This tiny model is for debugging. It is randomly initialized with the config adapted from google/gemma-3-27b-it.
Example usage:
from transformers import pipeline
model_id = "yujiepan/gemma-3-tiny-random"
pipe = pipeline(
"image-text-to-text", model=model_id, device="cuda",
trust_remote_code=True, max_new_tokens=3,
)
messages = [
{
"role": "system",
"content": [{"type": "text", "text": "You are a helpful assistant."}]
},
{
"role": "user",
"content": [
{"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
{"type": "text", "text": "What animal is on the candy?"}
]
}
]
output = pipe(text=messages, max_new_tokens=5)
print(output)
Codes to create this repo:
import torch
from transformers import (
AutoConfig,
AutoModelForCausalLM,
AutoProcessor,
AutoTokenizer,
Gemma3ForConditionalGeneration,
GenerationConfig,
pipeline,
set_seed,
)
source_model_id = "google/gemma-3-27b-it"
save_folder = "/tmp/yujiepan/gemma-3-tiny-random"
processor = AutoProcessor.from_pretrained(
source_model_id, trust_remote_code=True,
)
processor.save_pretrained(save_folder)
config = AutoConfig.from_pretrained(
source_model_id, trust_remote_code=True,
)
config.text_config.hidden_size = 32
config.text_config.intermediate_size = 128
config.text_config.head_dim = 32
config.text_config.num_attention_heads = 1
config.text_config.num_key_value_heads = 1
config.text_config.num_hidden_layers = 2
config.text_config.sliding_window_pattern = 2
config.vision_config.hidden_size = 32
config.vision_config.num_hidden_layers = 2
config.vision_config.num_attention_heads = 1
config.vision_config.intermediate_size = 128
model = Gemma3ForConditionalGeneration(
config,
).to(torch.bfloat16)
for layer in model.language_model.model.layers:
print(layer.is_sliding)
model.generation_config = GenerationConfig.from_pretrained(
source_model_id, trust_remote_code=True,
)
set_seed(42)
with torch.no_grad():
for name, p in sorted(model.named_parameters()):
torch.nn.init.normal_(p, 0, 0.5)
print(name, p.shape)
model.save_pretrained(save_folder)