcessor):
-        """Initialize chat state."""
+        # mlx-community/LFM2.5-Audio-1.5B-bf16
 
-    def new_turn(self, role: str):
-        """Start a new turn (user/assistant/system)."""
+        This model was converted to MLX format from [`LiquidAI/LFM2.5-Audio-1.5B`](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B) using mlx-audio version **0.2.10**.
 
-    def end_turn(self):
-        """End the current turn."""
+        Refer to the [original model card](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B) for more details on the model.
 
-    def add_text(self, text: str):
-        """Add text to current turn."""
+        ## Use with mlx-audio
 
-    def add_audio(self, audio: mx.array, sample_rate: int):
-        """Add audio to current turn."""
-```
+        ```bash
+        pip install -U mlx-audio
+        ```
 
-## License
+        ### CLI Example:
+        ```bash
+        python -m mlx_audio.sts.generate --model mlx-community/LFM2.5-Audio-1.5B-bf16 --audio "audio.wav"
+        ```
 
-This implementation follows the license terms of the original LFM2.5-Audio model.
-See [LiquidAI/LFM2.5-Audio-1.5B](https://huggingface.co/LiquidAI/LFM2.5-Audio-1.5B) for details.
+        ### Python Example:
+        ```python
+        from mlx_audio.sts.utils import load_model
 
+model = load_model("mlx-community/LFM2.5-Audio-1.5B-bf16")
+# Usage depends on the specific STS model type
+# See model documentation for details
+        ```

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9a7c8eb3d154134af4386b242c01b5027d9f464328211433fe03b942cb3d5323
-size 2940723623
+oid sha256:261c5cd8db83e10bd63ad2c0b9bdfad488ac77bfa1c72b9482e4c06c59b3ab43
+size 2940723367

model.safetensors.index.json

@@ -10,6 +10,9 @@
         "audio_adapter.layers.1.weight": "model.safetensors",
         "audio_adapter.layers.3.bias": "model.safetensors",
         "audio_adapter.layers.3.weight": "model.safetensors",
+        "audio_embedding.embedding.weight": "model.safetensors",
+        "audio_embedding.embedding_norm.weight": "model.safetensors",
+        "audio_embedding.to_logits.weight": "model.safetensors",
         "audio_encoder.layers.0.attn.k_proj.bias": "model.safetensors",
         "audio_encoder.layers.0.attn.k_proj.weight": "model.safetensors",
         "audio_encoder.layers.0.attn.out_proj.bias": "model.safetensors",
@@ -673,113 +676,110 @@
         "audio_encoder.layers.9.ff2_norm.weight": "model.safetensors",
         "audio_encoder.layers.9.final_norm.bias": "model.safetensors",
         "audio_encoder.layers.9.final_norm.weight": "model.safetensors",
-        "audio_encoder.subsampling.conv.0.bias": "model.safetensors",
-        "audio_encoder.subsampling.conv.0.weight": "model.safetensors",
-        "audio_encoder.subsampling.conv.2.bias": "model.safetensors",
-        "audio_encoder.subsampling.conv.2.weight": "model.safetensors",
-        "audio_encoder.subsampling.conv.3.bias": "model.safetensors",
-        "audio_encoder.subsampling.conv.3.weight": "model.safetensors",
-        "audio_encoder.subsampling.conv.5.bias": "model.safetensors",
-        "audio_encoder.subsampling.conv.5.weight": "model.safetensors",
-        "audio_encoder.subsampling.conv.6.bias": "model.safetensors",
-        "audio_encoder.subsampling.conv.6.weight": "model.safetensors",
-        "audio_encoder.subsampling.out_proj.bias": "model.safetensors",
-        "audio_encoder.subsampling.out_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.0.attn.k_norm.scale": "model.safetensors",
+        "audio_encoder.pre_encode.conv.0.bias": "model.safetensors",
+        "audio_encoder.pre_encode.conv.0.weight": "model.safetensors",
+        "audio_encoder.pre_encode.conv.2.bias": "model.safetensors",
+        "audio_encoder.pre_encode.conv.2.weight": "model.safetensors",
+        "audio_encoder.pre_encode.conv.3.bias": "model.safetensors",
+        "audio_encoder.pre_encode.conv.3.weight": "model.safetensors",
+        "audio_encoder.pre_encode.conv.5.bias": "model.safetensors",
+        "audio_encoder.pre_encode.conv.5.weight": "model.safetensors",
+        "audio_encoder.pre_encode.conv.6.bias": "model.safetensors",
+        "audio_encoder.pre_encode.conv.6.weight": "model.safetensors",
+        "audio_encoder.pre_encode.out.bias": "model.safetensors",
+        "audio_encoder.pre_encode.out.weight": "model.safetensors",
+        "audio_head.depthformer.blocks.0.attn.k_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.0.attn.k_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.0.attn.o_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.0.attn.q_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.0.attn.q_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.0.attn.q_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.0.attn.v_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.0.attn_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.0.attn_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.0.ffn.w1.weight": "model.safetensors",
         "audio_head.depthformer.blocks.0.ffn.w2.weight": "model.safetensors",
         "audio_head.depthformer.blocks.0.ffn.w3.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.0.ffn_norm.scale": "model.safetensors",
-        "audio_head.depthformer.blocks.1.attn.k_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.0.ffn_norm.weight": "model.safetensors",
+        "audio_head.depthformer.blocks.1.attn.k_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.1.attn.k_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.1.attn.o_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.1.attn.q_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.1.attn.q_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.1.attn.q_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.1.attn.v_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.1.attn_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.1.attn_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.1.ffn.w1.weight": "model.safetensors",
         "audio_head.depthformer.blocks.1.ffn.w2.weight": "model.safetensors",
         "audio_head.depthformer.blocks.1.ffn.w3.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.1.ffn_norm.scale": "model.safetensors",
-        "audio_head.depthformer.blocks.2.attn.k_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.1.ffn_norm.weight": "model.safetensors",
+        "audio_head.depthformer.blocks.2.attn.k_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.2.attn.k_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.2.attn.o_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.2.attn.q_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.2.attn.q_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.2.attn.q_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.2.attn.v_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.2.attn_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.2.attn_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.2.ffn.w1.weight": "model.safetensors",
         "audio_head.depthformer.blocks.2.ffn.w2.weight": "model.safetensors",
         "audio_head.depthformer.blocks.2.ffn.w3.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.2.ffn_norm.scale": "model.safetensors",
-        "audio_head.depthformer.blocks.3.attn.k_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.2.ffn_norm.weight": "model.safetensors",
+        "audio_head.depthformer.blocks.3.attn.k_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.3.attn.k_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.3.attn.o_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.3.attn.q_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.3.attn.q_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.3.attn.q_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.3.attn.v_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.3.attn_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.3.attn_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.3.ffn.w1.weight": "model.safetensors",
         "audio_head.depthformer.blocks.3.ffn.w2.weight": "model.safetensors",
         "audio_head.depthformer.blocks.3.ffn.w3.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.3.ffn_norm.scale": "model.safetensors",
-        "audio_head.depthformer.blocks.4.attn.k_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.3.ffn_norm.weight": "model.safetensors",
+        "audio_head.depthformer.blocks.4.attn.k_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.4.attn.k_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.4.attn.o_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.4.attn.q_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.4.attn.q_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.4.attn.q_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.4.attn.v_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.4.attn_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.4.attn_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.4.ffn.w1.weight": "model.safetensors",
         "audio_head.depthformer.blocks.4.ffn.w2.weight": "model.safetensors",
         "audio_head.depthformer.blocks.4.ffn.w3.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.4.ffn_norm.scale": "model.safetensors",
-        "audio_head.depthformer.blocks.5.attn.k_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.4.ffn_norm.weight": "model.safetensors",
+        "audio_head.depthformer.blocks.5.attn.k_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.5.attn.k_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.5.attn.o_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.5.attn.q_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.5.attn.q_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.5.attn.q_proj.weight": "model.safetensors",
         "audio_head.depthformer.blocks.5.attn.v_proj.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.5.attn_norm.scale": "model.safetensors",
+        "audio_head.depthformer.blocks.5.attn_norm.weight": "model.safetensors",
         "audio_head.depthformer.blocks.5.ffn.w1.weight": "model.safetensors",
         "audio_head.depthformer.blocks.5.ffn.w2.weight": "model.safetensors",
         "audio_head.depthformer.blocks.5.ffn.w3.weight": "model.safetensors",
-        "audio_head.depthformer.blocks.5.ffn_norm.scale": "model.safetensors",
-        "audio_head.in_proj.bias": "model.safetensors",
-        "audio_head.in_proj.weight": "model.safetensors",
-        "audio_in_emb.embedding.weight": "model.safetensors",
-        "audio_in_emb.embedding_norm.scale": "model.safetensors",
-        "audio_in_emb.to_logits.weight": "model.safetensors",
-        "depth_embeddings.embeddings.0.embedding.weight": "model.safetensors",
-        "depth_embeddings.embeddings.0.embedding_norm.scale": "model.safetensors",
-        "depth_embeddings.embeddings.0.to_logits.weight": "model.safetensors",
-        "depth_embeddings.embeddings.1.embedding.weight": "model.safetensors",
-        "depth_embeddings.embeddings.1.embedding_norm.scale": "model.safetensors",
-        "depth_embeddings.embeddings.1.to_logits.weight": "model.safetensors",
-        "depth_embeddings.embeddings.2.embedding.weight": "model.safetensors",
-        "depth_embeddings.embeddings.2.embedding_norm.scale": "model.safetensors",
-        "depth_embeddings.embeddings.2.to_logits.weight": "model.safetensors",
-        "depth_embeddings.embeddings.3.embedding.weight": "model.safetensors",
-        "depth_embeddings.embeddings.3.embedding_norm.scale": "model.safetensors",
-        "depth_embeddings.embeddings.3.to_logits.weight": "model.safetensors",
-        "depth_embeddings.embeddings.4.embedding.weight": "model.safetensors",
-        "depth_embeddings.embeddings.4.embedding_norm.scale": "model.safetensors",
-        "depth_embeddings.embeddings.4.to_logits.weight": "model.safetensors",
-        "depth_embeddings.embeddings.5.embedding.weight": "model.safetensors",
-        "depth_embeddings.embeddings.5.embedding_norm.scale": "model.safetensors",
-        "depth_embeddings.embeddings.5.to_logits.weight": "model.safetensors",
-        "depth_embeddings.embeddings.6.embedding.weight": "model.safetensors",
-        "depth_embeddings.embeddings.6.embedding_norm.scale": "model.safetensors",
-        "depth_embeddings.embeddings.6.to_logits.weight": "model.safetensors",
-        "depth_embeddings.embeddings.7.embedding.weight": "model.safetensors",
-        "depth_embeddings.embeddings.7.embedding_norm.scale": "model.safetensors",
-        "depth_embeddings.embeddings.7.to_logits.weight": "model.safetensors",
+        "audio_head.depthformer.blocks.5.ffn_norm.weight": "model.safetensors",
+        "depth_embeddings.0.embedding.weight": "model.safetensors",
+        "depth_embeddings.0.embedding_norm.weight": "model.safetensors",
+        "depth_embeddings.0.to_logits.weight": "model.safetensors",
+        "depth_embeddings.1.embedding.weight": "model.safetensors",
+        "depth_embeddings.1.embedding_norm.weight": "model.safetensors",
+        "depth_embeddings.1.to_logits.weight": "model.safetensors",
+        "depth_embeddings.2.embedding.weight": "model.safetensors",
+        "depth_embeddings.2.embedding_norm.weight": "model.safetensors",
+        "depth_embeddings.2.to_logits.weight": "model.safetensors",
+        "depth_embeddings.3.embedding.weight": "model.safetensors",
+        "depth_embeddings.3.embedding_norm.weight": "model.safetensors",
+        "depth_embeddings.3.to_logits.weight": "model.safetensors",
+        "depth_embeddings.4.embedding.weight": "model.safetensors",
+        "depth_embeddings.4.embedding_norm.weight": "model.safetensors",
+        "depth_embeddings.4.to_logits.weight": "model.safetensors",
+        "depth_embeddings.5.embedding.weight": "model.safetensors",
+        "depth_embeddings.5.embedding_norm.weight": "model.safetensors",
+        "depth_embeddings.5.to_logits.weight": "model.safetensors",
+        "depth_embeddings.6.embedding.weight": "model.safetensors",
+        "depth_embeddings.6.embedding_norm.weight": "model.safetensors",
+        "depth_embeddings.6.to_logits.weight": "model.safetensors",
+        "depth_embeddings.7.embedding.weight": "model.safetensors",
+        "depth_embeddings.7.embedding_norm.weight": "model.safetensors",
+        "depth_embeddings.7.to_logits.weight": "model.safetensors",
+        "depth_linear.bias": "model.safetensors",
+        "depth_linear.weight": "model.safetensors",
         "lfm.embed_tokens.weight": "model.safetensors",
         "lfm.embedding_norm.weight": "model.safetensors",
         "lfm.layers.0.conv.conv.weight": "model.safetensors",