Update usage section with proper inference example

README.md

@@ -26,6 +26,27 @@ python sample.py
 
 The checkpoint downloads automatically on first run.
 
+## Usage
+
+```python
+from huggingface_hub import hf_hub_download
+from sample import load_model, sample
+
+# Download checkpoint (~4GB, cached after first run)
+ckpt_path = hf_hub_download(
+    repo_id="justkitting/gpt-diffusion-nanogptsettings",
+    filename="pg19_step_50000_final.pth",
+)
+
+# Load model (uses EMA weights automatically)
+device = "cuda"
+model, config = load_model(ckpt_path, device)
+
+# Generate text
+text = sample(model, config, steps=128, length=512, device=device)
+print(text)
+```
+
 ## Results
 
 | Metric | Value |
@@ -92,18 +113,6 @@ The `.pth` file contains:
 - `config` — model config dict
 - `chunk_idx` — data chunk index
 
-## Usage
-
-```python
-import torch
-from model import GPT, GPTConfig
-
-ckpt = torch.load("pg19_step_50000_final.pth", map_location="cpu")
-config = GPTConfig(**ckpt["config"])
-model = GPT(config)
-model.load_state_dict(ckpt["ema_model"])  # use EMA weights for best quality
-```
-
 ## License
 
 MIT