add copyright header

arguments.py

@@ -1,11 +1,72 @@
+
+# QRotaryTraining - A novel method for fully training all parameters of large
+# language models (llms) while using less device memory than traditional methods.
+# Copyright (C) 2024 Carl Philipp Klemm
+#
+# This file is part of QRotaryTraining.
+#
+# QRotaryTraining is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# QRotaryTraining is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with QRotaryTraining.  If not, see <http://www.gnu.org/licenses/>.
+
 from dataclasses import dataclass, field
-from typing import Optional
+from typing import Optional, Self
+from enum import Enum
+
+
+class DatasetType(Enum):
+    TEXT = 1
+    S2S = 2
+    HUB = 3
+    CHAT = 4
+
+    @staticmethod
+    def to_string(dtype: Self) -> str:
+        if dtype == DatasetType.TEXT:
+            return "text"
+        elif dtype == DatasetType.S2S:
+            return "s2s"
+        elif dtype == DatasetType.HUB:
+            return "hub"
+        elif dtype == DatasetType.CHAT:
+            return "chat"
+        return "invalid"
+
+    @staticmethod
+    def from_string(string: str):
+        if string == str(DatasetType.TEXT):
+            return DatasetType.TEXT
+        elif string == str(DatasetType.S2S):
+            return DatasetType.S2S
+        elif string == str(DatasetType.HUB):
+            return DatasetType.HUB
+        elif string == str(DatasetType.CHAT):
+            return DatasetType.CHAT
+        return None
+
+    def __str__(self):
+        return DatasetType.to_string(self)
 
 
 @dataclass
 class DataArguments:
     dataset: str = field(
-        metadata={"help": "A json file (s2s) or text file with the dataset to train on"}
+        metadata={"help": "The dataset to train on"}
+    )
+    dataset_type: str = field(
+        default="text", metadata={"help": f"The type of dataset, set to one of {[e for e in DatasetType]}"}
+    )
+    dataset_chat_template: str | None = field(
+        default=None, metadata={"help": "overrides the chat template to be the one set here"}
     )
     eval_dataset_size: int = field(
         default=512, metadata={"help": "Size of validation dataset."}
@@ -26,10 +87,6 @@ class DataArguments:
         default=False,
         metadata={"help": "If this is set the dataset is assumed to be a name of a hf-hub dataset"}
     )
-    block_size: int = field(
-        default=512,
-        metadata={"help": "size of the blocks the text is split into for training"},
-    )
 
 
 @dataclass
@@ -65,8 +122,9 @@ class TrainingArguments():
     )
     resume: bool = field(default=False, metadata={"help": 'Resume from previous checkpoint'})
     ddp_find_unused_parameters: bool = field(default=True, metadata={"help": 'set if trainer should try to find unused parameters'})
-    output_dir: str = field(default='./output', metadata={"help": 'The output dir for logs and checkpoints'})
+    output_dir: str = field(default='./output', metadata={"help": 'The output dir for checkpoints'})
     per_device_train_batch_size: int = field(default=1, metadata={"help": 'The training batch size per GPU. Increase for better speed.'})
+    per_device_eval_batch_size: int = field(default=1, metadata={"help": 'The eval batch size per GPU. Increase for better speed.'})
     gradient_accumulation_steps: int = field(default=16, metadata={"help": 'How many gradients to accumulate before to perform an optimizer step'})
     epochs: int = field(default=3, metadata={"help": 'How many epochs to train for'})
     weight_decay: float = field(default=0.0, metadata={"help": 'The L2 weight decay rate of AdamW'})
@@ -82,6 +140,7 @@ class TrainingArguments():
                                    metadata={"help": 'Learning rate schedule. Constant a bit better than cosine, and has advantage for analysis'})
     warmup_steps: float = field(default=0, metadata={"help": 'number of steps to do a warmup for'})
     logging_steps: int = field(default=10, metadata={"help": 'The frequency of update steps after which to log the loss'})
+    logging_dir: str = field(default='./log', metadata={"help": 'The output dir for logs'})
     group_by_length: bool = field(default=False,
                                   metadata={"help": 'Group sequences into batches with same length. Saves memory and speeds up training considerably.'})
     save_steps: int = field(default=250, metadata={"help": 'How often to save a model'})
@@ -92,5 +151,5 @@ class TrainingArguments():
     max_instant_params: int = field(default=0, metadata={"help": "Maximum amount of paramters to optimize per step in millions"})
     churn_percent: int = field(default=100, metadata={"help": "The percentage of active parameters to replace when changeing active parameters"})
     eval_steps: int = field(default=-1, metadata={"help": "Number of optimization steps after wich to compute the evaluation loss"})
-    eval_prompt: str = field(default=None, metadata={"help": "A prompt to used during eval to check if the model is learning"})
+    eval_prompt: str | None = field(default=None, metadata={"help": "A prompt to used during eval to check if the model is learning"})
     reshufle_steps: int = field(default=50, metadata={"help": "Number of steps to take before changing the active parameters"})

datamodules.py

@@ -1,27 +1,49 @@
+
+# QRotaryTraining - A novel method for fully training all parameters of large
+# language models (llms) while using less device memory than traditional methods.
+# Copyright (C) 2024 Carl Philipp Klemm
+#
+# This file is part of QRotaryTraining.
+#
+# QRotaryTraining is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# QRotaryTraining is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with QRotaryTraining.  If not, see <http://www.gnu.org/licenses/>.
+
 import copy
 import torch
 import typing
 import datasets
 import itertools
 import transformers
+import os
 from dataclasses import dataclass
 from torch.nn.utils.rnn import pad_sequence
+from tqdm import tqdm
 
-from arguments import DataArguments
+from arguments import DataArguments, DatasetType
 
 IGNORE_INDEX = -100
 
 
-def group_texts(examples, block_size: int):
+def group_texts(examples, source_max_len: int):
     # Concatenate all texts.
     concatenated_examples = {k: list(itertools.chain(*examples[k])) for k in examples.keys()}
     total_length = len(concatenated_examples[list(examples.keys())[0]])
     # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
     # customize this part to your needs.
-    if total_length >= block_size:
-        total_length = (total_length // block_size) * block_size
+    if total_length >= source_max_len:
+        total_length = (total_length // source_max_len) * source_max_len
     # Split by chunks of max_len.
-    result = {k: [t[i: i + block_size] for i in range(0, total_length, block_size)] for k, t in concatenated_examples.items()}
+    result = {k: [t[i: i + source_max_len] for i in range(0, total_length, source_max_len)] for k, t in concatenated_examples.items()}
     result["labels"] = result["input_ids"].copy()
     return result
 
@@ -135,7 +157,7 @@ def create_data_module_s2s(tokenizer: transformers.PreTrainedTokenizer, data_arg
             eval_dataset = dataset['eval']
         else:
             print('Splitting train dataset in train and validation according to `eval_dataset_size`')
-            dataset = dataset.train_test_split(
+            dataset = dataset['train'].train_test_split(
                 test_size=data_args.eval_dataset_size, shuffle=True, seed=42
             )
             eval_dataset = dataset['test']
@@ -175,7 +197,7 @@ def create_data_module_hub(tokenizer: transformers.PreTrainedTokenizer, data_arg
             eval_dataset = dataset['eval']
         else:
             print('Splitting train dataset in train and validation according to `eval_dataset_size`')
-            dataset = dataset.train_test_split(
+            dataset = dataset['train'].train_test_split(
                 test_size=data_args.eval_dataset_size, shuffle=True, seed=42
             )
             eval_dataset = dataset['test']
@@ -198,14 +220,15 @@ def create_data_module_hub(tokenizer: transformers.PreTrainedTokenizer, data_arg
     )
 
 
-def create_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args: DataArguments, do_train: bool, do_eval: bool, do_predict: bool) -> typing.Dict:
+def create_data_module_txt(tokenizer: transformers.PreTrainedTokenizer,
+                           data_args: DataArguments, do_train: bool, do_eval: bool, do_predict: bool) -> typing.Dict:
     try:
         dataset = datasets.load_dataset('text', data_files={'train': [data_args.dataset]})
     except FileNotFoundError as ex:
         raise ValueError(f"Error loading dataset from {data_args.dataset}, {ex}")
 
-    if data_args.block_size > tokenizer.model_max_length:
-        raise ValueError(f"Block size of {data_args.block_size} is larger than the maximum size supported by the model: {tokenizer.model_max_length}")
+    if data_args.source_max_len > tokenizer.model_max_length:
+        raise ValueError(f"Max source length of {data_args.source_max_len} is larger than the maximum size supported by the model: {tokenizer.model_max_length}")
 
     def add_newline_fn(example):
         example['text'] = example['text'] + '\n'
@@ -218,10 +241,7 @@ def create_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args: D
             eval_dataset = dataset['eval']
         else:
             print('Splitting train dataset in train and validation according to `eval_dataset_size`')
-            breakpoint()
-            dataset = dataset['train'].train_test_split(
-                test_size=data_args.eval_dataset_size, shuffle=True, seed=42
-            )
+            dataset = dataset['train'].train_test_split(test_size=data_args.eval_dataset_size, shuffle=False)
             eval_dataset = dataset['test']
 
     if 'train' in dataset:
@@ -233,14 +253,14 @@ def create_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args: D
         lambda example: tokenizer(example['text']),
         batched=True,
         remove_columns='text',
-        num_proc=32,
+        num_proc=os.cpu_count(),
         load_from_cache_file=True)
     train_dataset_tokenized = train_dataset_tokenized.map(
-        lambda example: group_texts(example, data_args.block_size),
+        lambda example: group_texts(example, data_args.source_max_len),
         batched=True,
-        num_proc=32,
+        num_proc=max(1, min(os.cpu_count(), int(len(train_dataset_tokenized['input_ids']) / (data_args.source_max_len * 10)))),
         load_from_cache_file=True,
-        desc=f"Grouping texts in chunks of {data_args.block_size}")
+        desc=f"Grouping texts in chunks of {data_args.source_max_len}")
 
     eval_dataset_tokenized = None
     if eval_dataset is not None:
@@ -248,13 +268,18 @@ def create_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args: D
             lambda example: tokenizer(example['text']),
             batched=True,
             remove_columns='text',
-            num_proc=32)
+            num_proc=os.cpu_count())
         eval_dataset_tokenized = eval_dataset_tokenized.map(
-            lambda example: group_texts(example, data_args.block_size),
+            lambda example: group_texts(example, data_args.source_max_len),
             batched=True,
-            num_proc=32,
+            num_proc=max(1, min(os.cpu_count(), int(len(eval_dataset_tokenized['input_ids']) / (data_args.source_max_len * 10)))),
             load_from_cache_file=True,
-            desc=f"Grouping texts in chunks of {data_args.block_size}")
+            desc=f"Grouping texts in chunks of {data_args.source_max_len}")
+
+    for ids in train_dataset_tokenized['input_ids']:
+        assert len(ids) == data_args.source_max_len
+    for ids in eval_dataset_tokenized['input_ids']:
+        assert len(ids) == data_args.source_max_len
 
     return dict(
         train_dataset=train_dataset_tokenized if do_train else None,
@@ -262,3 +287,84 @@ def create_data_module(tokenizer: transformers.PreTrainedTokenizer, data_args: D
         predict_dataset=eval_dataset_tokenized if do_predict else None,
         data_collator=transformers.default_data_collator
     )
+
+
+def create_data_module_chat(tokenizer, data_args, do_train, do_eval, do_predict):
+    try:
+        dataset = datasets.Dataset.from_json(path_or_paths=data_args.dataset)
+    except FileNotFoundError as ex:
+        raise ValueError(f"Error loading dataset from {data_args.dataset}, {ex}")
+
+    if data_args.dataset_chat_template is not None:
+        tokenizer.chat_template = data_args.dataset_chat_template
+
+    target_len = data_args.source_max_len * 0.5
+    grouped_chats = list()
+    last_len = 0
+    for row in tqdm(dataset, desc="Grouping chat messages"):
+        content_length = len(tokenizer(row['content'])['input_ids'])
+        if last_len + content_length <= target_len and len(grouped_chats) > 0:
+            grouped_chats[-1]['chat'].append(row)
+            last_len += content_length
+        else:
+            last_len = 0
+            grouped_chats.append({'chat': [row]})
+    dataset = datasets.Dataset.from_list(grouped_chats)
+    dataset = dataset.map(lambda x: {"text": tokenizer.apply_chat_template(x["chat"], tokenize=False, add_generation_prompt=False)})
+    dataset.remove_columns('chat')
+
+    eval_dataset = None
+    if do_eval or do_predict:
+        print('Splitting train dataset in train and validation according to `eval_dataset_size`')
+        dataset_split = dataset.train_test_split(test_size=data_args.eval_dataset_size, shuffle=True)
+        train_dataset = dataset_split["train"]
+        eval_dataset = dataset_split["test"]
+
+    data_collator = DataCollatorForCausalLMText(
+        tokenizer=tokenizer,
+        max_len=data_args.source_max_len,
+    )
+    return dict(
+        train_dataset=train_dataset if do_train else None,
+        eval_dataset=eval_dataset,
+        predict_dataset=eval_dataset,
+        data_collator=data_collator
+    )
+
+
+def get_data_loaders(tokenizer, data_args: DataArguments, batch_size: int, eval_batch_size: int,
+                     do_train: bool, do_eval: bool, do_predict: bool = False):
+    data_type = DatasetType.from_string(data_args.dataset_type)
+    if data_type == DatasetType.S2S:
+        print("Loading dataset in s2s mode")
+        data_module = create_data_module_s2s(tokenizer, data_args, do_train, do_eval, do_predict)
+    elif data_type == DatasetType.HUB:
+        print("Loading dataset from hub, expecting alpaca style")
+        data_module = create_data_module_hub(tokenizer, data_args, do_train, do_eval, do_predict)
+    elif data_type == DatasetType.TEXT:
+        print("Loading dataset in txt mode")
+        data_module = create_data_module_txt(tokenizer, data_args, do_train, do_eval, do_predict)
+    elif data_type == DatasetType.CHAT:
+        print("Loading dataset in chat mode")
+        data_module = create_data_module_chat(tokenizer, data_args, do_train, do_eval, do_predict)
+    else:
+        raise RuntimeError("Unkown dataset type")
+
+    train_dataloader = None
+    eval_dataloader = None
+
+    if do_train:
+        train_dataloader = torch.utils.data.DataLoader(
+            data_module['train_dataset'],
+            shuffle=True,
+            collate_fn=data_module['data_collator'],
+            batch_size=batch_size
+        )
+    if do_eval:
+        eval_dataloader = torch.utils.data.DataLoader(
+            data_module['eval_dataset'],
+            shuffle=True,
+            collate_fn=data_module['data_collator'],
+            batch_size=eval_batch_size
+        )
+    return train_dataloader, eval_dataloader

dyntrainmodel.py

@@ -1,3 +1,23 @@
+
+# QRotaryTraining - A novel method for fully training all parameters of large
+# language models (llms) while using less device memory than traditional methods.
+# Copyright (C) 2024 Carl Philipp Klemm
+#
+# This file is part of QRotaryTraining.
+#
+# QRotaryTraining is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# QRotaryTraining is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with QRotaryTraining.  If not, see <http://www.gnu.org/licenses/>.
+
 from transformers import AutoModelForCausalLM
 import torch
 from utils import replace_module
@@ -68,7 +88,9 @@ class LinearGroup:
 
 class DyntrainModel:
     def __init__(self, model_name_or_path: str, cache_dir: str | None, quantize: bool,
-                 target_active_params: int, reshuffle_fraction: float, gradient_checkpointing: bool, trust_remote_code: bool = False):
+                 target_active_params: int, train_static_params: bool,
+                 reshuffle_fraction: float, gradient_checkpointing: bool,
+                 trust_remote_code: bool = False):
         self.model = AutoModelForCausalLM.from_pretrained(
             model_name_or_path,
             cache_dir=cache_dir,
@@ -82,6 +104,7 @@ class DyntrainModel:
             raise RuntimeError("reshuffle_percent must be between 0.1 and 1.0")
         self.devices = list[torch.device]()
         self.inital_reshufle = True
+        self.train_static_params = train_static_params
 
         if gradient_checkpointing:
             self.model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": False})
@@ -167,8 +190,14 @@ class DyntrainModel:
     def staticParameterCount(self) -> int:
         return sum(p.numel() for p in self.staticParameters())
 
+    def activeDynamicParameterCount(self) -> int:
+        return sum(p.numel() for p in self.dynamicParameters() if p.requires_grad)
+
     def activeParameterCount(self) -> int:
-        total_params = self.dynamicParameters() + self.staticParameters()
+        if self.train_static_params:
+            total_params = self.dynamicParameters() + self.staticParameters()
+        else:
+            total_params = self.dynamicParameters()
         return sum(p.numel() for p in total_params if p.requires_grad)
 
     def getDistanceAndErrorSample(self) -> (torch.Tensor, torch.Tensor):
@@ -187,7 +216,7 @@ class DyntrainModel:
         params = self.activeParameterCount()
 
         if params >= self.target_active_params:
-            RuntimeError("Insuficant active parameters to suffle active")
+            raise RuntimeError("Insuficant active parameters to suffle active")
         while params < self.target_active_params and len(self.frozen_linear_groups) > 0:
             i = randint(0, len(self.frozen_linear_groups) - 1)
             group = self.frozen_linear_groups.pop(i)
@@ -199,7 +228,7 @@ class DyntrainModel:
 
         active_params = self.activeParameterCount()
 
-        assert self.target_active_params * 1.3 > active_params and self.target_active_params * 0.7 < active_params
+        assert self.target_active_params * 1.4 > active_params and self.target_active_params * 0.6 < active_params
 
     def activeParamtersByDevice(self) -> list[int]:
         out = [0] * len(self.devices)
@@ -213,7 +242,7 @@ class DyntrainModel:
         for i, count in enumerate(active_counts):
             memory = torch.cuda.get_device_properties(self.devices[i]).total_memory
             if i == 0:
-                memory = int(memory * 0.8)
+                memory = int(memory * 0.5)
             bits_per_param.append(count / memory)
 
         max_index, max_bits_per_param = max(enumerate(active_counts), key=lambda x: x[1])
@@ -223,7 +252,7 @@ class DyntrainModel:
             if group.getDevice() is self.devices[max_index]:
                 memory = torch.cuda.get_device_properties(self.devices[max_index]).total_memory
                 if max_index == 0:
-                    memory = int(memory * 0.8)
+                    memory = int(memory * 0.5)
                 swing = group.paramCount() / memory
                 if max_bits_per_param - swing > min_bits_per_param + swing:
                     group.inplaceTo(device=self.devices[min_index])

gpl-3.0.txt

@@ -0,0 +1,674 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<https://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<https://www.gnu.org/licenses/why-not-lgpl.html>.

modules.py

@@ -1,3 +1,23 @@
+
+# QRotaryTraining - A novel method for fully training all parameters of large
+# language models (llms) while using less device memory than traditional methods.
+# Copyright (C) 2024 Carl Philipp Klemm
+#
+# This file is part of QRotaryTraining.
+#
+# QRotaryTraining is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# QRotaryTraining is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with QRotaryTraining.  If not, see <http://www.gnu.org/licenses/>.
+
 import torch
 import bitsandbytes as bnb
 import torch.multiprocessing as multiprocessing
@@ -108,7 +128,7 @@ class DynamicConvertingLinear(Linear):
 
 class DynamicQantizedLinear(Linear):
     def __init__(self, in_features: int, out_features: int, bias: bool, active_device: torch.device, cold_device: torch.device,
-                 output_dtype=None, compute_dtype=None, output_device=None):
+                 output_dtype=None, compute_dtype=None, output_device=None, cold_dtype=torch.float32):
         super().__init__(in_features, out_features, bias, cold_device, torch.float32)
         self.active_device = active_device
         self.cold_device = cold_device
@@ -120,8 +140,8 @@ class DynamicQantizedLinear(Linear):
         self.bias_quantized = None
         self.bias_state = None
         self.block_size = 128
-        self.quant_type = 'nf4'
-        self.weight_start = self.weight.clone().detach()
+        #self.weight_start = self.weight.clone().detach()
+        self.cold_dtype = cold_dtype
 
     @classmethod
     def fromLinear(cls, in_module: torch.nn.Linear, active_device: torch.device = torch.device("cuda:0"), cold_device: torch.device = torch.device("cpu"),
@@ -131,19 +151,19 @@ class DynamicQantizedLinear(Linear):
                          compute_dtype=compute_dtype, output_device=output_device)
         new_module.weight = torch.nn.Parameter(in_module.weight.to(torch.float32).to(cold_device))
         new_module.bias = torch.nn.Parameter(in_module.bias.to(torch.float32).to(cold_device)) if new_module.bias is not None else None
-        new_module.weight_start = new_module.weight.clone().detach()
+        #new_module.weight_start = new_module.weight.clone().detach()
         return new_module
 
     def compress(self) -> None:
-        weight = self.weight.contiguous().to(torch.float16).cuda(self.active_device)
+        weight = self.weight.contiguous().to(torch.float16).to(self.active_device)
         self.weight_quantized, self.weight_state = bnb.functional.quantize_blockwise(weight, blocksize=self.block_size)
         if self.bias is not None:
-            bias = self.bias.contiguous().to(torch.float16).cuda(self.active_device)
+            bias = self.bias.contiguous().to(torch.float16).to(self.active_device)
             self.bias_quantized, self.bias_state = bnb.functional.quantize_blockwise(bias, blocksize=self.block_size)
 
         frozen = self.isFrozen()
-        self.weight = torch.nn.Parameter(self.weight.to(self.cold_device))
-        self.bias = torch.nn.Parameter(self.bias.to(self.cold_device)) if self.bias is not None else None
+        self.weight = torch.nn.Parameter(self.weight.to(self.cold_dtype).to(self.cold_device))
+        self.bias = torch.nn.Parameter(self.bias.to(self.cold_dtype).to(self.cold_device)) if self.bias is not None else None
         self.setFrozen(frozen, False)
 
     def decompress(self) -> None:
@@ -151,16 +171,16 @@ class DynamicQantizedLinear(Linear):
         self.weight_state = None
         self.bias_quantized = None
         self.bias_state = None
-        self.weight_start = self.weight.clone().detach().to(self.cold_device)
-        self.weight = torch.nn.Parameter(self.weight.to(self.active_device))
+        #self.weight_start = self.weight.clone().detach().to(self.cold_device)
+        self.weight = torch.nn.Parameter(self.weight.to(self.active_device).to(torch.float32))
         if self.bias_quantized:
-            self.bias = torch.nn.Parameter(self.bias.to(self.active_device))
+            self.bias = torch.nn.Parameter(self.bias.to(self.active_device).to(torch.float32))
 
     def getDistanceAndError(self) -> tuple[torch.Tensor, torch.Tensor]:
         original_weight = self.weight.contiguous().to(self.active_device).to(torch.float16)
         quantized_original_weight, quantized_original_state = bnb.functional.quantize_blockwise(original_weight, blocksize=self.block_size)
         dequantized_original_weight = bnb.functional.dequantize_blockwise(quantized_original_weight, quantized_original_state).to(original_weight.dtype)
-        distance = (self.weight_start - self.weight.to(self.cold_device)).to(torch.float32)
+        distance = torch.zeros((2)) #(self.weight_start - self.weight.to(self.cold_device)).to(torch.float32)
         error = (dequantized_original_weight - original_weight).to(torch.float32)
         return (distance, error)
 

tokenizer.py

@@ -1,3 +1,23 @@
+
+# QRotaryTraining - A novel method for fully training all parameters of large
+# language models (llms) while using less device memory than traditional methods.
+# Copyright (C) 2024 Carl Philipp Klemm
+#
+# This file is part of QRotaryTraining.
+#
+# QRotaryTraining is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# QRotaryTraining is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with QRotaryTraining.  If not, see <http://www.gnu.org/licenses/>.
+
 import transformers
 
 from arguments import ModelArguments
@@ -30,13 +50,13 @@ def smart_tokenizer_and_embedding_resize(
 
 
 def get_tokenizer(model, cache_dir, model_args: ModelArguments):
-    print(f'Tokenizer: {model_args.tokenizer if model_args.tokenizer is not None else model_args.model_name_or_path}')
+    tokenizer_path = model_args.tokenizer if model_args.tokenizer is not None else model_args.model_name_or_path
+    print(f'Tokenizer: {tokenizer_path}')
     tokenizer = transformers.AutoTokenizer.from_pretrained(
-        model_args.tokenizer if model_args.tokenizer is not None else model_args.model_name_or_path,
+        tokenizer_path,
         cache_dir=cache_dir,
         padding_side="right",
         use_fast=False,
-        eos_token="[EOS]",
         tokenizer_type='llama' if 'llama' in model_args.model_name_or_path else None,
         trust_remote_code=model_args.trust_remote_code
     )

train_dynamic.py

@@ -1,6 +1,24 @@
-import transformers
-from transformers import get_scheduler
 
+# QRotaryTraining - A novel method for fully training all parameters of large
+# language models (llms) while using less device memory than traditional methods.
+# Copyright (C) 2024 Carl Philipp Klemm
+#
+# This file is part of QRotaryTraining.
+#
+# QRotaryTraining is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# QRotaryTraining is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with QRotaryTraining.  If not, see <http://www.gnu.org/licenses/>.
+
+import transformers
 import torch
 from torch.utils import tensorboard
 import os
@@ -8,9 +26,10 @@ import shutil
 import math
 from tqdm.auto import tqdm
 import gc
+import sys
 
 from arguments import DataArguments, ModelArguments, TrainingArguments
-from datamodules import create_data_module_s2s, create_data_module, create_data_module_hub
+from datamodules import get_data_loaders
 from tokenizer import get_tokenizer
 
 from dyntrainmodel import DyntrainModel
@@ -19,7 +38,16 @@ from dyntrainmodel import DyntrainModel
 def save_model(model, global_step: int, output_dir: str, max_checkpoints: int = 0):
     output_chkpt_dir = f"step_{global_step}" if global_step >= 0 else ""
     output_dir = os.path.join(output_dir, output_chkpt_dir)
+
+    print(f"saveing model to {output_chkpt_dir}")
+
+    temperature = model.generation_config.temperature
+    top_p = model.generation_config.top_p
+    model.generation_config.temperature = None
+    model.generation_config.top_p = None
     model.save_pretrained(output_dir)
+    model.generation_config.temperature = temperature
+    model.generation_config.top_p = top_p
 
     if max_checkpoints > 0:
         files = [f for f in os.listdir(output_dir) if os.path.isdir(os.path.join(output_dir, f)) and f.startswith("step_")]
@@ -57,37 +85,85 @@ def get_optimizer(dyamic_parameters: list[torch.nn.Parameter], static_parameters
     return optimizer
 
 
+def move_optimizer_param(param, device: torch.device, device_map: dict):
+    if isinstance(param, torch.Tensor):
+        move_device = device if device is not None else device_map[id(param)]
+        assert device is not None or move_device != torch.device("cpu")
+        old_device = param.device
+        param.data = param.data.to(move_device)
+        if param._grad is not None:
+            param._grad.data = param._grad.data.to(move_device)
+        if device is not None and id(param) not in device_map:
+            device_map[id(param)] = old_device
+            assert old_device != torch.device("cpu")
+    elif isinstance(param, dict):
+        for subparam in param.values():
+            move_optimizer_param(subparam, device, device_map)
+
+
+def suspend_optimizer(optimizer) -> dict:
+    device_map = dict()
+    for param in optimizer.state.values():
+        move_optimizer_param(param, torch.device("cpu"), device_map)
+    return device_map
+
+
+def resume_optimizer(optimizer, device_map: dict):
+    for param in optimizer.state.values():
+        move_optimizer_param(param, None, device_map)
+
+
 def evaluate(model: DyntrainModel, tokenizer,
              dataloader: torch.utils.data.DataLoader, globalstep: int,
-             log_writer: tensorboard.SummaryWriter, eval_prompt: str = None):
-    print("*** Eval ***")
-    loss = torch.zeros((1), device="cuda:0")
-    model.model.eval()
-    for batch in dataloader:
-        for key in batch:
-            batch[key] = batch[key].to("cuda:0")
-        outputs = model.model(**batch)
-        loss += outputs.loss
-    loss = loss / len(dataloader)
-    log_writer.add_scalar("Loss/Eval", loss, globalstep)
-    print(f"Eval Loss {loss.item()}")
-    return loss.item()
+             log_writer: tensorboard.SummaryWriter, eval_prompt: str | None = None):
+    with torch.no_grad():
+        loss = torch.zeros((1), device="cuda:0")
+        model.model.eval()
 
-    if eval_prompt is not None:
-        input_ids = tokenizer(eval_prompt, return_tensors="pt").input_ids.to(model.devices[0])
-        attention_mask = torch.ones(input_ids.shape, device=model.devices[0], requires_grad=False)
-        outputs = model.generate(input_ids, attention_mask=attention_mask, do_sample=True, temperature=1, max_new_tokens=100)
-        response_decoded = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
-        print(f"Eval generation: response_decoded")
-        log_writer.add_text("Text/Eval", response_decoded, globalstep)
+        for batch in tqdm(dataloader, desc="Doing eval"):
+            for key in batch:
+                batch[key] = batch[key].to("cuda:0")
+            outputs = model.model(**batch)
+            loss += outputs.loss
+        loss = loss / len(dataloader)
+        log_writer.add_scalar("Loss/Eval", loss, globalstep)
+        print(f"Eval Loss {loss.item()}")
+
+        if eval_prompt is not None:
+            input_ids = tokenizer(eval_prompt, return_tensors="pt").input_ids.to(model.devices[0])
+            attention_mask = torch.ones(input_ids.shape, device=model.devices[0], requires_grad=False)
+            outputs = model.model.generate(input_ids, attention_mask=attention_mask, do_sample=True, temperature=1,
+                                           max_new_tokens=100, min_new_tokens=100)
+            response_decoded = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
+            print(f"Eval generation: {response_decoded}")
+            log_writer.add_text("Text/Eval", response_decoded, globalstep)
+        model.model.train()
+
+
+def max_vram_allocated():
+    max_vram_alloc = 0
+    for i in range(0, torch.cuda.device_count()):
+        max_vram_alloc = max(torch.cuda.memory_allocated(i), max_vram_alloc)
+    return max_vram_alloc
+
+
+def min_vram_allocated():
+    max_vram_alloc = sys.maxsize
+    for i in range(0, torch.cuda.device_count()):
+        max_vram_alloc = min(torch.cuda.memory_allocated(i), max_vram_alloc)
+    return max_vram_alloc
 
 
 def train(model_args: ModelArguments, data_args: DataArguments, training_args: TrainingArguments):
-    log_writer = tensorboard.SummaryWriter()
+    log_writer = tensorboard.SummaryWriter(log_dir=training_args.logging_dir)
 
-    model = DyntrainModel(model_args.model_name_or_path, training_args.cache_dir, target_active_params=int(training_args.max_instant_params * 1e6),
-                          reshuffle_fraction=training_args.churn_percent / 100.0, gradient_checkpointing=True, trust_remote_code=True,
-                          quantize=model_args.quantize)
+    model = DyntrainModel(model_args.model_name_or_path, training_args.cache_dir,
+                          quantize=model_args.quantize,
+                          target_active_params=int(training_args.max_instant_params * 1e6),
+                          train_static_params=training_args.train_non_linear_layers,
+                          reshuffle_fraction=training_args.churn_percent / 100.0,
+                          gradient_checkpointing=True,
+                          trust_remote_code=True)
     devices = list(torch.device(i) for i in range(0, torch.cuda.device_count()))
     model.toDevices(devices)
     model.reshuffleActive()
@@ -96,32 +172,15 @@ def train(model_args: ModelArguments, data_args: DataArguments, training_args: T
     paramter_count = sum(p.numel() for p in model.model.parameters())
     active_paramter_count = sum(p.numel() for p in model.model.parameters() if p.requires_grad)
     static_parameter_count = model.staticParameterCount() if training_args.train_non_linear_layers else 0
-    print(f"Training model with {paramter_count / 1e6}m parameters and {active_paramter_count / 1e6}m"
+    print(f"Training model with {paramter_count / 1e6}m parameters and {active_paramter_count / 1e6}m "
           f"instantanous active paramters of which {static_parameter_count} are static")
 
     tokenizer = get_tokenizer(model.model, training_args.cache_dir, model_args)
 
-    if data_args.dataset.endswith("json"):
-        print("Loading dataset in s2s mode")
-        data_module = create_data_module_s2s(tokenizer, data_args, training_args.do_train, training_args.do_eval, False)
-    elif data_args.data_from_hub:
-        data_module = create_data_module_hub(tokenizer, data_args, training_args.do_train, training_args.do_eval, False)
-    else:
-        print("Loading dataset in txt mode")
-        data_module = create_data_module(tokenizer, data_args, training_args.do_train, training_args.do_eval, False)
-    dataset = {k: v for k, v in data_module.items() if k != 'predict_dataset'}
-    train_dataloader = torch.utils.data.DataLoader(
-        dataset['train_dataset'],
-        shuffle=True,
-        collate_fn=dataset['data_collator'],
-        batch_size=training_args.per_device_train_batch_size
-    ) if dataset['train_dataset'] is not None else None
-    eval_dataloader = torch.utils.data.DataLoader(
-        dataset['eval_dataset'],
-        shuffle=True,
-        collate_fn=dataset['data_collator'],
-        batch_size=training_args.per_device_train_batch_size
-    ) if dataset['eval_dataset'] is not None else None
+    train_dataloader, eval_dataloader = get_data_loaders(tokenizer, data_args,
+                                                         training_args.per_device_train_batch_size,
+                                                         training_args.per_device_eval_batch_size,
+                                                         training_args.do_train, training_args.do_eval)
 
     dynamic_param_ratio = (model.staticParameterCount() + model.dynamicParameterCount()) / model.dynamicParameterCount()
     steps_per_epoch = math.ceil(len(train_dataloader) / training_args.gradient_accumulation_steps) if train_dataloader is not None else 1
@@ -135,7 +194,7 @@ def train(model_args: ModelArguments, data_args: DataArguments, training_args: T
                               training_args.adam_epsilon,
                               training_args.adam8bit)
 
-    lr_scheduler = get_scheduler(
+    lr_scheduler = transformers.get_scheduler(
         name=training_args.lr_scheduler_type,
         optimizer=optimizer,
         num_warmup_steps=training_args.warmup_steps,
@@ -147,13 +206,11 @@ def train(model_args: ModelArguments, data_args: DataArguments, training_args: T
         global_step = 0
         model.model.train()
         for epoch in range(0, training_args.epochs):
-            model.model.train()
             print("*** Train ***")
-            print(f'Vram used for model before training starts: {torch.cuda.memory_allocated()/(1024.0*1024.0)}')
+            print(f'Vram used for model before training starts: {torch.cuda.memory_allocated()/(1024.0**3):.2f}')
             for step, batch in enumerate(train_dataloader):
                 for key in batch:
                     batch[key] = batch[key].to("cuda:0")
-
                 outputs = model.model(**batch)
                 loss = outputs.loss / training_args.gradient_accumulation_steps
                 loss.backward()
@@ -164,48 +221,54 @@ def train(model_args: ModelArguments, data_args: DataArguments, training_args: T
                     optimizer.step()
                     lr_scheduler.step()
 
+                    progress_bar.set_postfix_str(f"Loss: {loss.item():.2f} Max: {max_vram_allocated()/(1024.0**3):.2f}GB"
+                                                 f" Min: {min_vram_allocated()/(1024.0**3):.2f}GB")
+
                     model.model.zero_grad()
 
-                    if global_step % 5 == 0:
-                        print(f"Train Loss {loss.item()}")
+                    if global_step > 0:
+                        if global_step % training_args.reshufle_steps == 0 and training_args.max_instant_params != 0:
+                            print("Reshuffleing")
+                            lr_scheduler.optimizer = None
+                            del optimizer
+                            # distance, error = model.getDistanceAndErrorSample()
+                            # log_writer.add_histogram("Distances/Train", distance, max_bins=50)
+                            # log_writer.add_histogram("Errors/Train", error, max_bins=50)
 
-                    if global_step % training_args.reshufle_steps == 0 and training_args.max_instant_params != 0:
-                        print("Reshuffleing")
-                        lr_scheduler.optimizer = None
-                        del optimizer
-                        # distance, error = model.getDistanceAndErrorSample()
-                        # log_writer.add_histogram("Distances/Train", distance, max_bins=50)
-                        # log_writer.add_histogram("Errors/Train", error, max_bins=50)
+                            model.reshuffleActive()
+                            model.balanceActive()
+                            log_writer.add_scalar("Parameters/train", model.activeParameterCount(), global_step)
+                            optimizer = get_optimizer(model.dynamicParameters(),
+                                                      model.staticParameters() if training_args.train_non_linear_layers else None,
+                                                      training_args.learning_rate,
+                                                      training_args.learning_rate / dynamic_param_ratio,
+                                                      training_args.weight_decay,
+                                                      training_args.adam_epsilon,
+                                                      training_args.adam8bit)
+                            lr_scheduler.optimizer = optimizer
 
-                        model.reshuffleActive()
-                        model.balanceActive()
-                        log_writer.add_scalar("Parameters/train", model.activeParameterCount(), global_step)
-                        optimizer = get_optimizer(model.dynamicParameters(),
-                                                  model.staticParameters() if training_args.train_non_linear_layers else None,
-                                                  training_args.learning_rate,
-                                                  training_args.learning_rate / dynamic_param_ratio,
-                                                  training_args.weight_decay,
-                                                  training_args.adam_epsilon,
-                                                  training_args.adam8bit)
-                        lr_scheduler.optimizer = optimizer
+                        if global_step % training_args.save_steps == 0:
+                            save_model(model.model, global_step, training_args.output_dir, training_args.max_checkpoints)
+                        if training_args.eval_steps > 0 and global_step % training_args.eval_steps == 0:
+                            device_map = suspend_optimizer(optimizer)
+                            evaluate(model, tokenizer, eval_dataloader, global_step, log_writer, training_args.eval_prompt)
+                            resume_optimizer(optimizer, device_map)
 
                     global_step += 1
                     progress_bar.update()
 
-                if global_step > 0:
-                    if global_step % training_args.save_steps == 0:
-                        save_model(model.model, global_step, training_args.output_dir, training_args.max_checkpoints)
-                    if training_args.eval_steps > 0 and global_step % training_args.save_steps == 0:
-                        evaluate(model, eval_dataloader, global_step, log_writer, training_args.eval_prompt)
                 if training_args.flush_allocator:
                     gc.collect()
                     torch.cuda.empty_cache()
             if training_args.do_eval and training_args.eval_steps == -1:
-                evaluate(model, eval_dataloader, global_step, log_writer, training_args.eval_prompt)
+                device_map = suspend_optimizer(optimizer)
+                evaluate(model, tokenizer, eval_dataloader, global_step, log_writer, training_args.eval_prompt)
+                resume_optimizer(optimizer, device_map)
+
+    del optimizer
 
-    # Evaluation
     if training_args.do_eval:
-        evaluate(model, eval_dataloader, global_step, log_writer, training_args.eval_prompt)
+        evaluate(model, tokenizer, eval_dataloader, global_step, log_writer, training_args.eval_prompt)
 
     save_model(model.model, global_step, training_args.output_dir)
 

tune.sh

@@ -0,0 +1,54 @@
+#!/bin/sh
+#
+# QRotaryTraining - A novel method for fully training all parameters of large
+# language models (llms) while using less device memory than traditional methods.
+# Copyright (C) 2024 Carl Philipp Klemm
+#
+# This file is part of QRotaryTraining.
+#
+# QRotaryTraining is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# QRotaryTraining is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with QRotaryTraining.  If not, see <http://www.gnu.org/licenses/>.
+#
+
+BASE_DIR=$(dirname "$0")
+VENV_DIR=$(venvget)
+
+export MAX_JOBS=48
+
+export ROCR_VISIBLE_DEVICES="1,2"
+source $VENV_DIR/bin/activate
+
+python $SCRIPTS/train_dyamic/train_dynamic.py \
+	--model_name_or_path "huggyllama/llama-7b" \
+	--dataset "tatsu-lab/alpaca" \
+	--dataset_type "hub" \
+	--eval_dataset_size 200 \
+	--source_max_len 1024 \
+	--do_train \
+	--do_eval \
+	--eval_steps 100 \
+	--reshufle_steps 50 \
+	--per_device_train_batch_size 2 \
+	--per_device_eval_batch_size 1 \
+	--gradient_checkpointing True \
+	--gradient_accumulation_steps 4 \
+	--epochs 3 \
+	--logging_dir $BASE_DIR/log \
+	--logging_steps 5 \
+	--learning_rate 1e-6 \
+	--save_steps 500 \
+	--output_dir $BASE_DIR/llama-7b-quant \
+	--adam8bit \
+	--churn_percent 100\
+	--max_instant_params 3000 \
+	--quantize

utils.py

@@ -1,3 +1,23 @@
+
+# QRotaryTraining - A novel method for fully training all parameters of large
+# language models (llms) while using less device memory than traditional methods.
+# Copyright (C) 2024 Carl Philipp Klemm
+#
+# This file is part of QRotaryTraining.
+#
+# QRotaryTraining is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# QRotaryTraining is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with QRotaryTraining.  If not, see <http://www.gnu.org/licenses/>.
+
 from peft.utils import _get_submodules
 import torch