optimum-intel-internal-testing/tiny-random-lfm2-moe

Here is a code to create this tiny model:

import os
import torch

from transformers import AutoTokenizer, AutoConfig, Lfm2MoeForCausalLM

# # === Step 1: Define tiny model config ===
model_id = "LiquidAI/LFM2-24B-A2B"
config = AutoConfig.from_pretrained(model_id)

config.num_hidden_layers = 3
config.layer_types = [
    "full_attention",
    "full_attention",
    "conv",
]
config.num_attention_heads = 4
config.num_key_value_heads = 4
config.hidden_size = 16
config.num_dense_layers = 1
config.moe_intermediate_size = 16
config.intermediate_size = 16

# === Step 2: Create model from config ===
model = Lfm2MoeForCausalLM(config)

# === Step 3: Load or create tokenizer ===
tokenizer = AutoTokenizer.from_pretrained(model_id)

# === Step 4: Save model and tokenizer ===
output_dir = "./lfm2_moe"
os.makedirs(output_dir, exist_ok=True)
model.save_pretrained(output_dir, safe_serialization=False)
tokenizer.save_pretrained(output_dir)