qwen3-5lyr-tiny-random

Model Card for Model ID

Code used to create this, 5 layer version of https://huggingface.co/tiny-random/qwen3

This tiny model is for debugging. It is randomly initialized with the config adapted from Qwen/Qwen3-32B.

import torch

from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    GenerationConfig,
    pipeline,
    set_seed,
)

source_model_id = "Qwen/Qwen3-32B"
save_folder = "/tmp/tiny-random/qwen3-5lyr"

tokenizer = AutoTokenizer.from_pretrained(
    source_model_id, trust_remote_code=True,
)
tokenizer.save_pretrained(save_folder)

config = AutoConfig.from_pretrained(
    source_model_id, trust_remote_code=True,
)
config._name_or_path = source_model_id
config.hidden_size = 64
config.intermediate_size = 128
config.head_dim = 32
config.num_key_value_heads = 1
config.num_attention_heads = 2
config.num_hidden_layers = 5 # modified from https://huggingface.co/tiny-random/qwen3
config.max_window_layers = 1
config.tie_word_embeddings = True
model = AutoModelForCausalLM.from_config(
    config,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)
model.generation_config = GenerationConfig.from_pretrained(
    source_model_id, trust_remote_code=True,
)
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.save_pretrained(save_folder)