gemma-2-tiny-random

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from google/gemma-2-27b-it.

Example usage:

from transformers import pipeline
model_id = "yujiepan/gemma-2-tiny-random"
pipe = pipeline('text-generation', model=model_id, device='cuda', dtype="bfloat16")
print(pipe('Hello World!'))

Codes to create this repo:

import json
from pathlib import Path

import accelerate
import torch
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoProcessor,
    GenerationConfig,
    set_seed,
)

source_model_id = "google/gemma-2-27b-it"
save_folder = "/tmp/yujiepan/gemma-2-tiny-random"

processor = AutoProcessor.from_pretrained(
    source_model_id, trust_remote_code=True)
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['hidden_size'] = 8
config_json['intermediate_size'] = 64
config_json['num_attention_heads'] = 8
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 4
config_json['head_dim'] = 32
config_json['tie_word_embeddings'] = True
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 = AutoModelForCausalLM.from_config(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()
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.1)
        print(name, p.shape)
model.save_pretrained(save_folder)
print(model)

Printing the model:

Gemma2ForCausalLM(
  (model): Gemma2Model(
    (embed_tokens): Embedding(256000, 8, padding_idx=0)
    (layers): ModuleList(
      (0-1): 2 x Gemma2DecoderLayer(
        (self_attn): Gemma2Attention(
          (q_proj): Linear(in_features=8, out_features=256, bias=False)
          (k_proj): Linear(in_features=8, out_features=128, bias=False)
          (v_proj): Linear(in_features=8, out_features=128, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
        )
        (mlp): Gemma2MLP(
          (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): Gemma2RMSNorm((8,), eps=1e-06)
        (post_attention_layernorm): Gemma2RMSNorm((8,), eps=1e-06)
        (pre_feedforward_layernorm): Gemma2RMSNorm((8,), eps=1e-06)
        (post_feedforward_layernorm): Gemma2RMSNorm((8,), eps=1e-06)
      )
    )
    (norm): Gemma2RMSNorm((8,), eps=1e-06)
    (rotary_emb): Gemma2RotaryEmbedding()
  )
  (lm_head): Linear(in_features=8, out_features=256000, bias=False)
)