README.md

metadata

license: other
language:
  - zh
  - en
tags:
  - chatglm
  - glm-4v
  - quantization
  - auto-gptq
  - 4bit

Usage

This model is quantized using AutoGPTQ for THUDM/glm-4v-9b.

Use pip install AutoGPTQ (required)

(The quantization script will be released later)

pip install auto-gptq

Since the original auto-gptq library does not support the quantization of chatglm models, manual import (hack) is required.

from auto_gptq.modeling._base import BaseGPTQForCausalLM
from auto_gptq.modeling._const import SUPPORTED_MODELS

from auto_gptq.modeling.auto import GPTQ_CAUSAL_LM_MODEL_MAP

class ChatGLMGPTQForCausalLM(BaseGPTQForCausalLM):
    layer_type = ["GLMBlock", "TransformerLayer", "GLU"]

    layers_block_names = ["transformer.encoder.layers", 
                            "transformer.vision.transformer.layers", 
                            "transformer.vision.linear_proj"]
        
    outside_layer_modules = ["transformer.output_layer"]
    
    inside_layer_modules = [
        ["self_attention.query_key_value", "self_attention.dense", "mlp.dense_h_to_4h", "mlp.dense_4h_to_h"],
        ["attention.query_key_value", "attention.dense", "mlp.fc1", "mlp.fc2"],
        ["linear_proj", "dense_h_to_4h", "gate_proj", "dense_4h_to_h"],
    ]

GPTQ_CAUSAL_LM_MODEL_MAP['chatglm'] = ChatGLMGPTQForCausalLM
SUPPORTED_MODELS = SUPPORTED_MODELS.append('chatglm')

The complete model import code is as follows:

Load model

import os

import json
import random
import time

import torch
import datasets
from transformers import AutoTokenizer, AutoModelForCausalLM
from auto_gptq import AutoGPTQForCausalLM


from auto_gptq.modeling._base import BaseGPTQForCausalLM
from auto_gptq.modeling._const import SUPPORTED_MODELS

from auto_gptq.modeling.auto import GPTQ_CAUSAL_LM_MODEL_MAP

class ChatGLMGPTQForCausalLM(BaseGPTQForCausalLM):
    layer_type = ["GLMBlock", "TransformerLayer", "GLU"]

    layers_block_names = ["transformer.encoder.layers", 
                            "transformer.vision.transformer.layers", 
                            "transformer.vision.linear_proj"]
        
    outside_layer_modules = ["transformer.output_layer"]
    
    inside_layer_modules = [
        ["self_attention.query_key_value", "self_attention.dense", "mlp.dense_h_to_4h", "mlp.dense_4h_to_h"],
        ["attention.query_key_value", "attention.dense", "mlp.fc1", "mlp.fc2"],
        ["linear_proj", "dense_h_to_4h", "gate_proj", "dense_4h_to_h"],
    ]

GPTQ_CAUSAL_LM_MODEL_MAP['chatglm'] = ChatGLMGPTQForCausalLM
SUPPORTED_MODELS = SUPPORTED_MODELS.append('chatglm')

device = 'cuda:0'
quantized_model_dir = 'alexwww94/glm-4v-9b-gptq'
trust_remote_code = True

tokenizer = AutoTokenizer.from_pretrained(
    quantized_model_dir,
    trust_remote_code=trust_remote_code,
)

model = AutoGPTQForCausalLM.from_quantized(
    quantized_model_dir,
    device=device,
    trust_remote_code=trust_remote_code,
    torch_dtype=torch.float16,
    use_cache=True,
    inject_fused_mlp=True,
    inject_fused_attention=True,
)

You can also load the model using HuggingFace Transformers, but it will slow down inference.

import os

import json
import random
import time

import torch
import datasets
from transformers import AutoTokenizer, AutoModelForCausalLM

device = 'cuda:0'
quantized_model_dir = 'alexwww94/glm-4v-9b-gptq-4bit'
trust_remote_code = True

tokenizer = AutoTokenizer.from_pretrained(
    quantized_model_dir,
    trust_remote_code=trust_remote_code,
)

model = AutoModelForCausalLM.from_pretrained(
    quantized_model_dir,
    torch_dtype=torch.float16,
    device_map="auto",
    low_cpu_mem_usage=True,
    trust_remote_code=trust_remote_code,
    use_cache=True
).eval()

inference test

Load the CogVLM-SFT-311K-subset-gptq dataset as test data, which is a dataset for quantization.

dataset = datasets.load_dataset('alexwww94/CogVLM-SFT-311K-subset-gptq')

for example in dataset['single']:
    # prompt = "为什么马会被围栏限制在一个区域内？"
    prompt = json.loads(example['labels_zh'])['conversations'][0]
    answer = json.loads(example['labels_zh'])['conversations'][1]
    image = example['image']
    print(f"prompt: {prompt}")
    print("-" * 42)
    print(f"golden: {answer}")
    print("-" * 42)

    start = time.time()

    prompt.update({'image': image})
    inputs = tokenizer.apply_chat_template([prompt],
                                    add_generation_prompt=True, tokenize=True, return_tensors="pt",
                                    return_dict=True, dtyp=torch.bfloat16)  # chat mode
    inputs = inputs.to(device)
    inputs['images'] = inputs['images'].half()

    gen_kwargs = {"max_length": 2500, "do_sample": True, "top_k": 1}
    with torch.inference_mode():
        outputs = model.generate(**inputs, **gen_kwargs)
        outputs = outputs[:, inputs['input_ids'].shape[1]:]
        generated_text = tokenizer.decode(outputs[0]).split('<|endoftext|>')[0]

    end = time.time()
    print(f"quant: {generated_text}")
    num_new_tokens = len(tokenizer(generated_text)["input_ids"])
    print(f"generate {num_new_tokens} tokens using {end-start: .4f}s, {num_new_tokens / (end - start)} tokens/s.")
    print("=" * 42)

    # break

metrics

(to be released later)