feat: add full MPS (Apple Silicon) support

CLAUDE.md

@@ -0,0 +1,99 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Project Overview
+
+Dia is a 1.6B parameter text-to-speech (TTS) model by Nari Labs that generates realistic dialogue from text transcripts. It supports multi-speaker dialogue via `[S1]`/`[S2]` tags, voice cloning via audio prompts, and non-verbal generation (laughter, coughing, etc.).
+
+## Common Commands
+
+```bash
+# Install package in development mode
+pip install -e .
+
+# Run with uv (recommended)
+uv run python example/simple.py
+
+# Lint check
+ruff check
+
+# Format check
+ruff format --check --diff
+
+# Run Gradio web UI
+python app.py
+
+# Run CLI
+python cli.py --text "[S1] Hello world."
+```
+
+### Example Scripts
+
+```bash
+python example/simple.py           # Basic generation
+python example/voice_clone.py      # Voice cloning
+python example/simple_batch.py     # Batch generation
+python example/benchmark.py        # Performance benchmarking
+python example/simple-mac.py       # Mac-specific (uses float32)
+python example/simple-cpu.py       # CPU-only
+```
+
+## Architecture
+
+### Core Module Structure (`/dia/`)
+
+- **model.py**: Main `Dia` class - handles model loading, text encoding, audio generation
+- **config.py**: Pydantic configuration classes (`DiaConfig`, `EncoderConfig`, `DecoderConfig`)
+- **layers.py**: Neural network layers (transformer blocks, attention, MLP)
+- **audio.py**: Audio delay pattern utilities for multi-channel generation
+- **state.py**: Inference state management (KV caches, encoder/decoder states)
+
+### Encoder-Decoder Transformer
+
+**Encoder** (12 layers, 1024 hidden, 16 heads):
+- Processes byte-level text tokens (vocab size 256)
+- Max 1024 position embeddings
+
+**Decoder** (18 layers, 2048 hidden, 16 heads with 4 KV heads):
+- Generates 9-channel audio tokens (vocab size 1028)
+- Uses cross-attention to encoder outputs
+- Max 3072 position embeddings
+
+### Audio Processing
+
+- Uses Descript Audio Codec (DAC) for tokenization/detokenization
+- Sample rate: 44100 Hz
+- Delay pattern applied across 9 audio channels for causal generation
+- Key functions: `apply_audio_delay()`, `revert_audio_delay()` in `audio.py`
+
+### Generation Flow
+
+1. Text → byte tokens (with speaker tags)
+2. Encoder pass → context embeddings
+3. Optional audio prompt → DAC encode → delay pattern
+4. Autoregressive decoder generation with CFG
+5. DAC decode → waveform output
+
+### Key Generation Parameters
+
+- `cfg_scale`: Classifier-Free Guidance scale (default 3.0)
+- `temperature`: Sampling temperature (default 1.8)
+- `top_p`: Nucleus sampling (default 0.90)
+- `top_k`: Top-K filtering
+- `seed`: For reproducibility
+
+## Device Support
+
+- **CUDA**: Optimized with float16/bfloat16
+- **MPS** (Apple Silicon): Uses float32
+- **CPU**: Supported, uses float32
+
+Model auto-selects device in `Dia.from_pretrained()`.
+
+## Code Style
+
+- Line length: 119 characters
+- Linter: Ruff
+- Type hints used throughout
+- Pydantic for configuration validation