from transformers import AutoTokenizer, AutoProcessor, AutoModelForCausalLM
from qwen_vl_utils import process_vision_info
import torch

# fix flash_attn_varlen_func, see https://github.com/EvolvingLMMs-Lab/LLaVA-OneVision-1.5/pull/33/files
from flash_attn.flash_attn_interface import flash_attn_varlen_func
import transformers
transformers.modeling_flash_attention_utils.flash_attn_varlen_func = flash_attn_varlen_func

model_id = "yujiepan/llava-onevision-1.5-tiny-random"
model = AutoModelForCausalLM.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="cuda", trust_remote_code=True
)
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=32)
generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

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,
    AutoModelForImageTextToText,
    set_seed,
)
# fix flash_attn_varlen_func, see https://github.com/EvolvingLMMs-Lab/LLaVA-OneVision-1.5/pull/33/files
from flash_attn.flash_attn_interface import flash_attn_varlen_func
import transformers
transformers.modeling_flash_attention_utils.flash_attn_varlen_func = flash_attn_varlen_func

source_model_id = "lmms-lab/LLaVA-OneVision-1.5-8B-Instruct"
save_folder = "/tmp/yujiepan/llava-onevision-1.5-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['text_config'].update({
    'head_dim': 32,
    'hidden_size': 8,
    'intermediate_size': 64,
    'num_hidden_layers': 2,
    'num_attention_heads': 8,
    'num_key_value_heads': 4,
    'layer_types': ['full_attention'] * 2,
    'max_window_layers': 2,
})
config_json['vision_config'].update(
    {
        'depth': 2,
        'intermediate_size': 256,
        'embed_dim': 32 * 4,
        'hidden_size': 32 * 4,
        'text_hidden_size': 8,
        'num_heads': 4,
        'num_hidden_layers': 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)
torch.set_default_dtype(torch.bfloat16)
model = AutoModelForCausalLM.from_config(config, trust_remote_code=True).to(torch.bfloat16)
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:

text

LLaVAOneVision1_5_ForConditionalGeneration(
  (model): LLaVAOneVision1_5_Model(
    (visual): RiceTransformerPretrainedModel(
      (patch_embed): RicePatchEmbed(
        (proj): Conv2d(3, 128, kernel_size=(14, 14), stride=(14, 14), bias=False)
      )
      (rotary_pos_emb): RiceRotaryEmbedding()
      (pre_layernorm): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
      (blocks): ModuleList(
        (0-1): 2 x RiceBlock(
          (norm1): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          (norm2): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
          (attn): RiceSdpaAttention(
            (qkv): Linear(in_features=128, out_features=384, bias=True)
            (proj): Linear(in_features=128, out_features=128, bias=True)
          )
          (mlp): RiceMlp(
            (fc1): Linear(in_features=128, out_features=256, bias=True)
            (act): GELUActivation()
            (fc2): Linear(in_features=256, out_features=128, bias=True)
          )
        )
      )
      (merger): RicePatchMerger(
        (ln_q): LayerNorm((128,), eps=1e-05, elementwise_affine=True)
        (mlp): Sequential(
          (0): Linear(in_features=512, out_features=512, bias=True)
          (1): GELU(approximate='none')
          (2): Linear(in_features=512, out_features=8, bias=True)
        )
      )
    )
    (language_model): LLaVAOneVision1_5_TextModel(
      (embed_tokens): Embedding(151936, 8)
      (layers): ModuleList(
        (0-1): 2 x LLaVAOneVision1_5_DecoderLayer(
          (self_attn): LLaVAOneVision1_5_SdpaAttention(
            (q_proj): Linear(in_features=8, out_features=256, bias=False)
            (k_proj): Linear(in_features=8, out_features=128, bias=False)
            (v_proj): Linear(in_features=8, out_features=128, bias=False)
            (o_proj): Linear(in_features=256, out_features=8, bias=False)
            (q_norm): LLaVAOneVision1_5_RMSNorm((32,), eps=1e-06)
            (k_norm): LLaVAOneVision1_5_RMSNorm((32,), eps=1e-06)
          )
          (mlp): LLaVAOneVision1_5_MLP(
            (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): LLaVAOneVision1_5_RMSNorm((8,), eps=1e-06)
          (post_attention_layernorm): LLaVAOneVision1_5_RMSNorm((8,), eps=1e-06)
        )
      )
      (norm): LLaVAOneVision1_5_RMSNorm((8,), eps=1e-06)
      (rotary_emb): LLaVAOneVision1_5_RotaryEmbedding()
    )
  )
  (lm_head): Linear(in_features=8, out_features=151936, bias=False)
)