Update README.md

README.md

@@ -1,3 +1,55 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+---
+
+```python
+from transformers import (
+    AutoTokenizer,
+    Gemma4ForConditionalGeneration,
+)
+
+
+def generate_vlm_model(output_dir="./tiny-random-gemma4-moe"):
+    model_tr = Gemma4ForConditionalGeneration.from_pretrained("google/gemma-4-26B-A4B-it")
+    config = model_tr.config
+
+    config.text_config.global_head_dim = 4
+    config.text_config.head_dim = 4
+    config.text_config.hidden_size = 8
+    config.text_config.hidden_size_per_layer_input = 0
+    config.text_config.num_hidden_layers = 2
+    config.text_config.layer_types = ["sliding_attention", "full_attention"]
+    config.text_config.num_kv_shared_layers = 0
+    config.text_config.intermediate_size = 32
+    config.text_config.dtype = "float32"
+
+    config.vision_config.head_dim = 4
+    config.vision_config.hidden_size = 8
+    config.vision_config.intermediate_size = 32
+    config.vision_config.num_hidden_layers = 1
+    config.vision_config.num_key_value_heads = 2
+
+    model = Gemma4ForConditionalGeneration(config)
+    model.eval()
+
+    model.save_pretrained(output_dir)
+
+    tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-26B-A4B-it")
+    tokenizer.save_pretrained(output_dir)
+
+    # Estimate safetensors size
+    import os
+
+    safetensors_path = os.path.join(output_dir, "model.safetensors")
+    if os.path.exists(safetensors_path):
+        size_mb = os.path.getsize(safetensors_path) / (1024 * 1024)
+        print(f"  model.safetensors size: {size_mb:.1f} MB")
+
+    print(f"  VLM model saved to {output_dir}")
+    return model
+
+
+if __name__ == "__main__":
+    generate_vlm_model()
+
+```