wassname/gemma3-5lyr-tiny-random

tiny random Gemma 3 model (5 layers) for tunix

"""Create a tiny random Gemma 3 model (5 layers) and upload to HuggingFace.

One-off script. The model is text-only with random weights, intended for
fast smoke tests with tunix/JAX 

Usage: uv run python scripts/create_tiny_gemma3.py
"""

import torch
from transformers import (
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)
from transformers.models.gemma3 import Gemma3ForCausalLM, Gemma3TextConfig
from huggingface_hub import HfApi

source_model_id = "google/gemma-3-1b-it"
repo_id = "wassname/gemma3-5lyr-tiny-random"
save_folder = "/tmp/tiny-random/gemma3-5lyr"

# Tokenizer from source (same vocab)
tokenizer = AutoTokenizer.from_pretrained(source_model_id)
tokenizer.save_pretrained(save_folder)

# Tiny text-only config matching tunix ModelConfig in model.py
config = Gemma3TextConfig(
    vocab_size=262144,
    hidden_size=64,
    intermediate_size=128,
    num_hidden_layers=5,
    num_attention_heads=2,
    head_dim=32,
    num_key_value_heads=1,
    sliding_window=512,
    tie_word_embeddings=True,
)
config._name_or_path = source_model_id

model = Gemma3ForCausalLM(config).to(torch.bfloat16)

# Random init
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.generation_config = GenerationConfig.from_pretrained(source_model_id)
model.save_pretrained(save_folder)

# Upload
api = HfApi()
api.create_repo(repo_id, exist_ok=True)
api.upload_folder(folder_path=save_folder, repo_id=repo_id)
print(f"Uploaded to https://huggingface.co/{repo_id}")