Upload 8 files

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [2024] [Guangji Bai]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -1,3 +1,104 @@
----
-license: apache-2.0
----
+# *SparseLLM*: Towards Global Pruning of LLMs
+
+This repository contains the code for our **NeurIPS 2024** paper "[*SparseLLM*: Towards Global Pruning for Pre-trained Language Models](https://arxiv.org/abs/2402.17946)".
+
+## Updates
+
+- <span style="color:green;">&#x2705;</span> *SparseLLM* code for both **OPT** and **LLaMA** models is now available.
+- <span style="color:green;">&#x2705;</span> More model types and functionalities will be added soon.
+
+
+## Dependencies
+
+This project requires the following core dependencies:
+
+- `Python`: tested on v3.10.14
+- `PyTorch`: tested on v2.4.1 with CUDA 12.2 
+- `Transformers`: tested on v4.45.1
+- `Datasets`: tested on v3.0.1
+- `numpy`: tested on v2.1.1
+- `pandas`: tested on v2.2.3
+- `huggingface_hub`: tested on v0.25.1
+- `wandb`: tested on v0.18.2 (for experiment tracking)
+
+## Usage
+
+The scripts directory contains all the bash commands to replicate the main results in our NeurIPS 2024 paper. 
+
+### Example for Pruning OPT:
+
+Below is an example command for pruning the OPT-125M model using SparseLLM, to achieve 70% sparsity.
+
+```
+python opt_main.py \
+    --model facebook/opt-125m \
+    --dataset c4 \
+    --sparsity 0.7 \
+```
+
+We provide a quick overview of the key arguments:
+
+- `--model`: The identifier for the model on the Hugging Face model hub.
+- `--dataset`: The dataset to use for evaluation. We support datasets like `c4`, `wikitext2`, and `ptb`.
+- `--sparsity`: The desired sparsity level (percentage of weights to be pruned).
+
+**Remark:** OPT-350M is currently not supported by our method, due to potential numerical stability issue.
+
+### Example for Pruning LLaMA-2:
+
+For **LLaMA-2** models, use the llama_main.py file and specify the model path as `meta-llama/Llama-2-7b-hf`. Here is an example command for pruning LLaMA-2-7B:
+
+```
+python llama_main.py \
+    --model meta-llama/Llama-2-7b-hf \
+    --dataset c4 \
+    --sparsity 0.7 \
+```
+
+### Available Sparsity Methods
+
+We support the following pruning methods for both **OPT** and **LLaMA** models:
+
+- **Unstructured**: Pruning individual weights without any specific pattern.
+- **Semi-Structured N:M Sparsity**: For semi-structured pruning, use the following sparsity types:
+  - `--sparsity_type 2:4`: Prune 2 out of every 4 weights.
+  - `--sparsity_type 4:8`: Prune 4 out of every 8 weights.
+
+```
+python opt_main.py \
+    --model facebook/opt-125m \
+    --dataset c4 \
+    --prunen 2 \
+    --prunem 4 \
+```
+
+Similarly, for **LLaMA-2-7B** semi-structured pruning:
+
+```
+python llama_main.py \
+    --model meta-llama/Llama-2-7b-hf \
+    --dataset c4 \
+    --prunen 2 \
+    --prunem 4 \
+```
+
+## Reference
+
+If you find this code useful in your research, please consider citing:
+
+```bibtex
+@article{bai2024gradient,
+  title={Gradient-Free Adaptive Global Pruning for Pre-trained Language Models},
+  author={Bai, Guangji and Li, Yijiang and Ling, Chen and Kim, Kibaek and Zhao, Liang},
+  journal={arXiv preprint arXiv:2402.17946},
+  year={2024}
+}
+```
+We sincerely appreciate it 😊
+
+## Disclaimer
+
+1. This repository is built upon the work introduced in the papers [SparseGPT](https://arxiv.org/abs/2301.00774) and [Wanda](https://arxiv.org/abs/2306.11695).
+2. *SparseLLM* aims to advance the research on improving fully local pruning methods for large language models (LLMs). Due to the iterative alternating optimization nature of *SparseLLM*, its running time will be longer than that of one-shot pruning methods (roughly number of iteration times) such as SparseGPT or Wanda. Additionally, the performance and numerical stability of the alternating optimization process can be sensitive to the initialization of hyperparameters.
+3. *SparseLLM* relies on auxiliary variables to achieve subproblem decomposition, which inevitably introduces additional memory overhead. For larger models like LLaMA-2-7B and beyond, we used a smaller calibration data size (e.g., 64 or 32) to ensure the code could run on an A100 40GB GPU. We are actively working on optimizing the GPU memory consumption and improving the efficiency of the code to support larger models and data sizes more effectively.
+

datautils.py

@@ -0,0 +1,106 @@
+# DISCLAIMER: This file is a modified version of the original SparseGPT data loader. The original SparseGPT data loader can be found in [SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot].
+
+import random
+
+import numpy as np
+import torch
+from datasets import load_dataset
+from transformers import AutoTokenizer, LlamaTokenizer
+
+
+def set_seed(seed):
+    np.random.seed(seed)
+    torch.random.manual_seed(seed)
+
+def get_tokenizer(model):
+    if "llama" in model.lower():
+        tokenizer = LlamaTokenizer.from_pretrained(model, use_fast=False)
+        # fix for transformer 4.28.0.dev0 compatibility
+        if tokenizer.bos_token_id != 1 or tokenizer.eos_token_id != 2:
+            try:
+                tokenizer.bos_token_id = 1
+                tokenizer.eos_token_id = 2
+            except AttributeError:
+                pass
+    else:
+        tokenizer = AutoTokenizer.from_pretrained(model, use_fast=False)
+    return tokenizer
+
+def get_wikitext2(nsamples, seed, seqlen, model, tokenizer):
+    
+    traindata = load_dataset('wikitext', 'wikitext-2-raw-v1', split='train')
+    testdata = load_dataset('wikitext', 'wikitext-2-raw-v1', split='test')
+
+    trainenc = tokenizer(" ".join(traindata['text']), return_tensors='pt')
+    testenc = tokenizer("\n\n".join(testdata['text']), return_tensors='pt')
+
+    random.seed(seed)
+    trainloader = []
+    for _ in range(nsamples):
+        i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1)
+        j = i + seqlen
+        inp = trainenc.input_ids[:, i:j]
+        tar = inp.clone()
+        tar[:, :-1] = -100
+        trainloader.append((inp, tar))
+    return trainloader, testenc
+
+def get_ptb(nsamples, seed, seqlen, model, tokenizer):
+    traindata = load_dataset('ptb_text_only', 'penn_treebank', split='train')
+    testdata = load_dataset('ptb_text_only', 'penn_treebank', split='test')
+
+    trainenc = tokenizer(" ".join(traindata['sentence']), return_tensors='pt')
+    testenc = tokenizer(" ".join(testdata['sentence']), return_tensors='pt')
+
+    random.seed(seed)
+    trainloader = []
+    for _ in range(nsamples):
+        i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1)
+        j = i + seqlen
+        inp = trainenc.input_ids[:, i:j]
+        tar = inp.clone()
+        tar[:, :-1] = -100
+        trainloader.append((inp, tar))
+    return trainloader, testenc
+
+def get_c4(nsamples, seed, seqlen, model, tokenizer):
+    traindata = load_dataset(
+        'allenai/c4', data_files={'train': 'en/c4-train.00000-of-01024.json.gz'}, split='train'
+    )
+    valdata = load_dataset(
+        'allenai/c4', data_files={'validation': 'en/c4-validation.00000-of-00008.json.gz'}, split='validation'
+    )
+
+    random.seed(seed)
+    trainloader = []
+    for _ in range(nsamples):
+        while True:
+            i = random.randint(0, len(traindata) - 1)
+            trainenc = tokenizer(traindata[i]['text'], return_tensors='pt')
+            if trainenc.input_ids.shape[1] > seqlen:
+                break
+        i = random.randint(0, trainenc.input_ids.shape[1] - seqlen - 1)
+        j = i + seqlen
+        inp = trainenc.input_ids[:, i:j]
+        tar = inp.clone()
+        tar[:, :-1] = -100
+        trainloader.append((inp, tar))
+
+    valenc = tokenizer(' '.join(valdata[:1100]['text']), return_tensors='pt')
+    valenc = valenc.input_ids[:, :(256 * seqlen)]
+
+    class TokenizerWrapper:
+        def __init__(self, input_ids):
+            self.input_ids = input_ids
+    valenc = TokenizerWrapper(valenc)
+
+    return trainloader, valenc
+
+def get_loaders(name, nsamples=128, seed=0, seqlen=2048, model=''):
+    tokenizer = get_tokenizer(model)
+    if 'wikitext2' in name:
+        return get_wikitext2(nsamples, seed, seqlen, model, tokenizer)
+    if 'ptb' in name:
+        return get_ptb(nsamples, seed, seqlen, model, tokenizer)
+    if 'c4' in name:
+        return get_c4(nsamples, seed, seqlen, model, tokenizer)

llama_main.py

@@ -0,0 +1,46 @@
+import argparse
+from model_utils import get_llama, llama_sparsellm, llama_eval
+from datautils import get_loaders
+import torch
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    # Arguments parsing
+    parser.add_argument("--model", type=str, default='meta-llama/Llama-2-7b-hf', help="LLaMA model to load")
+    parser.add_argument("--dataset", type=str, choices=["wikitext2", "ptb", "c4"], default="c4", help="Dataset for calibration.")
+    parser.add_argument("--seed", type=int, default=0, help="Seed for sampling calibration data.")
+    parser.add_argument("--nsamples", type=int, default=32, help="Number of calibration data samples.")
+    parser.add_argument("--percdamp", type=float, default=0.01, help="Percent of Hessian diagonal for dampening.")
+    parser.add_argument("--sparsity", type=float, default=0.5, help="Target sparsity.")
+    parser.add_argument("--prunen", type=int, default=0, help="N for N:M pruning.")
+    parser.add_argument("--prunem", type=int, default=0, help="M for N:M pruning.")
+    parser.add_argument("--blocksize", type=int, default=128, help="Blocksize for adaptive mask selection.")
+    parser.add_argument("--gmp", action="store_true", help="Run GMP baseline.")
+    parser.add_argument("--wbits", type=int, default=16, help="Quantization bits.")
+    parser.add_argument("--minlayer", type=int, default=-1, help="Prune layers with id >= this.")
+    parser.add_argument("--maxlayer", type=int, default=1000, help="Prune layers with id < this.")
+    parser.add_argument("--prune_only", type=str, default="", help="Prune only layers containing this text.")
+    parser.add_argument("--invert", action="store_true", help="Invert subset.")
+    parser.add_argument("--save", type=str, default="", help="Path to save model.")
+    parser.add_argument("--true-sequential", action="store_true", help="Run in true sequential mode.")
+    parser.add_argument("--log_wandb", action="store_true", help="Log to W&B.")
+
+    args = parser.parse_args()
+
+    model = get_llama(args)
+    model.eval()
+    dataloader, testloader = get_loaders(args.dataset, nsamples=args.nsamples, seed=args.seed, model=args.model, seqlen=model.seqlen)
+
+    if (args.sparsity or args.prunen) and not args.gmp:
+        llama_sparsellm(model, dataloader, torch.device('cuda'), args)
+
+    for dataset in ['wikitext2', 'ptb', 'c4']:
+        dataloader, testloader = get_loaders(dataset, seed=args.seed, model=args.model, seqlen=model.seqlen)
+        llama_eval(model, testloader, torch.device('cuda'), args, dataset)
+
+    if args.save:
+        model.save_pretrained(args.save)
+
+if __name__ == '__main__':
+    main()

model_utils.py

@@ -0,0 +1,831 @@
+# This file will contain functions related to the model such as loading the model, SparseLLM pruning, and evaluation.
+
+import torch
+import torch.nn as nn
+from pruning_utils import *
+from quant import *
+import math
+from transformers import OPTForCausalLM, LlamaForCausalLM
+
+def get_opt(args):
+    def skip(*args, **kwargs):
+        pass
+    torch.nn.init.kaiming_uniform_ = skip
+    torch.nn.init.uniform_ = skip
+    torch.nn.init.normal_ = skip
+    model = OPTForCausalLM.from_pretrained(args.model, torch_dtype='auto')
+    model.seqlen = model.config.max_position_embeddings
+    return model
+
+def get_llama(args):
+    def skip(*args, **kwargs):
+        pass
+    torch.nn.init.kaiming_uniform_ = skip
+    torch.nn.init.uniform_ = skip
+    torch.nn.init.normal_ = skip
+    model = LlamaForCausalLM.from_pretrained(args.model, torch_dtype='auto')
+    model.seqlen = 2048
+    return model
+
+@torch.no_grad()
+def opt_sparsellm(model, dataloader, dev, args):
+    print('Starting ...')
+
+    use_cache = model.config.use_cache
+    model.config.use_cache = False
+    layers = model.model.decoder.layers
+
+    model.model.decoder.embed_tokens = model.model.decoder.embed_tokens.to(dev) 
+    model.model.decoder.embed_positions = model.model.decoder.embed_positions.to(dev)
+    if hasattr(model.model.decoder, 'project_out') and model.model.decoder.project_out:
+        model.model.decoder.project_out = model.model.decoder.project_out.to(dev) 
+    if hasattr(model.model.decoder, 'project_in') and model.model.decoder.project_in:
+        model.model.decoder.project_in = model.model.decoder.project_in.to(dev) 
+    layers[0] = layers[0].to(dev)
+
+    dtype = next(iter(model.parameters())).dtype
+    inps = torch.zeros(
+        (args.nsamples, model.seqlen, model.config.hidden_size), dtype=dtype, device=dev
+    )
+    cache = {'i': 0, 'attention_mask': None}
+
+    class Catcher(nn.Module):
+        def __init__(self, module):
+            super().__init__()
+            self.module = module
+        def forward(self, inp, **kwargs):
+            inps[cache['i']] = inp
+            cache['i'] += 1
+            cache['attention_mask'] = kwargs['attention_mask']
+            raise ValueError
+    layers[0] = Catcher(layers[0])
+    for batch in dataloader:
+        try:
+            model(batch[0].to(dev))
+        except ValueError:
+            pass
+    layers[0] = layers[0].module
+
+    layers[0] = layers[0].cpu()
+    model.model.decoder.embed_tokens = model.model.decoder.embed_tokens.cpu()
+    model.model.decoder.embed_positions = model.model.decoder.embed_positions.cpu()
+    if hasattr(model.model.decoder, 'project_out') and model.model.decoder.project_out:
+        model.model.decoder.project_out = model.model.decoder.project_out.cpu()
+    if hasattr(model.model.decoder, 'project_in') and model.model.decoder.project_in:
+        model.model.decoder.project_in = model.model.decoder.project_in.cpu()
+    torch.cuda.empty_cache()
+
+    outs = torch.zeros_like(inps)
+    attention_mask = cache['attention_mask']
+
+    print('Ready.')
+
+    for i in range(len(layers)):
+        layer = layers[i].to(dev)
+
+        subset = find_layers(layer)
+        
+        gpts = {}
+        for name in subset:
+            if (not (args.minlayer <= i < args.maxlayer and args.prune_only in name)) == (not args.invert):
+              continue
+            gpts[name] = SparseGPT_OPT(subset[name])
+            if args.wbits < 16:
+                gpts[name].quantizer = Quantizer()
+                gpts[name].quantizer.configure(
+                    args.wbits, perchannel=True, sym=False, mse=False
+                )
+
+        def add_batch(name):
+            def tmp(_, inp, out):
+                gpts[name].add_batch(inp[0].data, out.data, name)
+            return tmp
+        handles = []
+        for name in gpts:
+            handles.append(subset[name].register_forward_hook(add_batch(name)))
+        for j in range(args.nsamples):
+            outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask)[0]
+        for h in handles:
+            h.remove()
+
+        target_layer_names = ['fc1', 'fc2']
+
+        for name in gpts:
+            if name not in target_layer_names:   
+                print(i, name)
+                print('Pruning ...')
+                # Prune the layer
+                sparsity = args.sparsity
+                gpts[name].fasterprune(
+                    sparsity, prunen=args.prunen, prunem=args.prunem, percdamp=args.percdamp, blocksize=args.blocksize
+                )
+                gpts[name].free()
+
+        # Adjust hyperparameters as needed
+        alpha = 5.0
+        beta = 5.0
+        gamma = 5.0
+
+        # Define the number of optimization steps
+        opt_epochs = 10
+
+        # Get the inputs and outputs which are constants here
+        X_list = gpts['fc1'].batch_inp
+        Y_list = gpts['fc2'].batch_out
+        X = torch.stack(X_list, dim=0)
+        Y = torch.stack(Y_list, dim=0)
+        # Reshape to 2D
+        X, Y = X.reshape((-1, X.size(-1))).T, Y.reshape((-1, Y.size(-1))).T
+
+        # free memory 
+        X_list, Y_list = None, None
+        gpts['fc1'].batch_inp.clear()
+        gpts['fc2'].batch_out.clear()
+
+        hidden_z_list = gpts['fc1'].batch_out
+        z = torch.stack(hidden_z_list, dim=0)
+        hidden_z_list = None
+        gpts['fc1'].batch_out.clear()
+        hidden_p_list = gpts['fc2'].batch_inp
+        p = torch.stack(hidden_p_list, dim=0)
+        hidden_p_list = None
+        gpts['fc2'].batch_inp.clear()
+
+        # Initialize auxiliary variables z and p
+        z = z.reshape((-1, z.size(-1))).T.to(dev)
+        p = p.reshape((-1, p.size(-1))).T.to(dev)
+
+        torch.cuda.empty_cache()
+
+        # Pre-compute the pinverse of X and cache it to save computational cost
+        Xinv = torch.pinverse(X.to(dtype=torch.float32)).half()
+
+        for opt_step in range(opt_epochs):
+
+            ##############
+            # optimize W
+            ##############
+
+            if opt_step > 0:   # for the first step, no need for updating W
+
+                # Update the weight matrix of fc1
+                bias = subset['fc1'].bias.unsqueeze(1).expand(-1, z.size(-1))
+                # Calculate the weight matrix
+                weight_matrix_1 = torch.matmul(z - bias, Xinv)
+                # assign the new parameters to gpts class
+                gpts['fc1'].layer.weight.copy_(weight_matrix_1)
+                del bias, weight_matrix_1
+
+                # Update the weight matrix of fc2
+                pinv = torch.pinverse(p.to(dtype=torch.float32)).half()
+                bias = subset['fc2'].bias.unsqueeze(1).expand(-1, Y.size(-1))
+                # Calculate the weight matrix
+                weight_matrix_2 = torch.matmul(Y - bias, pinv)
+                # assign the new parameters to gpts class
+                gpts['fc2'].layer.weight.copy_(weight_matrix_2)
+
+                del bias, weight_matrix_2, pinv
+                torch.cuda.empty_cache()
+
+            ##############
+            # prune W
+            ##############
+
+            # modify gpts[name].H to be our auxiliary variable
+            if opt_step > 0:   # for the first step, no need for updating H    
+            
+                tmp_H = torch.zeros_like(gpts['fc2'].H)
+                tmp_p = p.T.reshape((args.nsamples, -1, p.size(0)))
+                tmp_nsamples = 0
+                for j in range(args.nsamples):
+                    tmp_inp = tmp_p[j].unsqueeze(0)
+                    tmp = tmp_inp.shape[0]
+                    if isinstance(gpts['fc2'].layer, nn.Linear) or isinstance(gpts['fc2'].layer, transformers.Conv1D):
+                        if len(tmp_inp.shape) == 3:
+                            tmp_inp = tmp_inp.reshape((-1, tmp_inp.shape[-1]))
+                        tmp_inp = tmp_inp.t()
+                    tmp_H *= tmp_nsamples / (tmp_nsamples + tmp)
+                    tmp_nsamples += tmp
+                    tmp_inp = math.sqrt(2 / tmp_nsamples) * tmp_inp.float()
+                    tmp_H += tmp_inp.matmul(tmp_inp.t())
+                gpts['fc2'].H.copy_(tmp_H)
+                del tmp_H, tmp_p
+                torch.cuda.empty_cache()
+
+            for name in target_layer_names:
+                print(i, name)
+                print('Pruning ...')
+                sparsity = args.sparsity
+                gpts[name].fasterprune(
+                    sparsity, prunen=args.prunen, prunem=args.prunem, percdamp=args.percdamp, blocksize=args.blocksize
+                )
+
+            ##############
+            # optimize p
+            ##############
+
+            # Activation inverse
+            next_weight = subset['fc2'].weight
+            m1 = beta * torch.matmul(next_weight.T, next_weight)
+            m2 = gamma * torch.eye(m1.shape[0], device=m1.device)
+            av = torch.inverse(m1 + m2).to(dtype=torch.float16)
+
+            del m1, m2
+            torch.cuda.empty_cache()
+
+            # Calculate ReLU
+            layer_nl_output = nn.functional.relu(z)
+
+            # Activation formulate
+            bias = subset['fc2'].bias.unsqueeze(1).expand(-1, Y.size(-1))
+            m3 = beta * torch.matmul(next_weight.T, Y - bias)
+            m4 = gamma * layer_nl_output
+            af = m3 + m4
+
+            p = torch.matmul(av, af)
+
+            del layer_nl_output, next_weight, av, m3, m4, af, bias
+            torch.cuda.empty_cache()
+
+            ##############
+            # optimize z
+            ##############
+
+            w = subset['fc1'].weight
+            bias = subset['fc1'].bias.unsqueeze(1).expand(-1, z.size(-1))
+            m = torch.matmul(w, X) + bias
+            sol1 = (gamma * p + alpha * m) / (gamma + alpha)
+            sol2 = m
+            del w, bias
+            torch.cuda.empty_cache()
+
+            z1 = torch.zeros_like(p)
+            z2 = torch.zeros_like(p)
+
+            chunk_size = 500  # Choose an appropriate size based on your memory constraints
+            # Assuming the first dimension is the one to be chunked
+            for k in range(0, sol1.size(0), chunk_size):
+                chunk = slice(k, k + chunk_size)
+                
+                # Apply the condition and assignment for the chunk
+                z1_chunk = z1[chunk]
+                sol1_chunk = sol1[chunk]
+                z1_chunk[sol1_chunk >= 0.] = sol1_chunk[sol1_chunk >= 0.]
+                z1[chunk] = z1_chunk
+
+                z2_chunk = z2[chunk]
+                sol2_chunk = sol2[chunk]
+                z2_chunk[sol2_chunk <= 0.] = sol2_chunk[sol2_chunk <= 0.]
+                z2[chunk] = z2_chunk
+
+            del z1_chunk, z2_chunk, sol1_chunk, sol2_chunk, sol1, sol2
+            torch.cuda.empty_cache()
+
+            for k in range(0, z1.size(0), chunk_size):
+                chunk = slice(k, k + chunk_size)
+                
+                # Compute fz_1 and fz_2 for the current chunk
+                fz_1_chunk = gamma * torch.square(p[chunk] - nn.functional.relu(z1[chunk])) + alpha * torch.square(z1[chunk] - m[chunk])
+                fz_2_chunk = gamma * torch.square(p[chunk] - nn.functional.relu(z2[chunk])) + alpha * torch.square(z2[chunk] - m[chunk])
+
+                # Determine indices for z1 and z2 for the current chunk
+                index_z1_chunk = fz_1_chunk <= fz_2_chunk
+                index_z2_chunk = fz_2_chunk < fz_1_chunk
+
+                # Update z for the current chunk
+                z[chunk][index_z1_chunk] = z1[chunk][index_z1_chunk]
+                z[chunk][index_z2_chunk] = z2[chunk][index_z2_chunk]
+
+            # Clear memory if necessary
+            del fz_1_chunk, fz_2_chunk, index_z1_chunk, index_z2_chunk, z1, z2, m, chunk
+            torch.cuda.empty_cache()
+
+        for name in target_layer_names:
+            gpts[name].free()
+
+        for j in range(args.nsamples):
+            outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask)[0]
+
+        layers[i] = layer.cpu()
+        del layer
+        torch.cuda.empty_cache()
+
+        inps, outs = outs, inps
+
+    model.config.use_cache = use_cache
+
+
+@torch.no_grad()
+def llama_sparsellm(model, dataloader, dev, args):
+    print("Starting...")
+
+    use_cache = model.config.use_cache
+    model.config.use_cache = False
+    layers = model.model.layers
+
+    model.model.embed_tokens = model.model.embed_tokens.to(dev)
+    model.model.norm = model.model.norm.to(dev)
+    layers[0] = layers[0].to(dev)
+
+    dtype = next(iter(model.parameters())).dtype
+    inps = torch.zeros(
+        (args.nsamples, model.seqlen, model.config.hidden_size), dtype=dtype, device=dev
+    )
+    cache = {"i": 0, "attention_mask": None}
+
+    class Catcher(nn.Module):
+        def __init__(self, module):
+            super().__init__()
+            self.module = module
+
+        def forward(self, inp, **kwargs):
+            inps[cache["i"]] = inp
+            cache["i"] += 1
+            cache["attention_mask"] = kwargs["attention_mask"]
+            raise ValueError
+
+    layers[0] = Catcher(layers[0])
+    for batch in dataloader:
+        try:
+            model(batch[0].to(dev))
+        except ValueError:
+            pass
+    layers[0] = layers[0].module
+
+    layers[0] = layers[0].cpu()
+    model.model.embed_tokens = model.model.embed_tokens.cpu()
+    model.model.norm = model.model.norm.cpu()
+    torch.cuda.empty_cache()
+
+    outs = torch.zeros_like(inps)
+    attention_mask = cache["attention_mask"]
+
+    print("Ready.")
+
+    for i in range(len(layers)):
+        layer = layers[i].to(dev)
+        full = find_layers(layer)
+
+        if args.true_sequential:
+            sequential = [
+                ["self_attn.k_proj", "self_attn.v_proj", "self_attn.q_proj"],
+                ["self_attn.o_proj"],
+                ["mlp.up_proj", "mlp.gate_proj"],
+                ["mlp.down_proj"],
+            ]
+        else:
+            sequential = [list(full.keys())]
+
+        for names in sequential:
+            subset = {n: full[n] for n in names}
+
+            gpts = {}
+            for name in subset:
+                if (
+                    not (args.minlayer <= i < args.maxlayer and args.prune_only in name)
+                ) == (not args.invert):
+                    continue
+                gpts[name] = SparseGPT_LlaMA(subset[name])
+                if args.wbits < 16:
+                    gpts[name].quantizer = Quantizer()
+                    gpts[name].quantizer.configure(
+                        args.wbits, perchannel=True, sym=False, mse=False
+                    )
+
+            def add_batch(name):
+                def tmp(_, inp, out):
+                    gpts[name].add_batch(inp[0].data, out.data, name)
+
+                return tmp
+
+            handles = []
+            for name in subset:
+                handles.append(subset[name].register_forward_hook(add_batch(name)))
+            for j in range(args.nsamples):
+                outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask)[0]
+            for h in handles:
+                h.remove()
+
+            target_layer_names = ["mlp.up_proj", "mlp.gate_proj", "mlp.down_proj"]
+
+            for name in subset:
+                if name not in target_layer_names:   
+                    print(i, name)
+                    print("Pruning ...")
+                    sparsity = args.sparsity
+                    gpts[name].fasterprune(
+                        sparsity,
+                        prunen=args.prunen,
+                        prunem=args.prunem,
+                        percdamp=args.percdamp,
+                        blocksize=args.blocksize,
+                    )
+                    gpts[name].free()
+
+            # Adjust hyperparameters as needed
+            alpha = 5.0
+            beta = 5.0
+            gamma = 5.0
+
+            # Define the number of global pruning epochs
+            opt_epochs = 8  # This might need to be adjusted
+
+            # Get the inputs and outputs which are constants here
+            X_list = gpts['mlp.up_proj'].batch_inp
+            Y_list = gpts['mlp.down_proj'].batch_out
+            X = torch.stack(X_list, dim=0)
+            Y = torch.stack(Y_list, dim=0)
+            # Reshape to 2D
+            X, Y = X.reshape((-1, X.size(-1))).T, Y.reshape((-1, Y.size(-1))).T
+
+            # free memory 
+            X_list, Y_list = None, None
+            gpts['mlp.up_proj'].batch_inp.clear()
+            gpts['mlp.down_proj'].batch_out.clear()
+
+            # Get the hidden variables and their initialization
+            # z: output of 'mlp.up_proj'
+            hidden_z_list = gpts['mlp.up_proj'].batch_out
+            z = torch.stack(hidden_z_list, dim=0)
+            hidden_z_list = None
+            gpts['mlp.up_proj'].batch_out.clear()
+            # p: input of 'mlp.down_proj'
+            hidden_p_list = gpts['mlp.down_proj'].batch_inp
+            p = torch.stack(hidden_p_list, dim=0)
+            hidden_p_list = None
+            gpts['mlp.down_proj'].batch_inp.clear()
+            # s: output of 'mlp.gate_proj'
+            hidden_s_list = gpts['mlp.gate_proj'].batch_out
+            s = torch.stack(hidden_s_list, dim=0)
+            hidden_s_list = None
+            gpts['mlp.gate_proj'].batch_out.clear()
+
+            # Reshape auxiliary variables
+            z = z.reshape((-1, z.size(-1))).T.to(dev)
+            p = p.reshape((-1, p.size(-1))).T.to(dev)
+            s = s.reshape((-1, s.size(-1))).T.to(dev)
+
+            torch.cuda.empty_cache()
+
+            # Pre-compute the pinverse of X and cache it to save computational cost
+            Xinv = torch.pinverse(X.to(dtype=torch.float32)).half()
+
+            # list to store training losses
+            training_loss = {'Y_p_loss': [], 'p_z_loss': [], 'z_X_loss': [], 'train_loss': []}
+
+            for opt_step in range(opt_epochs):
+
+                ##############
+                # optimize W
+                ##############
+
+                if opt_step > 0:   # for the first step, no need for updating W
+
+                    # Update the weight matrix of mlp.up_project
+                    # Calculate the weight matrix
+                    weight_matrix_1 = torch.matmul(z, Xinv)
+                    # assign the new parameters to gpts class
+                    gpts['mlp.up_proj'].layer.weight.copy_(weight_matrix_1)
+                    del weight_matrix_1
+
+                    # Update the weight matrix of mlp.down_proj
+                    pinv = torch.pinverse(p.to(dtype=torch.float32)).half()
+                    # Calculate the weight matrix
+                    weight_matrix_2 = torch.matmul(Y, pinv)
+                    # assign the new parameters to gpts class
+                    gpts['mlp.down_proj'].layer.weight.copy_(weight_matrix_2)
+                    del weight_matrix_2, pinv
+
+                    # Update the weight matrix of mlp.gate_project
+                    # Calculate the weight matrix
+                    weight_matrix_3 = torch.matmul(s, Xinv)
+                    # assign the new parameters to gpts class
+                    gpts['mlp.gate_proj'].layer.weight.copy_(weight_matrix_3)
+                    del weight_matrix_3
+
+                    torch.cuda.empty_cache()
+
+                ##############
+                # prune W
+                ##############
+
+                # modify gpts[name].H to be our auxiliary variable
+                if opt_step > 0:   # for the first step, no need for updating H    
+                
+                    tmp_H = torch.zeros_like(gpts['mlp.down_proj'].H)
+                    tmp_p = p.T.reshape((args.nsamples, -1, p.size(0)))
+                    tmp_nsamples = 0
+                    for j in range(args.nsamples):
+                        tmp_inp = tmp_p[j].unsqueeze(0)
+                        tmp = tmp_inp.shape[0]
+                        if isinstance(gpts['mlp.down_proj'].layer, nn.Linear) or isinstance(gpts['mlp.down_proj'].layer, transformers.Conv1D):
+                            if len(tmp_inp.shape) == 3:
+                                tmp_inp = tmp_inp.reshape((-1, tmp_inp.shape[-1]))
+                            tmp_inp = tmp_inp.t()
+                        tmp_H *= tmp_nsamples / (tmp_nsamples + tmp)
+                        tmp_nsamples += tmp
+                        tmp_inp = math.sqrt(2 / tmp_nsamples) * tmp_inp.float()
+                        tmp_H += tmp_inp.matmul(tmp_inp.t())
+                    gpts['mlp.down_proj'].H.copy_(tmp_H)
+                    del tmp_H, tmp_p
+                    torch.cuda.empty_cache()
+
+                for name in target_layer_names:
+                    print(i, name)
+                    print('Pruning ...')
+                    sparsity = args.sparsity
+                    gpts[name].fasterprune(
+                        sparsity,
+                        prunen=args.prunen,
+                        prunem=args.prunem,
+                        percdamp=args.percdamp,
+                        blocksize=args.blocksize,
+                    )
+
+                ##############
+                # optimize p
+                ##############
+
+                # Activation inverse
+                next_weight = subset['mlp.down_proj'].weight
+                m1 = beta * torch.matmul(next_weight.T, next_weight)
+                m2 = gamma * torch.eye(m1.shape[0], device=m1.device)
+                av = torch.inverse(m1 + m2).to(dtype=torch.float16)
+
+                del m1, m2
+                torch.cuda.empty_cache()
+
+                # Calculate SwiGLU output
+                layer_nl_output = nn.functional.silu(s) * z
+
+                # Activation formulate
+                m3 = beta * torch.matmul(next_weight.T, Y)
+                m4 = gamma * layer_nl_output
+                af = m3 + m4
+
+                p = torch.matmul(av, af)
+
+                del layer_nl_output, next_weight, av, m3, m4, af
+                torch.cuda.empty_cache()
+
+                ##############
+                # optimize z
+                ##############
+
+                w = subset['mlp.up_proj'].weight
+                m = torch.matmul(w, X)
+                swish = nn.functional.silu(s)
+                z = (m + swish * p) / (swish ** 2 + 1)    
+
+                del w, m, swish
+                torch.cuda.empty_cache()
+
+                ##############
+                # optimize s
+                ##############
+
+                w = subset['mlp.gate_proj'].weight
+                # convert the layer's weight tensor to float32 and enable grad
+                w = w.to(dtype=torch.float32).requires_grad_(True)
+
+                s_update_epochs = 2
+                s_learning_rate = 0.01
+                for _ in range(s_update_epochs):
+
+                    batch_size = 1000  # Choose an appropriate batch size based on your memory constraints
+                    # s: [hidden_d, n_samples]
+                    for k in range(0, s.size(-1), batch_size):
+                        chunk = slice(k, k + batch_size)
+
+                        # get the "mini-batch" for each tensor and turn on autograd
+                        X_batch = X[:,chunk].to(dtype=torch.float32).requires_grad_(True)
+                        z_batch = z[:,chunk].to(dtype=torch.float32).requires_grad_(True)
+                        p_batch = p[:,chunk].to(dtype=torch.float32).requires_grad_(True)
+                        s_batch = s[:,chunk].to(dtype=torch.float32).requires_grad_(True)
+
+                        with torch.enable_grad():   # temporarily turn on the Pytorch computational graph functionality
+
+                            loss_s = alpha * torch.norm(s_batch - torch.matmul(w, X_batch))**2
+                            loss_s += gamma * torch.norm(p_batch - nn.functional.silu(s_batch) * z_batch)**2
+
+                        loss_s.backward()
+                        s_batch -= s_learning_rate * s_batch.grad                    
+                        s_batch.grad.zero_()
+                        s[:,chunk] = s_batch.detach().to(dtype=torch.float16)
+
+                s_batch, X_batch, z_batch, p_batch, w = s_batch.detach(), X_batch.detach(), z_batch.detach(), p_batch.detach(), w.detach()
+                del w, loss_s, s_batch, X_batch, z_batch, p_batch
+                torch.cuda.empty_cache()
+
+                # compute and save the training loss after each epoch
+                tmp_training_loss = nn.functional.mse_loss(torch.matmul(subset['mlp.down_proj'].weight, 
+                                                                        nn.functional.silu(torch.matmul(subset['mlp.gate_proj'].weight, X)) 
+                                                                        * torch.matmul(subset['mlp.up_proj'].weight, X)), Y)
+                training_loss['train_loss'].append(tmp_training_loss.item())
+
+        for j in range(args.nsamples):
+            outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask)[0]
+
+        layers[i] = layer.cpu()
+        del layer
+        del gpts
+        torch.cuda.empty_cache()
+
+        inps, outs = outs, inps
+
+    model.config.use_cache = use_cache
+
+
+@torch.no_grad()
+def opt_eval(model, testenc, dev, args, dataset: str):
+    print('Evaluating ...')
+
+    testenc = testenc.input_ids
+    nsamples = testenc.numel() // model.seqlen
+
+    use_cache = model.config.use_cache
+    model.config.use_cache = False
+    layers = model.model.decoder.layers
+
+    model.model.decoder.embed_tokens = model.model.decoder.embed_tokens.to(dev)
+    model.model.decoder.embed_positions = model.model.decoder.embed_positions.to(dev)
+    if hasattr(model.model.decoder, 'project_out') and model.model.decoder.project_out:
+        model.model.decoder.project_out = model.model.decoder.project_out.to(dev) 
+    if hasattr(model.model.decoder, 'project_in') and model.model.decoder.project_in:
+        model.model.decoder.project_in = model.model.decoder.project_in.to(dev) 
+    layers[0] = layers[0].to(dev)
+
+    dtype = next(iter(model.parameters())).dtype
+    inps = torch.zeros(
+        (nsamples, model.seqlen, model.config.hidden_size), dtype=dtype, device=dev
+    )
+    cache = {'i': 0, 'attention_mask': None}
+
+    class Catcher(nn.Module):
+        def __init__(self, module):
+            super().__init__()
+            self.module = module
+        def forward(self, inp, **kwargs):
+            inps[cache['i']] = inp
+            cache['i'] += 1
+            cache['attention_mask'] = kwargs['attention_mask']
+            raise ValueError
+    layers[0] = Catcher(layers[0])
+    for i in range(nsamples):
+        batch = testenc[:, (i * model.seqlen):((i + 1) * model.seqlen)].to(dev)
+        try:
+            model(batch)
+        except ValueError:
+            pass
+    layers[0] = layers[0].module
+
+    layers[0] = layers[0].cpu()
+    model.model.decoder.embed_tokens = model.model.decoder.embed_tokens.cpu()
+    model.model.decoder.embed_positions = model.model.decoder.embed_positions.cpu()
+    if hasattr(model.model.decoder, 'project_out') and model.model.decoder.project_out:
+        model.model.decoder.project_out = model.model.decoder.project_out.cpu()
+    if hasattr(model.model.decoder, 'project_in') and model.model.decoder.project_in:
+        model.model.decoder.project_in = model.model.decoder.project_in.cpu()
+    torch.cuda.empty_cache()
+
+    outs = torch.zeros_like(inps)
+    attention_mask = cache['attention_mask']
+
+    for i in range(len(layers)):
+        print(i)
+        layer = layers[i].to(dev)
+
+        if args.gmp:
+            subset = find_layers(layer)
+            for name in subset:
+                W = subset[name].weight.data
+                thresh = torch.sort(torch.abs(W.flatten()))[0][int(W.numel() * args.sparsity)]
+                W.data[torch.abs(W.data) <= thresh] = 0
+
+        for j in range(nsamples):
+            outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask)[0]
+        layers[i] = layer.cpu()
+        del layer
+        torch.cuda.empty_cache()
+        inps, outs = outs, inps
+
+    if model.model.decoder.final_layer_norm is not None:
+        model.model.decoder.final_layer_norm = model.model.decoder.final_layer_norm.to(dev)
+    if model.model.decoder.project_out is not None:
+        model.model.decoder.project_out = model.model.decoder.project_out.to(dev)
+    model.lm_head = model.lm_head.to(dev)
+
+    testenc = testenc.to(dev)
+    nlls = []
+    for i in range(nsamples):
+        hidden_states = inps[i].unsqueeze(0)
+        if model.model.decoder.final_layer_norm is not None:
+            hidden_states = model.model.decoder.final_layer_norm(hidden_states)
+        if model.model.decoder.project_out is not None:
+            hidden_states = model.model.decoder.project_out(hidden_states)
+        lm_logits = model.lm_head(hidden_states)
+        shift_logits = lm_logits[:, :-1, :].contiguous()
+        shift_labels = testenc[
+            :, (i * model.seqlen):((i + 1) * model.seqlen)
+        ][:, 1:]
+        loss_fct = nn.CrossEntropyLoss()
+        loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+        neg_log_likelihood = loss.float() * model.seqlen
+        nlls.append(neg_log_likelihood)
+    ppl = torch.exp(torch.stack(nlls).sum() / (nsamples * model.seqlen))
+    print(f"Perplexity: {ppl.item():3f}")
+
+    model.config.use_cache = use_cache
+
+
+
+@torch.no_grad()
+def llama_eval(model, testenc, dev, args, dataset: str):
+    print("Evaluating ...")
+
+    testenc = testenc.input_ids
+    nsamples = testenc.numel() // model.seqlen
+
+    use_cache = model.config.use_cache
+    model.config.use_cache = False
+    layers = model.model.layers
+
+    model.model.embed_tokens = model.model.embed_tokens.to(dev)
+    layers[0] = layers[0].to(dev)
+
+    dtype = next(iter(model.parameters())).dtype
+    inps = torch.zeros(
+        (nsamples, model.seqlen, model.config.hidden_size), dtype=dtype, device=dev
+    )
+    cache = {"i": 0, "attention_mask": None}
+
+    class Catcher(nn.Module):
+        def __init__(self, module):
+            super().__init__()
+            self.module = module
+
+        def forward(self, inp, **kwargs):
+            inps[cache["i"]] = inp
+            cache["i"] += 1
+            cache["attention_mask"] = kwargs["attention_mask"]
+            raise ValueError
+
+    layers[0] = Catcher(layers[0])
+    for i in range(nsamples):
+        batch = testenc[:, (i * model.seqlen) : ((i + 1) * model.seqlen)].to(dev)
+        try:
+            model(batch)
+        except ValueError:
+            pass
+    layers[0] = layers[0].module
+
+    layers[0] = layers[0].cpu()
+    model.model.embed_tokens = model.model.embed_tokens.cpu()
+    torch.cuda.empty_cache()
+
+    outs = torch.zeros_like(inps)
+    attention_mask = cache["attention_mask"]
+
+    for i in range(len(layers)):
+        print(i)
+        layer = layers[i].to(dev)
+
+        if args.gmp:
+            subset = find_layers(layer)
+            for name in subset:
+                W = subset[name].weight.data
+                thresh = torch.sort(torch.abs(W.flatten()))[0][
+                    int(W.numel() * args.sparsity)
+                ]
+                W.data[torch.abs(W.data) <= thresh] = 0
+
+        for j in range(nsamples):
+            outs[j] = layer(inps[j].unsqueeze(0), attention_mask=attention_mask)[0]
+        layers[i] = layer.cpu()
+        del layer
+        torch.cuda.empty_cache()
+        inps, outs = outs, inps
+
+    if model.model.norm is not None:
+        model.model.norm = model.model.norm.to(dev)
+    model.lm_head = model.lm_head.to(dev)
+
+    testenc = testenc.to(dev)
+    nlls = []
+    for i in range(nsamples):
+        hidden_states = inps[i].unsqueeze(0)
+        if model.model.norm is not None:
+            hidden_states = model.model.norm(hidden_states)
+        lm_logits = model.lm_head(hidden_states)
+        shift_logits = lm_logits[:, :-1, :].contiguous()
+        shift_labels = testenc[:, (i * model.seqlen) : ((i + 1) * model.seqlen)][:, 1:]
+        loss_fct = nn.CrossEntropyLoss()
+        loss = loss_fct(
+            shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
+        )
+        neg_log_likelihood = loss.float() * model.seqlen
+        nlls.append(neg_log_likelihood)
+    ppl = torch.exp(torch.stack(nlls).sum() / (nsamples * model.seqlen))
+    print(f"Perplexity: {ppl.item():3f}")
+
+    model.config.use_cache = use_cache

opt_main.py

@@ -0,0 +1,45 @@
+import argparse
+from model_utils import get_opt, opt_sparsellm, opt_eval
+from datautils import get_loaders
+import torch
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    # Arguments parsing
+    parser.add_argument('--model', type=str, default='facebook/opt-125m', help='OPT model to load; pass `facebook/opt-X`.')
+    parser.add_argument('--dataset', type=str, choices=['wikitext2', 'ptb', 'c4'], default='c4', help='Where to extract calibration data from.')
+    parser.add_argument('--seed', type=int, default=0, help='Seed for sampling the calibration data.')
+    parser.add_argument('--nsamples', type=int, default=64, help='Number of calibration data samples.')
+    parser.add_argument('--percdamp', type=float, default=.01, help='Percent of the average Hessian diagonal to use for dampening.')
+    parser.add_argument('--sparsity', type=float, default=0.7, help='Target sparsity')
+    parser.add_argument('--prunen', type=int, default=0, help='N for N:M pruning.')
+    parser.add_argument('--prunem', type=int, default=0, help='M for N:M pruning.')
+    parser.add_argument('--blocksize', type=int, default=128, help='Blocksize to use for adaptive mask selection.')
+    parser.add_argument('--gmp', action='store_true', help='Whether to run the GMP baseline.')
+    parser.add_argument('--wbits', type=int, default=16, help='Whether to quantize as well.')
+    parser.add_argument('--minlayer', type=int, default=-1, help='Prune all layers with id >= this.')
+    parser.add_argument('--maxlayer', type=int, default=1000, help='Prune all layers with id < this.')
+    parser.add_argument('--prune_only', type=str, default='', help='Prune only layers that contain this text.')
+    parser.add_argument('--invert', action='store_true', help='Invert subset.')
+    parser.add_argument('--save', type=str, default='', help='Path to saved model.')
+    parser.add_argument('--log_wandb', action='store_true', help='Whether to log to wandb.')
+
+    args = parser.parse_args()
+
+    model = get_opt(args)
+    model.eval()
+    dataloader, testloader = get_loaders(args.dataset, nsamples=args.nsamples, seed=args.seed, model=args.model, seqlen=model.seqlen)
+
+    if (args.sparsity or args.prunen) and not args.gmp:
+        opt_sparsellm(model, dataloader, torch.device('cuda'), args)
+
+    for dataset in ['wikitext2', 'ptb', 'c4']:
+        dataloader, testloader = get_loaders(dataset, seed=args.seed, model=args.model, seqlen=model.seqlen)
+        opt_eval(model, testloader, torch.device('cuda'), args, dataset)
+
+    if args.save:
+        model.save_pretrained(args.save)
+
+if __name__ == '__main__':
+    main()

pruning_utils.py

@@ -0,0 +1,324 @@
+# This file will contain helper functions related to the pruning process, including any specialized pruning functions and the SparseGPT functionality.
+# DISCLAIMER: The SparseGPT class is a modified version of the original SparseGPT class. The original SparseGPT class can be found in [SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot].
+
+import math
+import time
+
+import torch
+import torch.nn as nn
+import transformers
+
+from quant import *
+
+# turned this flag to be True
+DEBUG = True
+
+torch.backends.cuda.matmul.allow_tf32 = False
+torch.backends.cudnn.allow_tf32 = False
+
+def find_layers(module, layers=[nn.Conv2d, nn.Linear], name=''):
+    if type(module) in layers:
+        return {name: module}
+    res = {}
+    for name1, child in module.named_children():
+        res.update(find_layers(
+            child, layers=layers, name=name + '.' + name1 if name != '' else name1
+        ))
+    return res
+
+class SparseGPT_OPT:
+
+    def __init__(self, layer):
+        self.layer = layer
+        self.dev = self.layer.weight.device
+        W = layer.weight.data.clone()
+        if isinstance(self.layer, nn.Conv2d):
+            W = W.flatten(1)
+        if isinstance(self.layer, transformers.Conv1D):
+            W = W.t()
+        self.rows = W.shape[0]
+        self.columns = W.shape[1]
+        self.H = torch.zeros((self.columns, self.columns), device=self.dev)
+        self.nsamples = 0
+        self.batch_inp = []
+        self.batch_out = []
+
+    def add_batch(self, inp, out, name, blocksize=1024):
+        if DEBUG:
+            self.inp1 = inp
+            self.out1 = out
+        if len(inp.shape) == 2:
+            inp = inp.unsqueeze(0)
+        ###### added code
+        if name == 'fc1' or name == 'fc2':
+            self.batch_inp.append(inp[0].clone().detach())
+            if len(out.shape) == 3:
+                out = out.squeeze(0)
+            self.batch_out.append(out.clone().detach())
+        ######
+        tmp = inp.shape[0]
+        if isinstance(self.layer, nn.Linear) or isinstance(self.layer, transformers.Conv1D):
+            if len(inp.shape) == 3:
+                inp = inp.reshape((-1, inp.shape[-1]))
+            inp = inp.t()
+        self.H *= self.nsamples / (self.nsamples + tmp)
+        self.nsamples += tmp
+        inp = math.sqrt(2 / self.nsamples) * inp.float()
+        self.H += inp.matmul(inp.t())
+
+    def fasterprune(
+        self, sparsity, prunen=0, prunem=0, blocksize=128, percdamp=.01
+    ):
+        W = self.layer.weight.data.clone()
+        if isinstance(self.layer, nn.Conv2d):
+            W = W.flatten(1)
+        if isinstance(self.layer, transformers.Conv1D):
+            W = W.t()
+        W = W.float()
+
+        if hasattr(self, 'quantizer'):
+            if not self.quantizer.ready():
+                self.quantizer.find_params(W, weight=True)
+
+        tick = time.time()
+
+        H = self.H
+        # del self.H 
+        dead = torch.diag(H) == 0
+        H[dead, dead] = 1
+        W[:, dead] = 0
+
+        Losses = torch.zeros(self.rows, device=self.dev)
+
+        damp = percdamp * torch.mean(torch.diag(H))
+        diag = torch.arange(self.columns, device=self.dev)
+        H[diag, diag] += damp
+        H = torch.linalg.cholesky(H)
+        H = torch.cholesky_inverse(H)
+        H = torch.linalg.cholesky(H, upper=True)
+        Hinv = H
+
+        mask = None
+
+        for i1 in range(0, self.columns, blocksize):
+            i2 = min(i1 + blocksize, self.columns)
+            count = i2 - i1
+
+            W1 = W[:, i1:i2].clone()
+            Q1 = torch.zeros_like(W1)
+            Err1 = torch.zeros_like(W1)
+            Losses1 = torch.zeros_like(W1)
+            Hinv1 = Hinv[i1:i2, i1:i2]
+
+            if prunen == 0: 
+                if mask is not None:
+                    mask1 = mask[:, i1:i2]
+                else:
+                    tmp = W1 ** 2 / (torch.diag(Hinv1).reshape((1, -1))) ** 2
+                    thresh = torch.sort(tmp.flatten())[0][int(tmp.numel() * sparsity)]
+                    mask1 = tmp <= thresh
+            else:
+                mask1 = torch.zeros_like(W1) == 1
+
+            for i in range(count):
+                w = W1[:, i]
+                d = Hinv1[i, i]
+
+                if prunen != 0 and i % prunem == 0:
+                    tmp = W1[:, i:(i + prunem)] ** 2 / (torch.diag(Hinv1)[i:(i + prunem)].reshape((1, -1))) ** 2
+                    mask1.scatter_(1, i + torch.topk(tmp, prunen, dim=1, largest=False)[1], True)
+
+                q = w.clone()
+                q[mask1[:, i]] = 0
+
+                if hasattr(self, 'quantizer'):
+                    q = quantize(
+                        q.unsqueeze(1), self.quantizer.scale, self.quantizer.zero, self.quantizer.maxq
+                    ).flatten()
+
+                Q1[:, i] = q
+                Losses1[:, i] = (w - q) ** 2 / d ** 2
+
+                err1 = (w - q) / d
+                W1[:, i:] -= err1.unsqueeze(1).matmul(Hinv1[i, i:].unsqueeze(0))
+                Err1[:, i] = err1
+
+            W[:, i1:i2] = Q1
+            Losses += torch.sum(Losses1, 1) / 2
+
+            W[:, i2:] -= Err1.matmul(Hinv[i1:i2, i2:])
+
+            # if DEBUG:
+            #     self.layer.weight.data[:, :i2] = W[:, :i2]
+            #     self.layer.weight.data[:, i2:] = W[:, i2:]
+            #     print(torch.sum((self.layer(self.inp1) - self.out1) ** 2))
+            #     print(torch.sum(Losses))
+
+        torch.cuda.synchronize()
+        print('time %.2f' % (time.time() - tick))
+        print('error', torch.sum(Losses).item())
+
+        if isinstance(self.layer, transformers.Conv1D):
+            W = W.t()
+        self.layer.weight.data = W.reshape(self.layer.weight.shape).to(self.layer.weight.data.dtype)
+        # if DEBUG:
+            # print(torch.sum((self.layer(self.inp1) - self.out1) ** 2))
+
+    def free(self):
+        if DEBUG:
+            self.inp1 = None
+            self.out1 = None
+        self.H = None
+        torch.cuda.empty_cache()
+
+
+
+
+class SparseGPT_LlaMA:
+
+    def __init__(self, layer):
+        self.layer = layer
+        self.dev = self.layer.weight.device
+        W = layer.weight.data.clone()
+        if isinstance(self.layer, nn.Conv2d):
+            W = W.flatten(1)
+        if isinstance(self.layer, transformers.Conv1D):
+            W = W.t()
+        self.rows = W.shape[0]
+        self.columns = W.shape[1]
+        self.H = torch.zeros((self.columns, self.columns), device=self.dev)
+        self.nsamples = 0
+        self.batch_inp = []
+        self.batch_out = []
+
+    def add_batch(self, inp, out, name, blocksize=1024):
+        if DEBUG:
+            self.inp1 = inp
+            self.out1 = out
+        if len(inp.shape) == 2:
+            inp = inp.unsqueeze(0)
+        ###### added code
+        if name == 'mlp.up_proj' or name == 'mlp.down_proj':
+            self.batch_inp.append(inp[0].clone().detach())
+            if len(out.shape) == 3:
+                out = out.squeeze(0)
+            self.batch_out.append(out.clone().detach())
+        if name == 'mlp.gate_proj':   # for this layer, we only store the outputs. for inputs, they are shared with 'mlp.up_proj'
+            if len(out.shape) == 3:
+                out = out.squeeze(0)
+            self.batch_out.append(out.clone().detach())
+        ######
+        tmp = inp.shape[0]
+        if isinstance(self.layer, nn.Linear) or isinstance(self.layer, transformers.Conv1D):
+            if len(inp.shape) == 3:
+                inp = inp.reshape((-1, inp.shape[-1]))
+            inp = inp.t()
+        self.H *= self.nsamples / (self.nsamples + tmp)
+        self.nsamples += tmp
+        inp = math.sqrt(2 / self.nsamples) * inp.float()
+        self.H += inp.matmul(inp.t())
+
+    def fasterprune(
+        self, sparsity, prunen=0, prunem=0, blocksize=128, percdamp=.01
+    ):
+        W = self.layer.weight.data.clone()
+        if isinstance(self.layer, nn.Conv2d):
+            W = W.flatten(1)
+        if isinstance(self.layer, transformers.Conv1D):
+            W = W.t()
+        W = W.float()
+
+        if hasattr(self, 'quantizer'):
+            if not self.quantizer.ready():
+                self.quantizer.find_params(W, weight=True)
+
+        tick = time.time()
+
+        H = self.H
+        # del self.H 
+        dead = torch.diag(H) == 0
+        H[dead, dead] = 1
+        W[:, dead] = 0
+
+        Losses = torch.zeros(self.rows, device=self.dev)
+
+        damp = percdamp * torch.mean(torch.diag(H))
+        diag = torch.arange(self.columns, device=self.dev)
+        H[diag, diag] += damp
+        H = torch.linalg.cholesky(H)
+        H = torch.cholesky_inverse(H)
+        H = torch.linalg.cholesky(H, upper=True)
+        Hinv = H
+
+        mask = None
+
+        for i1 in range(0, self.columns, blocksize):
+            i2 = min(i1 + blocksize, self.columns)
+            count = i2 - i1
+
+            W1 = W[:, i1:i2].clone()
+            Q1 = torch.zeros_like(W1)
+            Err1 = torch.zeros_like(W1)
+            Losses1 = torch.zeros_like(W1)
+            Hinv1 = Hinv[i1:i2, i1:i2]
+
+            if prunen == 0: 
+                if mask is not None:
+                    mask1 = mask[:, i1:i2]
+                else:
+                    tmp = W1 ** 2 / (torch.diag(Hinv1).reshape((1, -1))) ** 2
+                    thresh = torch.sort(tmp.flatten())[0][int(tmp.numel() * sparsity)]
+                    mask1 = tmp <= thresh
+            else:
+                mask1 = torch.zeros_like(W1) == 1
+
+            for i in range(count):
+                w = W1[:, i]
+                d = Hinv1[i, i]
+
+                if prunen != 0 and i % prunem == 0:
+                    tmp = W1[:, i:(i + prunem)] ** 2 / (torch.diag(Hinv1)[i:(i + prunem)].reshape((1, -1))) ** 2
+                    mask1.scatter_(1, i + torch.topk(tmp, prunen, dim=1, largest=False)[1], True)
+
+                q = w.clone()
+                q[mask1[:, i]] = 0
+
+                if hasattr(self, 'quantizer'):
+                    q = quantize(
+                        q.unsqueeze(1), self.quantizer.scale, self.quantizer.zero, self.quantizer.maxq
+                    ).flatten()
+
+                Q1[:, i] = q
+                Losses1[:, i] = (w - q) ** 2 / d ** 2
+
+                err1 = (w - q) / d
+                W1[:, i:] -= err1.unsqueeze(1).matmul(Hinv1[i, i:].unsqueeze(0))
+                Err1[:, i] = err1
+
+            W[:, i1:i2] = Q1
+            Losses += torch.sum(Losses1, 1) / 2
+
+            W[:, i2:] -= Err1.matmul(Hinv[i1:i2, i2:])
+
+            # if DEBUG:
+            #     self.layer.weight.data[:, :i2] = W[:, :i2]
+            #     self.layer.weight.data[:, i2:] = W[:, i2:]
+            #     print(torch.sum((self.layer(self.inp1) - self.out1) ** 2))
+            #     print(torch.sum(Losses))
+
+        torch.cuda.synchronize()
+        print('time %.2f' % (time.time() - tick))
+        print('error', torch.sum(Losses).item())
+
+        if isinstance(self.layer, transformers.Conv1D):
+            W = W.t()
+        self.layer.weight.data = W.reshape(self.layer.weight.shape).to(self.layer.weight.data.dtype)
+        # if DEBUG:
+            # print(torch.sum((self.layer(self.inp1) - self.out1) ** 2))
+
+    def free(self):
+        if DEBUG:
+            self.inp1 = None
+            self.out1 = None
+        self.H = None
+        torch.cuda.empty_cache()

quant.py

@@ -0,0 +1,126 @@
+import numpy as np
+import torch
+import torch.nn as nn
+
+
+def quantize(x, scale, zero, maxq):
+    q = torch.clamp(torch.round(x / scale) + zero, 0, maxq)
+    return scale * (q - zero)
+
+class Quantizer(nn.Module):
+
+    def __init__(self, shape=1):
+        super(Quantizer, self).__init__()
+        self.register_buffer('maxq', torch.tensor(0))
+        self.register_buffer('scale', torch.zeros(shape))
+        self.register_buffer('zero', torch.zeros(shape))
+
+    def configure(
+            self,
+            bits, perchannel=False, sym=True, 
+            mse=False, norm=2.4, grid=100, maxshrink=.8,
+            grouprows=1
+        ):
+        self.maxq = torch.tensor(2 ** bits - 1)
+        self.perchannel = perchannel
+        self.sym = sym
+        self.mse = mse
+        self.norm = norm
+        self.grid = grid
+        self.maxshrink = maxshrink 
+        self.grouprows = grouprows
+
+    def find_params(self, x, weight=False):
+        dev = x.device
+        self.maxq = self.maxq.to(dev)
+
+        shape = x.shape
+        if self.perchannel:
+            if weight:
+                x = x.flatten(1)
+                if self.grouprows > 1: 
+                    x = x.reshape((x.shape[0] // self.grouprows, -1))
+            else:
+                if len(shape) == 4:
+                    x = x.permute([1, 0, 2, 3])
+                    x = x.flatten(1)
+                if len(shape) == 3:
+                    x = x.reshape((-1, shape[-1])).t()
+                if len(shape) == 2:
+                    x = x.t()
+        else:
+            x = x.flatten().unsqueeze(0)
+
+        tmp = torch.zeros(x.shape[0], device=dev)
+        xmin = torch.minimum(x.min(1)[0], tmp)
+        xmax = torch.maximum(x.max(1)[0], tmp)
+
+        if self.sym:
+            xmax = torch.maximum(torch.abs(xmin), xmax)
+            tmp = xmin < 0
+            if torch.any(tmp):
+                xmin[tmp] = -xmax[tmp]
+        tmp = (xmin == 0) & (xmax == 0)
+        xmin[tmp] = -1
+        xmax[tmp] = +1
+
+        self.scale = (xmax - xmin) / self.maxq
+        if self.sym:
+            self.zero = torch.full_like(self.scale, (self.maxq + 1) / 2)
+        else:
+            self.zero = torch.round(-xmin / self.scale)
+
+        if self.mse:
+            best = torch.full([x.shape[0]], float('inf'), device=dev)
+            for i in range(int(self.maxshrink * self.grid)):
+                p = 1 - i / self.grid 
+                xmin1 = p * xmin
+                xmax1 = p * xmax
+                scale1 = (xmax1 - xmin1) / self.maxq
+                zero1 = torch.round(-xmin1 / scale1) if not self.sym else self.zero
+                q = quantize(x, scale1.unsqueeze(1), zero1.unsqueeze(1), self.maxq)
+                q -= x
+                q.abs_()
+                q.pow_(self.norm)
+                err = torch.sum(q, 1)
+                tmp = err < best
+                if torch.any(tmp):
+                    best[tmp] = err[tmp]
+                    self.scale[tmp] = scale1[tmp]
+                    self.zero[tmp] = zero1[tmp]
+        if not self.perchannel:
+            if weight:
+                tmp = shape[0]
+            else:
+                tmp = shape[1] if len(shape) != 3 else shape[2]
+            self.scale = self.scale.repeat(tmp)
+            self.zero = self.zero.repeat(tmp)
+
+        if weight:
+            if self.grouprows > 1:
+                self.scale = self.scale.unsqueeze(1).repeat(1, self.grouprows)
+                self.zero = self.zero.unsqueeze(1).repeat(1, self.grouprows)
+            shape = [-1] + [1] * (len(shape) - 1)
+            self.scale = self.scale.reshape(shape)
+            self.zero = self.zero.reshape(shape)
+            return
+        if len(shape) == 4:
+            self.scale = self.scale.reshape((1, -1, 1, 1))
+            self.zero = self.zero.reshape((1, -1, 1, 1))
+        if len(shape) == 3:
+            self.scale = self.scale.reshape((1, 1, -1))
+            self.zero = self.zero.reshape((1, 1, -1)) 
+        if len(shape) == 2:
+            self.scale = self.scale.unsqueeze(0)
+            self.zero = self.zero.unsqueeze(0)
+
+    def quantize(self, x):
+        if self.ready():
+            return quantize(x, self.scale, self.zero, self.maxq)
+        return x
+
+    def enabled(self):
+        return self.maxq > 0
+
+    def ready(self):
+        return torch.all(self.scale != 0)