import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "tiny-random/glm-ocr"
model = AutoModelForImageTextToText.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="cuda",
)
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)
inputs.pop("token_type_ids", None)
generated_ids = model.generate(**inputs, max_new_tokens=16)
output_text = processor.decode(generated_ids[0], skip_special_tokens=False)
print(output_text)

Codes to create this repo:

Click to expand

python

import json
from copy import deepcopy
from pathlib import Path

import accelerate
import torch
import torch.nn as nn
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoProcessor,
    GenerationConfig,
    GlmOcrForConditionalGeneration,
    set_seed,
)

source_model_id = "zai-org/GLM-OCR"
save_folder = "/tmp/tiny-random/glm-ocr"

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: dict = json.load(f)

config_json['text_config'].update({
    "head_dim": 32,
    "hidden_size": 8,
    "intermediate_size": 64,
    "num_attention_heads": 8,
    "num_hidden_layers": 2,
    "num_key_value_heads": 4,
    "rope_parameters": {
        "rope_type": "default",
        "mrope_section": [4, 4, 8],
        "partial_rotary_factor": 1.0,
        "rope_theta": 10000,
    },
})
config_json['vision_config'].update({
    "hidden_size": 32,
    "depth": 2,
    "num_heads": 1,
    "intermediate_size": 64,
    "out_hidden_size": 8,
})
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 = GlmOcrForConditionalGeneration(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()
set_seed(42)
n_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.numel() / n_params * 100, '%')
# MTP
set_seed(42)
config = config.get_text_config()
model.model.language_model.layers.append(nn.ModuleDict(dict(
    shared_head=nn.ModuleDict(dict(
        norm=nn.RMSNorm(config.hidden_size),
        head=deepcopy(model.model.language_model.embed_tokens),
    )),
    embed_tokens=deepcopy(model.model.language_model.embed_tokens),
    eh_proj=nn.Linear(config.hidden_size * 2,
                      config.hidden_size, bias=False),
    enorm=nn.RMSNorm(config.hidden_size),
    hnorm=nn.RMSNorm(config.hidden_size),
    input_layernorm=nn.RMSNorm(config.hidden_size),
    post_mlp_layernorm=nn.RMSNorm(config.hidden_size),
    post_attention_layernorm=nn.RMSNorm(config.hidden_size),
    post_self_attn_layernorm=nn.RMSNorm(config.hidden_size),
    self_attn=deepcopy(model.model.language_model.layers[1].self_attn),
    mlp=deepcopy(model.model.language_model.layers[1].mlp),
)))
# for i in range(1, len(model.model.language_model.layers)):
#     model.model.language_model.layers[i].mlp.gate.e_score_correction_bias = torch.rand_like(
#         model.model.language_model.layers[i].mlp.gate.e_score_correction_bias).float()
model.save_pretrained(save_folder)
print(model)

Printing the model:

Click to expand

text

GlmOcrForConditionalGeneration(
  (model): GlmOcrModel(
    (visual): GlmOcrVisionModel(
      (patch_embed): GlmOcrVisionPatchEmbed(
        (proj): Conv3d(3, 32, kernel_size=(2, 14, 14), stride=(2, 14, 14))
      )
      (rotary_pos_emb): GlmOcrVisionRotaryEmbedding()
      (blocks): ModuleList(
        (0-1): 2 x GlmOcrVisionBlock(
          (norm1): GlmOcrRMSNorm((32,), eps=1e-05)
          (norm2): GlmOcrRMSNorm((32,), eps=1e-05)
          (attn): GlmOcrVisionAttention(
            (qkv): Linear(in_features=32, out_features=96, bias=True)
            (proj): Linear(in_features=32, out_features=32, bias=True)
            (q_norm): GlmOcrRMSNorm((32,), eps=1e-05)
            (k_norm): GlmOcrRMSNorm((32,), eps=1e-05)
          )
          (mlp): GlmOcrVisionMlp(
            (gate_proj): Linear(in_features=32, out_features=64, bias=True)
            (up_proj): Linear(in_features=32, out_features=64, bias=True)
            (down_proj): Linear(in_features=64, out_features=32, bias=True)
            (act_fn): SiLUActivation()
          )
        )
      )
      (merger): GlmOcrVisionPatchMerger(
        (proj): Linear(in_features=8, out_features=8, bias=False)
        (post_projection_norm): LayerNorm((8,), eps=1e-05, elementwise_affine=True)
        (gate_proj): Linear(in_features=8, out_features=24, bias=False)
        (up_proj): Linear(in_features=8, out_features=24, bias=False)
        (down_proj): Linear(in_features=24, out_features=8, bias=False)
        (act1): GELU(approximate='none')
        (act_fn): SiLUActivation()
      )
      (downsample): Conv2d(32, 8, kernel_size=(2, 2), stride=(2, 2))
      (post_layernorm): GlmOcrRMSNorm((32,), eps=1e-05)
    )
    (language_model): GlmOcrTextModel(
      (embed_tokens): Embedding(59392, 8, padding_idx=59246)
      (layers): ModuleList(
        (0-1): 2 x GlmOcrTextDecoderLayer(
          (self_attn): GlmOcrTextAttention(
            (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)
          )
          (mlp): GlmOcrTextMLP(
            (gate_up_proj): Linear(in_features=8, out_features=128, bias=False)
            (down_proj): Linear(in_features=64, out_features=8, bias=False)
            (activation_fn): SiLUActivation()
          )
          (input_layernorm): GlmOcrRMSNorm((8,), eps=1e-05)
          (post_attention_layernorm): GlmOcrRMSNorm((8,), eps=1e-05)
          (post_self_attn_layernorm): GlmOcrRMSNorm((8,), eps=1e-05)
          (post_mlp_layernorm): GlmOcrRMSNorm((8,), eps=1e-05)
        )
        (2): ModuleDict(
          (shared_head): ModuleDict(
            (norm): RMSNorm((8,), eps=None, elementwise_affine=True)
            (head): Embedding(59392, 8, padding_idx=59246)
          )
          (embed_tokens): Embedding(59392, 8, padding_idx=59246)
          (eh_proj): Linear(in_features=16, out_features=8, bias=False)
          (enorm): RMSNorm((8,), eps=None, elementwise_affine=True)
          (hnorm): RMSNorm((8,), eps=None, elementwise_affine=True)
          (input_layernorm): RMSNorm((8,), eps=None, elementwise_affine=True)
          (post_mlp_layernorm): RMSNorm((8,), eps=None, elementwise_affine=True)
          (post_attention_layernorm): RMSNorm((8,), eps=None, elementwise_affine=True)
          (post_self_attn_layernorm): RMSNorm((8,), eps=None, elementwise_affine=True)
          (self_attn): GlmOcrTextAttention(
            (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)
          )
          (mlp): GlmOcrTextMLP(
            (gate_up_proj): Linear(in_features=8, out_features=128, bias=False)
            (down_proj): Linear(in_features=64, out_features=8, bias=False)
            (activation_fn): SiLUActivation()
          )
        )
      )
      (norm): GlmOcrRMSNorm((8,), eps=1e-05)
      (rotary_emb): GlmOcrTextRotaryEmbedding()
    )
  )
  (lm_head): Linear(in_features=8, out_features=59392, bias=False)
)