Upload 11 files

.gitattributes

@@ -1,35 +1,35 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,20 @@
+# Python-generated files
+__pycache__/
+*.py[oc]
+build/
+dist/
+wheels/
+*.egg-info
+
+# Virtual environments
+.venv
+
+.gradio
+
+**/*.pth
+**/*.mp3
+!example_prompt.mp3
+
+.ruff_cache
+.ipynb_checkpoints
+config.json

README.md

@@ -1,13 +1,14 @@
----
-title: 'Test1 '
-emoji: 🦀
-colorFrom: pink
-colorTo: pink
-sdk: gradio
-sdk_version: 5.27.0
-app_file: app.py
-pinned: false
-license: mit
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+---
+title: Dia 1.6B
+emoji: 👯
+colorFrom: purple
+colorTo: yellow
+python_version: 3.10
+sdk: gradio
+sdk_version: 5.25.2
+app_file: app.py
+pinned: false
+short_description: Generate realistic dialogue from a script, using Dia!
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,273 @@
+import tempfile
+import time
+from pathlib import Path
+from typing import Optional, Tuple
+import spaces
+
+import gradio as gr
+import numpy as np
+import soundfile as sf
+import torch
+
+from dia.model import Dia
+from transformers import pipeline
+
+# Load Nari model
+print("Loading Nari model...")
+try:
+    model = Dia.from_pretrained("nari-labs/Dia-1.6B", compute_dtype="float32")
+except Exception as e:
+    print(f"Error loading Nari model: {e}")
+    raise
+
+# Load summarization model
+print("Loading summarizer model...")
+try:
+    summarizer = pipeline("summarization", model="facebook/bart-large-cnn")
+except Exception as e:
+    print(f"Error loading summarizer: {e}")
+    summarizer = None
+
+
+@spaces.GPU
+def run_inference(
+    text_input: str,
+    audio_prompt_input: Optional[Tuple[int, np.ndarray]],
+    max_new_tokens: int,
+    cfg_scale: float,
+    temperature: float,
+    top_p: float,
+    cfg_filter_top_k: int,
+    speed_factor: float,
+    apply_summary: bool,
+):
+    """
+    Runs Nari inference using the globally loaded model and provided inputs.
+    Uses temporary files for text and audio prompt compatibility with inference.generate.
+    """
+    if not text_input or text_input.isspace():
+        raise gr.Error("Text input cannot be empty.")
+
+    temp_audio_prompt_path = None
+    output_audio = (44100, np.zeros(1, dtype=np.float32))
+
+    try:
+        # Optionally summarize text
+        if apply_summary and summarizer is not None:
+            print("Summarizing input text...")
+            summarized = summarizer(text_input, max_length=150, min_length=30, do_sample=False)
+            if summarized and isinstance(summarized, list):
+                text_input = summarized[0]["summary_text"]
+                print(f"Summarized Text: {text_input}")
+
+        # Process Audio Prompt
+        prompt_path_for_generate = None
+        if audio_prompt_input is not None:
+            sr, audio_data = audio_prompt_input
+            if audio_data is not None and audio_data.size != 0 and audio_data.max() != 0:
+                with tempfile.NamedTemporaryFile(mode="wb", suffix=".wav", delete=False) as f_audio:
+                    temp_audio_prompt_path = f_audio.name
+                    if np.issubdtype(audio_data.dtype, np.integer):
+                        max_val = np.iinfo(audio_data.dtype).max
+                        audio_data = audio_data.astype(np.float32) / max_val
+                    elif not np.issubdtype(audio_data.dtype, np.floating):
+                        try:
+                            audio_data = audio_data.astype(np.float32)
+                        except Exception as conv_e:
+                            raise gr.Error(f"Failed to convert audio prompt to float32: {conv_e}")
+
+                    if audio_data.ndim > 1:
+                        audio_data = np.mean(audio_data, axis=-1)
+                        audio_data = np.ascontiguousarray(audio_data)
+
+                    try:
+                        sf.write(temp_audio_prompt_path, audio_data, sr, subtype="FLOAT")
+                        prompt_path_for_generate = temp_audio_prompt_path
+                        print(f"Saved temporary audio prompt: {temp_audio_prompt_path}")
+                    except Exception as write_e:
+                        raise gr.Error(f"Failed to save audio prompt: {write_e}")
+
+        # Multi-Voice Handling
+        text_segments = split_by_speaker(text_input)
+        print(f"Detected {len(text_segments)} speaker segments.")
+
+        final_audio = []
+
+        start_time = time.time()
+
+        for idx, segment in enumerate(text_segments):
+            if not segment.strip():
+                continue
+            with torch.inference_mode():
+                output_audio_np = model.generate(
+                    segment,
+                    max_tokens=max_new_tokens,
+                    cfg_scale=cfg_scale,
+                    temperature=temperature,
+                    top_p=top_p,
+                    cfg_filter_top_k=cfg_filter_top_k,
+                    use_torch_compile=False,
+                    audio_prompt=prompt_path_for_generate,
+                )
+            if output_audio_np is not None:
+                final_audio.append(output_audio_np)
+
+        if final_audio:
+            output_audio_np = np.concatenate(final_audio)
+
+        end_time = time.time()
+        print(f"Generation completed in {end_time - start_time:.2f}s.")
+
+        # Resample for speed adjustment
+        output_sr = 44100
+        original_len = len(output_audio_np)
+        speed_factor = max(0.1, min(speed_factor, 5.0))
+        target_len = int(original_len / speed_factor)
+
+        if target_len != original_len and target_len > 0:
+            x_original = np.arange(original_len)
+            x_resampled = np.linspace(0, original_len - 1, target_len)
+            resampled_audio_np = np.interp(x_resampled, x_original, output_audio_np)
+            output_audio = (output_sr, resampled_audio_np.astype(np.float32))
+        else:
+            output_audio = (output_sr, output_audio_np)
+
+        # Convert float32 audio to int16 for Gradio
+        audio_for_gradio = np.clip(output_audio[1], -1.0, 1.0)
+        audio_for_gradio = (audio_for_gradio * 32767).astype(np.int16)
+        output_audio = (output_sr, audio_for_gradio)
+
+    except Exception as e:
+        import traceback
+        traceback.print_exc()
+        raise gr.Error(f"Inference failed: {e}")
+
+    finally:
+        if temp_audio_prompt_path and Path(temp_audio_prompt_path).exists():
+            try:
+                Path(temp_audio_prompt_path).unlink()
+                print(f"Deleted temporary audio prompt file: {temp_audio_prompt_path}")
+            except Exception as e:
+                print(f"Warning: {e}")
+
+    return output_audio
+
+
+def split_by_speaker(text: str):
+    """Split text into segments by speaker labels like [S1], [S2], etc."""
+    import re
+
+    segments = re.split(r'(?=\[S\d\])', text)
+    return [seg.strip() for seg in segments if seg.strip()]
+
+
+# --- Build Gradio UI ---
+css = """
+#col-container {max-width: 90%; margin-left: auto; margin-right: auto;}
+"""
+default_text = "[S1] Hello there! How are you? \n[S2] I'm great, thanks! And you? \n[S1] Doing well! (laughs)"
+
+example_txt_path = Path("./example.txt")
+if example_txt_path.exists():
+    try:
+        file_text = example_txt_path.read_text(encoding="utf-8").strip()
+        if file_text:
+            default_text = file_text
+    except Exception:
+        pass
+
+with gr.Blocks(css=css) as demo:
+    gr.Markdown("# Nari Text-to-Speech with Multi-Voice and Summarization")
+
+    with gr.Row(equal_height=False):
+        with gr.Column(scale=1):
+            text_input = gr.Textbox(
+                label="Input Text",
+                placeholder="Enter multi-speaker dialogue...",
+                value=default_text,
+                lines=8,
+            )
+            audio_prompt_input = gr.Audio(
+                label="Audio Prompt (Optional)",
+                show_label=True,
+                sources=["upload", "microphone"],
+                type="numpy",
+            )
+            with gr.Accordion("Advanced Settings", open=False):
+                max_new_tokens = gr.Slider(
+                    label="Max New Tokens",
+                    minimum=860,
+                    maximum=3072,
+                    value=model.config.data.audio_length,
+                    step=50,
+                )
+                cfg_scale = gr.Slider(
+                    label="CFG Scale",
+                    minimum=1.0,
+                    maximum=5.0,
+                    value=3.0,
+                    step=0.1,
+                )
+                temperature = gr.Slider(
+                    label="Temperature",
+                    minimum=1.0,
+                    maximum=1.5,
+                    value=1.3,
+                    step=0.05,
+                )
+                top_p = gr.Slider(
+                    label="Top P",
+                    minimum=0.8,
+                    maximum=1.0,
+                    value=0.95,
+                    step=0.01,
+                )
+                cfg_filter_top_k = gr.Slider(
+                    label="CFG Filter Top K",
+                    minimum=15,
+                    maximum=50,
+                    value=30,
+                    step=1,
+                )
+                speed_factor_slider = gr.Slider(
+                    label="Speed Factor",
+                    minimum=0.5,
+                    maximum=1.5,
+                    value=0.94,
+                    step=0.02,
+                )
+                apply_summary = gr.Checkbox(
+                    label="Summarize Input Text before Generation?",
+                    value=False,
+                )
+
+            run_button = gr.Button("Generate Audio", variant="primary")
+
+        with gr.Column(scale=1):
+            audio_output = gr.Audio(
+                label="Generated Audio",
+                type="numpy",
+                autoplay=False,
+            )
+
+    run_button.click(
+        fn=run_inference,
+        inputs=[
+            text_input,
+            audio_prompt_input,
+            max_new_tokens,
+            cfg_scale,
+            temperature,
+            top_p,
+            cfg_filter_top_k,
+            speed_factor_slider,
+            apply_summary,
+        ],
+        outputs=[audio_output],
+        api_name="generate_audio",
+    )
+
+# --- Launch ---
+if __name__ == "__main__":
+    print("Launching Gradio app...")
+    demo.launch()

dia/audio.py

@@ -0,0 +1,203 @@
+import typing as tp
+
+import torch
+
+
+def build_delay_indices(
+    B: int, T: int, C: int, delay_pattern: tp.List[int]
+) -> tp.Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Precompute (t_idx_BxTxC, indices_BTCx3) so that out[t, c] = in[t - delay[c], c].
+    Negative t_idx => BOS; t_idx >= T => PAD.
+    """
+    delay_arr = torch.tensor(delay_pattern, dtype=torch.int32)
+
+    t_idx_BxT = torch.broadcast_to(
+        torch.arange(T, dtype=torch.int32)[None, :],
+        [B, T],
+    )
+    t_idx_BxTx1 = t_idx_BxT[..., None]
+    t_idx_BxTxC = t_idx_BxTx1 - delay_arr.view(1, 1, C)
+
+    b_idx_BxTxC = torch.broadcast_to(
+        torch.arange(B, dtype=torch.int32).view(B, 1, 1),
+        [B, T, C],
+    )
+    c_idx_BxTxC = torch.broadcast_to(
+        torch.arange(C, dtype=torch.int32).view(1, 1, C),
+        [B, T, C],
+    )
+
+    # We must clamp time indices to [0..T-1] so gather_nd equivalent won't fail
+    t_clamped_BxTxC = torch.clamp(t_idx_BxTxC, 0, T - 1)
+
+    indices_BTCx3 = torch.stack(
+        [
+            b_idx_BxTxC.reshape(-1),
+            t_clamped_BxTxC.reshape(-1),
+            c_idx_BxTxC.reshape(-1),
+        ],
+        dim=1,
+    ).long()  # Ensure indices are long type for indexing
+
+    return t_idx_BxTxC, indices_BTCx3
+
+
+def apply_audio_delay(
+    audio_BxTxC: torch.Tensor,
+    pad_value: int,
+    bos_value: int,
+    precomp: tp.Tuple[torch.Tensor, torch.Tensor],
+) -> torch.Tensor:
+    """
+    Applies the delay pattern to batched audio tokens using precomputed indices,
+    inserting BOS where t_idx < 0 and PAD where t_idx >= T.
+
+    Args:
+        audio_BxTxC: [B, T, C] int16 audio tokens (or int32/float)
+        pad_value: the padding token
+        bos_value: the BOS token
+        precomp:  (t_idx_BxTxC, indices_BTCx3) from build_delay_indices
+
+    Returns:
+        result_BxTxC: [B, T, C] delayed audio tokens
+    """
+    device = audio_BxTxC.device  # Get device from input tensor
+    t_idx_BxTxC, indices_BTCx3 = precomp
+    t_idx_BxTxC = t_idx_BxTxC.to(device)  # Move precomputed indices to device
+    indices_BTCx3 = indices_BTCx3.to(device)
+
+    # Equivalent of tf.gather_nd using advanced indexing
+    # Ensure indices are long type if not already (build_delay_indices should handle this)
+    gathered_flat = audio_BxTxC[
+        indices_BTCx3[:, 0], indices_BTCx3[:, 1], indices_BTCx3[:, 2]
+    ]
+    gathered_BxTxC = gathered_flat.view(audio_BxTxC.shape)
+
+    # Create masks on the correct device
+    mask_bos = t_idx_BxTxC < 0  # => place bos_value
+    mask_pad = t_idx_BxTxC >= audio_BxTxC.shape[1]  # => place pad_value
+
+    # Create scalar tensors on the correct device
+    bos_tensor = torch.tensor(bos_value, dtype=audio_BxTxC.dtype, device=device)
+    pad_tensor = torch.tensor(pad_value, dtype=audio_BxTxC.dtype, device=device)
+
+    # If mask_bos, BOS; else if mask_pad, PAD; else original gather
+    # All tensors should now be on the same device
+    result_BxTxC = torch.where(
+        mask_bos, bos_tensor, torch.where(mask_pad, pad_tensor, gathered_BxTxC)
+    )
+
+    return result_BxTxC
+
+
+def build_revert_indices(
+    B: int, T: int, C: int, delay_pattern: tp.List[int]
+) -> tp.Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Precompute indices for the revert operation using PyTorch.
+
+    Returns:
+        A tuple (t_idx_BxTxC, indices_BTCx3) where:
+            - t_idx_BxTxC is a tensor of shape [B, T, C] computed as time indices plus the delay.
+            - indices_BTCx3 is a tensor of shape [B*T*C, 3] used for gathering, computed from:
+                batch indices, clamped time indices, and channel indices.
+    """
+    # Use default device unless specified otherwise; assumes inputs might define device later
+    device = None  # Or determine dynamically if needed, e.g., from a model parameter
+
+    delay_arr = torch.tensor(delay_pattern, dtype=torch.int32, device=device)
+
+    t_idx_BT1 = torch.broadcast_to(torch.arange(T, device=device).unsqueeze(0), [B, T])
+    t_idx_BT1 = t_idx_BT1.unsqueeze(-1)
+
+    t_idx_BxTxC = torch.minimum(
+        t_idx_BT1 + delay_arr.view(1, 1, C),
+        torch.tensor(T - 1, device=device),
+    )
+    b_idx_BxTxC = torch.broadcast_to(
+        torch.arange(B, device=device).view(B, 1, 1), [B, T, C]
+    )
+    c_idx_BxTxC = torch.broadcast_to(
+        torch.arange(C, device=device).view(1, 1, C), [B, T, C]
+    )
+
+    indices_BTCx3 = torch.stack(
+        [
+            b_idx_BxTxC.reshape(-1),
+            t_idx_BxTxC.reshape(-1),
+            c_idx_BxTxC.reshape(-1),
+        ],
+        axis=1,
+    ).long()  # Ensure indices are long type
+
+    return t_idx_BxTxC, indices_BTCx3
+
+
+def revert_audio_delay(
+    audio_BxTxC: torch.Tensor,
+    pad_value: int,
+    precomp: tp.Tuple[torch.Tensor, torch.Tensor],
+    T: int,
+) -> torch.Tensor:
+    """
+    Reverts a delay pattern from batched audio tokens using precomputed indices (PyTorch version).
+
+    Args:
+        audio_BxTxC: Input delayed audio tensor
+        pad_value: Padding value for out-of-bounds indices
+        precomp: Precomputed revert indices tuple containing:
+            - t_idx_BxTxC: Time offset indices tensor
+            - indices_BTCx3: Gather indices tensor for original audio
+        T: Original sequence length before padding
+
+    Returns:
+        Reverted audio tensor with same shape as input
+    """
+    t_idx_BxTxC, indices_BTCx3 = precomp
+    device = audio_BxTxC.device  # Get device from input tensor
+
+    # Move precomputed indices to the same device as audio_BxTxC if they aren't already
+    t_idx_BxTxC = t_idx_BxTxC.to(device)
+    indices_BTCx3 = indices_BTCx3.to(device)
+
+    # Using PyTorch advanced indexing (equivalent to tf.gather_nd or np equivalent)
+    gathered_flat = audio_BxTxC[
+        indices_BTCx3[:, 0], indices_BTCx3[:, 1], indices_BTCx3[:, 2]
+    ]
+    gathered_BxTxC = gathered_flat.view(
+        audio_BxTxC.size()
+    )  # Use .size() for robust reshaping
+
+    # Create pad_tensor on the correct device
+    pad_tensor = torch.tensor(pad_value, dtype=audio_BxTxC.dtype, device=device)
+    # Create T tensor on the correct device for comparison
+    T_tensor = torch.tensor(T, device=device)
+
+    result_BxTxC = torch.where(
+        t_idx_BxTxC >= T_tensor, pad_tensor, gathered_BxTxC
+    )  # Changed np.where to torch.where
+
+    return result_BxTxC
+
+
+@torch.no_grad()
+@torch.inference_mode()
+def decode(
+    model,
+    audio_codes,
+):
+    """
+    Decodes the given frames into an output audio waveform
+    """
+    if len(audio_codes) != 1:
+        raise ValueError(f"Expected one frame, got {len(audio_codes)}")
+
+    try:
+        audio_values = model.quantizer.from_codes(audio_codes)
+        audio_values = model.decode(audio_values[0])
+
+        return audio_values
+    except Exception as e:
+        print(f"Error in decode method: {str(e)}")
+        raise

dia/config.py

@@ -0,0 +1,197 @@
+"""Configuration management module for the Dia model.
+
+This module provides comprehensive configuration management for the Dia model,
+utilizing Pydantic for validation. It defines configurations for data processing,
+model architecture (encoder and decoder), and training settings.
+
+Key components:
+- DataConfig: Parameters for data loading and preprocessing.
+- EncoderConfig: Architecture details for the encoder module.
+- DecoderConfig: Architecture details for the decoder module.
+- ModelConfig: Combined model architecture settings.
+- TrainingConfig: Training hyperparameters and settings.
+- DiaConfig: Master configuration combining all components.
+"""
+
+import os
+from typing import Annotated
+
+from pydantic import BaseModel, BeforeValidator, Field
+
+
+class DataConfig(BaseModel, frozen=True):
+    """Configuration for data loading and preprocessing.
+
+    Attributes:
+        text_length: Maximum length of text sequences (must be multiple of 128).
+        audio_length: Maximum length of audio sequences (must be multiple of 128).
+        channels: Number of audio channels.
+        text_pad_value: Value used for padding text sequences.
+        audio_eos_value: Value representing the end of audio sequences.
+        audio_bos_value: Value representing the beginning of audio sequences.
+        audio_pad_value: Value used for padding audio sequences.
+        delay_pattern: List of delay values for each audio channel.
+    """
+
+    text_length: Annotated[int, BeforeValidator(lambda x: (x + 127) // 128 * 128)] = (
+        Field(gt=0, multiple_of=128)
+    )
+    audio_length: Annotated[int, BeforeValidator(lambda x: (x + 127) // 128 * 128)] = (
+        Field(gt=0, multiple_of=128)
+    )
+    channels: int = Field(default=9, gt=0, multiple_of=1)
+    text_pad_value: int = Field(default=0)
+    audio_eos_value: int = Field(default=1024)
+    audio_pad_value: int = Field(default=1025)
+    audio_bos_value: int = Field(default=1026)
+    delay_pattern: list[Annotated[int, Field(ge=0)]] = Field(
+        default_factory=lambda: [0, 8, 9, 10, 11, 12, 13, 14, 15]
+    )
+
+    def __hash__(self) -> int:
+        """Generate a hash based on all fields of the config."""
+        return hash(
+            (
+                self.text_length,
+                self.audio_length,
+                self.channels,
+                self.text_pad_value,
+                self.audio_pad_value,
+                self.audio_bos_value,
+                self.audio_eos_value,
+                tuple(self.delay_pattern),
+            )
+        )
+
+
+class EncoderConfig(BaseModel, frozen=True):
+    """Configuration for the encoder component of the Dia model.
+
+    Attributes:
+        n_layer: Number of transformer layers.
+        n_embd: Embedding dimension.
+        n_hidden: Hidden dimension size in the MLP layers.
+        n_head: Number of attention heads.
+        head_dim: Dimension per attention head.
+    """
+
+    n_layer: int = Field(gt=0)
+    n_embd: int = Field(gt=0)
+    n_hidden: int = Field(gt=0)
+    n_head: int = Field(gt=0)
+    head_dim: int = Field(gt=0)
+
+
+class DecoderConfig(BaseModel, frozen=True):
+    """Configuration for the decoder component of the Dia model.
+
+    Attributes:
+        n_layer: Number of transformer layers.
+        n_embd: Embedding dimension.
+        n_hidden: Hidden dimension size in the MLP layers.
+        gqa_query_heads: Number of query heads for grouped-query self-attention.
+        kv_heads: Number of key/value heads for grouped-query self-attention.
+        gqa_head_dim: Dimension per query head for grouped-query self-attention.
+        cross_query_heads: Number of query heads for cross-attention.
+        cross_head_dim: Dimension per cross-attention head.
+    """
+
+    n_layer: int = Field(gt=0)
+    n_embd: int = Field(gt=0)
+    n_hidden: int = Field(gt=0)
+    gqa_query_heads: int = Field(gt=0)
+    kv_heads: int = Field(gt=0)
+    gqa_head_dim: int = Field(gt=0)
+    cross_query_heads: int = Field(gt=0)
+    cross_head_dim: int = Field(gt=0)
+
+
+class ModelConfig(BaseModel, frozen=True):
+    """Main configuration container for the Dia model architecture.
+
+    Attributes:
+        encoder: Configuration for the encoder component.
+        decoder: Configuration for the decoder component.
+        src_vocab_size: Size of the source (text) vocabulary.
+        tgt_vocab_size: Size of the target (audio code) vocabulary.
+        dropout: Dropout probability applied within the model.
+        normalization_layer_epsilon: Epsilon value for normalization layers (e.g., LayerNorm).
+        weight_dtype: Data type for model weights (e.g., "float32", "bfloat16").
+        rope_min_timescale: Minimum timescale for Rotary Positional Embeddings (RoPE).
+        rope_max_timescale: Maximum timescale for Rotary Positional Embeddings (RoPE).
+    """
+
+    encoder: EncoderConfig
+    decoder: DecoderConfig
+    src_vocab_size: int = Field(default=128, gt=0)
+    tgt_vocab_size: int = Field(default=1028, gt=0)
+    dropout: float = Field(default=0.0, ge=0.0, lt=1.0)
+    normalization_layer_epsilon: float = Field(default=1.0e-5, ge=0.0)
+    weight_dtype: str = Field(default="float32", description="Weight precision")
+    rope_min_timescale: int = Field(
+        default=1, description="Timescale For global Attention"
+    )
+    rope_max_timescale: int = Field(
+        default=10_000, description="Timescale For global Attention"
+    )
+
+
+class TrainingConfig(BaseModel, frozen=True):
+    pass
+
+
+class DiaConfig(BaseModel, frozen=True):
+    """Master configuration for the Dia model.
+
+    Combines all sub-configurations into a single validated object.
+
+    Attributes:
+        version: Configuration version string.
+        model: Model architecture configuration.
+        training: Training process configuration (precision settings).
+        data: Data loading and processing configuration.
+    """
+
+    version: str = Field(default="1.0")
+    model: ModelConfig
+    # TODO: remove training. this is just for backwards-compatability
+    training: TrainingConfig
+    data: DataConfig
+
+    def save(self, path: str) -> None:
+        """Save the current configuration instance to a JSON file.
+
+        Ensures the parent directory exists and the file has a .json extension.
+
+        Args:
+            path: The target file path to save the configuration.
+
+        Raises:
+            ValueError: If the path is not a file with a .json extension.
+        """
+        os.makedirs(os.path.dirname(path), exist_ok=True)
+        config_json = self.model_dump_json(indent=2)
+        with open(path, "w") as f:
+            f.write(config_json)
+
+    @classmethod
+    def load(cls, path: str) -> "DiaConfig | None":
+        """Load and validate a Dia configuration from a JSON file.
+
+        Args:
+            path: The path to the configuration file.
+
+        Returns:
+            A validated DiaConfig instance if the file exists and is valid,
+            otherwise None if the file is not found.
+
+        Raises:
+            ValueError: If the path does not point to an existing .json file.
+            pydantic.ValidationError: If the JSON content fails validation against the DiaConfig schema.
+        """
+        try:
+            with open(path, "r") as f:
+                content = f.read()
+            return cls.model_validate_json(content)
+        except FileNotFoundError:
+            return None

dia/layers.py

@@ -0,0 +1,642 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch import Tensor
+from torch.nn import RMSNorm
+
+from .config import DiaConfig
+from .state import DecoderInferenceState, EncoderInferenceState, KVCache
+
+
+def _normalize_axes(axes: tuple[int, ...], ndim: int) -> tuple[int, ...]:
+    return tuple(ax if ax >= 0 else ndim + ax for ax in axes)
+
+
+class DenseGeneral(nn.Module):
+    """
+    PyTorch equivalent of flax.linen.DenseGeneral with shapes defined at init.
+
+    Stores weights (`kernel`) in the same layout as Jax and uses torch.tensordot
+    for the generalized matrix multiplication. Weight/bias shapes are calculated
+    and parameters created during initialization based on config.
+    `load_weights` validates shapes and copies data.
+
+    Attributes:
+        axis (Tuple[int, ...]): Input axis or axes to contract.
+        in_shapes (Tuple[int, ...]): Sizes of the input dimensions specified by `axis`.
+        out_features (Tuple[int, ...]): Shape of the output features (non-contracted dims).
+        use_bias (bool): Whether to add a bias term.
+        weight (nn.Parameter): The kernel parameter.
+        bias (Optional[nn.Parameter]): The bias parameter (if use_bias=True).
+    """
+
+    def __init__(
+        self,
+        in_shapes: tuple[int, ...],
+        out_features: tuple[int, ...],
+        axis: tuple[int, ...] = (-1,),
+        weight_dtype: torch.dtype | None = None,
+        device: torch.device | None = None,
+    ):
+        super().__init__()
+        self.in_shapes = in_shapes
+        self.out_features = out_features
+        self.axis = axis
+        self.kernel_shape = self.in_shapes + self.out_features
+
+        factory_kwargs = {"device": device, "dtype": weight_dtype}
+        self.weight = nn.Parameter(torch.empty(self.kernel_shape, **factory_kwargs))
+        self.register_parameter("bias", None)
+
+    def forward(self, inputs: Tensor) -> Tensor:
+        norm_axis = _normalize_axes(self.axis, inputs.ndim)
+        kernel_contract_axes = tuple(range(len(norm_axis)))
+
+        output = torch.tensordot(
+            inputs.to(self.weight.dtype),
+            self.weight,
+            dims=(norm_axis, kernel_contract_axes),
+        ).to(inputs.dtype)
+        return output
+
+
+class MlpBlock(nn.Module):
+    """MLP block using DenseGeneral."""
+
+    def __init__(
+        self, embed_dim: int, intermediate_dim: int, compute_dtype: torch.dtype
+    ):
+        super().__init__()
+        self.dtype = compute_dtype
+
+        self.wi_fused = DenseGeneral(
+            in_shapes=(embed_dim,),
+            out_features=(2, intermediate_dim),
+            axis=(-1,),
+            weight_dtype=compute_dtype,
+        )
+
+        self.wo = DenseGeneral(
+            in_shapes=(intermediate_dim,),
+            out_features=(embed_dim,),
+            axis=(-1,),
+            weight_dtype=compute_dtype,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """Forward pass."""
+        fused_x = self.wi_fused(x)
+
+        gate = fused_x[..., 0, :]
+        up = fused_x[..., 1, :]
+
+        hidden = torch.mul(F.silu(gate), up).to(self.dtype)
+
+        output = self.wo(hidden)
+        return output
+
+
+class RotaryEmbedding(nn.Module):
+    """Rotary Position Embedding (RoPE) implementation in PyTorch."""
+
+    def __init__(
+        self,
+        embedding_dims: int,
+        min_timescale: int = 1,
+        max_timescale: int = 10000,
+        dtype: torch.dtype = torch.float32,
+    ):
+        super().__init__()
+        if embedding_dims % 2 != 0:
+            raise ValueError("Embedding dim must be even for RoPE.")
+        self.embedding_dims = embedding_dims
+        self.min_timescale = min_timescale
+        self.max_timescale = max_timescale
+        self.dtype = dtype
+
+        half_embedding_dim = embedding_dims // 2
+        fraction = (2.0 * torch.arange(0, half_embedding_dim)) / embedding_dims
+        self.register_buffer(
+            "timescale",
+            self.min_timescale * (self.max_timescale / self.min_timescale) ** fraction,
+            persistent=False,
+        )
+
+    def extra_repr(self) -> str:
+        s = f"{self.timescale.shape}"
+        return s
+
+    def forward(self, inputs: torch.Tensor, position: torch.Tensor):
+        """Applies RoPE."""
+        position = position.unsqueeze(-1).unsqueeze(-1)
+        timescale = self.timescale.to(inputs.device)
+        sinusoid_inp = position / timescale
+        sin = torch.sin(sinusoid_inp).to(inputs.dtype)
+        cos = torch.cos(sinusoid_inp).to(inputs.dtype)
+        first_half, second_half = torch.chunk(inputs, 2, dim=-1)
+        first_part = first_half * cos - second_half * sin
+        second_part = second_half * cos + first_half * sin
+        return torch.cat((first_part, second_part), dim=-1)
+
+
+class Attention(nn.Module):
+    """Attention using DenseGeneral."""
+
+    def __init__(
+        self,
+        config: DiaConfig,
+        q_embed_dim: int,
+        kv_embed_dim: int,
+        num_query_heads: int,
+        num_kv_heads: int,
+        head_dim: int,
+        compute_dtype: torch.dtype,
+        is_cross_attn: bool = False,
+        out_embed_dim: int | None = None,
+    ):
+        super().__init__()
+        self.num_query_heads = num_query_heads
+        self.num_kv_heads = num_kv_heads
+        self.head_dim = head_dim
+        self.is_cross_attn = is_cross_attn
+        self.output_dim = out_embed_dim if out_embed_dim is not None else q_embed_dim
+        self.projected_query_dim = num_query_heads * head_dim
+        if num_query_heads % num_kv_heads != 0:
+            raise ValueError(
+                f"num_query_heads ({num_query_heads}) must be divisible by num_kv_heads ({num_kv_heads})"
+            )
+        self.num_gqa_groups = num_query_heads // num_kv_heads
+
+        # --- Projection Layers using DenseGeneral ---
+        self.q_proj = DenseGeneral(
+            in_shapes=(q_embed_dim,),
+            out_features=(num_query_heads, head_dim),
+            axis=(-1,),
+            weight_dtype=compute_dtype,
+        )
+        self.k_proj = DenseGeneral(
+            in_shapes=(kv_embed_dim,),
+            out_features=(num_kv_heads, head_dim),
+            axis=(-1,),
+            weight_dtype=compute_dtype,
+        )
+        self.v_proj = DenseGeneral(
+            in_shapes=(kv_embed_dim,),
+            out_features=(num_kv_heads, head_dim),
+            axis=(-1,),
+            weight_dtype=compute_dtype,
+        )
+        self.o_proj = DenseGeneral(
+            in_shapes=(num_query_heads, head_dim),
+            out_features=(self.output_dim,),
+            axis=(-2, -1),
+            weight_dtype=compute_dtype,
+        )
+
+        # --- Rotary Embedding ---
+        self.rotary_emb = RotaryEmbedding(
+            embedding_dims=self.head_dim,
+            min_timescale=config.model.rope_min_timescale,
+            max_timescale=config.model.rope_max_timescale,
+            dtype=compute_dtype,
+        )
+
+    def forward(
+        self,
+        Xq: torch.Tensor,  # (B, T, D) T = 1 in AR generation
+        Xkv: torch.Tensor,  # (B, S, E) S = 1 in AR generation
+        q_positions: torch.Tensor,  # (B, T)
+        kv_positions: torch.Tensor | None = None,  # (B, S)
+        attn_mask: torch.Tensor
+        | None = None,  # None in Decoder Self Attention, Valid mask in Others
+        cache: KVCache | None = None,  # None in Encoder, KVCache in Decoder
+        prefill: bool = False,
+        is_causal: bool = False,
+    ) -> tuple[torch.Tensor, tuple[torch.Tensor, torch.Tensor] | None]:
+        """
+        Performs attention calculation with optional KV caching.
+
+        Args:
+            Xq: Query tensor (B, T, D). T=1 during single-step decoding.
+            Xkv: Key/Value source tensor (B, S, E). S=1 during single-step decoding for self-attn.
+            q_positions: Positions for queries (B, T).
+            kv_positions: Positions for keys/values (B, S). If None, uses q_positions.
+            attn_mask: Attention mask.
+            cache: KVCache.
+            prefill: If True, use prefill mode.
+
+        Returns:
+            A tuple containing:
+            - output: The attention output tensor (B, T, output_dim).
+            - present_kv: The K/V state to be cached for the next step ((B, N, S_new, H), (B, N, S_new, H)). For self-attn, S_new = S_past + S. For cross-attn, S_new = S_kv.
+        """
+        if kv_positions is None:
+            kv_positions = q_positions
+        original_dtype = Xq.dtype
+
+        Xq_BxTxNxH = self.q_proj(Xq)
+        Xq_BxTxNxH = self.rotary_emb(Xq_BxTxNxH, position=q_positions)
+        Xq_BxNxTxH = Xq_BxTxNxH.transpose(1, 2)
+
+        attn_k: torch.Tensor | None = None
+        attn_v: torch.Tensor | None = None
+
+        if self.is_cross_attn:
+            attn_k, attn_v = cache.k, cache.v
+        else:
+            Xk_BxSxKxH = self.k_proj(Xkv)  # (B, S, K, H)
+            Xv_BxSxKxH = self.v_proj(Xkv)  # (B, S, K, H)
+            Xk_BxSxKxH = self.rotary_emb(
+                Xk_BxSxKxH, position=kv_positions
+            )  # (B, S, K, H)
+
+            Xk_BxKxSxH = Xk_BxSxKxH.transpose(1, 2)  # (B, K, S, H)
+            Xv_BxKxSxH = Xv_BxSxKxH.transpose(1, 2)  # (B, K, S, H)
+
+            if cache is None:
+                attn_k = Xk_BxKxSxH
+                attn_v = Xv_BxKxSxH
+            else:
+                if prefill:
+                    attn_k, attn_v = Xk_BxKxSxH, Xv_BxKxSxH
+                    cache.prefill(attn_k, attn_v)
+                else:
+                    attn_k, attn_v = cache.update(Xk_BxKxSxH, Xv_BxKxSxH)
+
+        attn_output = F.scaled_dot_product_attention(
+            Xq_BxNxTxH,
+            attn_k,
+            attn_v,
+            attn_mask=attn_mask,
+            scale=1.0,
+            enable_gqa=self.num_gqa_groups > 1,
+            is_causal=is_causal,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()  # (B, T, N, H)
+        output = self.o_proj(attn_output)
+
+        return output.to(original_dtype)
+
+
+class EncoderLayer(nn.Module):
+    """Transformer Encoder Layer using DenseGeneral."""
+
+    def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
+        super().__init__()
+        self.config = config
+        model_config = config.model
+        enc_config = config.model.encoder
+        embed_dim = enc_config.n_embd
+
+        self.pre_sa_norm = RMSNorm(
+            embed_dim,
+            eps=model_config.normalization_layer_epsilon,
+            dtype=torch.float32,
+        )
+        self.self_attention = Attention(
+            config,
+            q_embed_dim=embed_dim,
+            kv_embed_dim=embed_dim,
+            num_query_heads=enc_config.n_head,
+            num_kv_heads=enc_config.n_head,
+            head_dim=enc_config.head_dim,
+            compute_dtype=compute_dtype,
+            is_cross_attn=False,
+            out_embed_dim=embed_dim,
+        )
+        self.post_sa_norm = RMSNorm(
+            embed_dim,
+            eps=model_config.normalization_layer_epsilon,
+            dtype=torch.float32,
+        )
+        self.mlp = MlpBlock(
+            embed_dim=embed_dim,
+            intermediate_dim=enc_config.n_hidden,
+            compute_dtype=compute_dtype,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        state: EncoderInferenceState,
+    ) -> torch.Tensor:
+        residual = x
+        x_norm = self.pre_sa_norm(x)
+        sa_out = self.self_attention(
+            Xq=x_norm,
+            Xkv=x_norm,
+            q_positions=state.positions,
+            kv_positions=state.positions,
+            attn_mask=state.attn_mask,
+        )
+        x = residual + sa_out
+
+        residual = x
+        x_norm = self.post_sa_norm(x)
+        mlp_out = self.mlp(x_norm)
+        x = residual + mlp_out
+
+        return x
+
+
+class Encoder(nn.Module):
+    """Transformer Encoder Stack using DenseGeneral."""
+
+    def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
+        super().__init__()
+        self.config = config
+        model_config = config.model
+        enc_config = config.model.encoder
+
+        self.embedding = nn.Embedding(
+            model_config.src_vocab_size,
+            enc_config.n_embd,
+            dtype=compute_dtype,
+        )
+        self.layers = nn.ModuleList(
+            [EncoderLayer(config, compute_dtype) for _ in range(enc_config.n_layer)]
+        )
+        self.norm = RMSNorm(
+            enc_config.n_embd,
+            eps=model_config.normalization_layer_epsilon,
+            dtype=torch.float32,
+        )
+
+    def forward(
+        self,
+        x_ids: torch.Tensor,
+        state: EncoderInferenceState,
+    ) -> torch.Tensor:
+        x = self.embedding(x_ids)
+
+        for layer in self.layers:
+            x = layer(x, state)
+
+        x = self.norm(x)
+        return x
+
+
+class DecoderLayer(nn.Module):
+    """Transformer Decoder Layer using DenseGeneral."""
+
+    def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
+        super().__init__()
+        self.config = config
+        model_config = config.model
+        dec_config = config.model.decoder
+        enc_config = config.model.encoder
+        dec_embed_dim = dec_config.n_embd
+        enc_embed_dim = enc_config.n_embd
+
+        # Norms
+        self.pre_sa_norm = RMSNorm(
+            dec_embed_dim,
+            eps=model_config.normalization_layer_epsilon,
+            dtype=torch.float32,
+        )
+        self.pre_ca_norm = RMSNorm(
+            dec_embed_dim,
+            eps=model_config.normalization_layer_epsilon,
+            dtype=torch.float32,
+        )
+        self.pre_mlp_norm = RMSNorm(
+            dec_embed_dim,
+            eps=model_config.normalization_layer_epsilon,
+            dtype=torch.float32,
+        )
+
+        # Self-Attention (GQA) with Causal Masking
+        self.self_attention = Attention(
+            config,
+            q_embed_dim=dec_embed_dim,
+            kv_embed_dim=dec_embed_dim,
+            num_query_heads=dec_config.gqa_query_heads,
+            num_kv_heads=dec_config.kv_heads,
+            head_dim=dec_config.gqa_head_dim,
+            compute_dtype=compute_dtype,
+            is_cross_attn=False,
+            out_embed_dim=dec_embed_dim,
+        )
+        # Cross-Attention (MHA)
+        self.cross_attention = Attention(
+            config=config,
+            q_embed_dim=dec_embed_dim,
+            kv_embed_dim=enc_embed_dim,  # Note kv_embed_dim
+            num_query_heads=dec_config.cross_query_heads,
+            num_kv_heads=dec_config.cross_query_heads,
+            head_dim=dec_config.cross_head_dim,
+            compute_dtype=compute_dtype,
+            is_cross_attn=True,
+            out_embed_dim=dec_embed_dim,
+        )
+        # MLP
+        self.mlp = MlpBlock(
+            embed_dim=dec_embed_dim,
+            intermediate_dim=dec_config.n_hidden,
+            compute_dtype=compute_dtype,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        state: DecoderInferenceState,
+        self_attn_cache: KVCache | None = None,
+        cross_attn_cache: KVCache | None = None,
+        prefill: bool = False,
+    ) -> torch.Tensor:
+        residual = x
+        x_norm = self.pre_sa_norm(x)
+
+        sa_out = self.self_attention(
+            Xq=x_norm,  # (2, 1, D)
+            Xkv=x_norm,  # (2, 1, D)
+            q_positions=state.dec_positions,  # (2, 1)
+            kv_positions=state.dec_positions,  # (2, 1)
+            attn_mask=None,
+            cache=self_attn_cache,
+            prefill=prefill,
+            is_causal=prefill,
+        )
+
+        x = residual + sa_out
+
+        residual = x
+        x_norm = self.pre_ca_norm(x)
+        ca_out = self.cross_attention(
+            Xq=x_norm,
+            Xkv=state.enc_out,
+            q_positions=state.dec_positions,
+            kv_positions=state.enc_positions,
+            attn_mask=state.dec_cross_attn_mask,
+            cache=cross_attn_cache,
+        )
+        x = residual + ca_out
+
+        residual = x
+        x_norm = self.pre_mlp_norm(x)
+        mlp_out = self.mlp(x_norm)
+        x = residual + mlp_out
+
+        return x
+
+
+class Decoder(nn.Module):
+    """Transformer Decoder Stack using DenseGeneral."""
+
+    def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
+        super().__init__()
+        self.config = config
+        model_config = config.model
+        dec_config = config.model.decoder
+        data_config = config.data
+        self.num_channels = data_config.channels
+        self.num_layers = dec_config.n_layer
+
+        self.embeddings = nn.ModuleList(
+            [
+                nn.Embedding(
+                    model_config.tgt_vocab_size, dec_config.n_embd, dtype=compute_dtype
+                )
+                for _ in range(self.num_channels)
+            ]
+        )
+        self.layers = nn.ModuleList(
+            [
+                DecoderLayer(config=config, compute_dtype=compute_dtype)
+                for _ in range(self.num_layers)
+            ]
+        )
+
+        self.norm = RMSNorm(
+            dec_config.n_embd,
+            eps=model_config.normalization_layer_epsilon,
+            dtype=torch.float32,
+        )
+
+        self.logits_dense = DenseGeneral(
+            in_shapes=(dec_config.n_embd,),
+            out_features=(self.num_channels, model_config.tgt_vocab_size),
+            axis=(-1,),
+            weight_dtype=compute_dtype,
+        )
+
+    def precompute_cross_attn_cache(
+        self,
+        enc_out: torch.Tensor,  # (B, S, E)
+        enc_positions: torch.Tensor,  # (B, S)
+    ) -> list[KVCache]:
+        """
+        Computes the Key and Value tensors for cross-attention for each layer from the encoder output.
+        """
+        per_layer_kv_cache: list[KVCache] = []
+
+        for layer in self.layers:
+            cross_attn_module = layer.cross_attention
+            k_proj = cross_attn_module.k_proj(enc_out)
+            v_proj = cross_attn_module.v_proj(enc_out)
+
+            k_proj = cross_attn_module.rotary_emb(k_proj, position=enc_positions)
+            k = k_proj.transpose(1, 2)
+            v = v_proj.transpose(1, 2)
+
+            per_layer_kv_cache.append(KVCache.from_kv(k, v))
+
+        return per_layer_kv_cache
+
+    def decode_step(
+        self,
+        tgt_ids_Bx1xC: torch.Tensor,  # [B, 1, C]
+        state: DecoderInferenceState,
+    ) -> torch.Tensor:
+        """
+        Performs a single decoding step, managing KV caches layer by layer.
+
+        Returns:
+            A tuple containing:
+            - logits_Bx1xCV: The final output logits for the current step (B, 1, C*V), cast to float32.
+        """
+
+        x = None
+        for i in range(self.num_channels):
+            channel_tokens = tgt_ids_Bx1xC[..., i]
+            channel_embed = self.embeddings[i](channel_tokens)
+            x = channel_embed if x is None else x + channel_embed
+
+        for i, layer in enumerate(self.layers):
+            self_cache = state.self_attn_cache[i]
+            cross_cache = state.cross_attn_cache[i]
+            x = layer(
+                x,  # (2, 1, D)
+                state,
+                self_attn_cache=self_cache,
+                cross_attn_cache=cross_cache,
+            )
+
+        x = self.norm(x)
+        logits_Bx1xCxV = self.logits_dense(x)
+
+        return logits_Bx1xCxV.to(torch.float32)
+
+    def forward(
+        self, tgt_ids_BxTxC: torch.Tensor, state: DecoderInferenceState
+    ) -> torch.Tensor:
+        """
+        Forward pass for the Decoder stack, managing KV caches.
+
+        Args:
+            tgt_ids_BxTxC: Target token IDs (B, T, C).
+            encoder_out: Output from the encoder (B, S, E).
+            tgt_positions: Positions for target sequence (B, T).
+            src_positions: Positions for source sequence (B, S).
+            self_attn_mask: Mask for self-attention.
+            cross_attn_mask: Mask for cross-attention.
+            past_key_values: List containing the self-attention KV cache for each layer
+                             from the previous decoding step. `len(past_key_values)` should
+                             equal `num_layers`.
+            precomputed_cross_attn_kv: A single tuple containing the pre-computed K/V cache
+                                      derived from `encoder_out`. This is passed identically
+                                      to all layers.
+
+        Returns:
+            A tuple containing:
+            - logits: The final output logits (B, T, C * V), cast to float32.
+            - present_key_values: A list containing the updated self-attention KV cache
+                                 for each layer for the *current* decoding step.
+        """
+        _, _, num_channels_in = tgt_ids_BxTxC.shape
+        assert num_channels_in == self.num_channels, "Input channels mismatch"
+
+        # Embeddings
+        x = None
+        for i in range(self.num_channels):
+            channel_tokens = tgt_ids_BxTxC[..., i]
+            channel_embed = self.embeddings[i](channel_tokens)
+            x = channel_embed if x is None else x + channel_embed
+
+        for i, layer in enumerate(self.layers):
+            self_cache = state.self_attn_cache[i]
+            cross_cache = state.cross_attn_cache[i]
+            x = layer(
+                x,
+                state,
+                self_attn_cache=self_cache,
+                cross_attn_cache=cross_cache,
+                prefill=True,
+            )
+
+        # Final Norm
+        x = self.norm(x)
+        logits_BxTxCxV = self.logits_dense(x)
+
+        return logits_BxTxCxV.to(torch.float32)
+
+
+class DiaModel(nn.Module):
+    """PyTorch Dia Model using DenseGeneral."""
+
+    def __init__(self, config: DiaConfig, compute_dtype: torch.dtype):
+        super().__init__()
+        self.config = config
+        self.encoder = Encoder(config, compute_dtype)
+        self.decoder = Decoder(config, compute_dtype)

dia/model.py

@@ -0,0 +1,488 @@
+import time
+from enum import Enum
+
+import dac
+import numpy as np
+import torch
+import torchaudio
+from huggingface_hub import hf_hub_download
+
+from .audio import (
+    apply_audio_delay,
+    build_delay_indices,
+    build_revert_indices,
+    decode,
+    revert_audio_delay,
+)
+from .config import DiaConfig
+from .layers import DiaModel
+from .state import DecoderInferenceState, DecoderOutput, EncoderInferenceState
+
+
+DEFAULT_SAMPLE_RATE = 44100
+
+
+def _get_default_device():
+    if torch.cuda.is_available():
+        return torch.device("cuda")
+    elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
+        return torch.device("mps")
+    return torch.device("cpu")
+
+
+def _sample_next_token(
+    logits_BCxV: torch.Tensor,
+    temperature: float,
+    top_p: float,
+    cfg_filter_top_k: int | None = None,
+) -> torch.Tensor:
+    if temperature == 0.0:
+        return torch.argmax(logits_BCxV, dim=-1)
+
+    logits_BCxV = logits_BCxV / temperature
+    if cfg_filter_top_k is not None:
+        _, top_k_indices_BCxV = torch.topk(logits_BCxV, k=cfg_filter_top_k, dim=-1)
+        mask = torch.ones_like(logits_BCxV, dtype=torch.bool)
+        mask.scatter_(dim=-1, index=top_k_indices_BCxV, value=False)
+        logits_BCxV = logits_BCxV.masked_fill(mask, -torch.inf)
+
+    if top_p < 1.0:
+        probs_BCxV = torch.softmax(logits_BCxV, dim=-1)
+        sorted_probs_BCxV, sorted_indices_BCxV = torch.sort(
+            probs_BCxV, dim=-1, descending=True
+        )
+        cumulative_probs_BCxV = torch.cumsum(sorted_probs_BCxV, dim=-1)
+
+        sorted_indices_to_remove_BCxV = cumulative_probs_BCxV > top_p
+        sorted_indices_to_remove_BCxV[..., 1:] = sorted_indices_to_remove_BCxV[
+            ..., :-1
+        ].clone()
+        sorted_indices_to_remove_BCxV[..., 0] = 0
+
+        indices_to_remove_BCxV = torch.zeros_like(sorted_indices_to_remove_BCxV)
+        indices_to_remove_BCxV.scatter_(
+            dim=-1, index=sorted_indices_BCxV, src=sorted_indices_to_remove_BCxV
+        )
+        logits_BCxV = logits_BCxV.masked_fill(indices_to_remove_BCxV, -torch.inf)
+
+    final_probs_BCxV = torch.softmax(logits_BCxV, dim=-1)
+
+    sampled_indices_BC = torch.multinomial(final_probs_BCxV, num_samples=1)
+    sampled_indices_C = sampled_indices_BC.squeeze(-1)
+    return sampled_indices_C
+
+
+class ComputeDtype(str, Enum):
+    FLOAT32 = "float32"
+    FLOAT16 = "float16"
+    BFLOAT16 = "bfloat16"
+
+    def to_dtype(self) -> torch.dtype:
+        if self == ComputeDtype.FLOAT32:
+            return torch.float32
+        elif self == ComputeDtype.FLOAT16:
+            return torch.float16
+        elif self == ComputeDtype.BFLOAT16:
+            return torch.bfloat16
+        else:
+            raise ValueError(f"Unsupported compute dtype: {self}")
+
+
+class Dia:
+    def __init__(
+        self,
+        config: DiaConfig,
+        compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
+        device: torch.device | None = None,
+    ):
+        """Initializes the Dia model.
+
+        Args:
+            config: The configuration object for the model.
+            device: The device to load the model onto. If None, will automatically select the best available device.
+
+        Raises:
+            RuntimeError: If there is an error loading the DAC model.
+        """
+        super().__init__()
+        self.config = config
+        self.device = device if device is not None else _get_default_device()
+        if isinstance(compute_dtype, str):
+            compute_dtype = ComputeDtype(compute_dtype)
+        self.compute_dtype = compute_dtype.to_dtype()
+        self.model = DiaModel(config, self.compute_dtype)
+        self.dac_model = None
+
+    @classmethod
+    def from_local(
+        cls,
+        config_path: str,
+        checkpoint_path: str,
+        compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
+        device: torch.device | None = None,
+    ) -> "Dia":
+        """Loads the Dia model from local configuration and checkpoint files.
+
+        Args:
+            config_path: Path to the configuration JSON file.
+            checkpoint_path: Path to the model checkpoint (.pth) file.
+            device: The device to load the model onto. If None, will automatically select the best available device.
+
+        Returns:
+            An instance of the Dia model loaded with weights and set to eval mode.
+
+        Raises:
+            FileNotFoundError: If the config or checkpoint file is not found.
+            RuntimeError: If there is an error loading the checkpoint.
+        """
+        config = DiaConfig.load(config_path)
+        if config is None:
+            raise FileNotFoundError(f"Config file not found at {config_path}")
+
+        dia = cls(config, compute_dtype, device)
+
+        try:
+            state_dict = torch.load(checkpoint_path, map_location=dia.device)
+            dia.model.load_state_dict(state_dict)
+        except FileNotFoundError:
+            raise FileNotFoundError(f"Checkpoint file not found at {checkpoint_path}")
+        except Exception as e:
+            raise RuntimeError(
+                f"Error loading checkpoint from {checkpoint_path}"
+            ) from e
+
+        dia.model.to(dia.device)
+        dia.model.eval()
+        dia._load_dac_model()
+        return dia
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        model_name: str = "nari-labs/Dia-1.6B",
+        compute_dtype: str | ComputeDtype = ComputeDtype.FLOAT32,
+        device: torch.device | None = None,
+    ) -> "Dia":
+        """Loads the Dia model from a Hugging Face Hub repository.
+
+        Downloads the configuration and checkpoint files from the specified
+        repository ID and then loads the model.
+
+        Args:
+            model_name: The Hugging Face Hub repository ID (e.g., "NariLabs/Dia-1.6B").
+            device: The device to load the model onto. If None, will automatically select the best available device.
+
+        Returns:
+            An instance of the Dia model loaded with weights and set to eval mode.
+
+        Raises:
+            FileNotFoundError: If config or checkpoint download/loading fails.
+            RuntimeError: If there is an error loading the checkpoint.
+        """
+        config_path = hf_hub_download(repo_id=model_name, filename="config.json")
+        checkpoint_path = hf_hub_download(repo_id=model_name, filename="dia-v0_1.pth")
+        return cls.from_local(config_path, checkpoint_path, compute_dtype, device)
+
+    def _load_dac_model(self):
+        try:
+            dac_model_path = dac.utils.download()
+            dac_model = dac.DAC.load(dac_model_path).to(self.device)
+        except Exception as e:
+            raise RuntimeError("Failed to load DAC model") from e
+        self.dac_model = dac_model
+
+    def _prepare_text_input(self, text: str) -> torch.Tensor:
+        """Encodes text prompt, pads, and creates attention mask and positions."""
+        text_pad_value = self.config.data.text_pad_value
+        max_len = self.config.data.text_length
+
+        byte_text = text.encode("utf-8")
+        replaced_bytes = byte_text.replace(b"[S1]", b"\x01").replace(b"[S2]", b"\x02")
+        text_tokens = list(replaced_bytes)
+
+        current_len = len(text_tokens)
+        padding_needed = max_len - current_len
+        if padding_needed <= 0:
+            text_tokens = text_tokens[:max_len]
+            padded_text_np = np.array(text_tokens, dtype=np.uint8)
+        else:
+            padded_text_np = np.pad(
+                text_tokens,
+                (0, padding_needed),
+                mode="constant",
+                constant_values=text_pad_value,
+            ).astype(np.uint8)
+
+        src_tokens = (
+            torch.from_numpy(padded_text_np).to(torch.long).to(self.device).unsqueeze(0)
+        )  # [1, S]
+        return src_tokens
+
+    def _prepare_audio_prompt(
+        self, audio_prompt: torch.Tensor | None
+    ) -> tuple[torch.Tensor, int]:
+        num_channels = self.config.data.channels
+        audio_bos_value = self.config.data.audio_bos_value
+        audio_pad_value = self.config.data.audio_pad_value
+        delay_pattern = self.config.data.delay_pattern
+        max_delay_pattern = max(delay_pattern)
+
+        prefill = torch.full(
+            (1, num_channels),
+            fill_value=audio_bos_value,
+            dtype=torch.int,
+            device=self.device,
+        )
+
+        prefill_step = 1
+
+        if audio_prompt is not None:
+            prefill_step += audio_prompt.shape[0]
+            prefill = torch.cat([prefill, audio_prompt], dim=0)
+
+        delay_pad_tensor = torch.full(
+            (max_delay_pattern, num_channels),
+            fill_value=-1,
+            dtype=torch.int,
+            device=self.device,
+        )
+        prefill = torch.cat([prefill, delay_pad_tensor], dim=0)
+
+        delay_precomp = build_delay_indices(
+            B=1,
+            T=prefill.shape[0],
+            C=num_channels,
+            delay_pattern=delay_pattern,
+        )
+
+        prefill = apply_audio_delay(
+            audio_BxTxC=prefill.unsqueeze(0),
+            pad_value=audio_pad_value,
+            bos_value=audio_bos_value,
+            precomp=delay_precomp,
+        ).squeeze(0)
+
+        return prefill, prefill_step
+
+    def _prepare_generation(
+        self, text: str, audio_prompt: str | torch.Tensor | None, verbose: bool
+    ):
+        enc_input_cond = self._prepare_text_input(text)
+        enc_input_uncond = torch.zeros_like(enc_input_cond)
+        enc_input = torch.cat([enc_input_uncond, enc_input_cond], dim=0)
+
+        if isinstance(audio_prompt, str):
+            audio_prompt = self.load_audio(audio_prompt)
+        prefill, prefill_step = self._prepare_audio_prompt(audio_prompt)
+
+        if verbose:
+            print("generate: data loaded")
+
+        enc_state = EncoderInferenceState.new(self.config, enc_input_cond)
+        encoder_out = self.model.encoder(enc_input, enc_state)
+
+        dec_cross_attn_cache = self.model.decoder.precompute_cross_attn_cache(
+            encoder_out, enc_state.positions
+        )
+        dec_state = DecoderInferenceState.new(
+            self.config,
+            enc_state,
+            encoder_out,
+            dec_cross_attn_cache,
+            self.compute_dtype,
+        )
+        dec_output = DecoderOutput.new(self.config, self.device)
+        dec_output.prefill(prefill, prefill_step)
+
+        dec_step = prefill_step - 1
+        if dec_step > 0:
+            dec_state.prepare_step(0, dec_step)
+            tokens_BxTxC = (
+                dec_output.get_tokens_at(0, dec_step).unsqueeze(0).expand(2, -1, -1)
+            )
+            self.model.decoder.forward(tokens_BxTxC, dec_state)
+
+        return dec_state, dec_output
+
+    def _decoder_step(
+        self,
+        tokens_Bx1xC: torch.Tensor,
+        dec_state: DecoderInferenceState,
+        cfg_scale: float,
+        temperature: float,
+        top_p: float,
+        cfg_filter_top_k: int,
+    ) -> torch.Tensor:
+        audio_eos_value = self.config.data.audio_eos_value
+        logits_Bx1xCxV = self.model.decoder.decode_step(tokens_Bx1xC, dec_state)
+
+        logits_last_BxCxV = logits_Bx1xCxV[:, -1, :, :]
+        uncond_logits_CxV = logits_last_BxCxV[0, :, :]
+        cond_logits_CxV = logits_last_BxCxV[1, :, :]
+
+        logits_CxV = cond_logits_CxV + cfg_scale * (cond_logits_CxV - uncond_logits_CxV)
+        logits_CxV[:, audio_eos_value + 1 :] = -torch.inf
+        logits_CxV[1:, audio_eos_value:] = -torch.inf
+
+        pred_C = _sample_next_token(
+            logits_CxV.float(),
+            temperature=temperature,
+            top_p=top_p,
+            cfg_filter_top_k=cfg_filter_top_k,
+        )
+        return pred_C
+
+    def _generate_output(self, generated_codes: torch.Tensor) -> np.ndarray:
+        num_channels = self.config.data.channels
+        seq_length = generated_codes.shape[0]
+        delay_pattern = self.config.data.delay_pattern
+        audio_pad_value = self.config.data.audio_pad_value
+        max_delay_pattern = max(delay_pattern)
+
+        revert_precomp = build_revert_indices(
+            B=1,
+            T=seq_length,
+            C=num_channels,
+            delay_pattern=delay_pattern,
+        )
+
+        codebook = revert_audio_delay(
+            audio_BxTxC=generated_codes.unsqueeze(0),
+            pad_value=audio_pad_value,
+            precomp=revert_precomp,
+            T=seq_length,
+        )[:, :-max_delay_pattern, :]
+
+        min_valid_index = 0
+        max_valid_index = 1023
+        invalid_mask = (codebook < min_valid_index) | (codebook > max_valid_index)
+        codebook[invalid_mask] = 0
+
+        audio = decode(self.dac_model, codebook.transpose(1, 2))
+
+        return audio.squeeze().cpu().numpy()
+
+    def load_audio(self, audio_path: str) -> torch.Tensor:
+        audio, sr = torchaudio.load(audio_path, channels_first=True)  # C, T
+        if sr != DEFAULT_SAMPLE_RATE:
+            audio = torchaudio.functional.resample(audio, sr, DEFAULT_SAMPLE_RATE)
+        audio = audio.to(self.device).unsqueeze(0)  # 1, C, T
+        audio_data = self.dac_model.preprocess(audio, DEFAULT_SAMPLE_RATE)
+        _, encoded_frame, _, _, _ = self.dac_model.encode(audio_data)  # 1, C, T
+        return encoded_frame.squeeze(0).transpose(0, 1)
+
+    def save_audio(self, path: str, audio: np.ndarray):
+        import soundfile as sf
+
+        sf.write(path, audio, DEFAULT_SAMPLE_RATE)
+
+    @torch.inference_mode()
+    def generate(
+        self,
+        text: str,
+        max_tokens: int | None = None,
+        cfg_scale: float = 3.0,
+        temperature: float = 1.3,
+        top_p: float = 0.95,
+        use_torch_compile: bool = False,
+        cfg_filter_top_k: int = 35,
+        audio_prompt: str | torch.Tensor | None = None,
+        audio_prompt_path: str | None = None,
+        use_cfg_filter: bool | None = None,
+        verbose: bool = False,
+    ) -> np.ndarray:
+        audio_eos_value = self.config.data.audio_eos_value
+        audio_pad_value = self.config.data.audio_pad_value
+        delay_pattern = self.config.data.delay_pattern
+        max_tokens = self.config.data.audio_length if max_tokens is None else max_tokens
+        max_delay_pattern = max(delay_pattern)
+        self.model.eval()
+
+        if audio_prompt_path:
+            print("Warning: audio_prompt_path is deprecated. Use audio_prompt instead.")
+            audio_prompt = audio_prompt_path
+        if use_cfg_filter is not None:
+            print("Warning: use_cfg_filter is deprecated.")
+
+        if verbose:
+            total_start_time = time.time()
+
+        dec_state, dec_output = self._prepare_generation(text, audio_prompt, verbose)
+        dec_step = dec_output.prefill_step - 1
+
+        bos_countdown = max_delay_pattern
+        eos_detected = False
+        eos_countdown = -1
+
+        if use_torch_compile:
+            step_fn = torch.compile(self._decoder_step, mode="default")
+        else:
+            step_fn = self._decoder_step
+
+        if verbose:
+            print("generate: starting generation loop")
+            if use_torch_compile:
+                print(
+                    "generate: by using use_torch_compile=True, the first step would take long"
+                )
+            start_time = time.time()
+
+        while dec_step < max_tokens:
+            dec_state.prepare_step(dec_step)
+            tokens_Bx1xC = (
+                dec_output.get_tokens_at(dec_step).unsqueeze(0).expand(2, -1, -1)
+            )
+            pred_C = step_fn(
+                tokens_Bx1xC,
+                dec_state,
+                cfg_scale,
+                temperature,
+                top_p,
+                cfg_filter_top_k,
+            )
+
+            if (
+                not eos_detected and pred_C[0] == audio_eos_value
+            ) or dec_step == max_tokens - max_delay_pattern - 1:
+                eos_detected = True
+                eos_countdown = max_delay_pattern
+
+            if eos_countdown > 0:
+                step_after_eos = max_delay_pattern - eos_countdown
+                for i, d in enumerate(delay_pattern):
+                    if step_after_eos == d:
+                        pred_C[i] = audio_eos_value
+                    elif step_after_eos > d:
+                        pred_C[i] = audio_pad_value
+                eos_countdown -= 1
+
+            bos_countdown = max(0, bos_countdown - 1)
+            dec_output.update_one(pred_C, dec_step + 1, bos_countdown > 0)
+
+            if eos_countdown == 0:
+                break
+
+            dec_step += 1
+            if verbose and dec_step % 86 == 0:
+                duration = time.time() - start_time
+                print(
+                    f"generate step {dec_step}: speed={86 / duration:.3f} tokens/s, realtime factor={1 / duration:.3f}x"
+                )
+                start_time = time.time()
+
+        if dec_output.prefill_step >= dec_step + 1:
+            print("Warning: Nothing generated")
+            return None
+
+        generated_codes = dec_output.generated_tokens[
+            dec_output.prefill_step : dec_step + 1, :
+        ]
+
+        if verbose:
+            total_step = dec_step + 1 - dec_output.prefill_step
+            total_duration = time.time() - total_start_time
+            print(
+                f"generate: total step={total_step}, total duration={total_duration:.3f}s"
+            )
+
+        return self._generate_output(generated_codes)

dia/state.py

@@ -0,0 +1,234 @@
+from dataclasses import dataclass
+
+import torch
+
+from .config import DiaConfig
+
+
+def create_attn_mask(
+    q_padding_mask_1d: torch.Tensor,
+    k_padding_mask_1d: torch.Tensor,
+    device: torch.device,
+    is_causal: bool = False,
+) -> torch.Tensor:
+    """
+    Creates the attention mask (self or cross) mimicking JAX segment ID logic.
+    """
+    B1, Tq = q_padding_mask_1d.shape
+    B2, Tk = k_padding_mask_1d.shape
+    assert B1 == B2, "Query and key batch dimensions must match"
+
+    p_mask_q = q_padding_mask_1d.unsqueeze(2)  # Shape [B, Tq, 1]
+    p_mask_k = k_padding_mask_1d.unsqueeze(1)  # Shape [B, 1, Tk]
+
+    # Condition A: Non-padding query attends to non-padding key
+    non_pad_attends_non_pad = p_mask_q & p_mask_k  # Shape [B, Tq, Tk]
+
+    # Condition B: Padding query attends to padding key
+    pad_attends_pad = (~p_mask_q) & (~p_mask_k)  # Shape [B, Tq, Tk]
+
+    # Combine: True if padding status is compatible (both non-pad OR both pad)
+    mask = non_pad_attends_non_pad | pad_attends_pad  # Shape [B, Tq, Tk]
+
+    if is_causal:
+        assert Tq == Tk, (
+            "Causal mask requires query and key sequence lengths to be equal"
+        )
+        causal_mask_2d = torch.tril(
+            torch.ones((Tq, Tk), dtype=torch.bool, device=device)
+        )  # Shape [Tq, Tk]
+        causal_mask = mask & causal_mask_2d  # Shape [B, Tq, Tk]
+        return causal_mask.unsqueeze(1)  # Shape [B, 1, Tq, Tk]
+    else:
+        return mask.unsqueeze(1)  # Shape [B, 1, Tq, Tk]
+
+
+@dataclass
+class EncoderInferenceState:
+    """Parameters specifically for encoder inference."""
+
+    max_seq_len: int
+    device: torch.device
+    positions: torch.Tensor
+    padding_mask: torch.Tensor
+    attn_mask: torch.Tensor
+
+    @classmethod
+    def new(cls, config: DiaConfig, cond_src: torch.Tensor) -> "EncoderInferenceState":
+        """Creates EtorchrInferenceParams from DiaConfig and a device."""
+        device = cond_src.device
+
+        positions = (
+            torch.arange(config.data.text_length, device=device)
+            .to(torch.long)
+            .unsqueeze(0)
+            .expand(2, -1)
+        )
+        padding_mask = (cond_src != config.data.text_pad_value).to(device).expand(2, -1)
+        attn_mask = create_attn_mask(
+            padding_mask, padding_mask, device, is_causal=False
+        )
+
+        return cls(
+            max_seq_len=config.data.text_length,
+            device=device,
+            positions=positions,
+            padding_mask=padding_mask,
+            attn_mask=attn_mask,
+        )
+
+
+class KVCache:
+    def __init__(
+        self,
+        num_heads: int,
+        max_len: int,
+        head_dim: int,
+        dtype: torch.dtype,
+        device: torch.device,
+        k: torch.Tensor | None = None,
+        v: torch.Tensor | None = None,
+    ):
+        self.k = (
+            torch.zeros((2, num_heads, max_len, head_dim), dtype=dtype, device=device)
+            if k is None
+            else k
+        )
+        self.v = (
+            torch.zeros((2, num_heads, max_len, head_dim), dtype=dtype, device=device)
+            if v is None
+            else v
+        )
+        self.current_idx = torch.tensor(0)
+
+    @classmethod
+    def from_kv(cls, k: torch.Tensor, v: torch.Tensor) -> "KVCache":
+        return cls(
+            num_heads=k.shape[1],
+            max_len=k.shape[2],
+            head_dim=k.shape[3],
+            dtype=k.dtype,
+            device=k.device,
+            k=k,
+            v=v,
+        )
+
+    def update(
+        self, k: torch.Tensor, v: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        self.k[:, :, self.current_idx : self.current_idx + 1, :] = k
+        self.v[:, :, self.current_idx : self.current_idx + 1, :] = v
+        self.current_idx += 1
+        return self.k[:, :, : self.current_idx, :], self.v[:, :, : self.current_idx, :]
+
+    def prefill(
+        self, k: torch.Tensor, v: torch.Tensor
+    ) -> tuple[torch.Tensor, torch.Tensor]:
+        prefill_len = k.shape[2]
+        self.k[:, :, :prefill_len, :] = k
+        self.v[:, :, :prefill_len, :] = v
+        self.current_idx = prefill_len - 1
+
+
+@dataclass
+class DecoderInferenceState:
+    """Parameters specifically for decoder inference."""
+
+    device: torch.device
+    dtype: torch.dtype
+    enc_out: torch.Tensor
+    enc_positions: torch.Tensor
+    dec_positions: torch.Tensor
+    dec_cross_attn_mask: torch.Tensor
+    self_attn_cache: list[KVCache]
+    cross_attn_cache: list[KVCache]
+
+    @classmethod
+    def new(
+        cls,
+        config: DiaConfig,
+        enc_state: EncoderInferenceState,
+        enc_out: torch.Tensor,
+        dec_cross_attn_cache: list[KVCache],
+        compute_dtype: torch.dtype,
+    ) -> "DecoderInferenceState":
+        """Creates DecoderInferenceParams from DiaConfig and a device."""
+        device = enc_out.device
+        max_audio_len = config.data.audio_length
+
+        dec_positions = torch.full(
+            (2, 1), fill_value=0, dtype=torch.long, device=device
+        )
+        tgt_padding_mask = torch.ones((2, 1), dtype=torch.bool, device=device)
+        dec_cross_attn_mask = create_attn_mask(
+            tgt_padding_mask, enc_state.padding_mask, device, is_causal=False
+        )
+
+        self_attn_cache = [
+            KVCache(
+                config.model.decoder.kv_heads,
+                max_audio_len,
+                config.model.decoder.gqa_head_dim,
+                compute_dtype,
+                device,
+            )
+            for _ in range(config.model.decoder.n_layer)
+        ]
+
+        return cls(
+            device=device,
+            dtype=compute_dtype,
+            enc_out=enc_out,
+            enc_positions=enc_state.positions,
+            dec_positions=dec_positions,
+            dec_cross_attn_mask=dec_cross_attn_mask,
+            self_attn_cache=self_attn_cache,
+            cross_attn_cache=dec_cross_attn_cache,
+        )
+
+    def prepare_step(self, step_from: int, step_to: int | None = None) -> None:
+        if step_to is None:
+            step_to = step_from + 1
+        self.dec_positions = (
+            torch.arange(step_from, step_to, device=self.device)
+            .unsqueeze(0)
+            .expand(2, -1)
+        )
+
+
+@dataclass
+class DecoderOutput:
+    generated_tokens: torch.Tensor
+    prefill_step: int
+
+    @classmethod
+    def new(cls, config: DiaConfig, device: torch.device) -> "DecoderOutput":
+        max_audio_len = config.data.audio_length
+        return cls(
+            generated_tokens=torch.full(
+                (max_audio_len, config.data.channels),
+                fill_value=-1,
+                dtype=torch.int,
+                device=device,
+            ),
+            prefill_step=0,
+        )
+
+    def get_tokens_at(self, step_from: int, step_to: int | None = None) -> torch.Tensor:
+        if step_to is None:
+            step_to = step_from + 1
+        return self.generated_tokens[step_from:step_to, :]
+
+    def update_one(self, dec_out: torch.Tensor, step: int, apply_mask: bool = False):
+        if apply_mask:
+            mask = self.generated_tokens[step : step + 1, :] == -1
+            self.generated_tokens[step : step + 1, :] = torch.where(
+                mask, dec_out, self.generated_tokens[step : step + 1, :]
+            )
+        else:
+            self.generated_tokens[step : step + 1, :] = dec_out
+
+    def prefill(self, dec_out: torch.Tensor, prefill_step: int):
+        length = dec_out.shape[0]
+        self.generated_tokens[0:length, :] = dec_out
+        self.prefill_step = prefill_step

requirements.txt

@@ -0,0 +1,9 @@
+descript-audio-codec>=1.0.0
+gradio>=5.25.2
+huggingface-hub>=0.30.2
+numpy>=2.2.4
+pydantic>=2.11.3
+soundfile>=0.13.1
+torchaudio>=2.0.0
+torch>=2.0.0
+gradio-dialogue>=0.0.4