File size: 1,833 Bytes
dfc1f50 dea814f dfc1f50 dea814f dfc1f50 dea814f dfc1f50 dea814f dfc1f50 dea814f dfc1f50 dea814f dfc1f50 dea814f dfc1f50 dea814f dfc1f50 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 | ---
license: apache-2.0
---
```python
from transformers import (
AutoTokenizer,
Gemma4ForConditionalGeneration,
)
def generate_vlm_model(output_dir="./tiny-random-gemma4-moe"):
from transformers import AutoConfig, AutoProcessor, AutoTokenizer, Gemma4ForConditionalGeneration
config = AutoConfig.from_pretrained("google/gemma-4-26B-A4B-it")
# Text config
config.text_config.global_head_dim = 4
config.text_config.head_dim = 4
config.text_config.hidden_size = 32
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 = 64
config.text_config.dtype = "float32"
# MOE parameters scaled down to avoid CPU plugin crash on SPR
config.text_config.num_experts = 4
config.text_config.top_k_experts = 2
config.text_config.moe_intermediate_size = 64
config.text_config.num_attention_heads = 4
config.text_config.num_key_value_heads = 2
config.text_config.num_global_key_value_heads = 2
# Vision config
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(str(output_dir))
tokenizer = AutoTokenizer.from_pretrained("google/gemma-4-26B-A4B-it")
tokenizer.save_pretrained(str(output_dir))
processor = AutoProcessor.from_pretrained("google/gemma-4-26B-A4B-it")
processor.save_pretrained(str(output_dir))
return model
if __name__ == "__main__":
generate_vlm_model()
``` |