tiny-random/gemma-4-assistant

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from google/gemma-4-31B-it-assistant.

File path	Size
model.safetensors	4.3MB

Example usage:

import torch
from transformers import AutoModelForCausalLM, AutoProcessor, AutoModelForMultimodalLM

model_id = "tiny-random/gemma-4-assistant"
target_model_id = "tiny-random/gemma-4-moe"

processor = AutoProcessor.from_pretrained(target_model_id)
target_model = AutoModelForMultimodalLM.from_pretrained(
    target_model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)
assistant_model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG",
            },
            {"type": "text", "text": "What is shown in this image?"},
        ],
    },
    {
        "role": "assistant",
        "content": [{"type": "text", "text": "Dummy response for image"}],
    },
    {
        "role": "user",
        "content": [
            {
                "type": "video",
                "video": "https://github.com/bebechien/gemma/raw/refs/heads/main/videos/ForBiggerBlazes.mp4",
            },
            {"type": "text", "text": "Describe this video."},
        ],
    },
]
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    return_dict=True,
    return_tensors="pt",
    add_generation_prompt=True,
).to(target_model.device)
input_len = inputs["input_ids"].shape[-1]
print("input_len:", input_len)
outputs = target_model.generate(
    **inputs,
    assistant_model=assistant_model,
    max_new_tokens=32,
)
response = processor.decode(outputs[0], skip_special_tokens=False)
response = response.replace("<|image|>", "I")
response = response.replace("<|video|>", "V")
print(response)

Codes to create this repo:

Click to expand

import json
from pathlib import Path

import torch
from huggingface_hub import file_exists, hf_hub_download

from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoProcessor,
    AutoTokenizer,
    Gemma4AssistantForCausalLM,
    Gemma4ForConditionalGeneration,
    GenerationConfig,
    set_seed,
)

source_model_id = "google/gemma-4-31B-it-assistant"
save_folder = "/tmp/tiny-random/gemma-4-assistant"

processor = AutoProcessor.from_pretrained(source_model_id)
processor.save_pretrained(save_folder)

with open(
    hf_hub_download(source_model_id, filename="config.json", repo_type="model"),
    "r",
    encoding="utf-8",
) as f:
    config_json = json.load(f)

config_json["backbone_hidden_size"] = 8
config_json["text_config"].update(
    {
        "global_head_dim": 64,
        "head_dim": 32,
        "hidden_size": 8,
        "intermediate_size": 64,
        "layer_types": [
            "sliding_attention",
            "sliding_attention",
            "sliding_attention",
            "full_attention",
        ],
        "moe_intermediate_size": 32,
        "num_attention_heads": 8,
        "num_hidden_layers": 4,
        "num_key_value_heads": 4,
    }
)

with open(f"{save_folder}/config.json", "w", encoding="utf-8") as f:
    json.dump(config_json, f, indent=2)

config = AutoConfig.from_pretrained(
    save_folder,
    trust_remote_code=True,
)
print(config)

torch.set_default_dtype(torch.bfloat16)
model = Gemma4AssistantForCausalLM(config)
torch.set_default_dtype(torch.float32)
if file_exists(filename="generation_config.json", repo_id=source_model_id, repo_type="model"):
    model.generation_config = GenerationConfig.from_pretrained(
        source_model_id,
        trust_remote_code=True,
    )
set_seed(42)
model = model.cpu()
all_numels = sum(p.numel() for p in model.parameters())
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.2)
        print(name, p.shape, f"{p.numel() / all_numels * 100: .4f}%")
model.save_pretrained(save_folder)

Printing the model:

Click to expand

Gemma4AssistantForCausalLM(
  (model): Gemma4TextModel(
    (embed_tokens): Gemma4TextScaledWordEmbedding(262144, 8, padding_idx=0)
    (layers): ModuleList(
      (0-2): 3 x Gemma4TextDecoderLayer(
        (self_attn): Gemma4TextAttention(
          (q_proj): Linear(in_features=8, out_features=256, bias=False)
          (q_norm): Gemma4RMSNorm()
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
        )
        (mlp): Gemma4TextMLP(
          (gate_proj): Linear(in_features=8, out_features=64, bias=False)
          (up_proj): Linear(in_features=8, out_features=64, bias=False)
          (down_proj): Linear(in_features=64, out_features=8, bias=False)
          (act_fn): GELUTanh()
        )
        (input_layernorm): Gemma4RMSNorm()
        (post_attention_layernorm): Gemma4RMSNorm()
        (pre_feedforward_layernorm): Gemma4RMSNorm()
        (post_feedforward_layernorm): Gemma4RMSNorm()
      )
      (3): Gemma4TextDecoderLayer(
        (self_attn): Gemma4TextAttention(
          (q_proj): Linear(in_features=8, out_features=512, bias=False)
          (q_norm): Gemma4RMSNorm()
          (o_proj): Linear(in_features=512, out_features=8, bias=False)
        )
        (mlp): Gemma4TextMLP(
          (gate_proj): Linear(in_features=8, out_features=64, bias=False)
          (up_proj): Linear(in_features=8, out_features=64, bias=False)
          (down_proj): Linear(in_features=64, out_features=8, bias=False)
          (act_fn): GELUTanh()
        )
        (input_layernorm): Gemma4RMSNorm()
        (post_attention_layernorm): Gemma4RMSNorm()
        (pre_feedforward_layernorm): Gemma4RMSNorm()
        (post_feedforward_layernorm): Gemma4RMSNorm()
      )
    )
    (norm): Gemma4RMSNorm()
    (rotary_emb): Gemma4TextRotaryEmbedding()
  )
  (lm_head): Linear(in_features=8, out_features=262144, bias=False)
  (pre_projection): Linear(in_features=16, out_features=8, bias=False)
  (post_projection): Linear(in_features=8, out_features=8, bias=False)
)

Test environment:

• torch: 2.10.0+cu130
• transformers: 5.9.0