smollm3-tiny-random

This tiny model is for debugging. It is randomly initialized with the config adapted from HuggingFaceTB/SmolLM3-3B.

Example usage:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "yujiepan/smollm3-tiny-random"
device = "cuda"  # for GPU usage or "cpu" for CPU usage

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, trust_remote_code=True
).to(device)

# prepare the model input
prompt = "Give me a brief explanation of gravity in simple terms."
messages_think = [
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages_think,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate the output
generated_ids = model.generate(**model_inputs, max_new_tokens=200)

# Get and decode the output
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):]
print(tokenizer.decode(output_ids, skip_special_tokens=True))

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 = "HuggingFaceTB/SmolLM3-3B"
save_folder = "/tmp/yujiepan/smollm3-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'] = 64
config_json['intermediate_size'] = 128
config_json['num_attention_heads'] = 2
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 1
config_json['tie_word_embeddings'] = True
config_json['layer_types'] = None
config_json['no_rope_layer_interval'] = 2
config_json['use_sliding_window'] = True
config_json['sliding_window'] = 128
config_json['use_cache'] = True
config_json['layer_types'] = None
config_json['no_rope_layers'] = None
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()  # cpu is more stable for random initialization across machines
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.2)
        print(name, p.shape)
model.save_pretrained(save_folder)
print(model)

Printing the model:

SmolLM3ForCausalLM(
  (model): SmolLM3Model(
    (embed_tokens): Embedding(128256, 64, padding_idx=128004)
    (layers): ModuleList(
      (0-1): 2 x SmolLM3DecoderLayer(
        (self_attn): SmolLM3Attention(
          (q_proj): Linear(in_features=64, out_features=64, bias=False)
          (k_proj): Linear(in_features=64, out_features=32, bias=False)
          (v_proj): Linear(in_features=64, out_features=32, bias=False)
          (o_proj): Linear(in_features=64, out_features=64, bias=False)
        )
        (mlp): SmolLM3MLP(
          (gate_proj): Linear(in_features=64, out_features=128, bias=False)
          (up_proj): Linear(in_features=64, out_features=128, bias=False)
          (down_proj): Linear(in_features=128, out_features=64, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): SmolLM3RMSNorm((64,), eps=1e-06)
        (post_attention_layernorm): SmolLM3RMSNorm((64,), eps=1e-06)
      )
    )
    (norm): SmolLM3RMSNorm((64,), eps=1e-06)
    (rotary_emb): SmolLM3RotaryEmbedding()
  )
  (lm_head): Linear(in_features=64, out_features=128256, bias=False)
)