olmo-3-tiny-random

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from allenai/Olmo-3-32B-Think.

Example usage:

import os
import re

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "yujiepan/olmo-3-tiny-random"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, device_map="auto", torch_dtype=torch.bfloat16)
messages = [
    {"role": "user", "content": "How to make pasta?" * 1500},
]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
)['input_ids']
print(inputs.shape)
outputs = model.generate(inputs.to(
    model.device), max_new_tokens=32)
print(outputs)

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 = "allenai/Olmo-3-32B-Think"
save_folder = "/tmp/yujiepan/olmo-3-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['head_dim'] = 32  # vllm requirement
config_json['intermediate_size'] = 32
config_json['num_attention_heads'] = 8
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 4  # better support tensor parallel
config_json['tie_word_embeddings'] = False
config_json['layer_types'] = ['sliding_attention', 'full_attention']
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, trust_remote_code=True)
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,
    )
    model.generation_config.do_sample = 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, 0.1)
        print(name, p.shape)
model.save_pretrained(save_folder)

Printing the model:

Olmo3ForCausalLM(
  (model): Olmo3Model(
    (embed_tokens): Embedding(100278, 8, padding_idx=100277)
    (layers): ModuleList(
      (0-1): 2 x Olmo3DecoderLayer(
        (self_attn): Olmo3Attention(
          (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)
          (q_norm): Olmo3RMSNorm((256,), eps=1e-06)
          (k_norm): Olmo3RMSNorm((128,), eps=1e-06)
        )
        (mlp): Olmo3MLP(
          (gate_proj): Linear(in_features=8, out_features=32, bias=False)
          (up_proj): Linear(in_features=8, out_features=32, bias=False)
          (down_proj): Linear(in_features=32, out_features=8, bias=False)
          (act_fn): SiLUActivation()
        )
        (post_attention_layernorm): Olmo3RMSNorm((8,), eps=1e-06)
        (post_feedforward_layernorm): Olmo3RMSNorm((8,), eps=1e-06)
      )
    )
    (norm): Olmo3RMSNorm((8,), eps=1e-06)
    (rotary_emb): Olmo3RotaryEmbedding()
  )
  (lm_head): Linear(in_features=8, out_features=100278, bias=False)
)