bits: 4
dynamic: null
group_size: 128
desc_act: true
static_groups: false
sym: false
lm_head: false
true_sequential: true
quant_method: "gptq"
checkpoint_format: "gptq"
meta：
- quantizer: gptqmodel:1.1.0
- uri: https://github.com/modelcloud/gptqmodel
- damp_percent: 0.1
- damp_auto_increment: 0.0015
Benchmark

Performance of Quantized Models

Dataset	GPTQ-4bit	FP16
mmlu	65.20	65.9
cmmlu	69.23	69.9
arc_challenge	45.48	47.9(0)
Example:

python
# Copyright 2024-2025 ModelCloud.ai
# Copyright 2024-2025 qubitium@modelcloud.ai
# Contact: qubitium@modelcloud.ai, x.com/qubitium
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import torch
from datasets import load_dataset
from gptqmodel import  GPTQModel, QuantizeConfig,  BACKEND
from gptqmodel.models.base import BaseGPTQModel 
from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModel
from gptqmodel.models.auto import MODEL_MAP 
import torch.nn.functional as F 
import numpy as np





pretrained_model_id = '/home/chentianqi/model/GSAI-ML/LLaDA-8B-Base' # "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
quantized_model_id = "FunAGI/LLaDA-8B-Base-gptqmodel-4bit"



class LladaGPTQ(BaseGPTQModel):
    # Non-repeating layers at the root level: same level as `layers_node`
    # Excluding `layers_node`.
    base_modules = ["model.transformer.wte", "model.transformer.ln_f"]
    pre_lm_head_norm_module = "model.transformer.ln_f"
    lm_head = "model.transformer.ff_out"
    # Below describes all the repeating layers in this transformer model
    # `model.layers` is a node/module that hold all the repeating layers. The parent node for all n-layers.
    layers_node = "model.transformer.blocks"
    # Each repeating layer in `model.layers` is of type `LlamaDecoderLayer`
    layer_type = "LLaDALlamaBlock"
    # Inside each `LlamaDecoderLayer` layer are many internal modules
    # List them in the order executed in model forward() code
    # Many models have same execution order of: attention (q_k_v) projection, attention (output) projection, mlp (n) projections
    layer_modules = [
        ["attn_out", "k_proj", "v_proj", "q_proj"],
        ["ff_proj", "up_proj"],
        ["ff_out"],
    ]
MODEL_MAP ["llada"] = LladaGPTQ

# os.makedirs(quantized_model_dir, exist_ok=True)
def get_wikitext2(tokenizer, nsamples, seqlen):
    traindata = load_dataset("wikitext", "wikitext-2-raw-v1", split="train").filter(
        lambda x: len(x["text"]) >= seqlen)

    return [tokenizer(example["text"]) for example in traindata.select(range(nsamples))]


@torch.no_grad()
def calculate_avg_ppl(model, tokenizer):
    from gptqmodel.utils import Perplexity

    ppl = Perplexity(
        model=model,
        tokenizer=tokenizer,
        dataset_path="wikitext",
        dataset_name="wikitext-2-raw-v1",
        split="train",
        text_column="text",
    )

    all = ppl.calculate(n_ctx=512, n_batch=512)

    # average ppl
    avg = sum(all) / len(all)

    return avg

dynamic = { 
     
 } 

def add_gumbel_noise(logits, temperature):
    '''
    The Gumbel max is a method for sampling categorical distributions.
    According to arXiv:2409.02908, for MDM, low-precision Gumbel Max improves perplexity score but reduces generation quality.
    Thus, we use float64.
    '''
    logits = logits.to(torch.float64)
    noise = torch.rand_like(logits, dtype=torch.float64)
    gumbel_noise = (- torch.log(noise)) ** temperature
    return logits.exp() / gumbel_noise


def get_num_transfer_tokens(mask_index, steps):
    '''
    In the reverse process, the interval [0, 1] is uniformly discretized into steps intervals.
    Furthermore, because LLaDA employs a linear noise schedule (as defined in Eq. (8)),
    the expected number of tokens transitioned at each step should be consistent.

    This function is designed to precompute the number of tokens that need to be transitioned at each step.
    '''
    mask_num = mask_index.sum(dim=1, keepdim=True) #

    base = mask_num // steps
    remainder = mask_num % steps

    num_transfer_tokens = torch.zeros(mask_num.size(0), steps, device=mask_index.device, dtype=torch.int64) + base

    for i in range(mask_num.size(0)):
        num_transfer_tokens[i, :remainder[i]] += 1

    return num_transfer_tokens

def forward_process(batch, prompt_index, mask_id):
    b, l = batch.shape

    target_len = (l - prompt_index.sum()).item()
    k = torch.randint(1, target_len + 1, (), device=batch.device)

    x = torch.round(torch.linspace(float(k), k + (b - 1) * (target_len / b), steps=b, device=batch.device)).long()
    x = ((x - 1) % target_len) + 1
    assert x.min() >= 1 and x.max() <= target_len

    indices = torch.arange(target_len, device=batch.device).repeat(b, 1)
    is_mask = indices < x.unsqueeze(1)
    for i in range(b):
        is_mask[i] = is_mask[i][torch.randperm(target_len)]

    is_mask = torch.cat((torch.zeros(b, prompt_index.sum(), dtype=torch.bool, device=batch.device), is_mask), dim=1)
    noisy_batch = torch.where(is_mask, mask_id, batch)

    # Return the masked batch and the mask ratio
    return noisy_batch, (x / target_len).unsqueeze(1).repeat(1, l)


def get_logits(model, batch, prompt_index, cfg_scale, mask_id):
    if cfg_scale > 0.:
        assert len(prompt_index) == batch.shape[1]
        prompt_index = prompt_index.unsqueeze(0).repeat(batch.shape[0], 1)
        un_batch = batch.clone()
        un_batch[prompt_index] = mask_id
        batch = torch.cat([batch, un_batch])

    input = batch
    logits = model(input).logits

    if cfg_scale > 0.:
        logits, un_logits = torch.chunk(logits, 2, dim=0)
        logits = un_logits + (cfg_scale + 1) * (logits - un_logits)
    return logits



@ torch.no_grad()
def get_log_likelihood(model, prompt, answer, mc_num=128, batch_size=32, cfg_scale=0., mask_id=126336):
    '''
    Args:
        model: Mask predictor.
        prompt: A tensor of shape (l1).
        answer: A tensor of shape (l2).
        mc_num: Monte Carlo estimation times.
                As detailed in Appendix B.5. Since MMLU, CMMLU, and C-EVAL only require the likelihood of a single token, a
                single Monte Carlo estimate is sufficient for these benchmarks. For all other benchmarks, we find that 128
                Monte Carlo samples are adequate to produce stable results.
        batch_size: Mini batch size.
        cfg_scale: Unsupervised classifier-free guidance scale.
        mask_id: The toke id of [MASK] is 126336.
    '''

    seq = torch.concatenate([prompt, answer])[None, :]
    seq = seq.repeat((batch_size, 1)).to(model.device)
    prompt_index = torch.arange(seq.shape[1], device=model.device) < len(prompt)

    loss_ = []
    for _ in range(mc_num // batch_size):
  
        perturbed_seq, p_mask = forward_process(seq, prompt_index, mask_id)
        mask_index = perturbed_seq == mask_id

        logits = get_logits(model, perturbed_seq, prompt_index, cfg_scale, mask_id)

        loss = F.cross_entropy(logits[mask_index], seq[mask_index], reduction='none') / p_mask[mask_index]
        loss = loss.sum() / batch_size

        loss_.append(loss.item())

    return - sum(loss_) / len(loss_)








@ torch.no_grad()
def generate(model, prompt, steps=128, gen_length=128, block_length=128, temperature=0.,
             cfg_scale=0., remasking='low_confidence', mask_id=126336):
    '''
    Args:
        model: Mask predictor.
        prompt: A tensor of shape (1, l).
        steps: Sampling steps, less than or equal to gen_length.
        gen_length: Generated answer length.
        block_length: Block length, less than or equal to gen_length. If less than gen_length, it means using semi_autoregressive remasking.
        temperature: Categorical distribution sampling temperature.
        cfg_scale: Unsupervised classifier-free guidance scale.
        remasking: Remasking strategy. 'low_confidence' or 'random'.
        mask_id: The toke id of [MASK] is 126336.
    '''
    x = torch.full((1, prompt.shape[1] + gen_length), mask_id, dtype=torch.long).to(model.device)
    x[:, :prompt.shape[1]] = prompt.clone()

    prompt_index = (x != mask_id)

    assert gen_length % block_length == 0
    num_blocks = gen_length // block_length

    assert steps % num_blocks == 0
    steps = steps // num_blocks

    for num_block in range(num_blocks):
        block_mask_index = (x[:, prompt.shape[1] + num_block * block_length: prompt.shape[1] + (num_block + 1) * block_length:] == mask_id)
        num_transfer_tokens = get_num_transfer_tokens(block_mask_index, steps)
        for i in range(steps):

            mask_index = (x == mask_id)
            if cfg_scale > 0.:
                un_x = x.clone()
                un_x[prompt_index] = mask_id
                x_ = torch.cat([x, un_x], dim=0)
                logits = model(x_).logits
                logits, un_logits = torch.chunk(logits, 2, dim=0)
                logits = un_logits + (cfg_scale + 1) * (logits - un_logits)
            else:
                logits = model(x).logits

            logits_with_noise = add_gumbel_noise(logits, temperature=temperature)
            x0 = torch.argmax(logits_with_noise, dim=-1) # b, l

            if remasking == 'low_confidence':
                p = F.softmax(logits.to(torch.float64), dim=-1)
                x0_p = torch.squeeze(
                    torch.gather(p, dim=-1, index=torch.unsqueeze(x0, -1)), -1) # b, l
            elif remasking == 'random':
                x0_p = torch.rand((x0.shape[0], x0.shape[1]), device=x0.device)
            else:
                raise NotImplementedError(remasking)

            x0_p[:, prompt.shape[1] + (num_block + 1) * block_length:] = -np.inf

            x0 = torch.where(mask_index, x0, x)
            confidence = torch.where(mask_index, x0_p, -np.inf)

            transfer_index = torch.zeros_like(x0, dtype=torch.bool, device=x0.device)
            for j in range(confidence.shape[0]):
                _, select_index = torch.topk(confidence[j], k=num_transfer_tokens[j, i])
                transfer_index[j, select_index] = True
            x[transfer_index] = x0[transfer_index]

    return x

def main():
    tokenizer = AutoTokenizer.from_pretrained(pretrained_model_id, use_fast=False)

    traindataset = get_wikitext2(tokenizer, nsamples=128, seqlen=1024)

    quantize_config = QuantizeConfig(
        dynamic=dynamic,
        bits=8,  # quantize model to 4-bit
        group_size=128,  # it is recommended to set the value to 128,
        desc_act = True, 
        sym=False
    )
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    prompt = "Question: Lily can run 12 kilometers per hour for 4 hours. After that, she runs 6 kilometers per hour. How many kilometers can she run in 8 hours? The answer: "

    # # Add special tokens for the Instruct model. The Base model does not require the following two lines.
    # m = [{"role": "user", "content": prompt}, ]
    # prompt = tokenizer.apply_chat_template(m, add_generation_prompt=True, tokenize=False)

    input_ids = tokenizer(prompt)['input_ids']
    input_ids = torch.tensor(input_ids).to(device).unsqueeze(0)



    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    model = GPTQModel.load(quantized_model_id, device=device , trust_remote_code=True    )

    steps=128
    out = generate(model, input_ids, steps=steps , gen_length=128, block_length=32, temperature=0., cfg_scale=0., remasking='low_confidence')
    print("*"*30+ f"GPTQ-4bit Steps {steps}"+ "*"*30)
    print(input_ids.shape)
    print( tokenizer.batch_decode(out[:, input_ids.shape[1]:], skip_special_tokens=True)[0])
    del model 

    model =AutoModel.from_pretrained(pretrained_model_id, trust_remote_code=True     ).cuda()

    out = generate(model, input_ids, steps=steps , gen_length=128, block_length=32, temperature=0., cfg_scale=0., remasking='low_confidence')
    print("*"*30+ f"FP16 Steps {steps}"+ "*"*30)
    print(input_ids.shape)
    print( tokenizer.batch_decode(out[:, input_ids.shape[1]:], skip_special_tokens=True)[0])


if __name__ == "__main__":
    import logging

    logging.basicConfig(
        format="%(asctime)s %(levelname)s [%(name)s] %(message)s",
        level=logging.INFO,
        datefmt="%Y-%m-%d %H:%M:%S",
    )

    main()