├── utils ├── __init__.py ├── mask.py ├── structure_arb.py └── autosearch_arb.py ├── figs ├── qa.png ├── teaser.png ├── overview.png ├── wikitext2_opt.png ├── wikitext2_llama.png └── wikitext2_vicuna.png ├── requirements.txt ├── modelutils.py ├── datautils.py ├── README.md ├── eval_ppl_utils.py ├── bigptq_arb.py ├── LICENSE ├── run_arb.py └── binary_arb.py /utils/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /figs/qa.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZHITENGLI/ARB-LLM/HEAD/figs/qa.png -------------------------------------------------------------------------------- /figs/teaser.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZHITENGLI/ARB-LLM/HEAD/figs/teaser.png -------------------------------------------------------------------------------- /figs/overview.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZHITENGLI/ARB-LLM/HEAD/figs/overview.png -------------------------------------------------------------------------------- /figs/wikitext2_opt.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZHITENGLI/ARB-LLM/HEAD/figs/wikitext2_opt.png -------------------------------------------------------------------------------- /figs/wikitext2_llama.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZHITENGLI/ARB-LLM/HEAD/figs/wikitext2_llama.png -------------------------------------------------------------------------------- /figs/wikitext2_vicuna.png: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/ZHITENGLI/ARB-LLM/HEAD/figs/wikitext2_vicuna.png -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | transformers==4.35.0 2 | datasets==2.14.6 3 | numpy==1.24.3 4 | huggingface-hub==0.16.4 5 | exceptiongroup 6 | protobuf 7 | sentencepiece 8 | pyparsing 9 | charset-normalizer==2.0.4 10 | pyarrow==12.0.0 11 | -------------------------------------------------------------------------------- /modelutils.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | 5 | DEV = torch.device('cuda:0') 6 | 7 | 8 | def find_layers(module, layers=[nn.Conv2d, nn.Linear], name=''): 9 | if type(module) in layers: 10 | return {name: module} 11 | res = {} 12 | for name1, child in module.named_children(): 13 | res.update(find_layers( 14 | child, layers=layers, name=name + '.' + name1 if name != '' else name1 15 | )) 16 | return res 17 | -------------------------------------------------------------------------------- /utils/mask.py: -------------------------------------------------------------------------------- 1 | 2 | import torch 3 | 4 | ''' 5 | Generate the structural mask on the basis of the split border 6 | ''' 7 | def generate_structural_mask(origin_matrix, mask3, braq1_border): 8 | mask1_2 = ~mask3 9 | 10 | binary_group = torch.abs(origin_matrix*mask1_2) 11 | 12 | mask2 = binary_group >= braq1_border 13 | mask1 = binary_group < braq1_border 14 | 15 | mask1 = mask1 * mask1_2 16 | mask2 = mask2 * mask1_2 17 | 18 | return mask1, mask2 19 | 20 | def generate_multi_structural_mask(origin_matrix, mask3, braq1_border, braq2_border, braq3_border): 21 | mask1_2 = ~mask3 22 | 23 | binary_group = torch.abs(origin_matrix*mask1_2) 24 | 25 | mask4 = binary_group >= braq3_border 26 | mask1 = binary_group < braq1_border 27 | mask2 = (binary_group >= braq1_border) & (binary_group < braq2_border) 28 | mask3 = (binary_group >= braq2_border) & (binary_group < braq3_border) 29 | 30 | mask1 = mask1 * mask1_2 31 | mask2 = mask2 * mask1_2 32 | mask3 = mask3 * mask1_2 33 | mask4 = mask3 * mask1_2 34 | 35 | return mask1, mask2, mask3, mask4 36 | 37 | 38 | def generate_mask(origin_matrix, braq2_border, braq1_border): 39 | mask3 = torch.abs(origin_matrix) >= braq2_border 40 | mask1 = torch.abs(origin_matrix) <= braq1_border 41 | mask2 = (torch.abs(origin_matrix) > braq1_border) & (torch.abs(origin_matrix) < braq2_border) 42 | return mask1, mask2, mask3 -------------------------------------------------------------------------------- /utils/structure_arb.py: -------------------------------------------------------------------------------- 1 | import torch 2 | from utils.autosearch_arb import structural_searching_multip, structural_searching_multip_alternating_group, structural_searching_multip_alternating_group_x, structural_searching_multip_alternating_group_rc 3 | 4 | import logging 5 | logger = logging.getLogger() 6 | 7 | ''' 8 | Used to generate masks for minor structural 2-bit salient data and split major 1-bit normal data according to different metric. 9 | ''' 10 | def structural_guassian_distribution_multip_alternating_group_x(tmp, H=None, metric="magnitude", up_lim=30, num_p=1, inp=None, method='arb', order2_group=False): 11 | if metric == "hessian": 12 | target_weights = tmp ** 2 / (torch.diag(H).reshape((1, -1))) ** 2 13 | elif metric == "magnitude": 14 | target_weights = tmp 15 | else: 16 | raise NotImplementedError 17 | 18 | # print(f'debug', inp) 19 | if method == 'arb': 20 | optimal_split_list, mask_list = structural_searching_multip_alternating_group(target_weights, up_lim, num_p, inp, order2_group=order2_group) 21 | elif method == 'arb-x': 22 | optimal_split_list, mask_list = structural_searching_multip_alternating_group_x(target_weights, up_lim, num_p, inp, order2_group=order2_group) 23 | elif method == 'arb-rc': 24 | optimal_split_list, mask_list = structural_searching_multip_alternating_group_rc(target_weights, up_lim, num_p, inp, order2_group=order2_group) 25 | elif method == 'braq': 26 | optimal_split_list, mask_list = structural_searching_multip(target_weights, up_lim, num_p, order2_group=order2_group) 27 | 28 | # print(mask1.sum() / mask1.numel(), mask2.sum() / mask2.numel(), mask3.sum() / mask3.numel()) 29 | mask_ratio = [] 30 | for i in range(len(mask_list)): 31 | mask_ratio.append(mask_list[i].sum() / mask_list[i].numel()) 32 | 33 | ratios_info = ", ".join([f"mask{idx+1} ratio: {ratio:.2f}" for idx, ratio in enumerate(mask_ratio)]) 34 | logger.info(ratios_info) 35 | 36 | return mask_list 37 | -------------------------------------------------------------------------------- /datautils.py: -------------------------------------------------------------------------------- 1 | import random 2 | 3 | import numpy as np 4 | import torch 5 | from datasets import load_dataset, load_from_disk 6 | from transformers import AutoTokenizer, LlamaTokenizer 7 | import os 8 | 9 | 10 | def set_seed(seed): 11 | np.random.seed(seed) 12 | torch.random.manual_seed(seed) 13 | 14 | ''' 15 | Generate tokenizer and return it to preload datasets by converting them to embedded vectors instead of natural words 16 | ''' 17 | def get_tokenizer(model): 18 | if "llama" in model.lower(): 19 | if '3' in model: 20 | tokenizer = AutoTokenizer.from_pretrained(model, use_fast=False) 21 | else: 22 | tokenizer = LlamaTokenizer.from_pretrained(model, use_fast=False) 23 | # fix for transformer 4.28.0.dev0 compatibility 24 | if tokenizer.bos_token_id != 1 or tokenizer.eos_token_id != 2: 25 | try: 26 | tokenizer.bos_token_id = 1 27 | tokenizer.eos_token_id = 2 28 | except AttributeError: 29 | pass 30 | else: 31 | tokenizer = AutoTokenizer.from_pretrained(model, use_fast=False) 32 | return tokenizer 33 | 34 | def get_wikitext2(nsamples, seed, seqlen, model, tokenizer): 35 | 36 | traindata = load_dataset('wikitext', 'wikitext-2-raw-v1', split='train') 37 | testdata = load_dataset('wikitext', 'wikitext-2-raw-v1', split='test') 38 | 39 | # traindata = load_from_disk('/data/dataset/llm/wikitext/traindata') 40 | # testdata = load_from_disk('/data/dataset/llm/wikitext/testdata') 41 | 42 | trainenc = tokenizer(" ".join(traindata['text']), return_tensors='pt') 43 | testenc = tokenizer("\n\n".join(testdata['text']), return_tensors='pt') 44 | 45 | random.seed(seed) 46 | trainloader = [] 47 | for _ in range(nsamples): 48 | i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1) 49 | j = i + seqlen 50 | inp = trainenc.input_ids[:, i:j] 51 | tar = inp.clone() 52 | tar[:, :-1] = -100 53 | trainloader.append((inp, tar)) 54 | return trainloader, testenc 55 | 56 | def get_ptb(nsamples, seed, seqlen, model, tokenizer): 57 | traindata = load_dataset('ptb_text_only', 'penn_treebank', split='train') 58 | testdata = load_dataset('ptb_text_only', 'penn_treebank', split='test') 59 | 60 | # traindata = load_from_disk('/data/dataset/llm/ptb/traindata') 61 | # testdata = load_from_disk('/data/dataset/llm/ptb/testdata') 62 | 63 | trainenc = tokenizer(" ".join(traindata['sentence']), return_tensors='pt') 64 | testenc = tokenizer(" ".join(testdata['sentence']), return_tensors='pt') 65 | 66 | random.seed(seed) 67 | trainloader = [] 68 | for _ in range(nsamples): 69 | i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1) 70 | j = i + seqlen 71 | inp = trainenc.input_ids[:, i:j] 72 | tar = inp.clone() 73 | tar[:, :-1] = -100 74 | trainloader.append((inp, tar)) 75 | return trainloader, testenc 76 | 77 | class TokenizerWrapper: 78 | def __init__(self, input_ids): 79 | self.input_ids = input_ids 80 | 81 | def get_c4(nsamples, seed, seqlen, model, tokenizer): 82 | traindata = load_dataset( 83 | 'allenai/c4', 'allenai--c4', data_files={'train': 'en/c4-train.00000-of-01024.json.gz'}, split='train' 84 | ) 85 | valdata = load_dataset( 86 | 'allenai/c4', 'allenai--c4', data_files={'validation': 'en/c4-validation.00000-of-00008.json.gz'}, split='validation' 87 | ) 88 | 89 | # traindata = load_from_disk('/data/dataset/llm/c4/traindata') 90 | # valdata = load_from_disk('/data/dataset/llm/c4/valdata') 91 | 92 | random.seed(seed) 93 | trainloader = [] 94 | for _ in range(nsamples): 95 | while True: 96 | i = random.randint(0, len(traindata) - 1) 97 | trainenc = tokenizer(traindata[i]['text'], return_tensors='pt') 98 | if trainenc.input_ids.shape[1] > seqlen: 99 | break 100 | i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1) 101 | j = i + seqlen 102 | inp = trainenc.input_ids[:, i:j] 103 | tar = inp.clone() 104 | tar[:, :-1] = -100 105 | trainloader.append((inp, tar)) 106 | 107 | valenc = tokenizer(' '.join(valdata[:1100]['text']), return_tensors='pt') 108 | valenc = valenc.input_ids[:, :(256 * seqlen)] 109 | 110 | 111 | valenc = TokenizerWrapper(valenc) 112 | 113 | return trainloader, valenc 114 | 115 | def get_loaders(name, nsamples=128, seed=0, seqlen=2048, model=''): 116 | cache_file=f'cache/{name}_{nsamples}_{seed}_{seqlen}_{model}.pt' 117 | try: 118 | return torch.load(cache_file) 119 | except: 120 | pass 121 | 122 | tokenizer = get_tokenizer(model) 123 | 124 | if 'wikitext2' in name: 125 | loaders= get_wikitext2(nsamples, seed, seqlen, model, tokenizer) 126 | if 'ptb' in name: 127 | loaders= get_ptb(nsamples, seed, seqlen, model, tokenizer) 128 | if 'c4' in name: 129 | loaders= get_c4(nsamples, seed, seqlen, model, tokenizer) 130 | directory='/'.join(cache_file.split('/')[:-1]) 131 | if not os.path.exists(directory): 132 | os.makedirs(directory) 133 | 134 | torch.save(loaders,cache_file) 135 | return loaders 136 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # [ICLR'25] ARB-LLM: Alternating Refined Binarizations for Large Language Models 2 | 3 | [Zhiteng Li](https://zhitengli.github.io), Xianglong Yan, Tianao Zhang, [Haotong Qin](https://htqin.github.io/), Dong Xie, Jiang Tian, Zhongchao Shi, [Linghe Kong](https://www.cs.sjtu.edu.cn/~linghe.kong/), [Yulun Zhang](http://yulunzhang.com/), and [Xiaokang Yang](https://scholar.google.com/citations?user=yDEavdMAAAAJ), "ARB-LLM: Alternating Refined Binarizations for Large Language Models", ICLR, 2025 4 | 5 | [[arXiv](https://arxiv.org/pdf/2410.03129 6 | )] [[supplementary material](https://github.com/ZHITENGLI/ARB-LLM/releases/tag/v1)] 7 | 8 | 9 | #### 🔥🔥🔥 News 10 | 11 | - **2025-02-16:** Code is released. ⭐️⭐️⭐️ 12 | - **2025-01-23:** ARB-LLM is accepted at ICLR 2025. 🎉🎉🎉 13 | - **2024-10-03:** This repo is released. 14 | 15 | --- 16 | 17 | > **Abstract:** Large Language Models (LLMs) have greatly pushed forward advancements in natural language processing, yet their high memory and computational demands hinder practical deployment. Binarization, as an effective compression technique, can shrink model weights to just 1 bit, significantly reducing the high demands on computation and memory. However, current binarization methods struggle to narrow the distribution gap between binarized and full-precision weights, while also overlooking the column deviation in LLM weight distribution. To tackle these issues, we propose ARB-LLM, a novel 1-bit post-training quantization (PTQ) technique tailored for LLMs. To narrow the distribution shift between binarized and full-precision weights, we first design an alternating refined binarization (ARB) algorithm to progressively update the binarization parameters, which significantly reduces the quantization error. Moreover, considering the pivot role of calibration data and the column deviation in LLM weights, we further extend ARB to ARB-X and ARB-RC. In addition, we refine the weight partition strategy with column-group bitmap (CGB), which further enhance performance. Equipping ARB-X and ARB-RC with CGB, we obtain ARB-LLMX 18 | and ARB-LLMRC 19 | respectively, which significantly outperform state-of-the-art (SOTA) binarization methods for LLMs. 20 | As a binary PTQ method, our ARB-LLMRC 21 | is the first to surpass FP16 models of the same size. The code and models will be available at https://github.com/ZHITENGLI/ARB-LLM. 22 | 23 | ![](figs/overview.png) 24 | 25 | --- 26 | 27 | Figure 1 in the main paper demonstrates that our proposed ARB-LLMRC outperforms the previous state-of-the-art binary PTQ method, BiLLM, across all scales of the OPT model family. Furthermore, our binarized model surpasses full-precision models of similar size. For example, the memory footprint of the binarized OPT-13B is comparable to that of the full-precision OPT-2.7B, yet the binarized model achieves better performance. 28 | 29 |

30 | 31 |

32 | 33 | ## Dependencies 34 | 35 | ```bash 36 | # Clone the github repo and go to the default directory 'ARB-LLM'. 37 | git clone https://github.com/ZHITENGLI/ARB-LLM.git 38 | conda create -n arbllm python=3.11 39 | conda activate arbllm 40 | pip install torch torchvision torchaudio 41 | pip install -r requirements.txt 42 | ``` 43 | 44 | ## 🔗 Contents 45 | 46 | 1. [Post-training quantization and evaluation](#post-training-quantization) 47 | 2. [Results](#-results) 48 | 3. [Citation](#citation) 49 | 4. [Acknowledgements](#-acknowledgements) 50 | 51 | ## Post-training quantization with PPL evaluation 52 | 53 | ### Binarization for OPT families 54 | 55 | - ARB-X 56 | ```shell 57 | python3 run_arb.py facebook/opt-6.7b c4 arb-x --blocksize 128 --salient_metric hessian --device "cuda:0" --save --num_p 1 --order2_group 58 | ``` 59 | 60 | - ARB-RC 61 | ```shell 62 | python3 run_arb.py facebook/opt-6.7b c4 arb-rc --blocksize 128 --salient_metric hessian --device "cuda:0" --save --num_p 1 --order2_group 63 | ``` 64 | 65 | ### Binarization for LLaMA families 66 | 67 | - ARB-X 68 | ```shell 69 | python3 run_arb.py meta-llama/llama-2-7b-hf c4 arb-x --blocksize 128 --salient_metric hessian --device "cuda:0" --save --num_p 1 --order2_group 70 | ``` 71 | 72 | - ARB-RC 73 | ```shell 74 | python3 run_arb.py meta-llama/llama-2-7b-hf c4 arb-rc --blocksize 128 --salient_metric hessian --device "cuda:0" --save --num_p 1 --order2_group 75 | ``` 76 | 77 | ### Binarization for Vicuna families (Instruction Fine-tuning Models) 78 | 79 | - ARB-X 80 | ```shell 81 | python3 run_arb.py lmsys/vicuna-7b-v1.5 c4 arb-x --blocksize 128 --salient_metric hessian --device "cuda:0" --save --num_p 1 --order2_group 82 | ``` 83 | 84 | - ARB-RC 85 | ```shell 86 | python3 run_arb.py lmsys/vicuna-7b-v1.5 c4 arb-rc --blocksize 128 --salient_metric hessian --device "cuda:0" --save --num_p 1 --order2_group 87 | ``` 88 | 89 | ## Evaluation on zero-shot QA datasets 90 | 91 | We use [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) kit to evaluate performance on QA datasets. Please refer to their framework for evaluating quantized models. 92 | 93 | ## 🔎 Results 94 | 95 |
96 | ARB-LLM achieves superior perplexity performance on WikiText2 datasets. (click to expand) 97 | 98 | - OPT family 99 |

100 | 101 |

102 | 103 | - LLaMA, LLaMA-2 and LLaMA-3 families 104 |

105 | 106 |

107 | 108 | - Vicuna 7B and 13B 109 |

110 | 111 |

112 | 113 |
114 | 115 |
116 | ARB-LLM achieves superior average accuracy on 7 zero-shot QA datasets. (click to expand) 117 | 118 |

119 | 120 |

121 | 122 |
123 | 124 | ## Citation 125 | 126 | If you find the code helpful in your research or work, please cite the following paper. 127 | 128 | ``` 129 | @article{li2024arbllmalternatingrefinedbinarizations, 130 | title={ARB-LLM: Alternating Refined Binarizations for Large Language Models}, 131 | author={Zhiteng Li and Xianglong Yan and Tianao Zhang and Haotong Qin and Dong Xie and Jiang Tian and zhongchao shi and Linghe Kong and Yulun Zhang and Xiaokang Yang}, 132 | year={2024}, 133 | eprint={2410.03129}, 134 | archivePrefix={arXiv}, 135 | primaryClass={cs.CV}, 136 | url={https://arxiv.org/abs/2410.03129}, 137 | } 138 | ``` 139 | 140 | ## 💡 Acknowledgements 141 | 142 | This work is released under the Apache 2.0 license. 143 | The codes are based on [BiLLM](https://github.com/Aaronhuang-778/BiLLM). Please also follow their licenses. Thanks for their awesome works. 144 | -------------------------------------------------------------------------------- /eval_ppl_utils.py: -------------------------------------------------------------------------------- 1 | import time 2 | 3 | import torch 4 | import torch.nn as nn 5 | 6 | import logging 7 | logger = logging.getLogger() 8 | 9 | @torch.no_grad() 10 | def llama_eval(model, testenc, dev, dataset: str, log_wandb: bool = False): 11 | print("Evaluating ...") 12 | 13 | testenc = testenc.input_ids 14 | nsamples = testenc.numel() // model.seqlen 15 | 16 | use_cache = model.config.use_cache 17 | model.config.use_cache = False 18 | layers = model.model.layers 19 | 20 | model.model.embed_tokens = model.model.embed_tokens.to(dev) 21 | layers[0] = layers[0].to(dev) 22 | 23 | dtype = next(iter(model.parameters())).dtype 24 | inps = torch.zeros( 25 | (nsamples, model.seqlen, model.config.hidden_size), dtype=dtype, device=dev 26 | ) 27 | cache = {"i": 0, "attention_mask": None} 28 | 29 | class Catcher(nn.Module): 30 | def __init__(self, module): 31 | super().__init__() 32 | self.module = module 33 | 34 | def forward(self, inp, **kwargs): 35 | inps[cache["i"]] = inp 36 | cache["i"] += 1 37 | cache["attention_mask"] = kwargs["attention_mask"] 38 | raise ValueError 39 | 40 | layers[0] = Catcher(layers[0]) 41 | for i in range(nsamples): 42 | batch = testenc[:, (i * model.seqlen) : ((i + 1) * model.seqlen)].to(dev) 43 | try: 44 | model(batch) 45 | except ValueError: 46 | pass 47 | layers[0] = layers[0].module 48 | 49 | layers[0] = layers[0].cpu() 50 | model.model.embed_tokens = model.model.embed_tokens.cpu() 51 | torch.cuda.empty_cache() 52 | 53 | outs = torch.zeros_like(inps) 54 | attention_mask = cache["attention_mask"] 55 | 56 | for i in range(len(layers)): 57 | # print(i) 58 | layer = layers[i].to(dev) 59 | 60 | for j in range(nsamples): 61 | outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask)[0] 62 | layers[i] = layer.cpu() 63 | del layer 64 | torch.cuda.empty_cache() 65 | inps, outs = outs, inps 66 | 67 | if model.model.norm is not None: 68 | model.model.norm = model.model.norm.to(dev) 69 | model.lm_head = model.lm_head.to(dev) 70 | 71 | testenc = testenc.to(dev) 72 | nlls = [] 73 | for i in range(nsamples): 74 | hidden_states = inps[i].unsqueeze(0) 75 | if model.model.norm is not None: 76 | hidden_states = model.model.norm(hidden_states) 77 | lm_logits = model.lm_head(hidden_states) 78 | shift_logits = lm_logits[:, :-1, :].contiguous() 79 | shift_labels = testenc[:, (i * model.seqlen) : ((i + 1) * model.seqlen)][:, 1:] 80 | loss_fct = nn.CrossEntropyLoss() 81 | loss = loss_fct( 82 | shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1) 83 | ) 84 | neg_log_likelihood = loss.float() * model.seqlen 85 | nlls.append(neg_log_likelihood) 86 | ppl = torch.exp(torch.stack(nlls).sum() / (nsamples * model.seqlen)) 87 | print(f"Perplexity: {ppl.item():3f}") 88 | logger.info(f"{dataset}/Perplexity: {ppl.item():3f}") 89 | 90 | model.config.use_cache = use_cache 91 | 92 | @torch.no_grad() 93 | def opt_eval(model, testenc, dev, dataset: str, log_wandb: bool = False): 94 | print('Evaluating ...') 95 | 96 | testenc = testenc.input_ids 97 | nsamples = testenc.numel() // model.seqlen 98 | 99 | use_cache = model.config.use_cache 100 | model.config.use_cache = False 101 | layers = model.model.decoder.layers 102 | 103 | model.model.decoder.embed_tokens = model.model.decoder.embed_tokens.to(dev) 104 | model.model.decoder.embed_positions = model.model.decoder.embed_positions.to(dev) 105 | if hasattr(model.model.decoder, 'project_out') and model.model.decoder.project_out: 106 | model.model.decoder.project_out = model.model.decoder.project_out.to(dev) 107 | if hasattr(model.model.decoder, 'project_in') and model.model.decoder.project_in: 108 | model.model.decoder.project_in = model.model.decoder.project_in.to(dev) 109 | layers[0] = layers[0].to(dev) 110 | 111 | dtype = next(iter(model.parameters())).dtype 112 | inps = torch.zeros( 113 | (nsamples, model.seqlen, model.config.hidden_size), dtype=dtype, device=dev 114 | ) 115 | cache = {'i': 0, 'attention_mask': None} 116 | 117 | class Catcher(nn.Module): 118 | def __init__(self, module): 119 | super().__init__() 120 | self.module = module 121 | def forward(self, inp, **kwargs): 122 | inps[cache['i']] = inp 123 | cache['i'] += 1 124 | cache['attention_mask'] = kwargs['attention_mask'] 125 | raise ValueError 126 | layers[0] = Catcher(layers[0]) 127 | for i in range(nsamples): 128 | batch = testenc[:, (i * model.seqlen):((i + 1) * model.seqlen)].to(dev) 129 | try: 130 | model(batch) 131 | except ValueError: 132 | pass 133 | layers[0] = layers[0].module 134 | 135 | layers[0] = layers[0].cpu() 136 | model.model.decoder.embed_tokens = model.model.decoder.embed_tokens.cpu() 137 | model.model.decoder.embed_positions = model.model.decoder.embed_positions.cpu() 138 | if hasattr(model.model.decoder, 'project_out') and model.model.decoder.project_out: 139 | model.model.decoder.project_out = model.model.decoder.project_out.cpu() 140 | if hasattr(model.model.decoder, 'project_in') and model.model.decoder.project_in: 141 | model.model.decoder.project_in = model.model.decoder.project_in.cpu() 142 | torch.cuda.empty_cache() 143 | 144 | outs = torch.zeros_like(inps) 145 | attention_mask = cache['attention_mask'] 146 | 147 | for i in range(len(layers)): 148 | # print(i) 149 | layer = layers[i].to(dev) 150 | 151 | for j in range(nsamples): 152 | outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask)[0] 153 | layers[i] = layer.cpu() 154 | del layer 155 | torch.cuda.empty_cache() 156 | inps, outs = outs, inps 157 | 158 | if model.model.decoder.final_layer_norm is not None: 159 | model.model.decoder.final_layer_norm = model.model.decoder.final_layer_norm.to(dev) 160 | if model.model.decoder.project_out is not None: 161 | model.model.decoder.project_out = model.model.decoder.project_out.to(dev) 162 | model.lm_head = model.lm_head.to(dev) 163 | 164 | testenc = testenc.to(dev) 165 | nlls = [] 166 | for i in range(nsamples): 167 | hidden_states = inps[i].unsqueeze(0) 168 | if model.model.decoder.final_layer_norm is not None: 169 | hidden_states = model.model.decoder.final_layer_norm(hidden_states) 170 | if model.model.decoder.project_out is not None: 171 | hidden_states = model.model.decoder.project_out(hidden_states) 172 | lm_logits = model.lm_head(hidden_states) 173 | shift_logits = lm_logits[:, :-1, :].contiguous() 174 | shift_labels = testenc[ 175 | :, (i * model.seqlen):((i + 1) * model.seqlen) 176 | ][:, 1:] 177 | loss_fct = nn.CrossEntropyLoss() 178 | loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)) 179 | neg_log_likelihood = loss.float() * model.seqlen 180 | nlls.append(neg_log_likelihood) 181 | ppl = torch.exp(torch.stack(nlls).sum() / (nsamples * model.seqlen)) 182 | print(f"Perplexity: {ppl.item():3f}") 183 | print({f'{dataset}/perplexity': ppl.item()}) 184 | logger.info(f"Perplexity: {ppl.item():3f}") 185 | logger.info({f'{dataset}/perplexity': ppl.item()}) 186 | 187 | model.config.use_cache = use_cache -------------------------------------------------------------------------------- /bigptq_arb.py: -------------------------------------------------------------------------------- 1 | import math 2 | import time 3 | from exceptiongroup import catch 4 | import torch 5 | import torch.nn as nn 6 | import transformers 7 | from utils.structure_arb import structural_guassian_distribution_multip_alternating_group_x 8 | 9 | import logging 10 | logger = logging.getLogger() 11 | 12 | DEBUG = True 13 | 14 | torch.backends.cuda.matmul.allow_tf32 = False 15 | torch.backends.cudnn.allow_tf32 = False 16 | 17 | ''' 18 | BRAGPTQ is the meaning of GPTQ used Binary Residual Approximation in paper to realize 1-bit quantization 19 | BRAGPTQ uses structural mask to distinguish outliers and other data, and takes advantage of part of GPTQ to lower error 20 | ''' 21 | class BRAGPTQ: 22 | def __init__( 23 | self, layer, braq_quantizer,salient_metric, disable_gptq=False, method='arb', order2_group=False 24 | ): 25 | self.method = method 26 | self.order2_group = order2_group 27 | self.layer = layer 28 | self.dev = self.layer.weight.device 29 | W = layer.weight.data.clone() 30 | if isinstance(self.layer, nn.Conv2d): 31 | W = W.flatten(1) 32 | if isinstance(self.layer, transformers.Conv1D): 33 | W = W.t() 34 | self.rows = W.shape[0] 35 | self.columns = W.shape[1] 36 | self.H = torch.zeros((self.columns, self.columns), device=self.dev) 37 | self.nsamples = 0 38 | self.braq_quantizer = braq_quantizer 39 | self.salient_metric = salient_metric # "magnitude" or "hessian" 40 | self.disable_gptq = disable_gptq 41 | 42 | self.inp = [] 43 | # self.inp2 = torch.zeros(self.columns, self.columns, device=self.dev, dtype=torch.float16) 44 | 45 | def add_batch(self, inp, out, blocksize=1024): 46 | if DEBUG: 47 | self.inp1 = inp 48 | self.out1 = out 49 | 50 | # save memory 51 | # print(inp.shape) # [1,2048,4096] 52 | # print(inp[0].T.shape) 53 | # self.inp2 = self.inp2 + (inp[0].T @ inp[0]) 54 | 55 | if len(inp.shape) == 2: 56 | inp = inp.unsqueeze(0) 57 | 58 | if self.method == 'arb-x': 59 | self.inp.append(inp) 60 | 61 | tmp = inp.shape[0] 62 | if isinstance(self.layer, nn.Linear) or isinstance( 63 | self.layer, transformers.Conv1D 64 | ): 65 | if len(inp.shape) == 3: 66 | inp = inp.reshape((-1, inp.shape[-1])) 67 | inp = inp.t() 68 | self.H *= self.nsamples / (self.nsamples + tmp) 69 | self.nsamples += tmp 70 | inp = math.sqrt(2 / self.nsamples) * inp.float() 71 | self.H += inp.matmul(inp.t()) 72 | # breakpoint() 73 | 74 | def fasterquant(self, 75 | blocksize=128, 76 | percdamp=0.01, 77 | orders=(1,1,2), 78 | num_p=1, 79 | ): 80 | W = self.layer.weight.data.clone() 81 | if isinstance(self.layer, nn.Conv2d): 82 | W = W.flatten(1) 83 | if isinstance(self.layer, transformers.Conv1D): 84 | W = W.t() 85 | W = W.float() 86 | tick = time.time() 87 | 88 | H = self.H 89 | del self.H 90 | dead = torch.diag(H) == 0 91 | H[dead, dead] = 1 92 | W[:, dead] = 0 93 | 94 | Losses = torch.zeros(self.rows, device=self.dev) 95 | 96 | damp = percdamp * torch.mean(torch.diag(H)) 97 | diag = torch.arange(self.columns, device=self.dev) 98 | H[diag, diag] += damp 99 | H = torch.linalg.cholesky(H) 100 | H = torch.cholesky_inverse(H) 101 | H = torch.linalg.cholesky(H, upper=True) 102 | Hinv = H 103 | 104 | if self.method == 'arb-x': 105 | self.inp = torch.concat(self.inp) 106 | # print(self.inp.shape) 107 | 108 | for blocki, col_st in enumerate(range(0, self.columns, blocksize)): 109 | col_ed = min(col_st + blocksize, self.columns) 110 | n_cols = col_ed - col_st 111 | 112 | st = col_st 113 | ed = col_ed 114 | 115 | if self.method == 'arb-x': 116 | # S = torch.einsum('bki,bkj->ij', self.inp[:, :, st:ed], self.inp[:, :, st:ed]) 117 | S = torch.matmul(self.inp[:, :, st:ed].to(torch.float32).transpose(1, 2), self.inp[:, :, st:ed].to(torch.float32)).mean(dim=0) # avoid overflow 118 | else: 119 | S = None 120 | # S = self.inp2[st:ed, st:ed] 121 | # print(S==S2) 122 | 123 | if self.order2_group: 124 | num_mask = 2 * (num_p+1) 125 | orders = [2 for _ in range(num_p+1)] + [1 for _ in range(num_p+1)] 126 | else: 127 | num_mask = 1 + num_p + 1 128 | orders = [2] + [1 for _ in range(num_p+1)] 129 | mask = torch.zeros_like(W[:, st:ed], dtype=torch.bool).unsqueeze(0).repeat_interleave(num_mask, dim=0) 130 | mask_list = structural_guassian_distribution_multip_alternating_group_x(W[:, st:ed], H[st:ed, st:ed], self.salient_metric, 50, num_p, S, self.method, self.order2_group) 131 | for i in range(num_mask): 132 | mask[i] = mask_list[i] 133 | 134 | assert self.braq_quantizer.groupsize % blocksize == 0 135 | 136 | if self.disable_gptq: 137 | # RTN 138 | # print("RTN") 139 | w = W[:, col_st:col_ed] 140 | 141 | # from low to high group 142 | q_part_groups = [] 143 | for i in range(mask.shape[0]): 144 | q_part_groups.append(self.braq_quantizer.quantize(w, mask[i], order=orders[i])) 145 | 146 | q = torch.zeros_like(w) 147 | for j in range(mask.shape[0]): 148 | q += q_part_groups[j][:] * mask[j, :] 149 | W[:, col_st:col_ed] = q 150 | else: 151 | # shape of W1: [oc, n_cols] 152 | W1 = W[:, col_st:col_ed].clone() 153 | Q1 = torch.zeros_like(W1) 154 | Err1 = torch.zeros_like(W1) 155 | Losses1 = torch.zeros_like(W1) 156 | Hinv1 = Hinv[col_st:col_ed, col_st:col_ed] 157 | 158 | # old_q_part_groups = [] 159 | q_part_groups = [] 160 | 161 | for i in range(mask.shape[0]): 162 | q_part_groups.append(self.braq_quantizer.quantize(W1, mask[i], order=orders[i], S=S)) 163 | 164 | for i in range(n_cols): 165 | # shape of w: [oc, 1] 166 | w = W1[:, i] 167 | d = Hinv1[i, i] 168 | 169 | q = torch.zeros_like(w) 170 | for j in range(mask.shape[0]): 171 | q += q_part_groups[j][:, i] * mask[j, :, i] 172 | 173 | Q1[:, i] = q 174 | Losses1[:, i] = (w - q) ** 2 / d**2 175 | # breakpoint() 176 | 177 | err1 = (w - q) / d 178 | Err1[:, i] = err1 179 | 180 | W[:, col_st:col_ed] = Q1 181 | Losses += torch.sum(Losses1, 1) / 2 182 | 183 | W[:, col_ed:] -= Err1.matmul(Hinv[col_st:col_ed, col_ed:]) 184 | 185 | if DEBUG: 186 | self.layer.weight.data[:, :col_ed] = W[:, :col_ed] 187 | self.layer.weight.data[:, col_ed:] = W[:, col_ed:] 188 | x_error = torch.sum((self.layer(self.inp1) - self.out1) ** 2) 189 | # print(torch.sum(Losses)) 190 | 191 | torch.cuda.synchronize() 192 | # print("time %.2f" % (time.time() - tick)) 193 | # print("error", torch.sum(Losses).item()) 194 | times = time.time() - tick 195 | logger.info(f'time {times:.2f}') 196 | logger.info(f'error {torch.sum(Losses).item()}') 197 | logger.info(f'x error {x_error.item()}') 198 | 199 | if isinstance(self.layer, transformers.Conv1D): 200 | W = W.t() 201 | self.layer.weight.data = W.reshape(self.layer.weight.shape).to( 202 | self.layer.weight.data.dtype 203 | ) 204 | if DEBUG: 205 | print(torch.sum((self.layer(self.inp1) - self.out1) ** 2)) 206 | 207 | del mask 208 | # del mask1, mask2, mask3 209 | del mask_list 210 | if not self.disable_gptq: 211 | del W1, Q1, W, Err1, Losses1, Hinv1 212 | del H, Hinv, self.inp, S, q_part_groups 213 | # del H, Hinv, self.inp2, S, q_part_groups 214 | torch.cuda.empty_cache() 215 | return {"error": torch.sum(Losses).item()} 216 | 217 | def free(self): 218 | if DEBUG: 219 | self.inp1 = None 220 | self.out1 = None 221 | self.H = None 222 | torch.cuda.empty_cache() 223 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright 2025 ARB-LLM Authors 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. -------------------------------------------------------------------------------- /run_arb.py: -------------------------------------------------------------------------------- 1 | import os 2 | os.environ["HF_ENDPOINT"] = "https://hf-mirror.com" 3 | 4 | import time 5 | 6 | import torch 7 | import torch.nn as nn 8 | 9 | from bigptq_arb import BRAGPTQ 10 | from binary_arb import Binarization 11 | from modelutils import find_layers 12 | from datautils import get_tokenizer 13 | 14 | import logging 15 | 16 | 17 | def setup_logger(log_file): 18 | logger = logging.getLogger() 19 | logger.setLevel(logging.DEBUG) 20 | 21 | console_handler = logging.StreamHandler() 22 | console_handler.setLevel(logging.DEBUG) 23 | formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s') 24 | console_handler.setFormatter(formatter) 25 | 26 | file_handler = logging.FileHandler(log_file) 27 | file_handler.setLevel(logging.DEBUG) 28 | file_handler.setFormatter(formatter) 29 | 30 | logger.addHandler(console_handler) 31 | logger.addHandler(file_handler) 32 | 33 | return logger 34 | 35 | 36 | def get_model(model): 37 | import torch 38 | 39 | def skip(*args, **kwargs): 40 | pass 41 | 42 | torch.nn.init.kaiming_uniform_ = skip 43 | torch.nn.init.uniform_ = skip 44 | torch.nn.init.normal_ = skip 45 | if "opt" in model: 46 | from transformers import OPTForCausalLM 47 | 48 | model = OPTForCausalLM.from_pretrained(model, torch_dtype="auto") 49 | model.seqlen = model.config.max_position_embeddings 50 | elif "llama" in model: 51 | from transformers import LlamaForCausalLM 52 | 53 | model = LlamaForCausalLM.from_pretrained(model, torch_dtype="auto") 54 | model.seqlen = 2048 55 | return model 56 | 57 | 58 | ''' 59 | The function is employed to calibrate and quantize models layer by layer. 60 | ''' 61 | @torch.no_grad() 62 | def quant_sequential(model, dataloader, dev): 63 | print("Starting ...") 64 | 65 | for name, module in model.named_modules(): 66 | module.global_name = args.model + name 67 | 68 | use_cache = model.config.use_cache 69 | model.config.use_cache = False 70 | 71 | if "opt" in args.model: 72 | layers = model.model.decoder.layers 73 | model.model.decoder.embed_tokens = model.model.decoder.embed_tokens.to(dev) 74 | model.model.decoder.embed_positions = model.model.decoder.embed_positions.to( 75 | dev 76 | ) 77 | if ( 78 | hasattr(model.model.decoder, "project_out") 79 | and model.model.decoder.project_out 80 | ): 81 | model.model.decoder.project_out = model.model.decoder.project_out.to(dev) 82 | if ( 83 | hasattr(model.model.decoder, "project_in") 84 | and model.model.decoder.project_in 85 | ): 86 | model.model.decoder.project_in = model.model.decoder.project_in.to(dev) 87 | elif "llama" in args.model: 88 | layers = model.model.layers 89 | model.model.embed_tokens = model.model.embed_tokens.to(dev) 90 | model.model.norm = model.model.norm.to(dev) 91 | layers[0] = layers[0].to(dev) 92 | 93 | dtype = next(iter(model.parameters())).dtype 94 | inps = torch.zeros( 95 | (args.nsamples, model.seqlen, model.config.hidden_size), dtype=dtype, device=dev 96 | ) 97 | cache = {"i": 0, "attention_mask": None} 98 | 99 | class Catcher(nn.Module): 100 | def __init__(self, module): 101 | super().__init__() 102 | self.module = module 103 | 104 | def forward(self, inp, **kwargs): 105 | inps[cache["i"]] = inp 106 | cache["i"] += 1 107 | cache["attention_mask"] = kwargs["attention_mask"] 108 | raise ValueError 109 | 110 | layers[0] = Catcher(layers[0]) 111 | for batch in dataloader: 112 | try: 113 | model(batch[0].to(dev)) 114 | except ValueError: 115 | pass 116 | layers[0] = layers[0].module 117 | 118 | layers[0] = layers[0].cpu() 119 | if "opt" in args.model: 120 | model.model.decoder.embed_tokens = model.model.decoder.embed_tokens.cpu() 121 | model.model.decoder.embed_positions = model.model.decoder.embed_positions.cpu() 122 | if ( 123 | hasattr(model.model.decoder, "project_out") 124 | and model.model.decoder.project_out 125 | ): 126 | model.model.decoder.project_out = model.model.decoder.project_out.cpu() 127 | if ( 128 | hasattr(model.model.decoder, "project_in") 129 | and model.model.decoder.project_in 130 | ): 131 | model.model.decoder.project_in = model.model.decoder.project_in.cpu() 132 | elif "llama" in args.model: 133 | model.model.embed_tokens = model.model.embed_tokens.cpu() 134 | model.model.norm = model.model.norm.cpu() 135 | torch.cuda.empty_cache() 136 | 137 | outs = torch.zeros_like(inps) 138 | attention_mask = cache["attention_mask"] 139 | 140 | print("Ready.") 141 | 142 | for i in range(len(layers)): 143 | layer = layers[i].to(dev) 144 | 145 | subset = find_layers(layer) 146 | 147 | gptq = {} 148 | for name in subset: 149 | if ( 150 | not (args.minlayer <= i < args.maxlayer and args.quant_only in name) 151 | ) == (not args.invert): 152 | continue 153 | braq_quantizer = Binarization( 154 | subset[name].weight, 155 | method=args.low_quant_method, 156 | groupsize=groupsize, 157 | ) 158 | gptq[name] = BRAGPTQ( 159 | subset[name], 160 | braq_quantizer, 161 | salient_metric=args.salient_metric, 162 | disable_gptq=args.disable_gptq, 163 | method=args.low_quant_method, 164 | order2_group=args.order2_group, 165 | ) 166 | 167 | def add_batch(name): 168 | def tmp(_, inp, out): 169 | gptq[name].add_batch(inp[0].data, out.data) 170 | 171 | return tmp 172 | 173 | handles = [] 174 | for name in gptq: 175 | handles.append(subset[name].register_forward_hook(add_batch(name))) 176 | for j in range(args.nsamples): 177 | outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask)[0] 178 | for h in handles: 179 | h.remove() 180 | 181 | for name in gptq: 182 | # print(i, name) 183 | # print("Quantizing ...") 184 | logging.info(f'{i} {name}') 185 | logging.info("Quantizing ...") 186 | info = gptq[name].fasterquant( 187 | percdamp=args.percdamp, 188 | blocksize=args.blocksize, 189 | num_p=args.num_p, 190 | ) 191 | gptq[name].free() 192 | 193 | for j in range(args.nsamples): 194 | outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask)[0] 195 | 196 | # # debug 197 | # print('fp16', fp_outs.shape) # [128, 2048, 4096] 198 | # print('fp16', fp_outs[0][:5]) 199 | # print('billm', outs.shape) # [128, 2048, 4096] 200 | # print('billm', outs[0][:5]) 201 | # print('------------------') 202 | 203 | layers[i] = layer.cpu() 204 | del layer 205 | del gptq 206 | torch.cuda.empty_cache() 207 | 208 | inps, outs = outs, inps 209 | 210 | model.config.use_cache = use_cache 211 | 212 | 213 | if __name__ == "__main__": 214 | import argparse 215 | from datautils import * 216 | 217 | def list_of_ints(arg): 218 | return list(map(int, arg.split(','))) 219 | 220 | def list_of_floats(arg): 221 | return list(map(float, arg.split(','))) 222 | 223 | parser = argparse.ArgumentParser() 224 | 225 | parser.add_argument( 226 | "model", type=str, help="model to load; for example `huggyllama/llama-7b`." 227 | ) 228 | parser.add_argument( 229 | "dataset", 230 | type=str, 231 | choices=["wikitext2", "ptb", "c4"], 232 | help="Where to extract calibration data from.", 233 | ) 234 | parser.add_argument( 235 | "low_quant_method", 236 | type=str, 237 | choices=["arb", "arb-x", 'arb-rc', 'braq'], 238 | help="alternating refined binarization method", 239 | ) 240 | parser.add_argument( 241 | "--order2_group", 242 | action='store_true', 243 | help="division for salient weights", 244 | ) 245 | parser.set_defaults(order2_group=False) 246 | parser.add_argument("--load_quantized", action="store_true") 247 | parser.add_argument( 248 | "--seed", type=int, default=0, help="Seed for sampling the calibration data." 249 | ) 250 | parser.add_argument( 251 | "--nsamples", type=int, default=128, help="Number of calibration data samples." 252 | ) 253 | parser.add_argument( 254 | "--percdamp", 255 | type=float, 256 | default=0.01, 257 | help="Percent of the average Hessian diagonal to use for dampening.", 258 | ) 259 | parser.add_argument( 260 | "--blocksize", 261 | type=int, 262 | default=128, 263 | help="Blocksize to use for adaptive mask selection.", 264 | ) 265 | parser.add_argument( 266 | "--num_p", 267 | type=int, 268 | default=1, 269 | help="Number of division for non-salient weights", 270 | ) 271 | parser.add_argument( 272 | "--salient_metric", 273 | type=str, 274 | default="magnitude", 275 | choices=["magnitude", "hessian"], 276 | ) 277 | parser.add_argument( 278 | "--device", 279 | type=str, 280 | default="cuda:0", 281 | help="set the device to use for quantization.", 282 | ) 283 | parser.add_argument( 284 | "--disable_gptq", 285 | action="store_true", 286 | help="disable GPTQ for quantization.", 287 | ) 288 | parser.add_argument( 289 | "--minlayer", type=int, default=-1, help="Quant all layers with id >= this." 290 | ) 291 | parser.add_argument( 292 | "--maxlayer", type=int, default=1000, help="Quant all layers with id < this." 293 | ) 294 | parser.add_argument( 295 | "--quant_only", 296 | type=str, 297 | default="", 298 | help="Quant only layers that contain this text.", 299 | ) 300 | parser.add_argument("--invert", action="store_true", help="Invert subset.") 301 | parser.add_argument( 302 | "--save", 303 | action="store_true", 304 | ) 305 | parser.add_argument( 306 | "--log_wandb", action="store_true", help="Whether to log to wandb." 307 | ) 308 | parser.add_argument( 309 | "--tasks", 310 | type=str, 311 | default="", 312 | ) 313 | parser.add_argument( 314 | "--experiment", 315 | type=str, 316 | default="", 317 | ) 318 | parser.add_argument("--num_fewshot", type=int, default=0) 319 | parser.add_argument("--limit", type=int, default=-1) 320 | 321 | args = parser.parse_args() 322 | groupsize = args.blocksize 323 | 324 | device = args.device 325 | save_title = f"{args.model.split('/')[-1]}_{args.dataset}_{args.low_quant_method}_{groupsize}_{args.salient_metric}_nump_{args.num_p}_order2group_{args.order2_group}" 326 | save_file = "./output/" + save_title.replace("/", "_") + ".pt" 327 | if args.load_quantized: 328 | model = get_model(save_file) 329 | model.eval() 330 | 331 | else: # braq 332 | # log 333 | log_file = "./log/" + save_title.replace("/", "_") + f"_{args.experiment}" + ".log" 334 | log_path = os.path.dirname(log_file) 335 | if not os.path.exists(log_path): 336 | os.makedirs(log_path) 337 | logger = setup_logger(log_file) 338 | 339 | model = get_model(args.model) 340 | model.eval() 341 | tick = time.time() 342 | dataloader, testloader = get_loaders( 343 | args.dataset, 344 | nsamples=args.nsamples, 345 | seed=args.seed, 346 | model=args.model, 347 | seqlen=model.seqlen, 348 | ) 349 | # print(model) 350 | 351 | quant_sequential(model, dataloader, device) 352 | print("quantization time:", time.time() - tick, "s") 353 | print(f'Experiment: {args.experiment}') 354 | logger.info(f'Experiment: {args.experiment}') 355 | 356 | if args.save: 357 | save_path = os.path.dirname(save_file) 358 | if not os.path.exists(save_path): 359 | os.makedirs(save_path) 360 | model.save_pretrained(save_file) 361 | 362 | for dataset in ["wikitext2", "ptb", "c4"]: 363 | dataloader, testloader = get_loaders( 364 | dataset, seed=args.seed, seqlen=model.seqlen, model=args.model 365 | ) 366 | print(dataset) 367 | if "opt" in args.model: 368 | from eval_ppl_utils import opt_eval 369 | 370 | opt_eval(model, testloader, device, dataset, args.log_wandb) 371 | elif "llama" in args.model: 372 | from eval_ppl_utils import llama_eval 373 | 374 | llama_eval(model, testloader, device, dataset, args.log_wandb) 375 | 376 | -------------------------------------------------------------------------------- /utils/autosearch_arb.py: -------------------------------------------------------------------------------- 1 | from re import L 2 | import numpy as np 3 | from pyparsing import line 4 | import torch 5 | from binary_arb import high_order_residual, high_order_residual_alternating_order1, high_order_residual_alternating_mean, high_order_residual_alternating_order2_rc_nomean, high_order_residual_alternating_order1_rc_nomean 6 | from utils.mask import generate_structural_mask 7 | 8 | error_N = 2048*4096*128 9 | 10 | def error_computing(origin_matrix, quantized_matrix): 11 | mse = torch.mean((origin_matrix - quantized_matrix) ** 2) 12 | return mse 13 | 14 | def error_computing_x_all_accelerate(origin_matrix, quantized_matrix, S): 15 | # inps shape [128, 2048, 128] 16 | R = (origin_matrix - quantized_matrix).T 17 | P = torch.einsum('ik,jk->ij', R, R) 18 | return torch.sum(P * S) / error_N 19 | 20 | def calculate_percentage_and_variance_original(weights, abs_weights, bin_edges): 21 | percentages = [] 22 | variances = [] 23 | accum_percentages = [0] 24 | total_elements = abs_weights.numel() 25 | for i in range(len(bin_edges) - 1): 26 | bin_mask = (abs_weights >= bin_edges[i]) & (abs_weights < bin_edges[i + 1]) 27 | bin_weights = weights[bin_mask] 28 | percentages.append(bin_weights.numel() / total_elements * 100) 29 | accum_percentages.append(accum_percentages[-1] + percentages[-1]) 30 | variances.append(torch.var(bin_weights)) 31 | return percentages, variances, accum_percentages 32 | 33 | ''' 34 | Include main method to search the rate for 2-bit salient data columns and the optimal split for 1-bit data 35 | ''' 36 | def structural_searching_multip(origin_matrix, up_lim=30, num_p=1, order2_group=False): 37 | minimal_value = float('inf') 38 | minimal_value_0 = float('inf') 39 | 40 | true_counts = origin_matrix.abs().sum(dim=0) 41 | 42 | error = [] 43 | lines = [] 44 | # search for the optimal split for the first group, high order=2,, structured search 45 | _, top_braq_2_columns = torch.topk(true_counts, up_lim) 46 | for i in range(1, up_lim): 47 | mask3 = torch.full((origin_matrix.shape[0], origin_matrix.shape[1]), False).to(origin_matrix.device) 48 | mask3[:, top_braq_2_columns[:i]] = True 49 | group3 = high_order_residual(origin_matrix, mask3, order=2) 50 | group4 = high_order_residual(origin_matrix, ~mask3, order=2) 51 | 52 | 53 | quantize_error_0 = error_computing(origin_matrix, group4+group3) 54 | error.append(quantize_error_0.item()) 55 | lines.append(i) 56 | 57 | if quantize_error_0 < minimal_value_0: 58 | minimal_value_0 = quantize_error_0 59 | optimal_split_0 = i 60 | 61 | _, top_braq_2_columns = torch.topk(true_counts, optimal_split_0) 62 | mask3 = torch.full((origin_matrix.shape[0], origin_matrix.shape[1]), False).to(origin_matrix.device) 63 | mask3[:, top_braq_2_columns] = True 64 | group3 = high_order_residual(origin_matrix, mask3, order=2) 65 | 66 | mask_list = [mask3] 67 | optimal_split_list = [] 68 | for i in range(num_p): 69 | search_matrix = origin_matrix * (~mask3) 70 | 71 | flat_abs_tensor = torch.abs(search_matrix).view(-1) 72 | percentiles = torch.linspace(0.10, 0.90, 81).to(origin_matrix.device) 73 | percentile_values = torch.tensor( 74 | np.quantile(flat_abs_tensor.detach().cpu().numpy(), q=percentiles.cpu().numpy(), axis=None, keepdims=False) 75 | ).to(origin_matrix.device) 76 | 77 | # search for the optimal split for the second group, high order=1,, non-structured search 78 | for split_value in percentile_values: 79 | mask1, mask2 = generate_structural_mask(origin_matrix, mask3, split_value) 80 | group1 = high_order_residual(origin_matrix, mask1, order=1) 81 | group2 = high_order_residual(origin_matrix, mask2, order=1) 82 | 83 | quantize_error = error_computing(origin_matrix, group1+group2+group3) 84 | if quantize_error < minimal_value: 85 | minimal_value = quantize_error 86 | optimal_split = split_value 87 | optimal_group2 = group2 88 | best_mask2 = mask2 89 | best_mask1 = mask1 90 | 91 | mask_list.append(best_mask2) 92 | optimal_split_list.append(optimal_split) 93 | group3 = group3 + optimal_group2 94 | mask3 = mask3 | best_mask2 95 | 96 | mask_list.append(best_mask1) 97 | 98 | return optimal_split_list, mask_list 99 | 100 | def structural_searching_multip_alternating_group(origin_matrix, up_lim=30, num_p=1, inp=None, iter=0, order2_group=False): 101 | minimal_value = float('inf') 102 | minimal_value_0 = float('inf') 103 | 104 | true_counts = origin_matrix.abs().sum(dim=0) 105 | 106 | # error = [] 107 | # lines = [] 108 | # search for the optimal split for the first group, high order=2,, structured search 109 | _, top_braq_2_columns = torch.topk(true_counts, up_lim) 110 | for i in range(1, up_lim): 111 | mask3 = torch.full((origin_matrix.shape[0], origin_matrix.shape[1]), False).to(origin_matrix.device) 112 | mask3[:, top_braq_2_columns[:i]] = True 113 | group3 = high_order_residual(origin_matrix, mask3, order=2) # for fair comparison and accelerate 114 | group4 = high_order_residual(origin_matrix, ~mask3, order=2) # for fair comparison and accelerate 115 | 116 | quantize_error_0 = error_computing(origin_matrix, group4+group3) 117 | # error.append(quantize_error_0.item()) 118 | # lines.append(i) 119 | # print(quantize_error_0) 120 | 121 | if quantize_error_0 < minimal_value_0: 122 | minimal_value_0 = quantize_error_0 123 | optimal_split_0 = i 124 | 125 | 126 | _, top_braq_2_columns = torch.topk(true_counts, optimal_split_0) 127 | mask3 = torch.full((origin_matrix.shape[0], origin_matrix.shape[1]), False).to(origin_matrix.device) 128 | mask3[:, top_braq_2_columns] = True 129 | 130 | group3 = high_order_residual_alternating_mean(origin_matrix, mask3, order=2) 131 | 132 | mask_list = [] 133 | optimal_split_list = [] 134 | 135 | # 2nd order group 136 | if order2_group: 137 | mask0 = mask3.clone() 138 | minimal_value2 = float('inf') 139 | group0 = torch.zeros(origin_matrix.shape, device=origin_matrix.device) 140 | for i in range(num_p): 141 | search_matrix = origin_matrix * mask0 142 | 143 | flat_abs_tensor = torch.abs(search_matrix).view(-1) 144 | flat_abs_tensor_nonzero = flat_abs_tensor[flat_abs_tensor != 0] 145 | percentiles = torch.linspace(0.10, 0.90, 81).to(origin_matrix.device) 146 | percentile_values = torch.tensor( 147 | np.quantile(flat_abs_tensor_nonzero.detach().cpu().numpy(), q=percentiles.cpu().numpy(), axis=None, keepdims=False) 148 | ).to(origin_matrix.device) 149 | 150 | # search for the optimal split for the second group, high order=1,, non-structured search 151 | flag = False 152 | for split_value in percentile_values: 153 | mask4, mask5 = generate_structural_mask(origin_matrix, ~mask0, split_value) 154 | group1 = high_order_residual(origin_matrix, mask4, order=2) 155 | group2 = high_order_residual(origin_matrix, mask5, order=2) 156 | 157 | quantize_error = error_computing(origin_matrix, group1+group2+group0) 158 | if quantize_error < minimal_value2: 159 | minimal_value2 = quantize_error 160 | optimal_split = split_value 161 | optimal_group2 = group2 162 | best_mask4 = mask4 163 | best_mask5 = mask5 164 | flag = True 165 | 166 | if not flag: 167 | print(False, 2) 168 | optimal_split = percentile_values[0] 169 | best_mask4, best_mask5 = generate_structural_mask(origin_matrix, ~mask0, optimal_split) 170 | 171 | mask0 = mask0 & (~best_mask5) 172 | mask_list.append(best_mask5) 173 | group0 = group0 + optimal_group2 174 | 175 | mask_list.append(best_mask4) 176 | 177 | else: 178 | mask_list.append(mask3) 179 | 180 | # 1st order group 181 | for i in range(num_p): 182 | search_matrix = origin_matrix * (~mask3) 183 | 184 | flat_abs_tensor = torch.abs(search_matrix).view(-1) 185 | percentiles = torch.linspace(0.10, 0.90, 81).to(origin_matrix.device) 186 | percentile_values = torch.tensor( 187 | np.quantile(flat_abs_tensor.detach().cpu().numpy(), q=percentiles.cpu().numpy(), axis=None, keepdims=False) 188 | ).to(origin_matrix.device) 189 | 190 | # search for the optimal split for the second group, high order=1,, non-structured search 191 | flag = False 192 | for split_value in percentile_values: 193 | mask1, mask2 = generate_structural_mask(origin_matrix, mask3, split_value) 194 | 195 | group1 = high_order_residual(origin_matrix, mask1, order=1) 196 | group2 = high_order_residual(origin_matrix, mask2, order=1) 197 | 198 | quantize_error = error_computing(origin_matrix, group1+group2+group3) 199 | if quantize_error < minimal_value: 200 | minimal_value = quantize_error 201 | optimal_split = split_value 202 | best_mask2 = mask2 203 | best_mask1 = mask1 204 | flag = True 205 | 206 | if not flag: 207 | print(False) 208 | optimal_split = percentile_values[0] 209 | best_mask1, best_mask2 = generate_structural_mask(origin_matrix, mask3, optimal_split) 210 | 211 | optimal_group2 = high_order_residual_alternating_order1(origin_matrix, best_mask2, order=1) 212 | mask_list.append(best_mask2) 213 | optimal_split_list.append(optimal_split) 214 | group3 = group3 + optimal_group2 215 | mask3 = mask3 | best_mask2 216 | 217 | mask_list.append(best_mask1) 218 | 219 | return optimal_split_list, mask_list 220 | 221 | def structural_searching_multip_alternating_group_x(origin_matrix, up_lim=30, num_p=1, inp=None, iter=0, order2_group=False): 222 | minimal_value = float('inf') 223 | minimal_value_0 = float('inf') 224 | 225 | true_counts = origin_matrix.abs().sum(dim=0) 226 | 227 | # error = [] 228 | # lines = [] 229 | # search for the optimal split for the first group, high order=2,, structured search 230 | _, top_braq_2_columns = torch.topk(true_counts, up_lim) 231 | for i in range(1, up_lim): 232 | mask3 = torch.full((origin_matrix.shape[0], origin_matrix.shape[1]), False).to(origin_matrix.device) 233 | mask3[:, top_braq_2_columns[:i]] = True 234 | group3 = high_order_residual(origin_matrix, mask3, order=2) # for fair comparison and accelerate 235 | group4 = high_order_residual(origin_matrix, ~mask3, order=2) # for fair comparison and accelerate 236 | 237 | quantize_error_0 = error_computing(origin_matrix, group4+group3) 238 | # error.append(quantize_error_0.item()) 239 | # lines.append(i) 240 | # print(quantize_error_0) 241 | 242 | if quantize_error_0 < minimal_value_0: 243 | minimal_value_0 = quantize_error_0 244 | optimal_split_0 = i 245 | 246 | 247 | _, top_braq_2_columns = torch.topk(true_counts, optimal_split_0) 248 | mask3 = torch.full((origin_matrix.shape[0], origin_matrix.shape[1]), False).to(origin_matrix.device) 249 | mask3[:, top_braq_2_columns] = True 250 | 251 | group3 = high_order_residual_alternating_mean(origin_matrix, mask3, order=2) 252 | 253 | mask_list = [] 254 | optimal_split_list = [] 255 | 256 | # 2nd order group 257 | if order2_group: 258 | mask0 = mask3.clone() 259 | minimal_value2 = float('inf') 260 | group0 = torch.zeros(origin_matrix.shape, device=origin_matrix.device) 261 | for i in range(num_p): 262 | search_matrix = origin_matrix * mask0 263 | 264 | flat_abs_tensor = torch.abs(search_matrix).view(-1) 265 | flat_abs_tensor_nonzero = flat_abs_tensor[flat_abs_tensor != 0] 266 | percentiles = torch.linspace(0.10, 0.90, 81).to(origin_matrix.device) 267 | percentile_values = torch.tensor( 268 | np.quantile(flat_abs_tensor_nonzero.detach().cpu().numpy(), q=percentiles.cpu().numpy(), axis=None, keepdims=False) 269 | ).to(origin_matrix.device) 270 | 271 | # search for the optimal split for the second group, high order=1,, non-structured search 272 | flag = False 273 | for split_value in percentile_values: 274 | mask4, mask5 = generate_structural_mask(origin_matrix, ~mask0, split_value) 275 | group1 = high_order_residual(origin_matrix, mask4, order=2) 276 | group2 = high_order_residual(origin_matrix, mask5, order=2) 277 | 278 | quantize_error = error_computing(origin_matrix, group1+group2+group0) 279 | if quantize_error < minimal_value2: 280 | minimal_value2 = quantize_error 281 | optimal_split = split_value 282 | optimal_group2 = group2 283 | best_mask4 = mask4 284 | best_mask5 = mask5 285 | flag = True 286 | 287 | if not flag: 288 | print(False, 2) 289 | optimal_split = percentile_values[0] 290 | best_mask4, best_mask5 = generate_structural_mask(origin_matrix, ~mask0, optimal_split) 291 | 292 | mask0 = mask0 & (~best_mask5) 293 | mask_list.append(best_mask5) 294 | group0 = group0 + optimal_group2 295 | 296 | mask_list.append(best_mask4) 297 | 298 | else: 299 | mask_list.append(mask3) 300 | 301 | # 1st order group 302 | for i in range(num_p): 303 | search_matrix = origin_matrix * (~mask3) 304 | 305 | flat_abs_tensor = torch.abs(search_matrix).view(-1) 306 | percentiles = torch.linspace(0.10, 0.90, 81).to(origin_matrix.device) 307 | percentile_values = torch.tensor( 308 | np.quantile(flat_abs_tensor.detach().cpu().numpy(), q=percentiles.cpu().numpy(), axis=None, keepdims=False) 309 | ).to(origin_matrix.device) 310 | 311 | # search for the optimal split for the second group, high order=1,, non-structured search 312 | flag = False 313 | for split_value in percentile_values: 314 | mask1, mask2 = generate_structural_mask(origin_matrix, mask3, split_value) 315 | 316 | group1 = high_order_residual(origin_matrix, mask1, order=1) 317 | group2 = high_order_residual(origin_matrix, mask2, order=1) 318 | 319 | quantize_error = error_computing_x_all_accelerate(origin_matrix, group1+group2+group3, inp) 320 | if quantize_error < minimal_value: 321 | minimal_value = quantize_error 322 | optimal_split = split_value 323 | best_mask2 = mask2 324 | best_mask1 = mask1 325 | flag = True 326 | 327 | if not flag: 328 | print(False) 329 | optimal_split = percentile_values[0] 330 | best_mask1, best_mask2 = generate_structural_mask(origin_matrix, mask3, optimal_split) 331 | 332 | optimal_group2 = high_order_residual_alternating_order1(origin_matrix, best_mask2, order=1) # accelerate 333 | mask_list.append(best_mask2) 334 | optimal_split_list.append(optimal_split) 335 | group3 = group3 + optimal_group2 336 | mask3 = mask3 | best_mask2 337 | 338 | mask_list.append(best_mask1) 339 | 340 | return optimal_split_list, mask_list 341 | 342 | def structural_searching_multip_alternating_group_rc(origin_matrix, up_lim=30, num_p=1, inp=None, iter=0, order2_group=False): 343 | minimal_value = float('inf') 344 | minimal_value_0 = float('inf') 345 | 346 | true_counts = origin_matrix.abs().sum(dim=0) 347 | 348 | # error = [] 349 | # lines = [] 350 | # search for the optimal split for the first group, high order=2,, structured search 351 | _, top_braq_2_columns = torch.topk(true_counts, up_lim) 352 | for i in range(1, up_lim): 353 | mask3 = torch.full((origin_matrix.shape[0], origin_matrix.shape[1]), False).to(origin_matrix.device) 354 | mask3[:, top_braq_2_columns[:i]] = True 355 | group3 = high_order_residual(origin_matrix, mask3, order=2) # for fair comparison and accelerate 356 | group4 = high_order_residual(origin_matrix, ~mask3, order=2) # for fair comparison and accelerate 357 | 358 | quantize_error_0 = error_computing(origin_matrix, group4+group3) 359 | # error.append(quantize_error_0.item()) 360 | # lines.append(i) 361 | # print(quantize_error_0) 362 | 363 | if quantize_error_0 < minimal_value_0: 364 | minimal_value_0 = quantize_error_0 365 | optimal_split_0 = i 366 | 367 | 368 | _, top_braq_2_columns = torch.topk(true_counts, optimal_split_0) 369 | mask3 = torch.full((origin_matrix.shape[0], origin_matrix.shape[1]), False).to(origin_matrix.device) 370 | mask3[:, top_braq_2_columns] = True 371 | 372 | group3 = high_order_residual_alternating_order2_rc_nomean(origin_matrix, mask3, order=2) 373 | 374 | mask_list = [] 375 | optimal_split_list = [] 376 | 377 | # 2nd order group 378 | if order2_group: 379 | mask0 = mask3.clone() 380 | minimal_value2 = float('inf') 381 | group0 = torch.zeros(origin_matrix.shape, device=origin_matrix.device) 382 | for i in range(num_p): 383 | search_matrix = origin_matrix * mask0 384 | 385 | flat_abs_tensor = torch.abs(search_matrix).view(-1) 386 | flat_abs_tensor_nonzero = flat_abs_tensor[flat_abs_tensor != 0] 387 | percentiles = torch.linspace(0.10, 0.90, 81).to(origin_matrix.device) 388 | percentile_values = torch.tensor( 389 | np.quantile(flat_abs_tensor_nonzero.detach().cpu().numpy(), q=percentiles.cpu().numpy(), axis=None, keepdims=False) 390 | ).to(origin_matrix.device) 391 | 392 | # search for the optimal split for the second group, high order=1,, non-structured search 393 | flag = False 394 | for split_value in percentile_values: 395 | mask4, mask5 = generate_structural_mask(origin_matrix, ~mask0, split_value) 396 | group1 = high_order_residual(origin_matrix, mask4, order=2) 397 | group2 = high_order_residual(origin_matrix, mask5, order=2) 398 | 399 | quantize_error = error_computing(origin_matrix, group1+group2+group0) 400 | if quantize_error < minimal_value2: 401 | minimal_value2 = quantize_error 402 | optimal_split = split_value 403 | optimal_group2 = group2 404 | best_mask4 = mask4 405 | best_mask5 = mask5 406 | flag = True 407 | 408 | if not flag: 409 | print(False, 2) 410 | optimal_split = percentile_values[0] 411 | best_mask4, best_mask5 = generate_structural_mask(origin_matrix, ~mask0, optimal_split) 412 | 413 | mask0 = mask0 & (~best_mask5) 414 | mask_list.append(best_mask5) 415 | group0 = group0 + optimal_group2 416 | 417 | mask_list.append(best_mask4) 418 | 419 | else: 420 | mask_list.append(mask3) 421 | 422 | # 1st order group 423 | for i in range(num_p): 424 | search_matrix = origin_matrix * (~mask3) 425 | 426 | flat_abs_tensor = torch.abs(search_matrix).view(-1) 427 | percentiles = torch.linspace(0.10, 0.90, 81).to(origin_matrix.device) 428 | percentile_values = torch.tensor( 429 | np.quantile(flat_abs_tensor.detach().cpu().numpy(), q=percentiles.cpu().numpy(), axis=None, keepdims=False) 430 | ).to(origin_matrix.device) 431 | 432 | # search for the optimal split for the second group, high order=1,, non-structured search 433 | flag = False 434 | for split_value in percentile_values: 435 | mask1, mask2 = generate_structural_mask(origin_matrix, mask3, split_value) 436 | 437 | group1 = high_order_residual(origin_matrix, mask1, order=1) 438 | group2 = high_order_residual(origin_matrix, mask2, order=1) 439 | 440 | # quantize_error = error_computing_x_all_accelerate(origin_matrix, group1+group2+group3, inp) 441 | quantize_error = error_computing(origin_matrix, group1+group2+group3) 442 | if quantize_error < minimal_value: 443 | minimal_value = quantize_error 444 | optimal_split = split_value 445 | best_mask2 = mask2 446 | best_mask1 = mask1 447 | flag = True 448 | 449 | if not flag: 450 | print(False) 451 | optimal_split = percentile_values[0] 452 | best_mask1, best_mask2 = generate_structural_mask(origin_matrix, mask3, optimal_split) 453 | 454 | # optimal_group2 = high_order_residual_alternating_order1(origin_matrix, best_mask2, order=1) 455 | optimal_group2 = high_order_residual_alternating_order1_rc_nomean(origin_matrix, best_mask2, order=1, iter=0) 456 | mask_list.append(best_mask2) 457 | optimal_split_list.append(optimal_split) 458 | group3 = group3 + optimal_group2 459 | mask3 = mask3 | best_mask2 460 | 461 | mask_list.append(best_mask1) 462 | 463 | return optimal_split_list, mask_list 464 | 465 | def find_optimal_split(group_max, origin_matrix, border): 466 | optimal_split = None 467 | minimal_value = float('inf') 468 | searching_steps = torch.arange(0.1,0.8,0.01) 469 | searching_steps = searching_steps * group_max 470 | 471 | group3 = high_order_residual(origin_matrix, torch.abs(origin_matrix) > border, order=2) 472 | for split_value in searching_steps: 473 | 474 | group1 = high_order_residual(origin_matrix, (torch.abs(origin_matrix) > split_value) & (torch.abs(origin_matrix) <= border), order=1) 475 | group2 = high_order_residual(origin_matrix, torch.abs(origin_matrix) <= split_value, order=1) 476 | 477 | quantize_error = error_computing(origin_matrix, group1+group2+group3) 478 | if quantize_error < minimal_value: 479 | minimal_value = quantize_error 480 | optimal_split = split_value 481 | 482 | return optimal_split, minimal_value 483 | -------------------------------------------------------------------------------- /binary_arb.py: -------------------------------------------------------------------------------- 1 | from numpy import mean 2 | import torch 3 | import torch.nn as nn 4 | import torch.nn.functional as F 5 | import math 6 | index = 0 7 | @torch.no_grad() 8 | def part_mean(tensor, op='-'): 9 | non_zero = tensor*(tensor!=0) 10 | 11 | mean_val = non_zero.mean(-1).view(-1, 1) 12 | 13 | return mean_val 14 | 15 | @torch.no_grad() 16 | def high_order_residual(x, mask, order=2): 17 | sum_order = torch.zeros_like(x) 18 | new_matrix = x.clone() 19 | new_matrix = new_matrix * mask 20 | global index 21 | index += 1 22 | for od in range(order): 23 | residual = new_matrix - sum_order 24 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 25 | 26 | mean_tensor_all = torch.nanmean(masked_x_tensor, dim=1) 27 | mean_tensor_all = torch.where(torch.isnan(mean_tensor_all), torch.zeros_like(mean_tensor_all), mean_tensor_all) 28 | masked_x_tensor -= mean_tensor_all[:, None] 29 | scale_tensor_all = torch.nanmean(torch.abs(masked_x_tensor), dim=1) 30 | scale_tensor_all = torch.where(torch.isnan(scale_tensor_all), torch.zeros_like(scale_tensor_all), scale_tensor_all) 31 | 32 | binary= torch.sign(masked_x_tensor) 33 | binary *= scale_tensor_all[:, None] 34 | binary += mean_tensor_all[:, None] 35 | sum_order = sum_order + binary*mask 36 | 37 | return sum_order 38 | 39 | @torch.no_grad() 40 | def high_order_residual_rc(x, mask, order=2): 41 | sum_order = torch.zeros_like(x) 42 | new_matrix = x.clone() 43 | new_matrix = new_matrix * mask 44 | global index 45 | index += 1 46 | for od in range(order): 47 | residual = new_matrix - sum_order 48 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 49 | 50 | # mean row 51 | mean_tensor_all_r = torch.nanmean(masked_x_tensor, dim=1) 52 | mean_tensor_all_r = torch.where(torch.isnan(mean_tensor_all_r), torch.zeros_like(mean_tensor_all_r), mean_tensor_all_r) 53 | masked_x_tensor -= mean_tensor_all_r[:, None] 54 | # mean column 55 | mean_tensor_all_c = torch.nanmean(masked_x_tensor, dim=0) 56 | mean_tensor_all_c = torch.where(torch.isnan(mean_tensor_all_c), torch.zeros_like(mean_tensor_all_c), mean_tensor_all_c) 57 | masked_x_tensor -= mean_tensor_all_c[None, :] 58 | 59 | # alpha row 60 | scale_tensor_all_r = torch.nanmean(torch.abs(masked_x_tensor), dim=1) 61 | scale_tensor_all_r = torch.where(torch.isnan(scale_tensor_all_r), torch.zeros_like(scale_tensor_all_r), scale_tensor_all_r) 62 | # alpha column 63 | scale_tensor_all_c = torch.nanmean(torch.abs(masked_x_tensor / scale_tensor_all_r[:, None]), dim=0) 64 | scale_tensor_all_c = torch.where(torch.isnan(scale_tensor_all_c), torch.zeros_like(scale_tensor_all_c), scale_tensor_all_c) 65 | 66 | binary= torch.sign(masked_x_tensor) 67 | binary *= scale_tensor_all_r[:, None] 68 | binary *= scale_tensor_all_c[None, :] 69 | binary += mean_tensor_all_r[:, None] + mean_tensor_all_c[None, :] 70 | sum_order = sum_order + binary*mask 71 | 72 | return sum_order 73 | 74 | @torch.no_grad() 75 | def high_order_residual_alternating_order1(x, mask, order=2, iter=15): 76 | sum_order = torch.zeros_like(x) 77 | new_matrix = x.clone() 78 | new_matrix = new_matrix * mask 79 | global index 80 | index += 1 81 | for od in range(order): 82 | residual = new_matrix - sum_order 83 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 84 | 85 | mean_tensor_all = torch.nanmean(masked_x_tensor, dim=1) 86 | mean_tensor_all = torch.where(torch.isnan(mean_tensor_all), torch.zeros_like(mean_tensor_all), mean_tensor_all) 87 | masked_x_tensor -= mean_tensor_all[:, None] 88 | scale_tensor_all = torch.nanmean(torch.abs(masked_x_tensor), dim=1) 89 | scale_tensor_all = torch.where(torch.isnan(scale_tensor_all), torch.zeros_like(scale_tensor_all), scale_tensor_all) 90 | 91 | binary= torch.sign(masked_x_tensor) 92 | new_binary = binary.clone() 93 | binary *= scale_tensor_all[:, None] 94 | binary += mean_tensor_all[:, None] 95 | sum_order = sum_order + binary*mask 96 | 97 | # Alternating update 98 | refine_mean = mean_tensor_all.clone() 99 | sum_order_alternating = sum_order.clone() 100 | 101 | for k in range(iter): 102 | # 1. Fix alpha and B, update mean 103 | residual = new_matrix - sum_order_alternating 104 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 105 | mean_tensor_all = torch.nanmean(masked_x_tensor, dim=1) 106 | mean_tensor_all = torch.where(torch.isnan(mean_tensor_all), torch.zeros_like(mean_tensor_all), mean_tensor_all) 107 | refine_mean += mean_tensor_all.clone() 108 | 109 | # 2. Fix mean and B, update alpha 110 | new_alpha = 1. / (torch.sum(new_binary * mask * new_binary * mask, dim=1) + 1e-8) * torch.sum(new_binary * mask * (new_matrix - refine_mean[:, None] * mask), dim=1) 111 | 112 | # 3. Fix mean and alpha, update B 113 | new_binary = torch.sign(new_matrix - refine_mean[:, None] * mask) 114 | 115 | # Final refine results 116 | sum_order_alternating = torch.zeros_like(x) + (new_alpha[:, None] * new_binary + refine_mean[:, None]) * mask 117 | 118 | 119 | return sum_order_alternating 120 | 121 | @torch.no_grad() 122 | def high_order_residual_alternating_order1_x(x, mask, order=2, S=None, iter=15, iter2=15): 123 | sum_order = torch.zeros_like(x) 124 | new_matrix = x.clone() 125 | new_matrix = new_matrix * mask 126 | global index 127 | index += 1 128 | for od in range(order): 129 | residual = new_matrix - sum_order 130 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 131 | 132 | mean_tensor_all = torch.nanmean(masked_x_tensor, dim=1) 133 | mean_tensor_all = torch.where(torch.isnan(mean_tensor_all), torch.zeros_like(mean_tensor_all), mean_tensor_all) 134 | masked_x_tensor -= mean_tensor_all[:, None] 135 | scale_tensor_all = torch.nanmean(torch.abs(masked_x_tensor), dim=1) 136 | scale_tensor_all = torch.where(torch.isnan(scale_tensor_all), torch.zeros_like(scale_tensor_all), scale_tensor_all) 137 | 138 | binary= torch.sign(masked_x_tensor) 139 | new_binary = binary.clone() 140 | binary *= scale_tensor_all[:, None] 141 | binary += mean_tensor_all[:, None] 142 | sum_order = sum_order + binary*mask 143 | 144 | # Alternating update 145 | refine_mean = mean_tensor_all.clone() 146 | sum_order_alternating = sum_order.clone() 147 | new_alpha = scale_tensor_all.clone() 148 | 149 | for k in range(iter): 150 | # 1. Fix alpha and B, update mean 151 | residual = new_matrix - sum_order_alternating 152 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 153 | mean_tensor_all = torch.nanmean(masked_x_tensor, dim=1) 154 | mean_tensor_all = torch.where(torch.isnan(mean_tensor_all), torch.zeros_like(mean_tensor_all), mean_tensor_all) 155 | refine_mean += mean_tensor_all.clone() 156 | 157 | # 2. Fix mean and B, update alpha 158 | new_alpha = 1. / (torch.sum(new_binary * mask * new_binary * mask, dim=1) + 1e-8) * torch.sum(new_binary * mask * (new_matrix - refine_mean[:, None] * mask), dim=1) 159 | 160 | # 3. Fix mean and alpha, update B 161 | new_binary = torch.sign(new_matrix - refine_mean[:, None] * mask) 162 | 163 | # Final refine results 164 | sum_order_alternating = torch.zeros_like(x) + (new_alpha[:, None] * new_binary + refine_mean[:, None]) * mask 165 | 166 | MM = mask[:, :, None] * mask[:, None, :] 167 | refine_mean_den = torch.sum(S * MM, dim=(1,2), dtype=torch.float32) + 1e-10 168 | masked_B = new_binary * mask 169 | new_alpha_den = torch.sum(S * masked_B[:, :, None] * masked_B[:, None, :], dim=(1,2)) + 1e-10 170 | # diag_S = torch.diag(S) 171 | for kk in range(iter2): 172 | # X error update mean 173 | refine_mean = torch.sum(S * (new_matrix - new_alpha[:, None] * new_binary * mask)[:, :, None] * MM, dim=(1,2)) / refine_mean_den 174 | 175 | # X error update alpha 176 | new_alpha = torch.sum(S * masked_B[:, :, None] * (new_matrix - refine_mean[:, None] * mask)[:, None, :], dim=(1,2)) / new_alpha_den 177 | 178 | sum_order_alternating = torch.zeros_like(x) + (new_alpha[:, None] * new_binary + refine_mean[:, None]) * mask 179 | 180 | return sum_order_alternating 181 | 182 | @torch.no_grad() 183 | def high_order_residual_alternating_order2_rc_nomean(x, mask, order=2, iter=15): 184 | sum_order = torch.zeros_like(x) 185 | new_matrix = x.clone() 186 | new_matrix = new_matrix * mask 187 | global index 188 | index += 1 189 | binary_list = [] 190 | alpha_list_r = [] 191 | alpha_list_c = [] 192 | for od in range(order): 193 | residual = new_matrix - sum_order 194 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 195 | 196 | # alpha row 197 | scale_tensor_all_r = torch.nanmean(torch.abs(masked_x_tensor), dim=1) 198 | scale_tensor_all_r = torch.where(torch.isnan(scale_tensor_all_r), torch.zeros_like(scale_tensor_all_r), scale_tensor_all_r) 199 | alpha_list_r.append(scale_tensor_all_r.clone()) 200 | # alpha column 201 | scale_tensor_all_c = torch.nanmean(torch.abs(masked_x_tensor / scale_tensor_all_r[:, None]), dim=0) 202 | scale_tensor_all_c = torch.where(torch.isnan(scale_tensor_all_c), torch.zeros_like(scale_tensor_all_c), scale_tensor_all_c) 203 | alpha_list_c.append(scale_tensor_all_c.clone()) 204 | 205 | binary= torch.sign(masked_x_tensor) 206 | binary_list.append(binary.clone()) 207 | binary *= scale_tensor_all_r[:, None] 208 | binary *= scale_tensor_all_c[None, :] 209 | sum_order = sum_order + binary*mask 210 | 211 | # Alternating update 212 | sum_order_alternating = sum_order.clone() 213 | 214 | for k in range(iter): 215 | # 2-1. Fix mean, alpha column, and B, update alpha row 0 216 | W_tilde = new_matrix - (alpha_list_c[1][None, :] * alpha_list_r[1][:, None] * binary_list[1]) * mask 217 | alpha_c_B = alpha_list_c[0][None, :] * binary_list[0] * mask 218 | alpha_list_r[0] = torch.sum(alpha_c_B * W_tilde, dim=1) / (torch.sum(alpha_c_B * alpha_c_B, dim=1) + 1e-8) 219 | 220 | # 2-2. Fix mean, alpha row, and B, update alpha column 0 221 | alpha_r_B = alpha_list_r[0][:, None] * binary_list[0] * mask 222 | alpha_list_c[0] = torch.sum(alpha_r_B * W_tilde, dim=0) / (torch.sum(alpha_r_B * alpha_r_B, dim=0) + 1e-8) 223 | 224 | # 2-3. Fix mean, alpha column, and B, update alpha row 1 225 | W_tilde = new_matrix - (alpha_list_c[0][None, :] * alpha_list_r[0][:, None] * binary_list[0]) * mask 226 | alpha_c_B = alpha_list_c[1][None, :] * binary_list[1] * mask 227 | alpha_list_r[1] = torch.sum(alpha_c_B * W_tilde, dim=1) / (torch.sum(alpha_c_B * alpha_c_B, dim=1) + 1e-8) 228 | 229 | # 2-4. Fix mean, alpha row, and B, update alpha column 1 230 | alpha_r_B = alpha_list_r[1][:, None] * binary_list[1] * mask 231 | alpha_list_c[1] = torch.sum(alpha_r_B * W_tilde, dim=0) / (torch.sum(alpha_r_B * alpha_r_B, dim=0) + 1e-8) 232 | 233 | # 3. Fix mean and alpha, update B 234 | new_matrix_expanded = new_matrix.unsqueeze(-1) 235 | comb0 = alpha_list_r[0].reshape(-1, 1) @ alpha_list_c[0].reshape(1, -1) 236 | comb1 = alpha_list_r[1].reshape(-1, 1) @ alpha_list_c[1].reshape(1, -1) 237 | v = torch.stack([-comb0 - comb1, -comb0 + comb1, 238 | comb0 - comb1, comb0 + comb1], dim=2) 239 | 240 | min_indices = torch.argmin(torch.abs(new_matrix_expanded - v), dim=-1) 241 | 242 | binary_list[0] = torch.ones_like(min_indices) 243 | binary_list[0][(min_indices == 0) | (min_indices == 1)] = -1 244 | binary_list[1] = torch.ones_like(min_indices) 245 | binary_list[1][(min_indices == 0) | (min_indices == 2)] = -1 246 | 247 | # Final refine results 248 | sum_order_alternating = torch.zeros_like(x) + (alpha_list_c[0][None, :] * alpha_list_r[0][:, None] * binary_list[0] + alpha_list_c[1][None, :] * alpha_list_r[1][:, None] * binary_list[1]) * mask 249 | 250 | return sum_order_alternating 251 | 252 | @torch.no_grad() 253 | def high_order_residual_alternating_order1_rc_nomean(x, mask, order=2, iter=15): 254 | sum_order = torch.zeros_like(x) 255 | new_matrix = x.clone() 256 | new_matrix = new_matrix * mask 257 | global index 258 | index += 1 259 | for od in range(order): 260 | residual = new_matrix - sum_order 261 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 262 | 263 | # alpha row 264 | scale_tensor_all_r = torch.nanmean(torch.abs(masked_x_tensor), dim=1) 265 | scale_tensor_all_r = torch.where(torch.isnan(scale_tensor_all_r), torch.zeros_like(scale_tensor_all_r), scale_tensor_all_r) 266 | # alpha column 267 | scale_tensor_all_c = torch.nanmean(torch.abs(masked_x_tensor / scale_tensor_all_r[:, None]), dim=0) 268 | scale_tensor_all_c = torch.where(torch.isnan(scale_tensor_all_c), torch.zeros_like(scale_tensor_all_c), scale_tensor_all_c) 269 | 270 | binary= torch.sign(masked_x_tensor) 271 | new_binary = binary.clone() 272 | binary *= scale_tensor_all_r[:, None] 273 | binary *= scale_tensor_all_c[None, :] 274 | sum_order = sum_order + binary*mask 275 | 276 | # Alternating update 277 | sum_order_alternating = sum_order.clone() 278 | new_alpha_r = scale_tensor_all_r.clone() 279 | new_alpha_c = scale_tensor_all_c.clone() 280 | for k in range(iter): 281 | # 1-1. Fix mean, alpha column, and B, update alpha row 282 | alpha_c_B = new_alpha_c[None, :] * new_binary * mask 283 | new_alpha_r = torch.sum(alpha_c_B * new_matrix, dim=1) / (torch.sum(alpha_c_B * alpha_c_B, dim=1) + 1e-8) 284 | 285 | # 1-2. Fix mean, alpha row, and B, update alpha column 286 | alpha_r_B = new_alpha_r[:, None] * new_binary * mask 287 | new_alpha_c = torch.sum(alpha_r_B * new_matrix, dim=0) / (torch.sum(alpha_r_B * alpha_r_B, dim=0) + 1e-8) 288 | 289 | # Final refine results 290 | sum_order_alternating = torch.zeros_like(x) + new_alpha_c[None, :] * new_alpha_r[:, None] * new_binary * mask 291 | 292 | return sum_order_alternating 293 | 294 | @torch.no_grad() 295 | def high_order_residual_alternating_mean(x, mask, order=2, num_iters=15): 296 | sum_order = torch.zeros_like(x) 297 | new_matrix = x.clone() 298 | new_matrix = new_matrix * mask 299 | global index 300 | index += 1 301 | binary_list = [] 302 | alpha_list = [] 303 | refine_mean = torch.zeros(x.shape[0], device=x.device) 304 | for od in range(order): 305 | residual = new_matrix - sum_order 306 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 307 | 308 | mean_tensor_all = torch.nanmean(masked_x_tensor, dim=1) 309 | mean_tensor_all = torch.where(torch.isnan(mean_tensor_all), torch.zeros_like(mean_tensor_all), mean_tensor_all) 310 | refine_mean += mean_tensor_all.clone() 311 | masked_x_tensor -= mean_tensor_all[:, None] 312 | scale_tensor_all = torch.nanmean(torch.abs(masked_x_tensor), dim=1) 313 | scale_tensor_all = torch.where(torch.isnan(scale_tensor_all), torch.zeros_like(scale_tensor_all), scale_tensor_all) 314 | alpha_list.append(scale_tensor_all.clone()) 315 | 316 | binary = torch.sign(masked_x_tensor) 317 | binary_list.append(binary.clone()) 318 | binary *= scale_tensor_all[:, None] 319 | binary += mean_tensor_all[:, None] 320 | sum_order = sum_order + binary*mask 321 | 322 | new_matrix = x.clone() * mask 323 | sum_order_alternating = sum_order.clone() 324 | 325 | for k in range(num_iters): 326 | # 1. Fix alpha1, alpha2, B1, and B2, update mean 327 | residual = new_matrix - sum_order_alternating 328 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 329 | mean_tensor_all = torch.nanmean(masked_x_tensor, dim=1) 330 | mean_tensor_all = torch.where(torch.isnan(mean_tensor_all), torch.zeros_like(mean_tensor_all), mean_tensor_all) 331 | refine_mean += mean_tensor_all.clone() 332 | 333 | # 2. Fix mean, B1, and B2, update alpha1 and alpha2 334 | alpha_list[0] = 1. / (torch.sum(binary_list[0] * mask * binary_list[0] * mask, dim=1) + 1e-8) * torch.sum(binary_list[0] * mask * (new_matrix - refine_mean[:, None] * mask - alpha_list[1][:, None] * binary_list[1] * mask), dim=1) 335 | alpha_list[1] = 1. / (torch.sum(binary_list[1] * mask * binary_list[1] * mask, dim=1) + 1e-8) * torch.sum(binary_list[1] * mask * (new_matrix - refine_mean[:, None] * mask - alpha_list[0][:, None] * binary_list[0] * mask), dim=1) 336 | 337 | # 3. Fix mean, alpha1, and alpha2, update B1 and B2 338 | new_matrix_expanded = (new_matrix - refine_mean[:, None] * mask).unsqueeze(-1) 339 | v = torch.stack([-alpha_list[0] - alpha_list[1], -alpha_list[0] + alpha_list[1], 340 | alpha_list[0] - alpha_list[1], alpha_list[0] + alpha_list[1]], dim=1).unsqueeze(1) 341 | 342 | min_indices = torch.argmin(torch.abs(new_matrix_expanded - v), dim=-1) 343 | 344 | binary_list[0] = torch.ones_like(min_indices) 345 | binary_list[0][(min_indices == 0) | (min_indices == 1)] = -1 346 | binary_list[1] = torch.ones_like(min_indices) 347 | binary_list[1][(min_indices == 0) | (min_indices == 2)] = -1 348 | 349 | sum_order_alternating = torch.zeros_like(x) + (alpha_list[0][:, None] * binary_list[0] + alpha_list[1][:, None] * binary_list[1] + refine_mean[:, None]) * mask 350 | 351 | return sum_order_alternating 352 | 353 | @torch.no_grad() 354 | def high_order_residual_alternating_mean_x(x, mask, order=2, S=None, num_iters=15, iter2=15): 355 | sum_order = torch.zeros_like(x) 356 | new_matrix = x.clone() 357 | new_matrix = new_matrix * mask 358 | global index 359 | index += 1 360 | binary_list = [] 361 | alpha_list = [] 362 | refine_mean = torch.zeros(x.shape[0], device=x.device) 363 | for od in range(order): 364 | residual = new_matrix - sum_order 365 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 366 | 367 | mean_tensor_all = torch.nanmean(masked_x_tensor, dim=1) 368 | mean_tensor_all = torch.where(torch.isnan(mean_tensor_all), torch.zeros_like(mean_tensor_all), mean_tensor_all) 369 | refine_mean += mean_tensor_all.clone() 370 | masked_x_tensor -= mean_tensor_all[:, None] 371 | scale_tensor_all = torch.nanmean(torch.abs(masked_x_tensor), dim=1) 372 | scale_tensor_all = torch.where(torch.isnan(scale_tensor_all), torch.zeros_like(scale_tensor_all), scale_tensor_all) 373 | alpha_list.append(scale_tensor_all.clone()) 374 | 375 | binary = torch.sign(masked_x_tensor) 376 | binary_list.append(binary.clone()) 377 | binary *= scale_tensor_all[:, None] 378 | binary += mean_tensor_all[:, None] 379 | sum_order = sum_order + binary*mask 380 | 381 | new_matrix = x.clone() * mask 382 | sum_order_alternating = sum_order.clone() 383 | 384 | for k in range(num_iters): 385 | # 1. Fix alpha1, alpha2, B1, and B2, update mean 386 | residual = new_matrix - sum_order_alternating 387 | masked_x_tensor = torch.where(mask, residual, torch.tensor(float('nan'))) 388 | mean_tensor_all = torch.nanmean(masked_x_tensor, dim=1) 389 | mean_tensor_all = torch.where(torch.isnan(mean_tensor_all), torch.zeros_like(mean_tensor_all), mean_tensor_all) 390 | refine_mean += mean_tensor_all.clone() 391 | 392 | # 2. Fix mean, B1, and B2, update alpha1 and alpha2 393 | alpha_list[0] = 1. / (torch.sum(binary_list[0] * mask * binary_list[0] * mask, dim=1) + 1e-8) * torch.sum(binary_list[0] * mask * (new_matrix - refine_mean[:, None] * mask - alpha_list[1][:, None] * binary_list[1] * mask), dim=1) 394 | alpha_list[1] = 1. / (torch.sum(binary_list[1] * mask * binary_list[1] * mask, dim=1) + 1e-8) * torch.sum(binary_list[1] * mask * (new_matrix - refine_mean[:, None] * mask - alpha_list[0][:, None] * binary_list[0] * mask), dim=1) 395 | 396 | # 3. Fix mean, alpha1, and alpha2, update B1 and B2 397 | new_matrix_expanded = (new_matrix - refine_mean[:, None] * mask).unsqueeze(-1) 398 | v = torch.stack([-alpha_list[0] - alpha_list[1], -alpha_list[0] + alpha_list[1], 399 | alpha_list[0] - alpha_list[1], alpha_list[0] + alpha_list[1]], dim=1).unsqueeze(1) 400 | 401 | min_indices = torch.argmin(torch.abs(new_matrix_expanded - v), dim=-1) 402 | 403 | binary_list[0] = torch.ones_like(min_indices) 404 | binary_list[0][(min_indices == 0) | (min_indices == 1)] = -1 405 | binary_list[1] = torch.ones_like(min_indices) 406 | binary_list[1][(min_indices == 0) | (min_indices == 2)] = -1 407 | 408 | sum_order_alternating = torch.zeros_like(x) + (alpha_list[0][:, None] * binary_list[0] + alpha_list[1][:, None] * binary_list[1] + refine_mean[:, None]) * mask 409 | 410 | MM = mask[:, :, None] * mask[:, None, :] 411 | refine_mean_den = torch.sum(S * MM, dim=(1,2)) + 1e-10 412 | masked_B0 = binary_list[0] * mask 413 | new_alpha0_den = torch.sum(S * masked_B0[:, :, None] * masked_B0[:, None, :], dim=(1,2)) + 1e-10 414 | masked_B1 = binary_list[1] * mask 415 | new_alpha1_den = torch.sum(S * masked_B1[:, :, None] * masked_B1[:, None, :], dim=(1,2)) + 1e-10 416 | for kk in range(iter2): 417 | # X error update mean 418 | refine_mean = torch.sum(S * (new_matrix - (alpha_list[0][:, None] * binary_list[0] + alpha_list[1][:, None] * binary_list[1]) * mask)[:, :, None] * MM, dim=(1,2)) / refine_mean_den 419 | 420 | # X error update alpha 421 | masked_W_mu = new_matrix - refine_mean[:, None] * mask 422 | alpha_list[0] = torch.sum(S * masked_B0[:, :, None] * (masked_W_mu[:, None, :] - (alpha_list[1][:, None] * masked_B1)[:, None, :]), dim=(1,2)) / new_alpha0_den 423 | alpha_list[1] = torch.sum(S * masked_B1[:, :, None] * (masked_W_mu[:, None, :] - (alpha_list[0][:, None] * masked_B0)[:, None, :]), dim=(1,2)) / new_alpha1_den 424 | 425 | sum_order_alternating = torch.zeros_like(x) + (alpha_list[0][:, None] * binary_list[0] + alpha_list[1][:, None] * binary_list[1] + refine_mean[:, None]) * mask 426 | 427 | return sum_order_alternating 428 | 429 | @torch.no_grad() 430 | def normal_quantize(x, scale, zero, maxq): 431 | q = torch.clamp(torch.round(x / scale) + zero, 0, maxq) 432 | return scale * (q - zero) 433 | 434 | 435 | class Binarization(nn.Module): 436 | def __init__(self, weight, method="arb", groupsize=-1): 437 | super().__init__() 438 | oc,ic=weight.shape 439 | if groupsize==-1: 440 | groupsize=ic 441 | self.groupsize=groupsize 442 | self.n_groups=math.ceil(ic/groupsize) 443 | self.method=method 444 | self.mean = 0 445 | 446 | def quantize(self, w, mask, order=2, groupi=0, S=None): 447 | if self.method=="xnor": 448 | w_mean = self.mean[groupi] 449 | w = w - w_mean # oc, ic 450 | w = w.sign() 451 | w = w * self.scale[groupi] 452 | w+=w_mean 453 | elif self.method=="braq": # The method used in BiLLM 454 | w = high_order_residual(w, mask, order=order) 455 | 456 | # arb series 457 | elif self.method == "arb": 458 | if order == 2: 459 | w = high_order_residual_alternating_mean(w, mask, order=order) 460 | else: 461 | w = high_order_residual_alternating_order1(w, mask, order=order) 462 | elif self.method == 'arb-x': 463 | if order == 2: 464 | w = high_order_residual_alternating_mean_x(w, mask, order=order, S=S) 465 | else: 466 | w = high_order_residual_alternating_order1_x(w, mask, order=order, S=S) 467 | elif self.method == 'arb-rc': 468 | if order == 2: 469 | w = high_order_residual_alternating_order2_rc_nomean(w, mask, order=order) 470 | else: 471 | w = high_order_residual_alternating_order1_rc_nomean(w, mask, order=order) 472 | 473 | elif self.method=="sign": 474 | w=(w>0).float() 475 | w*=self.scale[groupi] 476 | elif self.method=="rtn": 477 | w=F.relu(w) 478 | w_int=(w/self.scale[groupi]).round().clamp(0,1) 479 | w=w_int*self.scale[groupi] 480 | elif self.method in ['2bit','4bit']: 481 | 482 | bits = int(self.method[0]) 483 | perchannel = True 484 | weight = True 485 | dev = w.device 486 | maxq = torch.tensor(2 ** bits - 1) 487 | scale = torch.zeros(1) 488 | zero = torch.zeros(1) 489 | 490 | if dev != scale.device: 491 | scale=scale.to(dev) 492 | zero=zero.to(dev) 493 | maxq=maxq.to(dev) 494 | 495 | x = w.clone() 496 | shape = x.shape 497 | 498 | if perchannel: 499 | if weight: 500 | x = x.flatten(1) 501 | else: 502 | if len(shape) == 4: 503 | x = x.permute([1, 0, 2, 3]) 504 | x = x.flatten(1) 505 | if len(shape) == 3: 506 | x = x.reshape((-1, shape[-1])).t() 507 | if len(shape) == 2: 508 | x = x.t() 509 | else: 510 | x = x.flatten().unsqueeze(0) 511 | tmp = torch.zeros(x.shape[0], device=dev) 512 | xmin = torch.minimum(x.min(1)[0], tmp) 513 | xmax = torch.maximum(x.max(1)[0], tmp) 514 | 515 | tmp = (xmin == 0) & (xmax == 0) 516 | xmin[tmp] = -1 517 | xmax[tmp] = +1 518 | scale = (xmax - xmin) / maxq 519 | zero = torch.round(-xmin / scale) 520 | if not perchannel: 521 | if weight: 522 | tmp = shape[0] 523 | else: 524 | tmp = shape[1] if len(shape) != 3 else shape[2] 525 | scale = scale.repeat(tmp) 526 | zero = zero.repeat(tmp) 527 | 528 | if weight: 529 | shape = [-1] + [1] * (len(shape) - 1) 530 | scale = scale.reshape(shape) 531 | zero = zero.reshape(shape) 532 | w = normal_quantize(w, scale, zero, maxq) 533 | 534 | elif self.method=="prune": 535 | return torch.zeros_like(w) 536 | return w 537 | --------------------------------------------------------------------------------