You can deploy this model and start earning money today!

tiny-random/
ernie-4.5-moe

Name: ernie-4.5-moe
Availability: InStock

baidu/ERNIE-4.5-21B-A3B-Thinking

finetune

endpoints_compatible

Views

No views yet

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from baidu/ERNIE-4.5-21B-A3B-Thinking.

Example usage:

python

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_id = "tiny-random/ernie-4.5-moe"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="bfloat16",
    trust_remote_code=True,
)
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,
    max_new_tokens=32,
)
print(tokenizer.decode(output[0], skip_special_tokens=False))

Codes to create this repo:

python

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 = "baidu/ERNIE-4.5-21B-A3B-Thinking"
save_folder = "/tmp/tiny-random/ernie-4.5-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'] = 8
config_json['intermediate_size'] = 32
config_json['moe_intermediate_size'] = 32
# config_json['moe_k'] = 6
config_json['head_dim'] = 32
config_json['num_attention_heads'] = 16
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 8
config_json['tie_word_embeddings'] = True
config_json['use_cache'] = 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,
    )
    model.generation_config.do_sample = True
    print(model.generation_config)
model = model.cpu()
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:

text

Ernie4_5_MoeForCausalLM(
  (model): Ernie4_5_MoeModel(
    (embed_tokens): Embedding(103424, 8, padding_idx=0)
    (layers): ModuleList(
      (0): Ernie4_5_MoeDecoderLayer(
        (self_attn): Ernie4_5_MoeAttention(
          (q_proj): Linear(in_features=8, out_features=512, bias=False)
          (k_proj): Linear(in_features=8, out_features=256, bias=False)
          (v_proj): Linear(in_features=8, out_features=256, bias=False)
          (o_proj): Linear(in_features=512, out_features=8, bias=False)
        )
        (mlp): Ernie4_5_MoeMLP(
          (gate_proj): Linear(in_features=8, out_features=32, bias=False)
          (up_proj): Linear(in_features=8, out_features=32, bias=False)
          (down_proj): Linear(in_features=32, out_features=8, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): Ernie4_5_MoeRMSNorm((8,), eps=1e-05)
        (post_attention_layernorm): Ernie4_5_MoeRMSNorm((8,), eps=1e-05)
      )
      (1): Ernie4_5_MoeDecoderLayer(
        (self_attn): Ernie4_5_MoeAttention(
          (q_proj): Linear(in_features=8, out_features=512, bias=False)
          (k_proj): Linear(in_features=8, out_features=256, bias=False)
          (v_proj): Linear(in_features=8, out_features=256, bias=False)
          (o_proj): Linear(in_features=512, out_features=8, bias=False)
        )
        (mlp): Ernie4_5_MoeSparseMoeBlock(
          (moe_statics): Ernie4_5_MoeStatics()
          (gate): Linear(in_features=8, out_features=64, bias=False)
          (experts): ModuleList(
            (0-63): 64 x Ernie4_5_MoeMLP(
              (gate_proj): Linear(in_features=8, out_features=32, bias=False)
              (up_proj): Linear(in_features=8, out_features=32, bias=False)
              (down_proj): Linear(in_features=32, out_features=8, bias=False)
              (act_fn): SiLU()
            )
          )
          (shared_experts): Ernie4_5_MoeMLP(
            (gate_proj): Linear(in_features=8, out_features=64, bias=False)
            (up_proj): Linear(in_features=8, out_features=64, bias=False)
            (down_proj): Linear(in_features=64, out_features=8, bias=False)
            (act_fn): SiLU()
          )
        )
        (input_layernorm): Ernie4_5_MoeRMSNorm((8,), eps=1e-05)
        (post_attention_layernorm): Ernie4_5_MoeRMSNorm((8,), eps=1e-05)
      )
    )
    (norm): Ernie4_5_MoeRMSNorm((8,), eps=1e-05)
    (rotary_emb): Ernie4_5_MoeRotaryEmbedding()
  )
  (lm_head): Linear(in_features=8, out_features=103424, bias=False)
)