lfm2-moe

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from LiquidAI/LFM2-8B-A1B.

Example usage:

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_id = "tiny-random/lfm2-moe"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="cuda",
    dtype="bfloat16",
    trust_remote_code=True,
    attn_implementation="flash_attention_2",
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Generate answer
prompt="What is AI?"
input_ids=tokenizer.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
).to(model.device)

output=model.generate(
    input_ids,
    do_sample=True,
    temperature=0.3,
    min_p=0.15,
    repetition_penalty=1.05,
    max_new_tokens=32,
)

print(tokenizer.decode(output[0], skip_special_tokens=False))

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 = "LiquidAI/LFM2-8B-A1B"
save_folder = "/tmp/tiny-random/lfm2-moe"

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['layer_types'] = ['conv', 'conv', 'full_attention']
config_json['moe_intermediate_size'] = 128
config_json['num_dense_layers'] = 2
config_json['num_attention_heads'] = 2
config_json['num_hidden_layers'] = 3
config_json['num_key_value_heads'] = 1
config_json['use_cache'] = True
# config_json['tie_word_embeddings'] = True
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, 0.1)
        print(name, p.shape)
model.save_pretrained(save_folder)
print(model)

Printing the model:

Lfm2MoeForCausalLM(
  (model): Lfm2MoeModel(
    (embed_tokens): Embedding(65536, 64, padding_idx=0)
    (layers): ModuleList(
      (0-1): 2 x Lfm2MoeDecoderLayer(
        (conv): Lfm2MoeShortConv(
          (conv): Conv1d(64, 64, kernel_size=(3,), stride=(1,), padding=(2,), groups=64, bias=False)
          (in_proj): Linear(in_features=64, out_features=192, bias=False)
          (out_proj): Linear(in_features=64, out_features=64, bias=False)
        )
        (feed_forward): Lfm2MoeMLP(
          (w1): Linear(in_features=64, out_features=128, bias=False)
          (w3): Linear(in_features=64, out_features=128, bias=False)
          (w2): Linear(in_features=128, out_features=64, bias=False)
        )
        (operator_norm): Lfm2MoeRMSNorm((64,), eps=1e-05)
        (ffn_norm): Lfm2MoeRMSNorm((64,), eps=1e-05)
      )
      (2): Lfm2MoeDecoderLayer(
        (self_attn): Lfm2MoeAttention(
          (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)
          (out_proj): Linear(in_features=64, out_features=64, bias=False)
          (q_layernorm): Lfm2MoeRMSNorm((32,), eps=1e-05)
          (k_layernorm): Lfm2MoeRMSNorm((32,), eps=1e-05)
        )
        (feed_forward): Lfm2MoeSparseMoeBlock(
          (gate): Linear(in_features=64, out_features=32, bias=False)
          (experts): Lfm2MoeExperts(
            (0-31): 32 x Lfm2MoeMLP(
              (w1): Linear(in_features=64, out_features=128, bias=False)
              (w3): Linear(in_features=64, out_features=128, bias=False)
              (w2): Linear(in_features=128, out_features=64, bias=False)
            )
          )
        )
        (operator_norm): Lfm2MoeRMSNorm((64,), eps=1e-05)
        (ffn_norm): Lfm2MoeRMSNorm((64,), eps=1e-05)
      )
    )
    (pos_emb): Lfm2MoeRotaryEmbedding()
    (embedding_norm): Lfm2MoeRMSNorm((64,), eps=1e-05)
  )
  (lm_head): Linear(in_features=64, out_features=65536, bias=False)
)