import torch
from transformers import AutoModelForCausalLM, AutoProcessor

model_id = "tiny-random/step3"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="cuda", torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
            {"type": "text", "text": "What's in this picture?"}
        ]
    },
]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
).to(model.device)
generate_ids = model.generate(**inputs, max_new_tokens=32, do_sample=False)
decoded = processor.decode(generate_ids[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=False)
print(decoded)

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,
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)

source_model_id = "stepfun-ai/step3"
save_folder = "/tmp/tiny-random/step3"

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

def rewrite_automap(filepath: str, source_model_id: str, overrides: dict = None):
    import json
    with open(filepath, 'r', encoding='utf-8') as f:
        config = json.load(f)
    for k, v in config['auto_map'].items():
        v = v.split('--')[-1]
        config['auto_map'][k] = f'{source_model_id}--{v}'
    if overrides is not None:
        config.update(overrides)
    with open(filepath, 'w', encoding='utf - 8') as f:
        json.dump(config, f, indent=2)

rewrite_automap(f'{save_folder}/processor_config.json', source_model_id)
rewrite_automap(f'{save_folder}/tokenizer_config.json', source_model_id)

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['architectures'] = ["Step3VLForConditionalGeneration"]
config_json['text_config'].update({
    "hidden_size": 32,
    "intermediate_size": 64,
    "num_hidden_layers": 2,
    "num_attention_heads": 2,
    "num_attention_groups": 1,
    "head_dim": 256,
    "share_q_dim": 512,
    "moe_layers_enum": "1",
    "moe_num_experts": 8,
    "moe_top_k": 3,
    "moe_intermediate_size": 64,
    "share_expert_dim": 64,
    # "tie_word_embeddings": True,
})
config_json['vision_config'].update({
    "hidden_size": 64,
    "output_hidden_size": 64,
    "intermediate_size": 128,
    "num_hidden_layers": 2,
    "num_attention_heads": 2
})

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)
# key_mapping = {
#     "^vision_model": "model.vision_model",
#     r"^model(?!\.(language_model|vision_model))": "model.language_model",
#     "vit_downsampler": "model.vit_downsampler",
#     "vit_downsampler2": "model.vit_downsampler2",
#     "vit_large_projector": "model.vit_large_projector",
# }
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)
rewrite_automap(f'{save_folder}/config.json', source_model_id)

for python_file in Path(save_folder).glob('*.py'):
    if python_file.name.startswith('modeling_') or python_file.name.startswith('configuration_') or python_file.name.endswith('.py'):
        python_file.unlink()

Printing the model:

text

Step3vForConditionalGeneration(
  (model): Step3vModel(
    (vision_model): StepCLIPVisionTransformer(
      (embeddings): StepCLIPVisionEmbeddings(
        (patch_embedding): Conv2d(3, 64, kernel_size=(14, 14), stride=(14, 14))
        (position_embedding): Embedding(2705, 64)
      )
      (transformer): StepCLIPEncoder(
        (layers): ModuleList(
          (0-1): 2 x StepCLIPEncoderLayer(
            (layer_norm1): LayerNorm((64,), eps=1e-06, elementwise_affine=True)
            (layer_norm2): LayerNorm((64,), eps=1e-06, elementwise_affine=True)
            (self_attn): StepCLIPAttention(
              (qkv_proj): Linear(in_features=64, out_features=192, bias=True)
              (out_proj): Linear(in_features=64, out_features=64, bias=True)
            )
            (mlp): StepCLIPMLP(
              (fc1): Linear(in_features=64, out_features=128, bias=True)
              (act): QuickGELUActivation()
              (fc2): Linear(in_features=128, out_features=64, bias=True)
            )
          )
        )
      )
    )
    (language_model): Step3Model(
      (embed_tokens): Embedding(128815, 32)
      (layers): ModuleList(
        (0): Step3vDecoderLayer(
          (self_attn): Step3vAttention(
            (q_proj): Linear(in_features=32, out_features=512, bias=False)
            (k_proj): Linear(in_features=32, out_features=256, bias=False)
            (v_proj): Linear(in_features=32, out_features=256, bias=False)
            (o_proj): Linear(in_features=512, out_features=32, bias=False)
            (inter_norm): Step3vRMSNorm((512,), eps=1e-05)
            (wq): Linear(in_features=512, out_features=512, bias=False)
          )
          (mlp): Step3vMLP(
            (gate_proj): Linear(in_features=32, out_features=64, bias=False)
            (up_proj): Linear(in_features=32, out_features=64, bias=False)
            (down_proj): Linear(in_features=64, out_features=32, bias=False)
            (act_fn): SiLU()
          )
          (input_layernorm): Step3vRMSNorm((32,), eps=1e-05)
          (post_attention_layernorm): Step3vRMSNorm((32,), eps=1e-05)
        )
        (1): Step3vDecoderLayer(
          (self_attn): Step3vAttention(
            (q_proj): Linear(in_features=32, out_features=512, bias=False)
            (k_proj): Linear(in_features=32, out_features=256, bias=False)
            (v_proj): Linear(in_features=32, out_features=256, bias=False)
            (o_proj): Linear(in_features=512, out_features=32, bias=False)
            (inter_norm): Step3vRMSNorm((512,), eps=1e-05)
            (wq): Linear(in_features=512, out_features=512, bias=False)
          )
          (moe): Step3vMoEMLP(
            (gate): Linear(in_features=32, out_features=8, bias=False)
            (up_proj): MoELinear()
            (gate_proj): MoELinear()
            (down_proj): MoELinear()
            (act_fn): SiLU()
          )
          (share_expert): Step3vMLP(
            (gate_proj): Linear(in_features=32, out_features=64, bias=False)
            (up_proj): Linear(in_features=32, out_features=64, bias=False)
            (down_proj): Linear(in_features=64, out_features=32, bias=False)
            (act_fn): SiLU()
          )
          (input_layernorm): Step3vRMSNorm((32,), eps=1e-05)
          (post_attention_layernorm): Step3vRMSNorm((32,), eps=1e-05)
        )
      )
      (norm): Step3vRMSNorm((32,), eps=1e-05)
      (rotary_emb): Step3vRotaryEmbedding()
    )
    (vit_downsampler): Conv2d(64, 64, kernel_size=(2, 2), stride=(2, 2))
    (vit_downsampler2): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
    (vit_large_projector): Linear(in_features=128, out_features=32, bias=False)
  )
  (lm_head): Linear(in_features=32, out_features=128815, bias=False)
)