This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from baidu/ERNIE-4.5-VL-424B-A47B-PT.

Example usage:

import numpy as np
import torch
import transformers
from PIL import Image
from transformers import AutoModel, AutoModelForCausalLM, AutoProcessor, AutoTokenizer

model_id = "yujiepan/ernie-4.5-vl-moe-tiny-random"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True,)
model = AutoModel.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    trust_remote_code=True,
)
model.add_image_preprocess(processor)
image = Image.fromarray(np.random.randint(0, 255, (224, 224, 3), dtype=np.uint8), 'RGB')
inputs = processor('What is this: <|IMAGE_START|><|image@placeholder|><|IMAGE_END|>', images=[image]).to('cuda')
# print(inputs)
generated_ids = model.generate(**inputs, max_new_tokens=4, use_cache=False)
output_text = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)
print(output_text)

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 = "baidu/ERNIE-4.5-VL-424B-A47B-PT"
save_folder = "/tmp/yujiepan/ernie-4.5-vl-moe-tiny-random"

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)
for k, v in config_json['auto_map'].items():
    config_json['auto_map'][k] = f'{source_model_id}--{v}'

config_json['hidden_size'] = 8
config_json['intermediate_size'] = 32
# config_json['head_dim'] = 32
config_json['num_attention_heads'] = 4
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 4
config_json['tie_word_embeddings'] = False
config_json['use_cache'] = True

config_json['pixel_hidden_size'] = 16
config_json['moe_layer_start_index'] = 1
config_json['moe_intermediate_size'] = [32, 32]
config_json['moe_num_experts'] = [32, 32]
config_json['vision_config']['depth'] = 2
config_json['vision_config']['embed_dim'] = 16
config_json['vision_config']['hidden_size'] = 16
config_json['vision_config']['num_heads'] = 1

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, 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,
    )
    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)

def modify_automap(path, source_model_id):
    import json
    with open(path, 'r', encoding='utf-8') as f:
        content = json.load(f)
    automap = {}
    if content.get('auto_map', None) is not None:
        for key, value in content.get('auto_map').items():
            if isinstance(value, str):
                value = source_model_id + '--' + value.split('--')[-1]
            else:
                value = [(source_model_id + '--' + v.split('--')[-1]) if '.' in str(v) else v for v in value]
            automap[key] = value
        with open(path, 'w', encoding='utf-8') as f:
            json.dump({**content, 'auto_map': automap}, f, indent=2)

modify_automap(f"{save_folder}/config.json", source_model_id)
modify_automap(f'{save_folder}/processor_config.json', source_model_id)
modify_automap(f'{save_folder}/preprocessor_config.json', source_model_id)
modify_automap(f'{save_folder}/tokenizer_config.json', source_model_id)
for python_file in Path(save_folder).glob('*.py'):
    python_file.unlink()

Printing the model:

Ernie4_5_VLMoeForConditionalGeneration(
  (model): Ernie4_5_Model(
    (embed_tokens): Embedding(103424, 8)
    (layers): ModuleList(
      (0): Ernie4_5_DecoderLayer(
        (self_attn): Ernie4_5_Attention(
          (q_proj): Linear(in_features=8, out_features=8, bias=False)
          (k_proj): Linear(in_features=8, out_features=8, bias=False)
          (v_proj): Linear(in_features=8, out_features=8, bias=False)
          (o_proj): Linear(in_features=8, out_features=8, bias=False)
          (rotary_emb): RopeEmbedding()
        )
        (mlp): Ernie4_5_MLP(
          (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)
        )
        (input_layernorm): RMSNorm()
        (post_attention_layernorm): RMSNorm()
        (residual_add1): FusedDropoutImpl(
          (dropout): Dropout(p=0.0, inplace=False)
        )
        (residual_add2): FusedDropoutImpl(
          (dropout): Dropout(p=0.0, inplace=False)
        )
      )
      (1): Ernie4_5_DecoderLayer(
        (self_attn): Ernie4_5_Attention(
          (q_proj): Linear(in_features=8, out_features=8, bias=False)
          (k_proj): Linear(in_features=8, out_features=8, bias=False)
          (v_proj): Linear(in_features=8, out_features=8, bias=False)
          (o_proj): Linear(in_features=8, out_features=8, bias=False)
          (rotary_emb): RopeEmbedding()
        )
        (mlp): MOEAllGatherLayerV2(
          (gate): TopKGate()
          (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)
            )
          )
          (moe_statics): MoEStatics()
        )
        (input_layernorm): RMSNorm()
        (post_attention_layernorm): RMSNorm()
        (residual_add1): FusedDropoutImpl(
          (dropout): Dropout(p=0.0, inplace=False)
        )
        (residual_add2): FusedDropoutImpl(
          (dropout): Dropout(p=0.0, inplace=False)
        )
      )
    )
    (norm): RMSNorm()
    (resampler_model): VariableResolutionResamplerModel(
      (spatial_linear): Sequential(
        (0): Linear(in_features=64, out_features=64, bias=True)
        (1): GELU(approximate='none')
        (2): Linear(in_features=64, out_features=64, bias=True)
        (3): LayerNorm((64,), eps=1e-06, elementwise_affine=True)
      )
      (temporal_linear): Sequential(
        (0): Linear(in_features=128, out_features=64, bias=True)
        (1): GELU(approximate='none')
        (2): Linear(in_features=64, out_features=64, bias=True)
        (3): LayerNorm((64,), eps=1e-06, elementwise_affine=True)
      )
      (mlp): Linear(in_features=64, out_features=8, bias=True)
      (after_norm): RMSNorm()
    )
  )
  (lm_head): Linear(in_features=8, out_features=103424, bias=False)
  (vision_model): DFNRopeVisionTransformerPreTrainedModel(
    (patch_embed): PatchEmbed(
      (proj): Linear(in_features=588, out_features=16, bias=False)
    )
    (rotary_pos_emb): VisionRotaryEmbedding()
    (blocks): ModuleList(
      (0-1): 2 x DFNRopeVisionBlock(
        (norm1): LayerNorm((16,), eps=1e-06, elementwise_affine=True)
        (norm2): LayerNorm((16,), eps=1e-06, elementwise_affine=True)
        (attn): VisionAttention(
          (qkv): Linear(in_features=16, out_features=48, bias=True)
          (proj): Linear(in_features=16, out_features=16, bias=True)
        )
        (mlp): VisionMlp(
          (fc1): Linear(in_features=16, out_features=64, bias=True)
          (act): QuickGELUActivation()
          (fc2): Linear(in_features=64, out_features=16, bias=True)
        )
      )
    )
    (ln): LayerNorm((16,), eps=1e-06, elementwise_affine=True)
  )
)