README.md

---
library_name: transformers
pipeline_tag: image-text-to-text
inference: true
widget:
  - text: Hello!
    example_title: Hello world
    group: Python
base_model:
- THUDM/GLM-4.1V-9B-Thinking
---

This tiny model is for debugging. It is randomly initialized with the config adapted from [THUDM/GLM-4.1V-9B-Thinking](https://huggingface.co/THUDM/GLM-4.1V-9B-Thinking).

### Example usage:

```python
import os
import re

import torch

from transformers import AutoProcessor, Glm4vForConditionalGeneration

model_id = "tiny-random/glm-4v"
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "url": "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
            },
            {
                "type": "text",
                "text": "describe this image"
            }
        ],
    }
]
processor = AutoProcessor.from_pretrained(model_id)
model = Glm4vForConditionalGeneration.from_pretrained(
    pretrained_model_name_or_path=model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
inputs = processor.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_dict=True,
    return_tensors="pt"
).to(model.device)
generated_ids = model.generate(**inputs, max_new_tokens=16)
output_text = processor.decode(generated_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False)
print(output_text)
```

### Codes to create this repo:

```python
import json
from pathlib import Path

import torch

import accelerate
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoProcessor,
    GenerationConfig,
    set_seed,
)
from transformers import AutoProcessor, Glm4vForConditionalGeneration

source_model_id = "THUDM/GLM-4.1V-9B-Thinking"
save_folder = "/tmp/tiny-random/glm-4v"

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'] = 64
config_json['intermediate_size'] = 128
config_json['num_attention_heads'] = 2
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 1
config_json['tie_word_embeddings'] = True
config_json['vision_config']['hidden_size'] = 64
config_json['vision_config']['depth'] = 2
config_json['vision_config']['num_heads'] = 2
config_json['vision_config']['intermediate_size'] = 128
config_json['vision_config']['out_hidden_size'] = 64
config_json['rope_scaling']['mrope_section'] = [2, 2, 4]

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 = Glm4vForConditionalGeneration(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,
    )
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)
```

### Printing the model:

```text
Glm4vForConditionalGeneration(
  (model): Glm4vModel(
    (visual): Glm4vVisionModel(
      (embeddings): Glm4vVisionEmbeddings(
        (position_embedding): Embedding(576, 64)
      )
      (patch_embed): Glm4vVisionPatchEmbed(
        (proj): Conv3d(3, 64, kernel_size=(2, 14, 14), stride=(2, 14, 14))
      )
      (rotary_pos_emb): Glm4vVisionRotaryEmbedding()
      (blocks): ModuleList(
        (0-1): 2 x Glm4vVisionBlock(
          (norm1): Glm4vRMSNorm((64,), eps=1e-05)
          (norm2): Glm4vRMSNorm((64,), eps=1e-05)
          (attn): Glm4vVisionAttention(
            (qkv): Linear(in_features=64, out_features=192, bias=False)
            (proj): Linear(in_features=64, out_features=64, bias=False)
          )
          (mlp): Glm4VisionMlp(
            (gate_proj): Linear(in_features=64, out_features=64, bias=False)
            (up_proj): Linear(in_features=64, out_features=64, bias=False)
            (down_proj): Linear(in_features=64, out_features=64, bias=False)
            (act_fn): SiLU()
          )
        )
      )
      (merger): Glm4vVisionPatchMerger(
        (proj): Linear(in_features=64, out_features=64, bias=False)
        (post_projection_norm): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
        (gate_proj): Linear(in_features=64, out_features=128, bias=False)
        (up_proj): Linear(in_features=64, out_features=128, bias=False)
        (down_proj): Linear(in_features=128, out_features=64, bias=False)
        (act1): GELU(approximate='none')
        (act_fn): SiLU()
      )
      (post_conv_layernorm): Glm4vRMSNorm((64,), eps=1e-05)
      (downsample): Conv2d(64, 64, kernel_size=(2, 2), stride=(2, 2))
      (post_layernorm): Glm4vRMSNorm((64,), eps=1e-05)
    )
    (language_model): Glm4vTextModel(
      (embed_tokens): Embedding(151552, 64, padding_idx=151329)
      (layers): ModuleList(
        (0-1): 2 x Glm4vTextDecoderLayer(
          (self_attn): Glm4vTextAttention(
            (q_proj): Linear(in_features=64, out_features=64, bias=True)
            (k_proj): Linear(in_features=64, out_features=32, bias=True)
            (v_proj): Linear(in_features=64, out_features=32, bias=True)
            (o_proj): Linear(in_features=64, out_features=64, bias=False)
          )
          (mlp): Glm4vTextMLP(
            (gate_up_proj): Linear(in_features=64, out_features=256, bias=False)
            (down_proj): Linear(in_features=128, out_features=64, bias=False)
            (activation_fn): SiLU()
          )
          (input_layernorm): Glm4vRMSNorm((64,), eps=1e-05)
          (post_attention_layernorm): Glm4vRMSNorm((64,), eps=1e-05)
          (post_self_attn_layernorm): Glm4vRMSNorm((64,), eps=1e-05)
          (post_mlp_layernorm): Glm4vRMSNorm((64,), eps=1e-05)
        )
      )
      (norm): Glm4vRMSNorm((64,), eps=1e-05)
      (rotary_emb): Glm4vTextRotaryEmbedding()
    )
  )
  (lm_head): Linear(in_features=64, out_features=151552, bias=False)
)
```