yujiepan/mamba-tiny-random

This model is randomly initialized, using the config from state-spaces/mamba-2.8b-hf but with smaller size.

Codes:

import os

import torch

import transformers
from huggingface_hub import create_repo, upload_folder

source_model_id = 'state-spaces/mamba-2.8b-hf'
tiny_random_name = 'mamba-tiny-random'
save_path = f'/tmp/yujiepan/{tiny_random_name}'
repo_id = f'yujiepan/{tiny_random_name}'

config = transformers.AutoConfig.from_pretrained(
    source_model_id, trust_remote_code=True)
config.hidden_size = 8
config.expand = 4
config.intermediate_size = 32
config.state_size = 8
config.num_hidden_layers = 2
config.n_layer = 2
config.torch_dtype = torch.bfloat16

model = transformers.AutoModelForCausalLM.from_config(
    config, torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)
model.generation_config = transformers.GenerationConfig.from_pretrained(
    source_model_id,
    trust_remote_code=True,
)

transformers.set_seed(42)
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        print(name, p.shape)
        torch.nn.init.uniform_(p, -0.5, 0.5)

model.save_pretrained(save_path)
tokenizer = transformers.AutoTokenizer.from_pretrained(
    source_model_id, trust_remote_code=True)

result = transformers.pipelines.pipeline(
    'text-generation',
    model=model, tokenizer=tokenizer,
    device='cuda',
    max_new_tokens=16,
)('Hello')
print(result)

model.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)

os.system(f'ls -alh {save_path}')
create_repo(repo_id, exist_ok=True)
upload_folder(repo_id=repo_id, folder_path=save_path)