Add Python conversion script to README

README.md

@@ -18,4 +18,62 @@ Compatible with RKLLM runtime version: 1.2.x
 
 Pretty much anything by these folks: [marty1885](https://github.com/marty1885) and [happyme531](https://huggingface.co/happyme531)
 
-Converted with instructions from [airockchip/rknn-llm #240](https://github.com/airockchip/rknn-llm/issues/240#issuecomment-2831806613)
+## Conversion Python script
+
+Based on instructions from [airockchip/rknn-llm #240](https://github.com/airockchip/rknn-llm/issues/240#issuecomment-2831806613)
+
+### `gemma-3-conversion.py`
+```
+from rkllm.api import RKLLM
+from transformers import Gemma3Processor, Gemma3ForConditionalGeneration
+import safetensors
+import torch
+
+# Unsloth version
+modelpath = 'unsloth/gemma-3-4b-it'
+
+model = Gemma3ForConditionalGeneration.from_pretrained(modelpath, device_map='cpu', torch_dtype=torch.bfloat16).eval()
+processor = Gemma3Processor.from_pretrained(modelpath, use_fast=True)
+
+model.language_model.save_pretrained('llm')
+processor.save_pretrained('llm')
+
+del model
+model = None
+del processor
+processor = None
+
+modelpath = 'llm'
+savepath = 'llm/gemma-3-4b-it-g128.rkllm'
+
+llm = RKLLM()
+
+ret = llm.load_huggingface(model=modelpath, device='cpu')
+if ret != 0:
+    print('Load model failed!')
+    exit(ret)
+
+
+ret = llm.build(
+        do_quantization=True, 
+        optimization_level=0, 
+        quantized_dtype='w8a8', 
+        # hybrid ratio of 25% gives a good balance
+        hybrid_rate=0.25,
+        max_context=4096 * 4,
+        quantized_algorithm='normal', 
+        target_platform='rk3588', 
+        num_npu_core=3, 
+        extra_qparams=None, 
+        dataset=None
+        )
+if ret != 0:
+    print('Build model failed!')
+    exit(ret)
+
+
+ret = llm.export_rkllm(savepath)
+if ret != 0:
+    print('Export model failed!')
+    exit(ret)
+```