First commit
						commit
						a13f526f87
					
				@ -0,0 +1,2 @@
 | 
			
		||||
data/
 | 
			
		||||
logs/
 | 
			
		||||
@ -0,0 +1,2 @@
 | 
			
		||||
*.pq filter=lfs diff=lfs merge=lfs -text
 | 
			
		||||
*.csv filter=lfs diff=lfs merge=lfs -text
 | 
			
		||||
@ -0,0 +1,219 @@
 | 
			
		||||
# Byte-compiled / optimized / DLL files
 | 
			
		||||
__pycache__/
 | 
			
		||||
*.py[codz]
 | 
			
		||||
*$py.class
 | 
			
		||||
 | 
			
		||||
# C extensions
 | 
			
		||||
*.so
 | 
			
		||||
 | 
			
		||||
# Distribution / packaging
 | 
			
		||||
.Python
 | 
			
		||||
build/
 | 
			
		||||
develop-eggs/
 | 
			
		||||
dist/
 | 
			
		||||
downloads/
 | 
			
		||||
eggs/
 | 
			
		||||
.eggs/
 | 
			
		||||
lib/
 | 
			
		||||
lib64/
 | 
			
		||||
parts/
 | 
			
		||||
sdist/
 | 
			
		||||
var/
 | 
			
		||||
wheels/
 | 
			
		||||
share/python-wheels/
 | 
			
		||||
*.egg-info/
 | 
			
		||||
.installed.cfg
 | 
			
		||||
*.egg
 | 
			
		||||
MANIFEST
 | 
			
		||||
 | 
			
		||||
# PyInstaller
 | 
			
		||||
#   Usually these files are written by a python script from a template
 | 
			
		||||
#   before PyInstaller builds the exe, so as to inject date/other infos into it.
 | 
			
		||||
*.manifest
 | 
			
		||||
*.spec
 | 
			
		||||
 | 
			
		||||
# Installer logs
 | 
			
		||||
pip-log.txt
 | 
			
		||||
pip-delete-this-directory.txt
 | 
			
		||||
 | 
			
		||||
# Unit test / coverage reports
 | 
			
		||||
htmlcov/
 | 
			
		||||
.tox/
 | 
			
		||||
.nox/
 | 
			
		||||
.coverage
 | 
			
		||||
.coverage.*
 | 
			
		||||
.cache
 | 
			
		||||
nosetests.xml
 | 
			
		||||
coverage.xml
 | 
			
		||||
*.cover
 | 
			
		||||
*.py.cover
 | 
			
		||||
.hypothesis/
 | 
			
		||||
.pytest_cache/
 | 
			
		||||
cover/
 | 
			
		||||
 | 
			
		||||
# Translations
 | 
			
		||||
*.mo
 | 
			
		||||
*.pot
 | 
			
		||||
 | 
			
		||||
# Django stuff:
 | 
			
		||||
*.log
 | 
			
		||||
local_settings.py
 | 
			
		||||
db.sqlite3
 | 
			
		||||
db.sqlite3-journal
 | 
			
		||||
 | 
			
		||||
# Flask stuff:
 | 
			
		||||
instance/
 | 
			
		||||
.webassets-cache
 | 
			
		||||
 | 
			
		||||
# Scrapy stuff:
 | 
			
		||||
.scrapy
 | 
			
		||||
 | 
			
		||||
# Sphinx documentation
 | 
			
		||||
docs/_build/
 | 
			
		||||
 | 
			
		||||
# PyBuilder
 | 
			
		||||
.pybuilder/
 | 
			
		||||
target/
 | 
			
		||||
 | 
			
		||||
# Jupyter Notebook
 | 
			
		||||
.ipynb_checkpoints
 | 
			
		||||
 | 
			
		||||
# IPython
 | 
			
		||||
profile_default/
 | 
			
		||||
ipython_config.py
 | 
			
		||||
 | 
			
		||||
# pyenv
 | 
			
		||||
#   For a library or package, you might want to ignore these files since the code is
 | 
			
		||||
#   intended to run in multiple environments; otherwise, check them in:
 | 
			
		||||
# .python-version
 | 
			
		||||
 | 
			
		||||
# pipenv
 | 
			
		||||
#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
 | 
			
		||||
#   However, in case of collaboration, if having platform-specific dependencies or dependencies
 | 
			
		||||
#   having no cross-platform support, pipenv may install dependencies that don't work, or not
 | 
			
		||||
#   install all needed dependencies.
 | 
			
		||||
# Pipfile.lock
 | 
			
		||||
 | 
			
		||||
# UV
 | 
			
		||||
#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
 | 
			
		||||
#   This is especially recommended for binary packages to ensure reproducibility, and is more
 | 
			
		||||
#   commonly ignored for libraries.
 | 
			
		||||
# uv.lock
 | 
			
		||||
 | 
			
		||||
# poetry
 | 
			
		||||
#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
 | 
			
		||||
#   This is especially recommended for binary packages to ensure reproducibility, and is more
 | 
			
		||||
#   commonly ignored for libraries.
 | 
			
		||||
#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
 | 
			
		||||
# poetry.lock
 | 
			
		||||
# poetry.toml
 | 
			
		||||
 | 
			
		||||
# pdm
 | 
			
		||||
#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
 | 
			
		||||
#   pdm recommends including project-wide configuration in pdm.toml, but excluding .pdm-python.
 | 
			
		||||
#   https://pdm-project.org/en/latest/usage/project/#working-with-version-control
 | 
			
		||||
# pdm.lock
 | 
			
		||||
# pdm.toml
 | 
			
		||||
.pdm-python
 | 
			
		||||
.pdm-build/
 | 
			
		||||
 | 
			
		||||
# pixi
 | 
			
		||||
#   Similar to Pipfile.lock, it is generally recommended to include pixi.lock in version control.
 | 
			
		||||
# pixi.lock
 | 
			
		||||
#   Pixi creates a virtual environment in the .pixi directory, just like venv module creates one
 | 
			
		||||
#   in the .venv directory. It is recommended not to include this directory in version control.
 | 
			
		||||
.pixi
 | 
			
		||||
 | 
			
		||||
# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
 | 
			
		||||
__pypackages__/
 | 
			
		||||
 | 
			
		||||
# Celery stuff
 | 
			
		||||
celerybeat-schedule
 | 
			
		||||
celerybeat.pid
 | 
			
		||||
 | 
			
		||||
# Redis
 | 
			
		||||
*.rdb
 | 
			
		||||
*.aof
 | 
			
		||||
*.pid
 | 
			
		||||
 | 
			
		||||
# RabbitMQ
 | 
			
		||||
mnesia/
 | 
			
		||||
rabbitmq/
 | 
			
		||||
rabbitmq-data/
 | 
			
		||||
 | 
			
		||||
# ActiveMQ
 | 
			
		||||
activemq-data/
 | 
			
		||||
 | 
			
		||||
# SageMath parsed files
 | 
			
		||||
*.sage.py
 | 
			
		||||
 | 
			
		||||
# Environments
 | 
			
		||||
.env
 | 
			
		||||
.envrc
 | 
			
		||||
.venv
 | 
			
		||||
env/
 | 
			
		||||
venv/
 | 
			
		||||
ENV/
 | 
			
		||||
env.bak/
 | 
			
		||||
venv.bak/
 | 
			
		||||
 | 
			
		||||
# Spyder project settings
 | 
			
		||||
.spyderproject
 | 
			
		||||
.spyproject
 | 
			
		||||
 | 
			
		||||
# Rope project settings
 | 
			
		||||
.ropeproject
 | 
			
		||||
 | 
			
		||||
# mkdocs documentation
 | 
			
		||||
/site
 | 
			
		||||
 | 
			
		||||
# mypy
 | 
			
		||||
.mypy_cache/
 | 
			
		||||
.dmypy.json
 | 
			
		||||
dmypy.json
 | 
			
		||||
 | 
			
		||||
# Pyre type checker
 | 
			
		||||
.pyre/
 | 
			
		||||
 | 
			
		||||
# pytype static type analyzer
 | 
			
		||||
.pytype/
 | 
			
		||||
 | 
			
		||||
# Cython debug symbols
 | 
			
		||||
cython_debug/
 | 
			
		||||
 | 
			
		||||
# PyCharm
 | 
			
		||||
#   JetBrains specific template is maintained in a separate JetBrains.gitignore that can
 | 
			
		||||
#   be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
 | 
			
		||||
#   and can be added to the global gitignore or merged into this file.  For a more nuclear
 | 
			
		||||
#   option (not recommended) you can uncomment the following to ignore the entire idea folder.
 | 
			
		||||
# .idea/
 | 
			
		||||
 | 
			
		||||
# Abstra
 | 
			
		||||
#   Abstra is an AI-powered process automation framework.
 | 
			
		||||
#   Ignore directories containing user credentials, local state, and settings.
 | 
			
		||||
#   Learn more at https://abstra.io/docs
 | 
			
		||||
.abstra/
 | 
			
		||||
 | 
			
		||||
# Visual Studio Code
 | 
			
		||||
#   Visual Studio Code specific template is maintained in a separate VisualStudioCode.gitignore 
 | 
			
		||||
#   that can be found at https://github.com/github/gitignore/blob/main/Global/VisualStudioCode.gitignore
 | 
			
		||||
#   and can be added to the global gitignore or merged into this file. However, if you prefer, 
 | 
			
		||||
#   you could uncomment the following to ignore the entire vscode folder
 | 
			
		||||
# .vscode/
 | 
			
		||||
 | 
			
		||||
# Ruff stuff:
 | 
			
		||||
.ruff_cache/
 | 
			
		||||
 | 
			
		||||
# PyPI configuration file
 | 
			
		||||
.pypirc
 | 
			
		||||
 | 
			
		||||
# Marimo
 | 
			
		||||
marimo/_static/
 | 
			
		||||
marimo/_lsp/
 | 
			
		||||
__marimo__/
 | 
			
		||||
 | 
			
		||||
# Streamlit
 | 
			
		||||
.streamlit/secrets.toml
 | 
			
		||||
 | 
			
		||||
logs/
 | 
			
		||||
data/
 | 
			
		||||
@ -0,0 +1,51 @@
 | 
			
		||||
FROM pytorch/pytorch:2.6.0-cuda12.6-cudnn9-runtime
 | 
			
		||||
ENV TZ=Europe/Samara
 | 
			
		||||
RUN ln -snf /usr/share/zoneinfo/$TZ /etc/localtime && echo $TZ > /etc/timezone
 | 
			
		||||
 | 
			
		||||
ARG USER
 | 
			
		||||
ARG GROUP
 | 
			
		||||
ARG UID
 | 
			
		||||
ARG GID
 | 
			
		||||
 | 
			
		||||
RUN apt update
 | 
			
		||||
RUN apt install sudo -y
 | 
			
		||||
RUN sed -i 's/^%sudo.*/%sudo   ALL=(ALL) NOPASSWD: ALL/' /etc/sudoers
 | 
			
		||||
 | 
			
		||||
RUN groupadd --gid ${GID} ${GROUP}
 | 
			
		||||
RUN useradd --shell /bin/bash --uid ${UID} --gid ${GID} -G sudo --create-home ${USER}
 | 
			
		||||
RUN mkdir /wd
 | 
			
		||||
RUN chown ${USER}:${GROUP} /wd
 | 
			
		||||
 | 
			
		||||
# SYSTEM CONFIGURATION
 | 
			
		||||
RUN apt install wget vim htop mc git tree -y
 | 
			
		||||
RUN apt-get install -y libssl-dev autoconf libtool make
 | 
			
		||||
RUN cd /usr/local/src && \
 | 
			
		||||
    wget https://curl.haxx.se/download/curl-7.88.1.zip && \
 | 
			
		||||
    unzip curl-7.88.1.zip && \
 | 
			
		||||
    cd curl-7.88.1 && \
 | 
			
		||||
    ./buildconf && \
 | 
			
		||||
    ./configure --with-ssl && \
 | 
			
		||||
    make && \
 | 
			
		||||
    make install && \
 | 
			
		||||
    cp /usr/local/bin/curl /usr/bin/curl && \
 | 
			
		||||
    ldconfig && \
 | 
			
		||||
    curl -V
 | 
			
		||||
RUN curl -fsSL https://code-server.dev/install.sh | sh
 | 
			
		||||
RUN /opt/conda/bin/conda install -n base ipykernel --update-deps --force-reinstall -y
 | 
			
		||||
 | 
			
		||||
USER ${USER}
 | 
			
		||||
 | 
			
		||||
# USER CONFIGURATION
 | 
			
		||||
RUN pip install schedulefree tensorboard opencv-python-headless scipy pandas matplotlib torchmetrics pyarrow einops nvitop
 | 
			
		||||
 | 
			
		||||
RUN openssl req -x509 -newkey rsa:4096 -keyout /home/${USER}/key.pem -out /home/${USER}/cert.pem -sha256 -nodes -days 365 -subj "/C=RU/ST=SamaraRegion/L=Samara/O=SSAU/OU=LIAV/CN=vscode.ssau.ru/"
 | 
			
		||||
RUN mkdir -p /home/${USER}/.config/code-server
 | 
			
		||||
RUN echo 'bind-addr: 0.0.0.0:8443' >> /home/${USER}/.config/code-server/config.yaml
 | 
			
		||||
RUN echo "cert: /home/${USER}/cert.pem" >> /home/${USER}/.config/code-server/config.yaml
 | 
			
		||||
RUN echo "cert-key: /home/${USER}/key.pem" >> /home/${USER}/.config/code-server/config.yaml
 | 
			
		||||
 | 
			
		||||
RUN code-server --install-extension ms-python.python 
 | 
			
		||||
 | 
			
		||||
ENV SHELL=/bin/bash
 | 
			
		||||
SHELL ["/bin/bash", "--login", "-i", "-c"]
 | 
			
		||||
WORKDIR /wd
 | 
			
		||||
@ -0,0 +1,8 @@
 | 
			
		||||
#!/bin/bash
 | 
			
		||||
CURDIRNAME=${PWD##*/}
 | 
			
		||||
 | 
			
		||||
docker build . -t ${USER}_${CURDIRNAME}_vscode \
 | 
			
		||||
	--build-arg USER=${USER} \
 | 
			
		||||
	--build-arg GROUP=${USER} \
 | 
			
		||||
	--build-arg UID=$(id -u ${USER}) \
 | 
			
		||||
	--build-arg GID=$(id -g ${USER})
 | 
			
		||||
@ -0,0 +1,115 @@
 | 
			
		||||
import torch
 | 
			
		||||
import torch.nn as nn
 | 
			
		||||
from torch.nn import functional as F
 | 
			
		||||
from einops import rearrange
 | 
			
		||||
import numpy as np
 | 
			
		||||
from torch.utils.tensorboard import SummaryWriter
 | 
			
		||||
from datetime import datetime
 | 
			
		||||
import sys
 | 
			
		||||
from pathlib import Path
 | 
			
		||||
torch.manual_seed(1337)
 | 
			
		||||
 | 
			
		||||
#################################### Model #########################################
 | 
			
		||||
 | 
			
		||||
 | 
			
		||||
# RoFormer: Enhanced Transformer with Rotary Position Embedding https://arxiv.org/abs/2104.09864
 | 
			
		||||
class RoPE(nn.Module):
 | 
			
		||||
    def __init__(self, dim, max_seq_len=512):
 | 
			
		||||
        super().__init__()
 | 
			
		||||
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
 | 
			
		||||
        t = torch.arange(max_seq_len).type_as(inv_freq)
 | 
			
		||||
        freqs = torch.einsum('i,j->ij', t, inv_freq)
 | 
			
		||||
        self.register_buffer('emb', torch.cat([freqs, freqs], dim=-1))
 | 
			
		||||
 | 
			
		||||
    def rotate_half(self, x):
 | 
			
		||||
        x1, x2 = x.chunk(2, dim=-1)
 | 
			
		||||
        return torch.cat((-x2, x1), dim=-1)
 | 
			
		||||
 | 
			
		||||
    def forward(self, x, offset=0):
 | 
			
		||||
        seq_len = x.size(1)
 | 
			
		||||
        emb = self.emb[offset:offset+seq_len].view(1, seq_len, -1)
 | 
			
		||||
        cos = emb.cos()
 | 
			
		||||
        sin = emb.sin()
 | 
			
		||||
        return (x * cos) + (self.rotate_half(x) * sin)
 | 
			
		||||
 | 
			
		||||
# Transformers without Normalization https://jiachenzhu.github.io/DyT/
 | 
			
		||||
class DyT(nn.Module):
 | 
			
		||||
    def __init__(self, num_features, alpha_init_value=0.5):
 | 
			
		||||
        super().__init__()
 | 
			
		||||
        self.alpha = nn.Parameter(torch.ones(1) * alpha_init_value)
 | 
			
		||||
        self.weight = nn.Parameter(torch.ones(num_features))
 | 
			
		||||
        self.bias = nn.Parameter(torch.zeros(num_features))
 | 
			
		||||
 | 
			
		||||
    def forward(self, x):
 | 
			
		||||
        x = torch.tanh(self.alpha * x)
 | 
			
		||||
        return x * self.weight + self.bias
 | 
			
		||||
 | 
			
		||||
# Attention Is All You Need https://arxiv.org/pdf/1706.03762v7
 | 
			
		||||
# NeoBERT: A Next-Generation BERT https://arxiv.org/html/2502.19587v1
 | 
			
		||||
class TransformerLayer(nn.Module):
 | 
			
		||||
    def __init__(self, h_dim, num_heads=4, dropout_rate = 0.1, max_seq_len = 128):
 | 
			
		||||
        super().__init__()
 | 
			
		||||
        self.__dict__.update({k:v for k,v in locals().items() if k != 'self'})
 | 
			
		||||
        self.q_proj = nn.Linear(h_dim, h_dim)
 | 
			
		||||
        self.k_proj = nn.Linear(h_dim, h_dim)
 | 
			
		||||
        self.v_proj = nn.Linear(h_dim, h_dim)
 | 
			
		||||
        self.o_proj = nn.Linear(h_dim, h_dim)
 | 
			
		||||
        self.ff1 = nn.Linear(h_dim, 4*h_dim)
 | 
			
		||||
        self.ff2 = nn.Linear(4*h_dim, h_dim)
 | 
			
		||||
        self.ln1 = DyT(h_dim)
 | 
			
		||||
        self.ln2 = DyT(h_dim)
 | 
			
		||||
        self.rope = RoPE(dim=h_dim//self.num_heads, max_seq_len=max_seq_len)
 | 
			
		||||
 | 
			
		||||
    def split_to_heads(self, x, B, T, H):
 | 
			
		||||
        if self.num_heads <= 1: return x
 | 
			
		||||
        return rearrange(x, 'b t (n h) -> (b n) t h', b=B, t=T, n=self.num_heads)
 | 
			
		||||
 | 
			
		||||
    def gather_heads(self, x, B, T, H):
 | 
			
		||||
        if self.num_heads <= 1: return x
 | 
			
		||||
        return rearrange(x, '(b n) t h -> b t (n h)', b=B, t=T, n=self.num_heads)
 | 
			
		||||
 | 
			
		||||
    def attention(self, x):
 | 
			
		||||
        q = self.rope(self.split_to_heads(self.q_proj(x), *x.shape))
 | 
			
		||||
        k = self.rope(self.split_to_heads(self.k_proj(x), *x.shape))
 | 
			
		||||
        v = self.split_to_heads(self.v_proj(x), *x.shape)
 | 
			
		||||
        scores = (q @ k.transpose(1, 2)) * (self.h_dim ** -0.5)
 | 
			
		||||
        tril = torch.tril(torch.ones(x.shape[1],x.shape[1])).to(self.q_proj.bias.device)
 | 
			
		||||
        scores = scores.masked_fill(tril == 0, float('-inf')) # encoder does not need this line
 | 
			
		||||
        attention = nn.functional.softmax(scores, dim=2)
 | 
			
		||||
        return self.o_proj(self.gather_heads(attention @ v, *x.shape))
 | 
			
		||||
 | 
			
		||||
    def forward(self, x):
 | 
			
		||||
        x = x + F.dropout1d(self.attention(self.ln1(x)), p=self.dropout_rate)
 | 
			
		||||
        x = x + F.dropout1d(self.ff2(F.gelu(self.ff1(self.ln2(x)))), p=self.dropout_rate)
 | 
			
		||||
        return x
 | 
			
		||||
 | 
			
		||||
class GPT2(nn.Module):
 | 
			
		||||
    def __init__(self, vocab_size, layers_num=1, h_dim=64, max_seq_len=128, num_heads=4, dropout_rate = 0.1, pixel_size=None):
 | 
			
		||||
        super().__init__()
 | 
			
		||||
        self.__dict__.update({k:v for k,v in locals().items() if k != 'self'})
 | 
			
		||||
        self.layers = nn.ModuleList([
 | 
			
		||||
            TransformerLayer(h_dim=self.h_dim, num_heads=self.num_heads, dropout_rate=self.dropout_rate, max_seq_len=self.max_seq_len) 
 | 
			
		||||
            for _ in range(layers_num)])
 | 
			
		||||
        self.tok_embeds = nn.Embedding(vocab_size, h_dim)
 | 
			
		||||
        self.pos_embeds = nn.Parameter(torch.randn(1, self.max_seq_len, h_dim))
 | 
			
		||||
        self.lm_head = nn.Linear(h_dim, vocab_size)
 | 
			
		||||
 | 
			
		||||
    def forward(self, x, targets=None):
 | 
			
		||||
        x = self.tok_embeds(x) + self.pos_embeds[:, :x.shape[1], :]
 | 
			
		||||
        for l in self.layers:
 | 
			
		||||
            x = l(x)
 | 
			
		||||
        logits = self.lm_head(x) # B,T,C
 | 
			
		||||
        loss = F.cross_entropy(rearrange(logits, "b t c -> b c t"), targets) if not targets is None else None
 | 
			
		||||
        return logits, loss
 | 
			
		||||
 | 
			
		||||
    # what is the purpose? autoregressive inference!
 | 
			
		||||
    def generate(self, start_idx, max_new_tokens):
 | 
			
		||||
        idx = start_idx
 | 
			
		||||
        for i in range(max_new_tokens):
 | 
			
		||||
            idx_cond = idx[:,-self.max_seq_len:]
 | 
			
		||||
            logits, loss = self(idx_cond)
 | 
			
		||||
            logits = logits[:,-1,:] # B, C
 | 
			
		||||
            probs = F.softmax(logits, dim=-1)
 | 
			
		||||
            idx_next = torch.multinomial(probs, num_samples=1).to(self.lm_head.bias.device)
 | 
			
		||||
            idx = torch.cat([idx, idx_next], dim=1)
 | 
			
		||||
        return idx
 | 
			
		||||
@ -0,0 +1,129 @@
 | 
			
		||||
import torch
 | 
			
		||||
import torch.nn as nn
 | 
			
		||||
from torch.nn import functional as F
 | 
			
		||||
from einops import rearrange
 | 
			
		||||
import numpy as np
 | 
			
		||||
from torch.utils.tensorboard import SummaryWriter
 | 
			
		||||
from datetime import datetime
 | 
			
		||||
import sys
 | 
			
		||||
from pathlib import Path
 | 
			
		||||
from char_gpt2 import GPT2
 | 
			
		||||
from optics_char_gpt2 import OpticGPT2
 | 
			
		||||
seed = 1337
 | 
			
		||||
torch.manual_seed(seed)
 | 
			
		||||
models = {'gpt2': GPT2, 'optic_gpt2': OpticGPT2}
 | 
			
		||||
 | 
			
		||||
batch_size = 50
 | 
			
		||||
max_iters = 40000
 | 
			
		||||
eval_interval = 300
 | 
			
		||||
learning_rate = 1e-3
 | 
			
		||||
device = 'cuda' if torch.cuda.is_available() else 'cpu'
 | 
			
		||||
eval_iters = 200
 | 
			
		||||
layers_num = 2
 | 
			
		||||
h_dim = 64
 | 
			
		||||
max_seq_len = 256
 | 
			
		||||
num_heads = 1
 | 
			
		||||
dropout_rate = 0.1
 | 
			
		||||
pixel_size = 3.6e-6
 | 
			
		||||
 | 
			
		||||
# CUDA_VISIBLE_DEVICES=1 python .src/main.py gpt2|optic_gpt2 ./data/wiki.train.tokens ./data/wiki.valid.tokens ./data/wiki.test.tokens comment
 | 
			
		||||
MODEL_CLASS = models[sys.argv[1]]
 | 
			
		||||
train_data_path = Path(sys.argv[2])
 | 
			
		||||
val_data_path = Path(sys.argv[3])
 | 
			
		||||
test_data_path = Path(sys.argv[4])
 | 
			
		||||
comment = f"{sys.argv[1]}_{train_data_path.name}_{sys.argv[5]}_{seed}"
 | 
			
		||||
 | 
			
		||||
logs_dir = f'logs/{datetime.now().date()}_{datetime.now().hour:02d}_{datetime.now().minute:02d}_{datetime.now().second:02d}_{comment}/'
 | 
			
		||||
writer = SummaryWriter(logs_dir)
 | 
			
		||||
script_snapshot_path = Path(logs_dir + Path(sys.argv[0]).name)
 | 
			
		||||
print("Logs dir:", logs_dir)
 | 
			
		||||
script_snapshot_path.write_bytes(Path(sys.argv[0]).read_bytes()) # copy this version of script
 | 
			
		||||
script_snapshot_path.chmod(0o400) # with read-only permission
 | 
			
		||||
#################################### Dataset #########################################
 | 
			
		||||
 | 
			
		||||
# wget https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt
 | 
			
		||||
with open(train_data_path, encoding='utf-8') as f:
 | 
			
		||||
    train_text = f.read()
 | 
			
		||||
 | 
			
		||||
with open(val_data_path, encoding='utf-8') as f:
 | 
			
		||||
    val_text = f.read()
 | 
			
		||||
 | 
			
		||||
with open(test_data_path, encoding='utf-8') as f:
 | 
			
		||||
    test_text = f.read()
 | 
			
		||||
 | 
			
		||||
text = train_text + val_text + test_text
 | 
			
		||||
 | 
			
		||||
chars = sorted(set(text))
 | 
			
		||||
vocab_size = len(chars)
 | 
			
		||||
 | 
			
		||||
wtoi = {w:i for i,w in enumerate(chars)}
 | 
			
		||||
itow = {i:w for i,w in enumerate(chars)}
 | 
			
		||||
 | 
			
		||||
encode = lambda s: [wtoi[w] for w in s]
 | 
			
		||||
decode = lambda idx: ''.join([itow[i] for i in idx])
 | 
			
		||||
 | 
			
		||||
def get_batch(data, seq_len, batch_size):
 | 
			
		||||
  ix = torch.randint(len(data)-seq_len, (batch_size, ))
 | 
			
		||||
  x = torch.stack([data[i:i+seq_len] for i in ix]).to(device)
 | 
			
		||||
  y = torch.stack([data[i+1:i+1+seq_len] for i in ix]).to(device)
 | 
			
		||||
  return x, y
 | 
			
		||||
 | 
			
		||||
train_data = torch.tensor(encode(train_text), dtype=torch.long)
 | 
			
		||||
val_data = torch.tensor(encode(val_text), dtype=torch.long)
 | 
			
		||||
test_data = torch.tensor(encode(test_text), dtype=torch.long)
 | 
			
		||||
 | 
			
		||||
@torch.no_grad()
 | 
			
		||||
def perplexity(model, data):
 | 
			
		||||
    stride = max(1, len(data) // 10000)
 | 
			
		||||
    losses = []
 | 
			
		||||
    for i in range(0, len(data)-max_seq_len-1, stride):
 | 
			
		||||
        x = data[i:(i+max_seq_len)].to(device)
 | 
			
		||||
        y = data[(i+1):(i+max_seq_len+1)].to(device)
 | 
			
		||||
        logits, loss = model(x[None,...], y[None,...])
 | 
			
		||||
        losses.append(loss.item())
 | 
			
		||||
        print(f"\rppl {i}/{len(data)-max_seq_len-1}", end="")
 | 
			
		||||
    return np.exp(np.mean(losses))
 | 
			
		||||
 | 
			
		||||
#################################### Model #########################################mo
 | 
			
		||||
def complete(m, start_idxs=[0], max_new_tokens=100):
 | 
			
		||||
  start_idx = torch.tensor([start_idxs]).to(device)
 | 
			
		||||
  generated_tokens = m.generate(start_idx=start_idx, max_new_tokens=max_new_tokens)
 | 
			
		||||
  return decode(generated_tokens[0].tolist())
 | 
			
		||||
 | 
			
		||||
m = MODEL_CLASS(vocab_size=vocab_size, h_dim=h_dim, max_seq_len=max_seq_len, num_heads=num_heads, pixel_size=pixel_size)
 | 
			
		||||
m = m.to(device)
 | 
			
		||||
#################################### Train #########################################
 | 
			
		||||
 | 
			
		||||
optimizer = torch.optim.AdamW(m.parameters(), lr=learning_rate, betas=(0.90, 0.95), weight_decay=0.01)
 | 
			
		||||
 | 
			
		||||
completion = complete(m, encode("\n"*max_seq_len), 2*max_seq_len)
 | 
			
		||||
print(completion)
 | 
			
		||||
writer.add_text('completions', completion, 0)
 | 
			
		||||
 | 
			
		||||
for i in range(max_iters):
 | 
			
		||||
  xb, yb = get_batch(train_data, seq_len=max_seq_len, batch_size=batch_size)
 | 
			
		||||
  logits, loss = m(xb, yb)
 | 
			
		||||
  optimizer.zero_grad(set_to_none=True)
 | 
			
		||||
  loss.backward()
 | 
			
		||||
  optimizer.step()
 | 
			
		||||
  writer.add_scalar('loss', loss.item(), i)
 | 
			
		||||
  print(f"\r{i}/{max_iters} {loss.item()}", end="")
 | 
			
		||||
  if i % 5000 == 0:
 | 
			
		||||
    ppl = perplexity(model=m, data=val_data)
 | 
			
		||||
    writer.add_scalar('val_perplexity', ppl.item(), i)
 | 
			
		||||
    print(f"\rPerplexity at {i}: {ppl}")
 | 
			
		||||
    writer.add_text('completions', complete(m, encode("\n"*max_seq_len), 2*max_seq_len), i)
 | 
			
		||||
 | 
			
		||||
ppl = perplexity(model=m, data=val_data)
 | 
			
		||||
print(f"\r{i+1}/{max_iters} {loss.item()}")
 | 
			
		||||
print(f"\rPerplexity at {i}: {ppl}")
 | 
			
		||||
writer.add_scalar('val_perplexity', ppl.item(), i+1)
 | 
			
		||||
writer.add_scalar('loss', loss.item(), i)
 | 
			
		||||
 | 
			
		||||
ppl = perplexity(model=m, data=test_data)
 | 
			
		||||
writer.add_scalar('test_perplexity', ppl.item(), i+1)
 | 
			
		||||
print(f"\rTest Perplexity at {i}: {ppl}")
 | 
			
		||||
 | 
			
		||||
completion = complete(m, encode("\n"*max_seq_len), 2*max_seq_len)
 | 
			
		||||
print(completion)
 | 
			
		||||
writer.add_text('completions', completion, i+1)
 | 
			
		||||
@ -0,0 +1,9 @@
 | 
			
		||||
__all__ = ["config",
 | 
			
		||||
           "propagator",
 | 
			
		||||
           "optical_mul"]
 | 
			
		||||
__version__ = "3.0.0"
 | 
			
		||||
 | 
			
		||||
from .config import Config
 | 
			
		||||
from . import propagator
 | 
			
		||||
from .optical_mul import OpticalMul
 | 
			
		||||
from .parallel import DataParallel
 | 
			
		||||
@ -0,0 +1,177 @@
 | 
			
		||||
import torch
 | 
			
		||||
import torch.nn as nn
 | 
			
		||||
from torch.nn import functional as F
 | 
			
		||||
from einops import rearrange
 | 
			
		||||
import optical_matrix_multiplication as omm
 | 
			
		||||
from optical_matrix_multiplication import propagator
 | 
			
		||||
import numpy as np
 | 
			
		||||
from torch.utils.tensorboard import SummaryWriter
 | 
			
		||||
from datetime import datetime
 | 
			
		||||
from pathlib import Path
 | 
			
		||||
import sys
 | 
			
		||||
torch.manual_seed(1337)
 | 
			
		||||
 | 
			
		||||
#################################### Model #########################################
 | 
			
		||||
 | 
			
		||||
def norm(matrix: torch.Tensor, max_val: float = 1) -> torch.Tensor:
 | 
			
		||||
    return matrix / (max_val + 1e-10)
 | 
			
		||||
 | 
			
		||||
def optics_matmul_shift(sim, tensor_1, tensor_2):
 | 
			
		||||
    tensor_1 = tensor_1[None,:,:,:]
 | 
			
		||||
    tensor_2 = tensor_2[None,:,:,:]
 | 
			
		||||
 | 
			
		||||
    if torch.min(tensor_1) >= 0 and torch.min(tensor_2) >= 0:
 | 
			
		||||
        max_abs = abs(max(torch.max(tensor_1), torch.max(tensor_2)))
 | 
			
		||||
        a, b =  norm(tensor_1, max_abs), norm(tensor_2, max_abs)
 | 
			
		||||
        return sim(a, b)[0,:,:,:] * max_abs **2
 | 
			
		||||
 | 
			
		||||
    min_abs = abs(min(torch.min(tensor_1), torch.min(tensor_2)))
 | 
			
		||||
    max_abs = abs(max(torch.max(tensor_1), torch.max(tensor_2))) + min_abs
 | 
			
		||||
 | 
			
		||||
    shift_a = min_abs * torch.ones(tensor_1.shape).to(tensor_1.device)
 | 
			
		||||
    shift_b = min_abs * torch.ones(tensor_2.shape).to(tensor_1.device)
 | 
			
		||||
    a_a_sh = tensor_1 + shift_a
 | 
			
		||||
    b_b_sh = tensor_2 + shift_b
 | 
			
		||||
 | 
			
		||||
    a_a_sh_norm, b_b_sh_norm = norm(a_a_sh, max_abs), norm(b_b_sh, max_abs)
 | 
			
		||||
    shift_a_norm, shift_b_norm = norm(shift_a, max_abs), norm(shift_b, max_abs) 
 | 
			
		||||
    a_a_sh_b_b_sh = sim(a_a_sh_norm, b_b_sh_norm)
 | 
			
		||||
    a_a_sh_b_sh = sim(a_a_sh_norm, shift_b_norm)
 | 
			
		||||
    a_sh_b_b_sh = sim(shift_a_norm, b_b_sh_norm)
 | 
			
		||||
    a_sh_b_sh = sim(shift_a_norm, shift_b_norm)
 | 
			
		||||
 | 
			
		||||
    return (a_a_sh_b_b_sh - a_a_sh_b_sh - a_sh_b_b_sh + a_sh_b_sh)[0,:,:,:] * max_abs ** 2
 | 
			
		||||
 | 
			
		||||
# RoFormer: Enhanced Transformer with Rotary Position Embedding https://arxiv.org/abs/2104.09864
 | 
			
		||||
class RoPE(nn.Module):
 | 
			
		||||
    def __init__(self, dim, max_seq_len=512):
 | 
			
		||||
        super().__init__()
 | 
			
		||||
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
 | 
			
		||||
        t = torch.arange(max_seq_len).type_as(inv_freq)
 | 
			
		||||
        freqs = torch.einsum('i,j->ij', t, inv_freq)
 | 
			
		||||
        self.register_buffer('emb', torch.cat([freqs, freqs], dim=-1))
 | 
			
		||||
 | 
			
		||||
    def rotate_half(self, x):
 | 
			
		||||
        x1, x2 = x.chunk(2, dim=-1)
 | 
			
		||||
        return torch.cat((-x2, x1), dim=-1)
 | 
			
		||||
 | 
			
		||||
    def forward(self, x, offset=0):
 | 
			
		||||
        seq_len = x.size(1)
 | 
			
		||||
        emb = self.emb[offset:offset+seq_len].view(1, seq_len, -1)
 | 
			
		||||
        cos = emb.cos()
 | 
			
		||||
        sin = emb.sin()
 | 
			
		||||
        return (x * cos) + (self.rotate_half(x) * sin)
 | 
			
		||||
 | 
			
		||||
# Transformers without Normalization https://jiachenzhu.github.io/DyT/
 | 
			
		||||
class DyT(nn.Module):
 | 
			
		||||
    def __init__(self, num_features, alpha_init_value=0.5):
 | 
			
		||||
        super().__init__()
 | 
			
		||||
        self.alpha = nn.Parameter(torch.ones(1) * alpha_init_value)
 | 
			
		||||
        self.weight = nn.Parameter(torch.ones(num_features))
 | 
			
		||||
        self.bias = nn.Parameter(torch.zeros(num_features))
 | 
			
		||||
 | 
			
		||||
    def forward(self, x):
 | 
			
		||||
        x = torch.tanh(self.alpha * x)
 | 
			
		||||
        return x * self.weight + self.bias
 | 
			
		||||
 | 
			
		||||
# Attention Is All You Need https://arxiv.org/pdf/1706.03762v7
 | 
			
		||||
# NeoBERT: A Next-Generation BERT https://arxiv.org/html/2502.19587v1
 | 
			
		||||
class TransformerLayer(nn.Module):
 | 
			
		||||
    def __init__(self, h_dim, sim_scores, sim_output, num_heads=4, dropout_rate = 0.1, max_seq_len = 128):
 | 
			
		||||
        super().__init__()
 | 
			
		||||
        self.__dict__.update({k:v for k,v in locals().items() if k != 'self'})
 | 
			
		||||
        self.q_proj = nn.Linear(h_dim, h_dim)
 | 
			
		||||
        self.k_proj = nn.Linear(h_dim, h_dim)
 | 
			
		||||
        self.v_proj = nn.Linear(h_dim, h_dim)
 | 
			
		||||
        self.o_proj = nn.Linear(h_dim, h_dim)
 | 
			
		||||
        self.ff1 = nn.Linear(h_dim, 4*h_dim)
 | 
			
		||||
        self.ff2 = nn.Linear(4*h_dim, h_dim)
 | 
			
		||||
        self.ln1 = DyT(h_dim)
 | 
			
		||||
        self.ln2 = DyT(h_dim)
 | 
			
		||||
        self.rope = RoPE(dim=h_dim//self.num_heads, max_seq_len=max_seq_len)
 | 
			
		||||
        self.k1 = nn.Parameter(torch.randn(1))
 | 
			
		||||
        self.k2 = nn.Parameter(torch.randn(1))
 | 
			
		||||
 | 
			
		||||
    def split_to_heads(self, x, B, T, H):
 | 
			
		||||
        if self.num_heads <= 1: return x
 | 
			
		||||
        return rearrange(x, 'b t (n h) -> (b n) t h', b=B, t=T, n=self.num_heads)
 | 
			
		||||
 | 
			
		||||
    def gather_heads(self, x, B, T, H):
 | 
			
		||||
        if self.num_heads <= 1: return x
 | 
			
		||||
        return rearrange(x, '(b n) t h -> b t (n h)', b=B, t=T, n=self.num_heads)
 | 
			
		||||
 | 
			
		||||
    def attention(self, x):
 | 
			
		||||
        q = self.rope(self.split_to_heads(self.q_proj(x), *x.shape))
 | 
			
		||||
        k = self.rope(self.split_to_heads(self.k_proj(x), *x.shape))
 | 
			
		||||
        v = self.split_to_heads(self.v_proj(x), *x.shape)
 | 
			
		||||
        scores = self.k1 * optics_matmul_shift(self.sim_scores, q, k.transpose(1, 2)) * (self.h_dim ** -0.5) 
 | 
			
		||||
        tril = torch.tril(torch.ones(x.shape[1],x.shape[1])).to(self.q_proj.bias.device)
 | 
			
		||||
        scores = scores.masked_fill(tril == 0, float('-inf'))
 | 
			
		||||
        attention = nn.functional.softmax(scores, dim=2)
 | 
			
		||||
        output = self.k2 * optics_matmul_shift(self.sim_output, attention, v)
 | 
			
		||||
        return self.o_proj(self.gather_heads(output, *x.shape))
 | 
			
		||||
 | 
			
		||||
    def forward(self, x):
 | 
			
		||||
        x = x + F.dropout1d(self.attention(self.ln1(x)), p=self.dropout_rate)
 | 
			
		||||
        x = x + F.dropout1d(self.ff2(F.gelu(self.ff1(self.ln2(x)))), p=self.dropout_rate)
 | 
			
		||||
        return x
 | 
			
		||||
 | 
			
		||||
class OpticGPT2(nn.Module):
 | 
			
		||||
    def __init__(self, vocab_size, layers_num=1, h_dim=64, max_seq_len=128, num_heads=4, dropout_rate = 0.1,
 | 
			
		||||
                 pixel_size = 3.6e-6):
 | 
			
		||||
        super().__init__()
 | 
			
		||||
        self.__dict__.update({k:v for k,v in locals().items() if k != 'self'})
 | 
			
		||||
        self.sim_scores = omm.OpticalMul(
 | 
			
		||||
                omm.Config(right_matrix_count_columns = max_seq_len,
 | 
			
		||||
                            right_matrix_count_rows = h_dim // num_heads,
 | 
			
		||||
                            right_matrix_width = pixel_size * max_seq_len,
 | 
			
		||||
                            right_matrix_height = pixel_size * (h_dim // num_heads),
 | 
			
		||||
                            min_height_gap = pixel_size,
 | 
			
		||||
                            right_matrix_split_x = 2,
 | 
			
		||||
                            right_matrix_split_y = 2,
 | 
			
		||||
                            left_matrix_split_x = 2,
 | 
			
		||||
                            left_matrix_split_y = 2,
 | 
			
		||||
                            result_matrix_split = 2,
 | 
			
		||||
                            distance = 0.01)
 | 
			
		||||
            )
 | 
			
		||||
 | 
			
		||||
        self.sim_output = omm.OpticalMul(
 | 
			
		||||
                omm.Config(right_matrix_count_columns = h_dim // num_heads,
 | 
			
		||||
                            right_matrix_count_rows = max_seq_len,
 | 
			
		||||
                            right_matrix_width = pixel_size * (h_dim // num_heads),
 | 
			
		||||
                            right_matrix_height = pixel_size * max_seq_len,
 | 
			
		||||
                            min_height_gap = pixel_size,
 | 
			
		||||
                            right_matrix_split_x = 2,
 | 
			
		||||
                            right_matrix_split_y = 2,
 | 
			
		||||
                            left_matrix_split_x = 2,
 | 
			
		||||
                            left_matrix_split_y = 2,
 | 
			
		||||
                            result_matrix_split = 2,
 | 
			
		||||
                            distance = 0.01)
 | 
			
		||||
            )
 | 
			
		||||
        self.layers = nn.ModuleList([
 | 
			
		||||
            TransformerLayer(h_dim=self.h_dim, sim_scores=self.sim_scores, sim_output=self.sim_output, num_heads=self.num_heads, 
 | 
			
		||||
                dropout_rate=self.dropout_rate, max_seq_len=self.max_seq_len) 
 | 
			
		||||
            for _ in range(layers_num)])
 | 
			
		||||
        self.tok_embeds = nn.Embedding(vocab_size, h_dim)
 | 
			
		||||
        self.pos_embeds = nn.Parameter(torch.randn(1, self.max_seq_len, h_dim))
 | 
			
		||||
        self.lm_head = nn.Linear(h_dim, vocab_size)
 | 
			
		||||
 | 
			
		||||
    def forward(self, x, targets=None):
 | 
			
		||||
        x = self.tok_embeds(x) + self.pos_embeds[:, :x.shape[1], :]
 | 
			
		||||
        for l in self.layers:
 | 
			
		||||
            x = l(x)
 | 
			
		||||
        logits = self.lm_head(x) # B,T,C
 | 
			
		||||
        loss = F.cross_entropy(rearrange(logits, "b t c -> b c t"), targets) if not targets is None else None
 | 
			
		||||
        return logits, loss
 | 
			
		||||
 | 
			
		||||
    # what is the purpose? autoregressive inference!
 | 
			
		||||
    def generate(self, start_idx, max_new_tokens):
 | 
			
		||||
        idx = start_idx
 | 
			
		||||
        for i in range(max_new_tokens):
 | 
			
		||||
            idx_cond = idx[:,-self.max_seq_len:]
 | 
			
		||||
            logits, loss = self(idx_cond)
 | 
			
		||||
            logits = logits[:,-1,:] # B, C
 | 
			
		||||
            probs = F.softmax(logits, dim=-1)
 | 
			
		||||
            idx_next = torch.multinomial(probs, num_samples=1).to(self.lm_head.bias.device)
 | 
			
		||||
            idx = torch.cat([idx, idx_next], dim=1)
 | 
			
		||||
        return idx
 | 
			
		||||
					Loading…
					
					
				
		Reference in New Issue