This tiny model is for debugging. It is randomly initialized with the config adapted from zai-org/GLM-4.5V.
Example usage:
import torch
from transformers import AutoProcessor, Glm4vMoeForConditionalGeneration
model_id = "tiny-random/glm-4v-moe"
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 = Glm4vMoeForConditionalGeneration.from_pretrained(
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)
inputs.pop("token_type_ids", None)
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:
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,
Glm4vForConditionalGeneration,
Glm4vMoeForConditionalGeneration,
set_seed,
)
from transformers.models.glm4v_moe.modeling_glm4v_moe import Glm4vMoeTextTopkRouter
source_model_id = "zai-org/GLM-4.5V"
save_folder = "/tmp/tiny-random/glm-4v-moe"
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({
"hidden_size": 32,
"head_dim": 32,
"intermediate_size": 128,
"first_k_dense_replace": 1,
"moe_intermediate_size": 64,
"num_attention_heads": 2,
"num_key_value_heads": 1,
"num_hidden_layers": 2, # one dense, one moe
"tie_word_embeddings": True,
})
config_json['text_config']['rope_scaling']['mrope_section'] = [2, 2, 4]
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'] = config_json['text_config']['hidden_size']
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 = Glm4vMoeForConditionalGeneration(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
num_params = sum(p.numel() for p in model.parameters())
with torch.no_grad():
for name, p in sorted(model.named_parameters()):
torch.nn.init.normal_(p, 0, 0.1)
print(name, p.shape, p.dtype, p.device, f'{p.numel() / num_params * 100: .2f}%')
for _, m in sorted(model.named_modules()):
if isinstance(m, Glm4vMoeTextTopkRouter):
assert 'e_score_correction_bias' in m.state_dict()
torch.nn.init.normal_(m.e_score_correction_bias, 0, 1)
model.save_pretrained(save_folder)
print(model)
Printing the model:
Glm4vMoeForConditionalGeneration(
(model): Glm4vMoeModel(
(visual): Glm4vMoeVisionModel(
(embeddings): Glm4vMoeVisionEmbeddings(
(position_embedding): Embedding(576, 64)
)
(patch_embed): Glm4vMoeVisionPatchEmbed(
(proj): Conv3d(3, 64, kernel_size=(2, 14, 14), stride=(2, 14, 14))
)
(rotary_pos_emb): Glm4vMoeVisionRotaryEmbedding()
(blocks): ModuleList(
(0-1): 2 x Glm4vMoeVisionBlock(
(norm1): Glm4vMoeRMSNorm((64,), eps=1e-05)
(norm2): Glm4vMoeRMSNorm((64,), eps=1e-05)
(attn): Glm4vMoeVisionAttention(
(qkv): Linear(in_features=64, out_features=192, bias=False)
(proj): Linear(in_features=64, out_features=64, bias=False)
)
(mlp): Glm4vMoeisionMlp(
(gate_proj): Linear(in_features=64, out_features=32, bias=False)
(up_proj): Linear(in_features=64, out_features=32, bias=False)
(down_proj): Linear(in_features=32, out_features=64, bias=False)
(act_fn): SiLU()
)
)
)
(merger): Glm4vMoeVisionPatchMerger(
(proj): Linear(in_features=32, out_features=32, bias=False)
(post_projection_norm): LayerNorm((32,), eps=1e-05, elementwise_affine=True)
(gate_proj): Linear(in_features=32, out_features=128, bias=False)
(up_proj): Linear(in_features=32, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=32, bias=False)
(act1): GELU(approximate='none')
(act_fn): SiLU()
)
(post_conv_layernorm): Glm4vMoeRMSNorm((64,), eps=1e-05)
(downsample): Conv2d(64, 32, kernel_size=(2, 2), stride=(2, 2))
(post_layernorm): Glm4vMoeRMSNorm((64,), eps=1e-05)
)
(language_model): Glm4vMoeTextModel(
(embed_tokens): Embedding(151552, 32, padding_idx=151329)
(layers): ModuleList(
(0): Glm4vMoeTextDecoderLayer(
(self_attn): Glm4vMoeTextAttention(
(q_proj): Linear(in_features=32, out_features=64, bias=True)
(k_proj): Linear(in_features=32, out_features=32, bias=True)
(v_proj): Linear(in_features=32, out_features=32, bias=True)
(o_proj): Linear(in_features=64, out_features=32, bias=False)
)
(mlp): Glm4vMoeTextMLP(
(gate_proj): Linear(in_features=32, out_features=128, bias=False)
(up_proj): Linear(in_features=32, out_features=128, bias=False)
(down_proj): Linear(in_features=128, out_features=32, bias=False)
(act_fn): SiLU()
)
(input_layernorm): Glm4vMoeTextRMSNorm((32,), eps=1e-05)
(post_attention_layernorm): Glm4vMoeTextRMSNorm((32,), eps=1e-05)
)
(1): Glm4vMoeTextDecoderLayer(
(self_attn): Glm4vMoeTextAttention(
(q_proj): Linear(in_features=32, out_features=64, bias=True)
(k_proj): Linear(in_features=32, out_features=32, bias=True)
(v_proj): Linear(in_features=32, out_features=32, bias=True)
(o_proj): Linear(in_features=64, out_features=32, bias=False)
)
(mlp): Glm4vMoeTextMoE(
(experts): ModuleList(
(0-127): 128 x Glm4vMoeTextMLP(
(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()
)
)
(gate): Glm4vMoeTextTopkRouter()
(shared_experts): Glm4vMoeTextMLP(
(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): Glm4vMoeTextRMSNorm((32,), eps=1e-05)
(post_attention_layernorm): Glm4vMoeTextRMSNorm((32,), eps=1e-05)
)
)
(norm): Glm4vMoeRMSNorm((32,), eps=1e-05)
(rotary_emb): Glm4vMoeTextRotaryEmbedding()
)
)
(lm_head): Linear(in_features=32, out_features=151552, bias=False)
)