tiny-random / hunyuan-moe

This tiny model is for debugging. It is randomly initialized with the config adapted from tencent/Hunyuan-A13B-Instruct .

Example usage:

import os
import re

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tiny-random/hunyuan-moe"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
# You may want to use bfloat16 and/or move to GPU here
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Write a short summary of the benefits of regular exercise"},
]
tokenized_chat = tokenizer.apply_chat_template(
    messages, tokenize=True, return_tensors="pt",
    enable_thinking=True,  # Toggle thinking mode (default: True)
)
outputs = model.generate(tokenized_chat.to(model.device), max_new_tokens=32)
output_text = tokenizer.decode(outputs[0])
print(output_text)

Codes to create this repo:

import json
from pathlib import Path

import torch

import accelerate
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)

source_model_id = "tencent/Hunyuan-A13B-Instruct"
save_folder = "/tmp/tiny-random/hunyuan-moe"

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

hf_hub_download(source_model_id, filename='hy.tiktoken', repo_type='model', local_dir=save_folder, local_dir_use_symlinks=False)
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)

for k, v in config_json['auto_map'].items():
    config_json['auto_map'][k] = f'{source_model_id}--{v}'
config_json['attention_head_dim'] = 32
config_json['hidden_size'] = 64
config_json['intermediate_size'] = 128
config_json['moe_intermediate_size'] = [128, 128]
config_json['moe_topk'] = [2, 2]
config_json['num_attention_heads'] = 2
config_json['num_experts'] = 8
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 1
config_json['num_shared_expert'] = [1, 1]
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)
automap = config_json['auto_map']
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,
    )
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, 0.2)
        print(name, p.shape)
model.save_pretrained(save_folder)
print(model)
with open(f"{save_folder}/config.json", "r", encoding='utf-8') as f:
    config_json = json.load(f)
    config_json['auto_map'] = automap
with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
    json.dump(config_json, f, indent=2)
for python_file in Path(save_folder).glob('*.py'):
    if python_file.name.startswith('modeling_') or python_file.name.startswith('configuration_'):
        python_file.unlink()

Printing the model:

HunYuanMoEV1ForCausalLM(
  (model): HunYuanModel(
    (embed_tokens): Embedding(128167, 64, padding_idx=127961)
    (layers): ModuleList(
      (0-1): 2 x HunYuanDecoderLayer(
        (self_attn): HunYuanSdpaAttention(
          (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)
          (o_proj): Linear(in_features=64, out_features=64, bias=False)
          (query_layernorm): HunYuanRMSNorm()
          (key_layernorm): HunYuanRMSNorm()
          (rotary_emb): HunYuanDynamicNTKAlphaRotaryEmbedding()
        )
        (mlp): HunYuanMoE(
          (shared_mlp): HunYuanMLP(
            (gate_proj): Linear(in_features=64, out_features=128, bias=False)
            (up_proj): Linear(in_features=64, out_features=128, bias=False)
            (down_proj): Linear(in_features=128, out_features=64, bias=False)
            (act_fn): SiLU()
          )
          (gate): HunYuanTopKGate(
            (wg): Linear(in_features=64, out_features=8, bias=False)
          )
          (experts): ModuleList(
            (0-7): 8 x HunYuanMLP(
              (gate_proj): Linear(in_features=64, out_features=128, bias=False)
              (up_proj): Linear(in_features=64, out_features=128, bias=False)
              (down_proj): Linear(in_features=128, out_features=64, bias=False)
              (act_fn): SiLU()
            )
          )
        )
        (input_layernorm): HunYuanRMSNorm()
        (post_attention_layernorm): HunYuanRMSNorm()
      )
    )
    (norm): HunYuanRMSNorm()
  )
  (lm_head): Linear(in_features=64, out_features=128167, bias=False)
)