From 8789d942a7e3be3331a95e76f0d17d82c26a990a Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E2=80=9CCortexelus=E2=80=9D?= <“emperorcj@gmail.com”>
Date: Wed, 15 Jul 2026 20:23:06 -0400
Subject: [PATCH] docs: MLX training/gradio + tflite gradio across the
 optimized READMEs

- optimized/README.md (new): index of the mlx / tflite / tensorRT runtimes with
  an inference / web-UI / training capability matrix + install one-liners.
- optimized/mlx/README.md: call out underfit as the recommended (dashboard) way
  to train on a Mac; the CLI stays the lower-level path.
- README.md: note the mlx/tflite web UIs + MLX LoRA training after the hardware
  table, plus a Docs row for MLX training.
---
 README.md               |  3 +++
 optimized/README.md     | 46 +++++++++++++++++++++++++++++++++++++++++
 optimized/mlx/README.md |  8 +++++++
 3 files changed, 57 insertions(+)
 create mode 100644 optimized/README.md

diff --git a/README.md b/README.md
index 668abde..5e4bd02 100644
--- a/README.md
+++ b/README.md
@@ -248,11 +248,14 @@ Optimized inference runtimes are available under [optimized/](optimized) — pic
 | [optimized/tensorRT](optimized/tensorRT) | Linux + NVIDIA GPU | CUDA/TensorRT | `curl -LsSf https://raw.githubusercontent.com/Stability-AI/stable-audio-3/main/optimized/tensorRT/bootstrap.sh \| bash` |
 
 
+Beyond inference, **[optimized/mlx](optimized/mlx)** and **[optimized/tflite](optimized/tflite)** each ship a **web UI** (`./sa3-gradio`), and **optimized/mlx** also does **LoRA training** on Apple Silicon (pure-MLX, no PyTorch) — see its [LoRA training](optimized/mlx/README.md#lora-training) section, or [underfit](https://github.com/dada-bots/underfit) for a full training dashboard built on it.
+
 ## Docs
 
 | Guide | Description |
 |-------|-------------|
 | [Inference Methods](docs/workflows/inference.md) | Overview of inference modes (text-to-audio, inpainting, etc.) |
+| [MLX LoRA training](optimized/mlx/README.md#lora-training) | Finetune on Apple Silicon (pure-MLX); powers underfit's Mac backend |
 | [LoRA Training](docs/workflows/lora.md) | Fine-tune with LoRA: setup, training loop, and checkpointing |
 | [Autoencoder Workflows](docs/workflows/autoencoder.md) | Encode and decode audio with the VAE directly |
 | [Prompting Guide](docs/guides/prompting.md) | Prompt and control signal reference |
diff --git a/optimized/README.md b/optimized/README.md
new file mode 100644
index 0000000..78aa2dd
--- /dev/null
+++ b/optimized/README.md
@@ -0,0 +1,46 @@
+# Optimized runtimes
+
+Platform-specific runtimes for **Stable Audio 3** — each a self-contained,
+minimal-dependency way to run SA3 on a given accelerator. Pick by hardware:
+
+| Runtime | Platforms | Accelerator | Inference | Web UI | Training |
+|---|---|---|---|---|---|
+| **[mlx](mlx)** | Apple Silicon Macs | Metal GPU (MLX) | ✅ | ✅ `./sa3-gradio` | ✅ LoRA (pure-MLX) |
+| **[tflite](tflite)** | macOS / Linux / Windows, x86 & ARM | CPU (LiteRT / XNNPACK) | ✅ | ✅ `./sa3-gradio` | LoRA *inference* only (weight-patching) |
+| **[tensorRT](tensorRT)** | Linux + NVIDIA GPU | CUDA / TensorRT | ✅ | — | — |
+
+All three read the same weights from
+[`stabilityai/stable-audio-3-optimized`](https://huggingface.co/stabilityai/stable-audio-3-optimized)
+(auto-downloaded on first use) and cover text-to-audio, audio-to-audio,
+inpainting, and CFG / negative-prompt guidance.
+
+## One-liners
+
+```bash
+# Apple Silicon (Metal GPU)
+curl -LsSf https://raw.githubusercontent.com/Stability-AI/stable-audio-3/main/optimized/mlx/bootstrap.sh | bash
+
+# Any CPU (macOS / Linux / Windows — needs Git Bash or WSL on Windows for curl|bash)
+curl -LsSf https://raw.githubusercontent.com/Stability-AI/stable-audio-3/main/optimized/tflite/bootstrap.sh | bash
+
+# Linux + NVIDIA
+curl -LsSf https://raw.githubusercontent.com/Stability-AI/stable-audio-3/main/optimized/tensorRT/bootstrap.sh | bash
+```
+
+## Highlights
+
+- **Web UI** — the **mlx** and **tflite** runtimes each ship a gradio app
+  (`./sa3-gradio`) with every generation mode wired (text-to-audio, CFG /
+  negative prompt, audio-to-audio, inpainting), tinted mel-spectrogram
+  previews, model / precision hot-swap, and LoRA support.
+- **LoRA training** — the **mlx** runtime has a complete pure-MLX LoRA trainer
+  (pre-encode → train → generate) that matches
+  [underfit](https://github.com/dada-bots/underfit)'s conventions and
+  checkpoint format, so it doubles as underfit's Apple-Silicon backend. See
+  [mlx → LoRA training](mlx/README.md#lora-training). For a full training
+  **dashboard** on a Mac, use underfit.
+- **LoRA inference** — every runtime loads `.safetensors` adapters (mlx/tflite
+  merge or patch them into the graph; per-adapter strength + sampling-step
+  gating on mlx).
+
+See each runtime's `README.md` for install, usage, flags, and benchmarks.
diff --git a/optimized/mlx/README.md b/optimized/mlx/README.md
index f9c4a7a..4b00ff8 100644
--- a/optimized/mlx/README.md
+++ b/optimized/mlx/README.md
@@ -200,6 +200,14 @@ Apple-Silicon backend). Three steps: **pre-encode → train → generate**. See
 `TRAINING_CONVENTIONS.md` for the complete convention inventory and the
 torch(MPS)-vs-MLX forward+backward parity results.
 
+> **Want a UI instead of the CLI? Use [underfit](https://github.com/dada-bots/underfit).**
+> It's a full LoRA-training dashboard that drives *this* MLX trainer as its
+> Apple-Silicon backend — dataset scanning/encoding, live loss + loss-by-timestep
+> curves, demo MP3s with spectrograms, per-checkpoint gradio launch, and
+> one-click model downloads. On a Mac it's the recommended way to train; follow
+> underfit's *Apple-Silicon quickstart*. The commands below are the lower-level
+> path — good for scripting, CI, or headless runs.
+
 **1. Pre-encode** your audio to SAME latents (once, offline — torch-free):
 
 ```bash