tiny-random/qwen3-vl

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from Qwen/Qwen3-VL-8B-Thinking.

Example usage:

import numpy as np
import torch
import transformers
from PIL import Image
from transformers import (
    AutoModel,
    AutoModelForCausalLM,
    AutoProcessor,
    AutoTokenizer,
    Qwen3VLForConditionalGeneration,
)

model_id = "tiny-random/qwen3-vl"
model = Qwen3VLForConditionalGeneration.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="cuda",
    attn_implementation="flash_attention_2",
)
processor = AutoProcessor.from_pretrained(model_id)
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
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt"
).to(model.device)

# 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:

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,
    # Qwen3VLMoeForConditionalGeneration,
    Qwen3VLForConditionalGeneration,
    set_seed,
)

source_model_id = "Qwen/Qwen3-VL-8B-Thinking"
save_folder = "/tmp/tiny-random/qwen3-vl"

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['text_config'].update({
    'head_dim': 32,
    'hidden_size': 8,
    'intermediate_size': 64,
    'moe_intermediate_size': 64,
    'num_hidden_layers': 2,
    'num_attention_heads': 8,
    'num_key_value_heads': 4,
})
config_json['text_config']['rope_scaling']['mrope_section'] = [8, 4, 4]
config_json['vision_config'].update(
    {
        'hidden_size': 32 * 4,
        'intermediate_size': 64,
        'num_heads': 4,
        'out_hidden_size': 8,
        'depth': 6,
        'deepstack_visual_indexes': [1, 3, 5],
    }
)
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 = Qwen3VLForConditionalGeneration(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)

Printing the model:

Qwen3VLForConditionalGeneration(
  (model): Qwen3VLModel(
    (visual): Qwen3VLVisionModel(
      (patch_embed): Qwen3VLVisionPatchEmbed(
        (proj): Conv3d(3, 128, kernel_size=(2, 16, 16), stride=(2, 16, 16))
      )
      (pos_embed): Embedding(2304, 128)
      (rotary_pos_emb): Qwen3VLVisionRotaryEmbedding()
      (blocks): ModuleList(
        (0-5): 6 x Qwen3VLVisionBlock(
          (norm1): LayerNorm((128,), eps=1e-06, elementwise_affine=True)
          (norm2): LayerNorm((128,), eps=1e-06, elementwise_affine=True)
          (attn): Qwen3VLVisionAttention(
            (qkv): Linear(in_features=128, out_features=384, bias=True)
            (proj): Linear(in_features=128, out_features=128, bias=True)
          )
          (mlp): Qwen3VLVisionMLP(
            (linear_fc1): Linear(in_features=128, out_features=64, bias=True)
            (linear_fc2): Linear(in_features=64, out_features=128, bias=True)
            (act_fn): PytorchGELUTanh()
          )
        )
      )
      (merger): Qwen3VLVisionPatchMerger(
        (norm): LayerNorm((128,), eps=1e-06, elementwise_affine=True)
        (linear_fc1): Linear(in_features=512, out_features=512, bias=True)
        (act_fn): GELU(approximate='none')
        (linear_fc2): Linear(in_features=512, out_features=8, bias=True)
      )
      (deepstack_merger_list): ModuleList(
        (0-2): 3 x Qwen3VLVisionPatchMerger(
          (norm): LayerNorm((512,), eps=1e-06, elementwise_affine=True)
          (linear_fc1): Linear(in_features=512, out_features=512, bias=True)
          (act_fn): GELU(approximate='none')
          (linear_fc2): Linear(in_features=512, out_features=8, bias=True)
        )
      )
    )
    (language_model): Qwen3VLTextModel(
      (embed_tokens): Embedding(151936, 8)
      (layers): ModuleList(
        (0-1): 2 x Qwen3VLTextDecoderLayer(
          (self_attn): Qwen3VLTextAttention(
            (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): Qwen3VLTextRMSNorm((32,), eps=1e-06)
            (k_norm): Qwen3VLTextRMSNorm((32,), eps=1e-06)
          )
          (mlp): Qwen3VLTextMLP(
            (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): Qwen3VLTextRMSNorm((8,), eps=1e-06)
          (post_attention_layernorm): Qwen3VLTextRMSNorm((8,), eps=1e-06)
        )
      )
      (norm): Qwen3VLTextRMSNorm((8,), eps=1e-06)
      (rotary_emb): Qwen3VLTextRotaryEmbedding()
    )
  )
  (lm_head): Linear(in_features=8, out_features=151936, bias=False)
)