Upload folder using huggingface_hub

.editorconfig

@@ -0,0 +1,42 @@
+# https://editorconfig.org/
+
+root = true
+
+[*]
+charset = utf-8
+end_of_line = lf
+indent_style = space
+indent_size = 4
+trim_trailing_whitespace = true
+insert_final_newline = true
+
+[*.py]
+indent_size = 4
+src_paths=evaluation
+
+[*.{yaml,yml,json}]
+indent_size = 2
+
+[*.md]
+indent_size = 2
+x-soft-wrap-text = true
+
+[*.rst]
+indent_size = 4
+x-soft-wrap-text = true
+
+[*.{bib,tex}]
+indent_size = 2
+
+[Makefile]
+indent_style = tab
+
+[*.sh]
+indent_style = tab
+
+[*.bat]
+end_of_line = crlf
+indent_style = tab
+
+[*.{cpp,h,cu,cuh}]
+indent_size = 2

.flake8

@@ -0,0 +1,40 @@
+[flake8]
+max-line-length = 120
+max-doc-length = 100
+select = B,C,E,F,W,Y,SIM
+ignore =
+    # E203: whitespace before ':'
+    # W503: line break before binary operator
+    # W504: line break after binary operator
+    # format by black
+    E203,W503,W504,
+    # E501: line too long
+    # W505: doc line too long
+    # too long docstring due to long example blocks
+    E501,W505,
+per-file-ignores =
+    # F401: module imported but unused
+    # intentionally unused imports
+    __init__.py: F401
+    # F401: module imported but unused
+    # F403: unable to detect undefined names
+    # F405: member mey be undefined, or defined from star imports
+    # members populated from optree
+    # E301: expected 1 blank line
+    # E302: expected 2 blank lines
+    # E305: expected 2 blank lines after class or function definition
+    # E701: multiple statements on one line (colon)
+    # E704: multiple statements on one line (def)
+    # format by black
+    *.pyi: E301,E302,E305,E701,E704
+exclude =
+    .git,
+    .vscode,
+    venv,
+    third-party,
+    __pycache__,
+    docs/source/conf.py,
+    build,
+    dist,
+    examples,
+    tests

.gitattributes

@@ -1,35 +1,11 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
+* text eol=lf
+*.ipynb linguist-detectable=false
+
+*.png binary
+*.jpg binary
+*.jpeg binary
+*.gif binary
+*.pdf binary
+images/sample.jpg filter=lfs diff=lfs merge=lfs -text
+images/training_pipelines.jpg filter=lfs diff=lfs merge=lfs -text
+quantized_files/model.safetensors filter=lfs diff=lfs merge=lfs -text

.github/workflows/lint.yml

@@ -0,0 +1,56 @@
+name: Lint
+
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+  # Allow to trigger the workflow manually
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: "${{ github.workflow }}-${{ github.ref }}"
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+env:
+  CUDA_VERSION: "11.7"
+
+jobs:
+  lint:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+        with:
+          submodules: "recursive"
+          fetch-depth: 1
+
+      - name: Set up Python 3.9
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.9"
+          update-environment: true
+
+      - name: Upgrade pip
+        run: |
+          python -m pip install --upgrade pip setuptools wheel
+
+      - name: Install DeepSeek-VL
+        env:
+          USE_FP16: "OFF"
+          TORCH_CUDA_ARCH_LIST: "Auto"
+        run: |
+          python -m pip install torch numpy pybind11
+          python -m pip install -vvv --no-build-isolation --editable '.[lint]'
+
+      - name: black
+        run: |
+          make black-format
+
+      - name: addlicense
+        run: |
+          make addlicense

.gitignore

@@ -0,0 +1,415 @@
+##### Python.gitignore #####
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+wheelhouse/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+*.whl
+
+# 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/
+docs/source/_build/
+_autosummary/
+
+# 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
+
+# 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
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# ruff
+.ruff_cache/
+
+# 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/
+
+
+##### macOS.gitignore #####
+# General
+.DS_Store
+.AppleDouble
+.LSOverride
+
+# Icon must end with two \r
+Icon
+
+# Thumbnails
+._*
+
+# Files that might appear in the root of a volume
+.DocumentRevisions-V100
+.fseventsd
+.Spotlight-V100
+.TemporaryItems
+.Trashes
+.VolumeIcon.icns
+.com.apple.timemachine.donotpresent
+
+# Directories potentially created on remote AFP share
+.AppleDB
+.AppleDesktop
+Network Trash Folder
+Temporary Items
+.apdisk
+
+
+##### Linux.gitignore #####
+*~
+
+# Temporary files which can be created if a process still has a handle open of a deleted file
+.fuse_hidden*
+
+# KDE directory preferences
+.directory
+
+# Linux trash folder which might appear on any partition or disk
+.Trash-*
+
+# .nfs files are created when an open file is removed but is still being accessed
+.nfs*
+
+
+##### Windows.gitignore #####
+# Windows thumbnail cache files
+Thumbs.db
+Thumbs.db:encryptable
+ehthumbs.db
+ehthumbs_vista.db
+
+# Dump file
+*.stackdump
+
+# Folder config file
+[Dd]esktop.ini
+
+# Recycle Bin used on file shares
+$RECYCLE.BIN/
+
+# Windows Installer files
+*.cab
+*.msi
+*.msix
+*.msm
+*.msp
+
+# Windows shortcuts
+*.lnk
+
+
+##### Archives.gitignore #####
+# It's better to unpack these files and commit the raw source because
+# git has its own built in compression methods.
+*.7z
+*.jar
+*.rar
+*.zip
+*.gz
+*.gzip
+*.tgz
+*.bzip
+*.bzip2
+*.bz2
+*.xz
+*.lzma
+*.cab
+*.xar
+
+# Packing-only formats
+*.iso
+*.tar
+
+# Package management formats
+*.dmg
+*.xpi
+*.gem
+*.egg
+*.deb
+*.rpm
+*.msi
+*.msm
+*.msp
+*.txz
+
+
+##### Xcode.gitignore #####
+# Xcode
+#
+# gitignore contributors: remember to update Global/Xcode.gitignore, Objective-C.gitignore & Swift.gitignore
+
+## User settings
+xcuserdata/
+
+## Compatibility with Xcode 8 and earlier (ignoring not required starting Xcode 9)
+*.xcscmblueprint
+*.xccheckout
+
+## Compatibility with Xcode 3 and earlier (ignoring not required starting Xcode 4)
+build/
+DerivedData/
+*.moved-aside
+*.pbxuser
+!default.pbxuser
+*.mode1v3
+!default.mode1v3
+*.mode2v3
+!default.mode2v3
+*.perspectivev3
+!default.perspectivev3
+
+## Gcc Patch
+/*.gcno
+
+
+##### JetBrains.gitignore #####
+# Covers JetBrains IDEs: IntelliJ, RubyMine, PhpStorm, AppCode, PyCharm, CLion, Android Studio and WebStorm
+# Reference: https://intellij-support.jetbrains.com/hc/en-us/articles/206544839
+
+# User settings
+.idea/*
+
+# User-specific stuff
+.idea/**/workspace.xml
+.idea/**/tasks.xml
+.idea/**/usage.statistics.xml
+.idea/**/dictionaries
+.idea/**/shelf
+
+# Generated files
+.idea/**/contentModel.xml
+
+# Sensitive or high-churn files
+.idea/**/dataSources/
+.idea/**/dataSources.ids
+.idea/**/dataSources.local.xml
+.idea/**/sqlDataSources.xml
+.idea/**/dynamic.xml
+.idea/**/uiDesigner.xml
+.idea/**/dbnavigator.xml
+
+# Gradle
+.idea/**/gradle.xml
+.idea/**/libraries
+
+# Gradle and Maven with auto-import
+# When using Gradle or Maven with auto-import, you should exclude module files,
+# since they will be recreated, and may cause churn. Uncomment if using
+# auto-import.
+# .idea/artifacts
+# .idea/compiler.xml
+# .idea/jarRepositories.xml
+# .idea/modules.xml
+# .idea/*.iml
+# .idea/modules
+# *.iml
+# *.ipr
+
+# CMake
+cmake-build-*/
+
+# Mongo Explorer plugin
+.idea/**/mongoSettings.xml
+
+# File-based project format
+*.iws
+
+# IntelliJ
+out/
+
+# mpeltonen/sbt-idea plugin
+.idea_modules/
+
+# JIRA plugin
+atlassian-ide-plugin.xml
+
+# Cursive Clojure plugin
+.idea/replstate.xml
+
+# Crashlytics plugin (for Android Studio and IntelliJ)
+com_crashlytics_export_strings.xml
+crashlytics.properties
+crashlytics-build.properties
+fabric.properties
+
+# Editor-based Rest Client
+.idea/httpRequests
+
+# Android studio 3.1+ serialized cache file
+.idea/caches/build_file_checksums.ser
+
+
+##### VisualStudioCode.gitignore #####
+.vscode/*
+# !.vscode/settings.json
+# !.vscode/tasks.json
+# !.vscode/launch.json
+!.vscode/extensions.json
+*.code-workspace
+
+# Local History for Visual Studio Code
+.history/
+
+
+##### Vim.gitignore #####
+# Swap
+.*.s[a-v][a-z]
+!*.svg  # comment out if you don't need vector files
+.*.sw[a-p]
+.s[a-rt-v][a-z]
+.ss[a-gi-z]
+.sw[a-p]
+
+# Session
+Session.vim
+Sessionx.vim
+
+# Temporary
+.netrwhist
+*~
+# Auto-generated tag files
+tags
+# Persistent undo
+[._]*.un~

.pre-commit-config.yaml

@@ -0,0 +1,75 @@
+# See https://pre-commit.com for more information
+# See https://pre-commit.com/hooks.html for more hooks
+ci:
+  skip: [pylint]
+  autofix_prs: true
+  autofix_commit_msg: "fix: [pre-commit.ci] auto fixes [...]"
+  autoupdate_commit_msg: "chore(pre-commit): [pre-commit.ci] autoupdate"
+  autoupdate_schedule: monthly
+default_stages: [commit, push, manual]
+repos:
+  - repo: https://github.com/pre-commit/pre-commit-hooks
+    rev: v4.5.0
+    hooks:
+      - id: check-symlinks
+      - id: destroyed-symlinks
+      - id: trailing-whitespace
+      - id: end-of-file-fixer
+      - id: check-yaml
+      - id: check-toml
+      - id: check-ast
+      - id: check-added-large-files
+      - id: check-merge-conflict
+      - id: check-executables-have-shebangs
+      - id: check-shebang-scripts-are-executable
+      - id: detect-private-key
+      - id: debug-statements
+      - id: double-quote-string-fixer
+  - repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.1.5
+    hooks:
+      - id: ruff
+        args: [--fix, --exit-non-zero-on-fix]
+  - repo: https://github.com/PyCQA/isort
+    rev: 5.12.0
+    hooks:
+      - id: isort
+  - repo: https://github.com/psf/black
+    rev: 23.11.0
+    hooks:
+      - id: black-jupyter
+  - repo: https://github.com/asottile/pyupgrade
+    rev: v3.15.0
+    hooks:
+      - id: pyupgrade
+        args: [--py38-plus] # sync with requires-python
+        exclude: |
+          (?x)(
+            ^images/
+          )
+  - repo: https://github.com/pycqa/flake8
+    rev: 6.1.0
+    hooks:
+      - id: flake8
+        additional_dependencies:
+          - flake8-bugbear
+          - flake8-comprehensions
+          - flake8-docstrings
+          - flake8-pyi
+          - flake8-simplify
+        exclude: |
+          (?x)(
+            ^images/
+          )
+  - repo: local
+    hooks:
+      - id: pylint
+        name: pylint
+        entry: pylint
+        language: system
+        types: [python]
+        require_serial: true
+        exclude: |
+          (?x)(
+            ^images/
+          )

.pylintrc

@@ -0,0 +1,629 @@
+[MAIN]
+
+# Analyse import fallback blocks. This can be used to support both Python 2 and
+# 3 compatible code, which means that the block might have code that exists
+# only in one or another interpreter, leading to false positives when analysed.
+analyse-fallback-blocks=no
+
+# Load and enable all available extensions. Use --list-extensions to see a list
+# all available extensions.
+#enable-all-extensions=
+
+# In error mode, messages with a category besides ERROR or FATAL are
+# suppressed, and no reports are done by default. Error mode is compatible with
+# disabling specific errors.
+#errors-only=
+
+# Always return a 0 (non-error) status code, even if lint errors are found.
+# This is primarily useful in continuous integration scripts.
+#exit-zero=
+
+# A comma-separated list of package or module names from where C extensions may
+# be loaded. Extensions are loading into the active Python interpreter and may
+# run arbitrary code.
+extension-pkg-allow-list=
+
+# A comma-separated list of package or module names from where C extensions may
+# be loaded. Extensions are loading into the active Python interpreter and may
+# run arbitrary code. (This is an alternative name to extension-pkg-allow-list
+# for backward compatibility.)
+extension-pkg-whitelist=
+
+# Return non-zero exit code if any of these messages/categories are detected,
+# even if score is above --fail-under value. Syntax same as enable. Messages
+# specified are enabled, while categories only check already-enabled messages.
+fail-on=
+
+# Specify a score threshold under which the program will exit with error.
+fail-under=10
+
+# Interpret the stdin as a python script, whose filename needs to be passed as
+# the module_or_package argument.
+#from-stdin=
+
+# Files or directories to be skipped. They should be base names, not paths.
+ignore=CVS,.vscode,.history
+
+# Add files or directories matching the regular expressions patterns to the
+# ignore-list. The regex matches against paths and can be in Posix or Windows
+# format. Because '\' represents the directory delimiter on Windows systems, it
+# can't be used as an escape character.
+ignore-paths=^images/$
+
+# Files or directories matching the regular expression patterns are skipped.
+# The regex matches against base names, not paths. The default value ignores
+# Emacs file locks
+ignore-patterns=^\.#
+
+# List of module names for which member attributes should not be checked
+# (useful for modules/projects where namespaces are manipulated during runtime
+# and thus existing member attributes cannot be deduced by static analysis). It
+# supports qualified module names, as well as Unix pattern matching.
+ignored-modules=
+
+# Python code to execute, usually for sys.path manipulation such as
+# pygtk.require().
+#init-hook=
+
+# Use multiple processes to speed up Pylint. Specifying 0 will auto-detect the
+# number of processors available to use, and will cap the count on Windows to
+# avoid hangs.
+jobs=0
+
+# Control the amount of potential inferred values when inferring a single
+# object. This can help the performance when dealing with large functions or
+# complex, nested conditions.
+limit-inference-results=100
+
+# List of plugins (as comma separated values of python module names) to load,
+# usually to register additional checkers.
+load-plugins=
+
+# Pickle collected data for later comparisons.
+persistent=yes
+
+# Minimum Python version to use for version dependent checks. Will default to
+# the version used to run pylint.
+py-version=3.8  # the lowest version we support (sync with requires-python in pyproject.toml)
+
+# Discover python modules and packages in the file system subtree.
+recursive=no
+
+# When enabled, pylint would attempt to guess common misconfiguration and emit
+# user-friendly hints instead of false-positive error messages.
+suggestion-mode=yes
+
+# Allow loading of arbitrary C extensions. Extensions are imported into the
+# active Python interpreter and may run arbitrary code.
+unsafe-load-any-extension=no
+
+# In verbose mode, extra non-checker-related info will be displayed.
+#verbose=
+
+
+[BASIC]
+
+# Naming style matching correct argument names.
+argument-naming-style=snake_case
+
+# Regular expression matching correct argument names. Overrides argument-
+# naming-style. If left empty, argument names will be checked with the set
+# naming style.
+#argument-rgx=
+
+# Naming style matching correct attribute names.
+attr-naming-style=snake_case
+
+# Regular expression matching correct attribute names. Overrides attr-naming-
+# style. If left empty, attribute names will be checked with the set naming
+# style.
+#attr-rgx=
+
+# Bad variable names which should always be refused, separated by a comma.
+bad-names=foo,
+          bar,
+          baz,
+          toto,
+          tutu,
+          tata
+
+# Bad variable names regexes, separated by a comma. If names match any regex,
+# they will always be refused
+bad-names-rgxs=
+
+# Naming style matching correct class attribute names.
+class-attribute-naming-style=any
+
+# Regular expression matching correct class attribute names. Overrides class-
+# attribute-naming-style. If left empty, class attribute names will be checked
+# with the set naming style.
+#class-attribute-rgx=
+
+# Naming style matching correct class constant names.
+class-const-naming-style=UPPER_CASE
+
+# Regular expression matching correct class constant names. Overrides class-
+# const-naming-style. If left empty, class constant names will be checked with
+# the set naming style.
+#class-const-rgx=
+
+# Naming style matching correct class names.
+class-naming-style=PascalCase
+
+# Regular expression matching correct class names. Overrides class-naming-
+# style. If left empty, class names will be checked with the set naming style.
+#class-rgx=
+
+# Naming style matching correct constant names.
+const-naming-style=UPPER_CASE
+
+# Regular expression matching correct constant names. Overrides const-naming-
+# style. If left empty, constant names will be checked with the set naming
+# style.
+#const-rgx=
+
+# Minimum line length for functions/classes that require docstrings, shorter
+# ones are exempt.
+docstring-min-length=-1
+
+# Naming style matching correct function names.
+function-naming-style=snake_case
+
+# Regular expression matching correct function names. Overrides function-
+# naming-style. If left empty, function names will be checked with the set
+# naming style.
+#function-rgx=
+
+# Good variable names which should always be accepted, separated by a comma.
+good-names=i,
+           j,
+           k,
+           ex,
+           Run,
+           _,
+           op,
+           fn,
+           f,
+           g,
+           p,
+           u,
+           t,
+           lr,
+           mu,
+           nu,
+           x,
+           y
+
+# Good variable names regexes, separated by a comma. If names match any regex,
+# they will always be accepted
+good-names-rgxs=
+
+# Include a hint for the correct naming format with invalid-name.
+include-naming-hint=no
+
+# Naming style matching correct inline iteration names.
+inlinevar-naming-style=any
+
+# Regular expression matching correct inline iteration names. Overrides
+# inlinevar-naming-style. If left empty, inline iteration names will be checked
+# with the set naming style.
+#inlinevar-rgx=
+
+# Naming style matching correct method names.
+method-naming-style=snake_case
+
+# Regular expression matching correct method names. Overrides method-naming-
+# style. If left empty, method names will be checked with the set naming style.
+#method-rgx=
+
+# Naming style matching correct module names.
+module-naming-style=snake_case
+
+# Regular expression matching correct module names. Overrides module-naming-
+# style. If left empty, module names will be checked with the set naming style.
+#module-rgx=
+
+# Colon-delimited sets of names that determine each other's naming style when
+# the name regexes allow several styles.
+name-group=
+
+# Regular expression which should only match function or class names that do
+# not require a docstring.
+no-docstring-rgx=^_
+
+# List of decorators that produce properties, such as abc.abstractproperty. Add
+# to this list to register other decorators that produce valid properties.
+# These decorators are taken in consideration only for invalid-name.
+property-classes=abc.abstractproperty
+
+# Regular expression matching correct type variable names. If left empty, type
+# variable names will be checked with the set naming style.
+#typevar-rgx=
+
+# Naming style matching correct variable names.
+variable-naming-style=snake_case
+
+# Regular expression matching correct variable names. Overrides variable-
+# naming-style. If left empty, variable names will be checked with the set
+# naming style.
+#variable-rgx=
+
+
+[CLASSES]
+
+# Warn about protected attribute access inside special methods
+check-protected-access-in-special-methods=no
+
+# List of method names used to declare (i.e. assign) instance attributes.
+defining-attr-methods=__init__,
+                      __new__,
+                      setUp,
+                      __post_init__
+
+# List of member names, which should be excluded from the protected access
+# warning.
+exclude-protected=_asdict,
+                  _fields,
+                  _replace,
+                  _source,
+                  _make
+
+# List of valid names for the first argument in a class method.
+valid-classmethod-first-arg=cls
+
+# List of valid names for the first argument in a metaclass class method.
+valid-metaclass-classmethod-first-arg=cls
+
+
+[DESIGN]
+
+# List of regular expressions of class ancestor names to ignore when counting
+# public methods (see R0903)
+exclude-too-few-public-methods=
+
+# List of qualified class names to ignore when counting class parents (see
+# R0901)
+ignored-parents=
+
+# Maximum number of arguments for function / method.
+max-args=5
+
+# Maximum number of attributes for a class (see R0902).
+max-attributes=7
+
+# Maximum number of boolean expressions in an if statement (see R0916).
+max-bool-expr=5
+
+# Maximum number of branch for function / method body.
+max-branches=12
+
+# Maximum number of locals for function / method body.
+max-locals=15
+
+# Maximum number of parents for a class (see R0901).
+max-parents=7
+
+# Maximum number of public methods for a class (see R0904).
+max-public-methods=20
+
+# Maximum number of return / yield for function / method body.
+max-returns=6
+
+# Maximum number of statements in function / method body.
+max-statements=50
+
+# Minimum number of public methods for a class (see R0903).
+min-public-methods=2
+
+
+[EXCEPTIONS]
+
+# Exceptions that will emit a warning when caught.
+overgeneral-exceptions=builtins.BaseException,
+                       builtins.Exception
+
+
+[FORMAT]
+
+# Expected format of line ending, e.g. empty (any line ending), LF or CRLF.
+expected-line-ending-format=
+
+# Regexp for a line that is allowed to be longer than the limit.
+ignore-long-lines=^\s*(# )?<?https?://\S+>?$
+
+# Number of spaces of indent required inside a hanging or continued line.
+indent-after-paren=4
+
+# String used as indentation unit. This is usually "    " (4 spaces) or "\t" (1
+# tab).
+indent-string='    '
+
+# Maximum number of characters on a single line.
+max-line-length=120
+
+# Maximum number of lines in a module.
+max-module-lines=1000
+
+# Allow the body of a class to be on the same line as the declaration if body
+# contains single statement.
+single-line-class-stmt=no
+
+# Allow the body of an if to be on the same line as the test if there is no
+# else.
+single-line-if-stmt=no
+
+
+[IMPORTS]
+
+# List of modules that can be imported at any level, not just the top level
+# one.
+allow-any-import-level=
+
+# Allow wildcard imports from modules that define __all__.
+allow-wildcard-with-all=no
+
+# Deprecated modules which should not be used, separated by a comma.
+deprecated-modules=
+
+# Output a graph (.gv or any supported image format) of external dependencies
+# to the given file (report RP0402 must not be disabled).
+ext-import-graph=
+
+# Output a graph (.gv or any supported image format) of all (i.e. internal and
+# external) dependencies to the given file (report RP0402 must not be
+# disabled).
+import-graph=
+
+# Output a graph (.gv or any supported image format) of internal dependencies
+# to the given file (report RP0402 must not be disabled).
+int-import-graph=
+
+# Force import order to recognize a module as part of the standard
+# compatibility libraries.
+known-standard-library=
+
+# Force import order to recognize a module as part of a third party library.
+known-third-party=enchant
+
+# Couples of modules and preferred modules, separated by a comma.
+preferred-modules=
+
+
+[LOGGING]
+
+# The type of string formatting that logging methods do. `old` means using %
+# formatting, `new` is for `{}` formatting.
+logging-format-style=old
+
+# Logging modules to check that the string format arguments are in logging
+# function parameter format.
+logging-modules=logging
+
+
+[MESSAGES CONTROL]
+
+# Only show warnings with the listed confidence levels. Leave empty to show
+# all. Valid levels: HIGH, CONTROL_FLOW, INFERENCE, INFERENCE_FAILURE,
+# UNDEFINED.
+confidence=HIGH,
+           CONTROL_FLOW,
+           INFERENCE,
+           INFERENCE_FAILURE,
+           UNDEFINED
+
+# Disable the message, report, category or checker with the given id(s). You
+# can either give multiple identifiers separated by comma (,) or put this
+# option multiple times (only on the command line, not in the configuration
+# file where it should appear only once). You can also use "--disable=all" to
+# disable everything first and then re-enable specific checks. For example, if
+# you want to run only the similarities checker, you can use "--disable=all
+# --enable=similarities". If you want to run only the classes checker, but have
+# no Warning level messages displayed, use "--disable=all --enable=classes
+# --disable=W".
+disable=duplicate-code,
+        consider-using-from-import
+
+# Enable the message, report, category or checker with the given id(s). You can
+# either give multiple identifier separated by comma (,) or put this option
+# multiple time (only on the command line, not in the configuration file where
+# it should appear only once). See also the "--disable" option for examples.
+enable=c-extension-no-member
+
+
+[METHOD_ARGS]
+
+# List of qualified names (i.e., library.method) which require a timeout
+# parameter e.g. 'requests.api.get,requests.api.post'
+timeout-methods=requests.api.delete,requests.api.get,requests.api.head,requests.api.options,requests.api.patch,requests.api.post,requests.api.put,requests.api.request
+
+
+[MISCELLANEOUS]
+
+# List of note tags to take in consideration, separated by a comma.
+notes=FIXME,
+      XXX,
+      TODO
+
+# Regular expression of note tags to take in consideration.
+notes-rgx=
+
+
+[REFACTORING]
+
+# Maximum number of nested blocks for function / method body
+max-nested-blocks=5
+
+# Complete name of functions that never returns. When checking for
+# inconsistent-return-statements if a never returning function is called then
+# it will be considered as an explicit return statement and no message will be
+# printed.
+never-returning-functions=sys.exit,argparse.parse_error
+
+
+[REPORTS]
+
+# Python expression which should return a score less than or equal to 10. You
+# have access to the variables 'fatal', 'error', 'warning', 'refactor',
+# 'convention', and 'info' which contain the number of messages in each
+# category, as well as 'statement' which is the total number of statements
+# analyzed. This score is used by the global evaluation report (RP0004).
+evaluation=max(0, 0 if fatal else 10.0 - ((float(5 * error + warning + refactor + convention) / statement) * 10))
+
+# Template used to display messages. This is a python new-style format string
+# used to format the message information. See doc for all details.
+msg-template=
+
+# Set the output format. Available formats are text, parseable, colorized, json
+# and msvs (visual studio). You can also give a reporter class, e.g.
+# mypackage.mymodule.MyReporterClass.
+#output-format=
+
+# Tells whether to display a full report or only the messages.
+reports=no
+
+# Activate the evaluation score.
+score=yes
+
+
+[SIMILARITIES]
+
+# Comments are removed from the similarity computation
+ignore-comments=yes
+
+# Docstrings are removed from the similarity computation
+ignore-docstrings=yes
+
+# Imports are removed from the similarity computation
+ignore-imports=yes
+
+# Signatures are removed from the similarity computation
+ignore-signatures=yes
+
+# Minimum lines number of a similarity.
+min-similarity-lines=4
+
+
+[SPELLING]
+
+# Limits count of emitted suggestions for spelling mistakes.
+max-spelling-suggestions=4
+
+# Spelling dictionary name. Available dictionaries: en_AU (hunspell), en_CA
+# (hunspell), en_GB (hunspell), en_US (hunspell), en_ZA (hunspell).
+spelling-dict=
+
+# List of comma separated words that should be considered directives if they
+# appear at the beginning of a comment and should not be checked.
+spelling-ignore-comment-directives=fmt: on,fmt: off,noqa:,noqa,nosec,isort:skip,mypy:
+
+# List of comma separated words that should not be checked.
+spelling-ignore-words=
+
+# A path to a file that contains the private dictionary; one word per line.
+spelling-private-dict-file=docs/source/spelling_wordlist.txt
+
+# Tells whether to store unknown words to the private dictionary (see the
+# --spelling-private-dict-file option) instead of raising a message.
+spelling-store-unknown-words=no
+
+
+[STRING]
+
+# This flag controls whether inconsistent-quotes generates a warning when the
+# character used as a quote delimiter is used inconsistently within a module.
+check-quote-consistency=no
+
+# This flag controls whether the implicit-str-concat should generate a warning
+# on implicit string concatenation in sequences defined over several lines.
+check-str-concat-over-line-jumps=no
+
+
+[TYPECHECK]
+
+# List of decorators that produce context managers, such as
+# contextlib.contextmanager. Add to this list to register other decorators that
+# produce valid context managers.
+contextmanager-decorators=contextlib.contextmanager
+
+# List of members which are set dynamically and missed by pylint inference
+# system, and so shouldn't trigger E1101 when accessed. Python regular
+# expressions are accepted.
+generated-members=numpy.*,
+                  torch.*
+
+# Tells whether missing members accessed in mixin class should be ignored. A
+# class is considered mixin if its name matches the mixin-class-rgx option.
+ignore-mixin-members=yes
+
+# Tells whether to warn about missing members when the owner of the attribute
+# is inferred to be None.
+ignore-none=yes
+
+# This flag controls whether pylint should warn about no-member and similar
+# checks whenever an opaque object is returned when inferring. The inference
+# can return multiple potential results while evaluating a Python object, but
+# some branches might not be evaluated, which results in partial inference. In
+# that case, it might be useful to still emit no-member and other checks for
+# the rest of the inferred objects.
+ignore-on-opaque-inference=yes
+
+# List of symbolic message names to ignore for Mixin members.
+ignored-checks-for-mixins=no-member,
+                          not-async-context-manager,
+                          not-context-manager,
+                          attribute-defined-outside-init
+
+# List of class names for which member attributes should not be checked (useful
+# for classes with dynamically set attributes). This supports the use of
+# qualified names.
+ignored-classes=optparse.Values,thread._local,_thread._local,argparse.Namespace
+
+# Show a hint with possible names when a member name was not found. The aspect
+# of finding the hint is based on edit distance.
+missing-member-hint=yes
+
+# The minimum edit distance a name should have in order to be considered a
+# similar match for a missing member name.
+missing-member-hint-distance=1
+
+# The total number of similar names that should be taken in consideration when
+# showing a hint for a missing member.
+missing-member-max-choices=1
+
+# Regex pattern to define which classes are considered mixins.
+mixin-class-rgx=.*[Mm]ixin
+
+# List of decorators that change the signature of a decorated function.
+signature-mutators=
+
+
+[VARIABLES]
+
+# List of additional names supposed to be defined in builtins. Remember that
+# you should avoid defining new builtins when possible.
+additional-builtins=
+
+# Tells whether unused global variables should be treated as a violation.
+allow-global-unused-variables=yes
+
+# List of names allowed to shadow builtins
+allowed-redefined-builtins=
+
+# List of strings which can identify a callback function by name. A callback
+# name must start or end with one of those strings.
+callbacks=cb_,
+          _cb
+
+# A regular expression matching the name of dummy variables (i.e. expected to
+# not be used).
+dummy-variables-rgx=_+$|(_[a-zA-Z0-9_]*[a-zA-Z0-9]+?$)|dummy|^ignored_|^unused_
+
+# Argument names that match this expression will be ignored.
+ignored-argument-names=_.*|^ignored_|^unused_
+
+# Tells whether we should check for unused import in __init__ files.
+init-import=no
+
+# List of qualified module names which can have objects that can redefine
+# builtins.
+redefining-builtins-modules=six.moves,past.builtins,future.builtins,builtins,io

LICENSE-CODE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 DeepSeek
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

LICENSE-MODEL

@@ -0,0 +1,91 @@
+DEEPSEEK LICENSE AGREEMENT
+
+Version 1.0, 23 October 2023
+
+Copyright (c) 2023 DeepSeek
+
+Section I: PREAMBLE
+
+Large generative models are being widely adopted and used, and have the potential to transform the way individuals conceive and benefit from AI or ML technologies.
+
+Notwithstanding the current and potential benefits that these artifacts can bring to society at large, there are also concerns about potential misuses of them, either due to their technical limitations or ethical considerations.
+
+In short, this license strives for both the open and responsible downstream use of the accompanying model. When it comes to the open character, we took inspiration from open source permissive licenses regarding the grant of IP rights. Referring to the downstream responsible use, we added use-based restrictions not permitting the use of the model in very specific scenarios, in order for the licensor to be able to enforce the license in case potential misuses of the Model may occur. At the same time, we strive to promote open and responsible research on generative models for content generation.
+
+Even though downstream derivative versions of the model could be released under different licensing terms, the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license (this license). We believe in the intersection between open and responsible AI development; thus, this agreement aims to strike a balance between both in order to enable responsible open-science in the field of AI.
+
+This License governs the use of the model (and its derivatives) and is informed by the model card associated with the model.
+
+NOW THEREFORE, You and DeepSeek agree as follows:
+
+1. Definitions
+"License" means the terms and conditions for use, reproduction, and Distribution as defined in this document.
+"Data" means a collection of information and/or content extracted from the dataset used with the Model, including to train, pretrain, or otherwise evaluate the Model. The Data is not licensed under this License.
+"Output" means the results of operating a Model as embodied in informational content resulting therefrom.
+"Model" means any accompanying machine-learning based assemblies (including checkpoints), consisting of learnt weights, parameters (including optimizer states), corresponding to the model architecture as embodied in the Complementary Material, that have been trained or tuned, in whole or in part on the Data, using the Complementary Material.
+"Derivatives of the Model" means all modifications to the Model, works based on the Model, or any other model which is created or initialized by transfer of patterns of the weights, parameters, activations or output of the Model, to the other model, in order to cause the other model to perform similarly to the Model, including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model.
+"Complementary Material" means the accompanying source code and scripts used to define, run, load, benchmark or evaluate the Model, and used to prepare data for training or evaluation, if any. This includes any accompanying documentation, tutorials, examples, etc, if any.
+"Distribution" means any transmission, reproduction, publication or other sharing of the Model or Derivatives of the Model to a third party, including providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access.
+"DeepSeek" (or "we") means Beijing DeepSeek Artificial Intelligence Fundamental Technology Research Co., Ltd., Hangzhou DeepSeek Artificial Intelligence Fundamental Technology Research Co., Ltd. and/or any of their affiliates.
+"You" (or "Your") means an individual or Legal Entity exercising permissions granted by this License and/or making use of the Model for whichever purpose and in any field of use, including usage of the Model in an end-use application - e.g. chatbot, translator, etc.
+"Third Parties" means individuals or legal entities that are not under common control with DeepSeek or You.
+
+Section II: INTELLECTUAL PROPERTY RIGHTS
+
+Both copyright and patent grants apply to the Model, Derivatives of the Model and Complementary Material. The Model and Derivatives of the Model are subject to additional terms as described in Section III.
+
+2. Grant of Copyright License. Subject to the terms and conditions of this License, DeepSeek hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model.
+
+3. Grant of Patent License. Subject to the terms and conditions of this License and where and as applicable, DeepSeek hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this paragraph) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Model and the Complementary Material, where such license applies only to those patent claims licensable by DeepSeek that are necessarily infringed by its contribution(s). If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Model and/or Complementary Material constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for the Model and/or works shall terminate as of the date such litigation is asserted or filed.
+
+
+Section III: CONDITIONS OF USAGE, DISTRIBUTION AND REDISTRIBUTION
+
+4. Distribution and Redistribution. You may host for Third Party remote access purposes (e.g. software-as-a-service), reproduce and distribute copies of the Model or Derivatives of the Model thereof in any medium, with or without modifications, provided that You meet the following conditions:
+a. Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model, and You shall give notice to subsequent users You Distribute to, that the Model or Derivatives of the Model are subject to paragraph 5. This provision does not apply to the use of Complementary Material.
+b. You must give any Third Party recipients of the Model or Derivatives of the Model a copy of this License;
+c. You must cause any modified files to carry prominent notices stating that You changed the files;
+d. You must retain all copyright, patent, trademark, and attribution notices excluding those notices that do not pertain to any part of the Model, Derivatives of the Model.
+e. You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions - respecting paragraph 4.a. – for use, reproduction, or Distribution of Your modifications, or for any such Derivatives of the Model as a whole, provided Your use, reproduction, and Distribution of the Model otherwise complies with the conditions stated in this License.
+
+5. Use-based restrictions. The restrictions set forth in Attachment A are considered Use-based restrictions. Therefore You cannot use the Model and the Derivatives of the Model for the specified restricted uses. You may use the Model subject to this License, including only for lawful purposes and in accordance with the License. Use may include creating any content with, finetuning, updating, running, training, evaluating and/or reparametrizing the Model. You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph (paragraph 5).
+
+6. The Output You Generate. Except as set forth herein, DeepSeek claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License.
+
+Section IV: OTHER PROVISIONS
+
+7. Updates and Runtime Restrictions. To the maximum extent permitted by law, DeepSeek reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License.
+
+8. Trademarks and related. Nothing in this License permits You to make use of DeepSeek’ trademarks, trade names, logos or to otherwise suggest endorsement or misrepresent the relationship between the parties; and any rights not expressly granted herein are reserved by DeepSeek.
+
+9. Personal information, IP rights and related. This Model may contain personal information and works with IP rights. You commit to complying with applicable laws and regulations in the handling of personal information and the use of such works. Please note that DeepSeek's license granted to you to use the Model does not imply that you have obtained a legitimate basis for processing the related information or works. As an independent personal information processor and IP rights user, you need to ensure full compliance with relevant legal and regulatory requirements when handling personal information and works with IP rights that may be contained in the Model, and are willing to assume solely any risks and consequences that may arise from that.
+
+10. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, DeepSeek provides the Model and the Complementary Material on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Model, Derivatives of the Model, and the Complementary Material and assume any risks associated with Your exercise of permissions under this License.
+
+11. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall DeepSeek be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Model and the Complementary Material (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if DeepSeek has been advised of the possibility of such damages.
+
+12. Accepting Warranty or Additional Liability. While redistributing the Model, Derivatives of the Model and the Complementary Material thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of DeepSeek, and only if You agree to indemnify, defend, and hold DeepSeek harmless for any liability incurred by, or claims asserted against, DeepSeek by reason of your accepting any such warranty or additional liability.
+
+13. If any provision of this License is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein.
+
+14. Governing Law and Jurisdiction. This agreement will be governed and construed under PRC laws without regard to choice of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this agreement. The courts located in the domicile of Hangzhou DeepSeek Artificial Intelligence Fundamental Technology Research Co., Ltd. shall have exclusive jurisdiction of any dispute arising out of this agreement.
+
+END OF TERMS AND CONDITIONS
+
+Attachment A
+
+Use Restrictions
+
+You agree not to use the Model or Derivatives of the Model:
+
+-	In any way that violates any applicable national or international law or regulation or infringes upon the lawful rights and interests of any third party;
+-	For military use in any way;
+-	For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
+-	To generate or disseminate verifiably false information and/or content with the purpose of harming others;
+-	To generate or disseminate inappropriate content subject to applicable regulatory requirements;
+-	To generate or disseminate personal identifiable information without due authorization or for unreasonable use;
+-	To defame, disparage or otherwise harass others;
+-	For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation;
+-	For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics;
+-	To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
+-	For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories.

Makefile

@@ -0,0 +1,99 @@
+print-%  : ; @echo $* = $($*)
+PROJECT_NAME   = DeepSeek-VL
+COPYRIGHT      = "DeepSeek."
+PROJECT_PATH   = deepseek_vl
+SHELL          = /bin/bash
+SOURCE_FOLDERS = deepseek_vl
+PYTHON_FILES   = $(shell find $(SOURCE_FOLDERS) -type f -name "*.py" -o -name "*.pyi") cli_chat.py inference.py
+COMMIT_HASH    = $(shell git log -1 --format=%h)
+PATH           := $(HOME)/go/bin:$(PATH)
+PYTHON         ?= $(shell command -v python3 || command -v python)
+PYTESTOPTS     ?=
+
+.PHONY: default
+default: install
+
+# Tools Installation
+
+check_pip_install = $(PYTHON) -m pip show $(1) &>/dev/null || (cd && $(PYTHON) -m pip install $(1) --upgrade)
+check_pip_install_extra = $(PYTHON) -m pip show $(1) &>/dev/null || (cd && $(PYTHON) -m pip install $(2) --upgrade)
+
+pylint-install:
+	$(call check_pip_install_extra,pylint,pylint[spelling])
+	$(call check_pip_install,pyenchant)
+
+flake8-install:
+	$(call check_pip_install,flake8)
+	$(call check_pip_install,flake8-bugbear)
+	$(call check_pip_install,flake8-comprehensions)
+	$(call check_pip_install,flake8-docstrings)
+	$(call check_pip_install,flake8-pyi)
+	$(call check_pip_install,flake8-simplify)
+
+py-format-install:
+	$(call check_pip_install,isort)
+	$(call check_pip_install_extra,black,black[jupyter])
+
+ruff-install:
+	$(call check_pip_install,ruff)
+
+mypy-install:
+	$(call check_pip_install,mypy)
+
+pre-commit-install:
+	$(call check_pip_install,pre-commit)
+	$(PYTHON) -m pre_commit install --install-hooks
+
+go-install:
+	# requires go >= 1.16
+	command -v go || (sudo apt-get install -y golang && sudo ln -sf /usr/lib/go/bin/go /usr/bin/go)
+
+addlicense-install: go-install
+	command -v addlicense || go install github.com/google/addlicense@latest
+
+addlicense: addlicense-install
+	addlicense -c $(COPYRIGHT) -ignore tests/coverage.xml -l mit -y 2023-$(shell date +"%Y") -check $(SOURCE_FOLDERS)
+
+# Python linters
+
+pylint: pylint-install
+	$(PYTHON) -m pylint $(PROJECT_PATH)
+
+flake8: flake8-install
+	$(PYTHON) -m flake8 --count --show-source --statistics
+
+py-format: py-format-install
+	$(PYTHON) -m isort --project $(PROJECT_PATH) --check $(PYTHON_FILES) && \
+	$(PYTHON) -m black --check $(PYTHON_FILES)
+
+black-format: py-format-install
+	$(PYTHON) -m black --check $(PYTHON_FILES)
+
+ruff: ruff-install
+	$(PYTHON) -m ruff check .
+
+ruff-fix: ruff-install
+	$(PYTHON) -m ruff check . --fix --exit-non-zero-on-fix
+
+mypy: mypy-install
+	$(PYTHON) -m mypy $(PROJECT_PATH) --install-types --non-interactive
+
+pre-commit: pre-commit-install
+	$(PYTHON) -m pre_commit run --all-files
+
+# Utility functions
+
+lint: ruff flake8 py-format mypy pylint addlicense
+
+format: py-format-install ruff-install addlicense-install
+	$(PYTHON) -m isort --project $(PROJECT_PATH) $(PYTHON_FILES)
+	$(PYTHON) -m black $(PYTHON_FILES)
+	addlicense -c $(COPYRIGHT) -ignore tests/coverage.xml -l mit -y 2023-$(shell date +"%Y") $(SOURCE_FOLDERS) cli_chat.py inference.py
+
+clean-py:
+	find . -type f -name  '*.py[co]' -delete
+	find . -depth -type d -name "__pycache__" -exec rm -r "{}" +
+	find . -depth -type d -name ".ruff_cache" -exec rm -r "{}" +
+	find . -depth -type d -name ".mypy_cache" -exec rm -r "{}" +
+
+clean: clean-py

README.md

@@ -1,3 +1,227 @@
----
-license: apache-2.0
----
+<!-- markdownlint-disable first-line-h1 -->
+<!-- markdownlint-disable html -->
+<!-- markdownlint-disable no-duplicate-header -->
+
+<div align="center">
+  <img src="images/logo.svg" width="60%" alt="DeepSeek LLM" />
+</div>
+<hr>
+<div align="center">
+
+  <a href="https://www.deepseek.com/" target="_blank">
+    <img alt="Homepage" src="images/badge.svg" />
+  </a>
+  <a href="https://huggingface.co/spaces/deepseek-ai/DeepSeek-VL-7B" target="_blank">
+    <img alt="Chat" src="https://img.shields.io/badge/🤖%20Chat-DeepSeek%20VL-536af5?color=536af5&logoColor=white" />
+  </a>
+  <a href="https://huggingface.co/deepseek-ai" target="_blank">
+    <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-DeepSeek%20AI-ffc107?color=ffc107&logoColor=white" />
+  </a>
+
+</div>
+
+
+<div align="center">
+
+  <a href="https://discord.gg/Tc7c45Zzu5" target="_blank">
+    <img alt="Discord" src="https://img.shields.io/badge/Discord-DeepSeek%20AI-7289da?logo=discord&logoColor=white&color=7289da" />
+  </a>
+  <a href="images/qr.jpeg" target="_blank">
+    <img alt="Wechat" src="https://img.shields.io/badge/WeChat-DeepSeek%20AI-brightgreen?logo=wechat&logoColor=white" />
+  </a>
+  <a href="https://twitter.com/deepseek_ai" target="_blank">
+    <img alt="Twitter Follow" src="https://img.shields.io/badge/Twitter-deepseek_ai-white?logo=x&logoColor=white" />
+  </a>
+
+</div>
+
+<div align="center">
+
+  <a href="LICENSE-CODE">
+    <img alt="Code License" src="https://img.shields.io/badge/Code_License-MIT-f5de53?&color=f5de53">
+  </a>
+  <a href="LICENSE-MODEL">
+    <img alt="Model License" src="https://img.shields.io/badge/Model_License-Model_Agreement-f5de53?&color=f5de53">
+  </a>
+</div>
+
+
+<p align="center">
+  <a href="#3-model-downloads"><b>📥 Model Download</b></a> |
+  <a href="#4-quick-start"><b>⚡ Quick Start</b></a> |
+  <a href="#5-license"><b>📜 License</b></a> |
+  <a href="#6-citation"><b>📖 Citation</b></a> <br>
+  <a href="https://arxiv.org/abs/2403.05525"><b>📄 Paper Link</b></a> |
+  <a href="https://huggingface.co/papers/2403.05525"><b>🤗 Huggingface Paper Link</b></a> |
+  <a href="https://huggingface.co/spaces/deepseek-ai/DeepSeek-VL-7B"><b>👁️ Demo</b></a>
+</p>
+
+
+## 1. Introduction
+
+Introducing DeepSeek-VL, an open-source Vision-Language (VL) Model designed for real-world vision and language understanding applications. DeepSeek-VL possesses general multimodal understanding capabilities, capable of processing logical diagrams, web pages, formula recognition, scientific literature, natural images, and embodied intelligence in complex scenarios.
+
+[DeepSeek-VL: Towards Real-World Vision-Language Understanding](https://arxiv.org/abs/2403.05525)
+
+Haoyu Lu*, Wen Liu*, Bo Zhang**, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, Yaofeng Sun, Chengqi Deng, Hanwei Xu, Zhenda Xie, Chong Ruan (*Equal Contribution, **Project Lead)
+
+![](https://github.com/deepseek-ai/DeepSeek-VL/blob/main/images/sample.jpg)
+
+## 2. Release
+
+<details>
+<summary>✅ <b>2024-03-14</b>: Demo for DeepSeek-VL-7B available on <a href="https://huggingface.co/spaces/deepseek-ai/DeepSeek-VL-7B">Hugging Face</a>.</summary>
+<br>Check out the gradio demo of DeepSeek-VL-7B at <a href="https://huggingface.co/spaces/deepseek-ai/DeepSeek-VL-7B">https://huggingface.co/spaces/deepseek-ai/DeepSeek-VL-7B</a>. Experience its capabilities firsthand!
+</details>
+
+
+<details>
+<summary>✅ <b>2024-03-13</b>: Support DeepSeek-VL gradio demo.
+
+</details>
+
+<details>
+<summary>✅ <b>2024-03-11</b>: DeepSeek-VL family released, including <code>DeepSeek-VL-7B-base</code>, <code>DeepSeek-VL-7B-chat</code>, <code>DeepSeek-VL-1.3B-base</code>, and <code>DeepSeek-VL-1.3B-chat</code>.</summary>
+<br>The release includes a diverse set of models tailored for various applications within the DeepSeek-VL family. The models come in two sizes: 7B and 1.3B parameters, each offering base and chat variants to cater to different needs and integration scenarios.
+
+</details>
+
+## 3. Model Downloads
+
+We release the DeepSeek-VL family, including 1.3B-base, 1.3B-chat, 7b-base and 7b-chat models, to the public.
+To support a broader and more diverse range of research within both academic and commercial communities.
+Please note that the use of this model is subject to the terms outlined in [License section](#5-license). Commercial usage is
+permitted under these terms.
+
+### Huggingface
+
+| Model                 | Sequence Length | Download                                                                    |
+|-----------------------|-----------------|-----------------------------------------------------------------------------|
+| DeepSeek-VL-1.3B-base | 4096            | [🤗 Hugging Face](https://huggingface.co/deepseek-ai/deepseek-vl-1.3b-base) |
+| DeepSeek-VL-1.3B-chat | 4096            | [🤗 Hugging Face](https://huggingface.co/deepseek-ai/deepseek-vl-1.3b-chat) |
+| DeepSeek-VL-7B-base   | 4096            | [🤗 Hugging Face](https://huggingface.co/deepseek-ai/deepseek-vl-7b-base)   |
+| DeepSeek-VL-7B-chat   | 4096            | [🤗 Hugging Face](https://huggingface.co/deepseek-ai/deepseek-vl-7b-chat)   |
+
+
+
+## 4. Quick Start
+
+### Installation
+
+On the basis of `Python >= 3.8` environment, install the necessary dependencies by running the following command:
+
+```shell
+pip install -e .
+```
+
+### Simple Inference Example
+
+```python
+import torch
+from transformers import AutoModelForCausalLM
+
+from deepseek_vl.models import VLChatProcessor, MultiModalityCausalLM
+from deepseek_vl.utils.io import load_pil_images
+
+
+# specify the path to the model
+model_path = "deepseek-ai/deepseek-vl-7b-chat"
+vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(model_path)
+tokenizer = vl_chat_processor.tokenizer
+
+vl_gpt: MultiModalityCausalLM = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True)
+vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()
+
+## single image conversation example
+conversation = [
+    {
+        "role": "User",
+        "content": "<image_placeholder>Describe each stage of this image.",
+        "images": ["./images/training_pipelines.jpg"],
+    },
+    {"role": "Assistant", "content": ""},
+]
+
+## multiple images (or in-context learning) conversation example
+# conversation = [
+#     {
+#         "role": "User",
+#         "content": "<image_placeholder>A dog wearing nothing in the foreground, "
+#                    "<image_placeholder>a dog wearing a santa hat, "
+#                    "<image_placeholder>a dog wearing a wizard outfit, and "
+#                    "<image_placeholder>what's the dog wearing?",
+#         "images": [
+#             "images/dog_a.png",
+#             "images/dog_b.png",
+#             "images/dog_c.png",
+#             "images/dog_d.png",
+#         ],
+#     },
+#     {"role": "Assistant", "content": ""}
+# ]
+
+# load images and prepare for inputs
+pil_images = load_pil_images(conversation)
+prepare_inputs = vl_chat_processor(
+    conversations=conversation,
+    images=pil_images,
+    force_batchify=True
+).to(vl_gpt.device)
+
+# run image encoder to get the image embeddings
+inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)
+
+# run the model to get the response
+outputs = vl_gpt.language_model.generate(
+    inputs_embeds=inputs_embeds,
+    attention_mask=prepare_inputs.attention_mask,
+    pad_token_id=tokenizer.eos_token_id,
+    bos_token_id=tokenizer.bos_token_id,
+    eos_token_id=tokenizer.eos_token_id,
+    max_new_tokens=512,
+    do_sample=False,
+    use_cache=True
+)
+
+answer = tokenizer.decode(outputs[0].cpu().tolist(), skip_special_tokens=True)
+print(f"{prepare_inputs['sft_format'][0]}", answer)
+```
+
+### CLI Chat
+```bash
+python cli_chat.py --model_path "deepseek-ai/deepseek-vl-7b-chat"
+
+# or local path
+python cli_chat.py --model_path "local model path"
+```
+
+### Gradio Demo
+```bash
+pip install -e .[gradio]
+
+python deepseek_vl/serve/app_deepseek.py
+```
+![](./images/gradio_demo.png)
+
+Have Fun!
+
+## 5. License
+
+This code repository is licensed under [the MIT License](https://github.com/deepseek-ai/DeepSeek-LLM/blob/HEAD/LICENSE-CODE). The use of DeepSeek-VL Base/Chat models is subject to [DeepSeek Model License](https://github.com/deepseek-ai/DeepSeek-LLM/blob/HEAD/LICENSE-MODEL). DeepSeek-VL series (including Base and Chat) supports commercial use.
+
+## 6. Citation
+
+```
+@misc{lu2024deepseekvl,
+      title={DeepSeek-VL: Towards Real-World Vision-Language Understanding},
+      author={Haoyu Lu and Wen Liu and Bo Zhang and Bingxuan Wang and Kai Dong and Bo Liu and Jingxiang Sun and Tongzheng Ren and Zhuoshu Li and Hao Yang and Yaofeng Sun and Chengqi Deng and Hanwei Xu and Zhenda Xie and Chong Ruan},
+      year={2024},
+      eprint={2403.05525},
+      archivePrefix={arXiv},
+      primaryClass={cs.AI}
+}
+```
+
+## 7. Contact
+
+If you have any questions, please raise an issue or contact us at [service@deepseek.com](mailto:service@deepseek.com).

cli_chat.py

@@ -0,0 +1,224 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+# -*- coding: utf-8 -*-
+
+import argparse
+import os
+import sys
+from threading import Thread
+
+import torch
+from PIL import Image
+from transformers import TextIteratorStreamer
+
+from deepseek_vl.utils.io import load_pretrained_model
+
+
+def load_image(image_file):
+    image = Image.open(image_file).convert("RGB")
+    return image
+
+
+def get_help_message(image_token):
+    help_msg = (
+        f"\t\t DeepSeek-VL-Chat is a chatbot that can answer questions based on the given image. Enjoy it! \n"
+        f"Usage: \n"
+        f"    1. type `exit` to quit. \n"
+        f"    2. type `{image_token}` to indicate there is an image. You can enter multiple images, "
+        f"e.g '{image_token} is a dot, {image_token} is a cat, and what is it in {image_token}?'. "
+        f"When you type `{image_token}`, the chatbot will ask you to input image file path. \n"
+        f"    4. type `help` to get the help messages. \n"
+        f"    5. type `new` to start a new conversation. \n"
+        f"    Here is an example, you can type: '<image_placeholder>Describe the image.'\n"
+    )
+
+    return help_msg
+
+
+@torch.inference_mode()
+def response(
+    args, conv, pil_images, tokenizer, vl_chat_processor, vl_gpt, generation_config
+):
+    prompt = conv.get_prompt()
+    prepare_inputs = vl_chat_processor.__call__(
+        prompt=prompt, images=pil_images, force_batchify=True
+    ).to(vl_gpt.device)
+
+    # run image encoder to get the image embeddings
+    inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)
+
+    streamer = TextIteratorStreamer(
+        tokenizer=tokenizer, skip_prompt=True, skip_special_tokens=True
+    )
+    generation_config["inputs_embeds"] = inputs_embeds
+    generation_config["attention_mask"] = prepare_inputs.attention_mask
+    generation_config["streamer"] = streamer
+
+    thread = Thread(target=vl_gpt.language_model.generate, kwargs=generation_config)
+    thread.start()
+
+    yield from streamer
+
+
+def get_user_input(hint: str):
+    user_input = ""
+    while user_input == "":
+        try:
+            user_input = input(f"{hint}")
+        except KeyboardInterrupt:
+            print()
+            continue
+        except EOFError:
+            user_input = "exit"
+
+    return user_input
+
+
+def chat(args, tokenizer, vl_chat_processor, vl_gpt, generation_config):
+    image_token = vl_chat_processor.image_token
+    help_msg = get_help_message(image_token)
+
+    while True:
+        print(help_msg)
+
+        pil_images = []
+        conv = vl_chat_processor.new_chat_template()
+        roles = conv.roles
+
+        while True:
+            # get user input
+            user_input = get_user_input(
+                f"{roles[0]} [{image_token} indicates an image]: "
+            )
+
+            if user_input == "exit":
+                print("Chat program exited.")
+                sys.exit(0)
+
+            elif user_input == "help":
+                print(help_msg)
+
+            elif user_input == "new":
+                os.system("clear")
+                pil_images = []
+                conv = vl_chat_processor.new_chat_template()
+                torch.cuda.empty_cache()
+                print("New conversation started.")
+
+            else:
+                conv.append_message(conv.roles[0], user_input)
+                conv.append_message(conv.roles[1], None)
+
+                # check if the user input is an image token
+                num_images = user_input.count(image_token)
+                cur_img_idx = 0
+
+                while cur_img_idx < num_images:
+                    try:
+                        image_file = input(
+                            f"({cur_img_idx + 1}/{num_images}) Input the image file path: "
+                        )
+                        image_file = (
+                            image_file.strip()
+                        )  # trim whitespaces around path, enables drop-in from for example Dolphin
+
+                    except KeyboardInterrupt:
+                        print()
+                        continue
+
+                    except EOFError:
+                        image_file = None
+
+                    if image_file and os.path.exists(image_file):
+                        pil_image = load_image(image_file)
+                        pil_images.append(pil_image)
+                        cur_img_idx += 1
+
+                    elif image_file == "exit":
+                        print("Chat program exited.")
+                        sys.exit(0)
+
+                    else:
+                        print(
+                            f"File error, `{image_file}` does not exist. Please input the correct file path."
+                        )
+
+                # get the answer by the model's prediction
+                answer = ""
+                answer_iter = response(
+                    args,
+                    conv,
+                    pil_images,
+                    tokenizer,
+                    vl_chat_processor,
+                    vl_gpt,
+                    generation_config,
+                )
+                sys.stdout.write(f"{conv.roles[1]}: ")
+                for char in answer_iter:
+                    answer += char
+                    sys.stdout.write(char)
+                    sys.stdout.flush()
+
+                sys.stdout.write("\n")
+                sys.stdout.flush()
+                conv.update_last_message(answer)
+                # conv.messages[-1][-1] = answer
+
+
+def main(args):
+    # setup
+    tokenizer, vl_chat_processor, vl_gpt = load_pretrained_model(args.model_path)
+    generation_config = dict(
+        pad_token_id=vl_chat_processor.tokenizer.eos_token_id,
+        bos_token_id=vl_chat_processor.tokenizer.bos_token_id,
+        eos_token_id=vl_chat_processor.tokenizer.eos_token_id,
+        max_new_tokens=args.max_gen_len,
+        use_cache=True,
+    )
+    if args.temperature > 0:
+        generation_config.update(
+            {
+                "do_sample": True,
+                "top_p": args.top_p,
+                "temperature": args.temperature,
+                "repetition_penalty": args.repetition_penalty,
+            }
+        )
+    else:
+        generation_config.update({"do_sample": False})
+
+    chat(args, tokenizer, vl_chat_processor, vl_gpt, generation_config)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model_path",
+        type=str,
+        default="deepseek-ai/deepseek-vl-7b-chat",
+        help="the huggingface model name or the local path of the downloaded huggingface model.",
+    )
+    parser.add_argument("--temperature", type=float, default=0.2)
+    parser.add_argument("--top_p", type=float, default=0.95)
+    parser.add_argument("--repetition_penalty", type=float, default=1.1)
+    parser.add_argument("--max_gen_len", type=int, default=512)
+    args = parser.parse_args()
+    main(args)

deepseek_vl/__init__.py

@@ -0,0 +1,31 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+
+# check if python version is above 3.10
+import sys
+
+if sys.version_info >= (3, 10):
+    print("Python version is above 3.10, patching the collections module.")
+    # Monkey patch collections
+    import collections
+    import collections.abc
+
+    for type_name in collections.abc.__all__:
+        setattr(collections, type_name, getattr(collections.abc, type_name))

deepseek_vl/models/__init__.py

@@ -0,0 +1,28 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from .image_processing_vlm import VLMImageProcessor
+from .modeling_vlm import MultiModalityCausalLM
+from .processing_vlm import VLChatProcessor
+
+__all__ = [
+    "VLMImageProcessor",
+    "VLChatProcessor",
+    "MultiModalityCausalLM",
+]

deepseek_vl/models/clip_encoder.py

@@ -0,0 +1,242 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from typing import Dict, List, Literal, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torchvision.transforms
+from einops import rearrange
+
+from deepseek_vl.models.sam import create_sam_vit
+from deepseek_vl.models.siglip_vit import create_siglip_vit
+
+
+class CLIPVisionTower(nn.Module):
+    def __init__(
+        self,
+        model_name: str = "siglip_large_patch16_384",
+        image_size: Union[Tuple[int, int], int] = 336,
+        select_feature: str = "patch",
+        select_layer: int = -2,
+        select_layers: list = None,
+        ckpt_path: str = "",
+        pixel_mean: Optional[List[float]] = None,
+        pixel_std: Optional[List[float]] = None,
+        **kwargs,
+    ):
+        super().__init__()
+
+        self.model_name = model_name
+        self.select_feature = select_feature
+        self.select_layer = select_layer
+        self.select_layers = select_layers
+
+        vision_tower_params = {
+            "model_name": model_name,
+            "image_size": image_size,
+            "ckpt_path": ckpt_path,
+            "select_layer": select_layer,
+        }
+        vision_tower_params.update(kwargs)
+        self.vision_tower, self.forward_kwargs = self.build_vision_tower(
+            vision_tower_params
+        )
+
+        if pixel_mean is not None and pixel_std is not None:
+            image_norm = torchvision.transforms.Normalize(
+                mean=pixel_mean, std=pixel_std
+            )
+        else:
+            image_norm = None
+
+        self.image_norm = image_norm
+
+    def build_vision_tower(self, vision_tower_params):
+        if self.model_name.startswith("siglip"):
+            self.select_feature = "same"
+            vision_tower = create_siglip_vit(**vision_tower_params)
+            forward_kwargs = dict()
+
+        elif self.model_name.startswith("sam"):
+            vision_tower = create_sam_vit(**vision_tower_params)
+            forward_kwargs = dict()
+
+        else:  # huggingface
+            from transformers import CLIPVisionModel
+
+            vision_tower = CLIPVisionModel.from_pretrained(**vision_tower_params)
+            forward_kwargs = dict(output_hidden_states=True)
+
+        return vision_tower, forward_kwargs
+
+    def feature_select(self, image_forward_outs):
+        if isinstance(image_forward_outs, torch.Tensor):
+            # the output has been the self.select_layer"s features
+            image_features = image_forward_outs
+        else:
+            image_features = image_forward_outs.hidden_states[self.select_layer]
+
+        if self.select_feature == "patch":
+            # if the output has cls_token
+            image_features = image_features[:, 1:]
+        elif self.select_feature == "cls_patch":
+            image_features = image_features
+        elif self.select_feature == "same":
+            image_features = image_features
+
+        else:
+            raise ValueError(f"Unexpected select feature: {self.select_feature}")
+        return image_features
+
+    def forward(self, images):
+        """
+
+        Args:
+            images (torch.Tensor): [b, 3, H, W]
+
+        Returns:
+            image_features (torch.Tensor): [b, n_patch, d]
+        """
+
+        if self.image_norm is not None:
+            images = self.image_norm(images)
+
+        image_forward_outs = self.vision_tower(images, **self.forward_kwargs)
+        image_features = self.feature_select(image_forward_outs)
+        return image_features
+
+
+class HybridVisionTower(nn.Module):
+    def __init__(
+        self,
+        high_res_cfg: Dict,
+        low_res_cfg: Dict,
+        freeze_high: bool = False,
+        freeze_low: bool = False,
+        concat_type: Literal["feature", "sequence", "add", "tuple"] = "tuple",
+        **ignore_kwargs,
+    ):
+        super().__init__()
+
+        self.vision_tower_high = CLIPVisionTower(**high_res_cfg)
+        self.vision_tower_low = CLIPVisionTower(**low_res_cfg)
+        self.low_res_size = low_res_cfg["image_size"]
+        self.concat_type = concat_type
+
+        self.high_layer_norm = nn.LayerNorm(high_res_cfg.get("output_dim", 1024))
+        self.low_layer_norm = nn.LayerNorm(low_res_cfg.get("output_dim", 1024))
+
+        if freeze_high:
+            for p_name, p in self.vision_tower_high.named_parameters():
+                p.requires_grad = False
+            self.vision_tower_high = self.vision_tower_high.eval()
+        else:
+            # train donwsamples and neck
+            for p_name, p in self.vision_tower_high.named_parameters():
+                if "downsamples" in p_name or "neck" in p_name:
+                    p.requires_grad = True
+                else:
+                    p.requires_grad = False
+
+        if freeze_low:
+            for p in self.vision_tower_low.parameters():
+                p.requires_grad = False
+            self.vision_tower_low = self.vision_tower_low.eval()
+
+        self.resize = torchvision.transforms.Resize(self.low_res_size, antialias=True)
+
+    def forward(self, images: torch.Tensor):
+        """
+
+        Args:
+            images (torch.Tensor): [bs, 3, H, W]
+
+        Returns:
+            res (torch.Tensor): [bs, t, c]
+        """
+
+        # [bs, c, h, w]
+        high_images = images
+
+        # [bs, c, h_low, w_low]
+        low_images = self.resize(images)
+
+        # separately run two vision towers
+        # run high_res vision tower
+        high_res = self.vision_tower_high(high_images)
+        # [bs, c, h, w] -> [bs, h*w, c]
+        high_res = rearrange(high_res, "b c h w -> b (h w) c")
+        # run low_res vision tower
+        low_res = self.vision_tower_low(low_images)
+
+        if self.concat_type == "feature":
+            images_features = torch.cat([high_res, low_res], dim=-1)
+        elif self.concat_type == "sequence":
+            images_features = torch.cat([high_res, low_res], dim=1)
+        elif self.concat_type == "add":
+            images_features = high_res + low_res
+        elif self.concat_type == "tuple":
+            images_features = (high_res, low_res)
+
+        else:
+            raise ValueError(
+                "Currently only support `feature`, `sequence`, `add` and `tuple` concat type."
+            )
+
+        return images_features
+
+
+if __name__ == "__main__":
+    image_size = 1024
+    x = torch.zeros(2, 3, image_size, image_size).bfloat16().cuda()
+
+    high_res_cfg = dict(
+        model_name="sam_b_downsample",
+        select_feature="same",
+        image_size=image_size,
+        pixel_mean=(0.48145466, 0.4578275, 0.40821073),
+        pixel_std=(0.26862954, 0.26130258, 0.27577711),
+        select_layer=-1,
+        ckpt_path="",
+    )
+
+    low_res_cfg = dict(
+        model_name="siglip_large_patch16_384",
+        select_feature="same",
+        image_size=384,
+        pixel_mean=(0.5, 0.5, 0.5),
+        pixel_std=(0.5, 0.5, 0.5),
+        select_layer=-1,
+        ckpt_path="",
+    )
+
+    net = (
+        HybridVisionTower(
+            high_res_cfg=high_res_cfg,
+            low_res_cfg=low_res_cfg,
+            freeze_high=True,
+            freeze_low=True,
+            concat_type="tuple",
+        )
+        .bfloat16()
+        .cuda()
+    )
+    high_x, low_x = net(x)
+    print(x.shape, high_x.shape, low_x.shape)

deepseek_vl/models/image_processing_vlm.py

@@ -0,0 +1,208 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from typing import List, Tuple, Union
+
+import numpy as np
+import torch
+import torchvision
+import torchvision.transforms.functional
+from PIL import Image
+from transformers import AutoImageProcessor, PretrainedConfig
+from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
+from transformers.image_utils import to_numpy_array
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+ImageType = Union[np.ndarray, torch.Tensor, Image.Image]
+IMAGENET_MEAN = (0.48145466, 0.4578275, 0.40821073)
+IMAGENET_STD = (0.26862954, 0.26130258, 0.27577711)
+IMAGENET_INCEPTION_MEAN = (0.5, 0.5, 0.5)
+IMAGENET_INCEPTION_STD = (0.5, 0.5, 0.5)
+
+
+def expand2square(pil_img, background_color):
+    width, height = pil_img.size
+    if width == height:
+        return pil_img
+    elif width > height:
+        result = Image.new(pil_img.mode, (width, width), background_color)
+        result.paste(pil_img, (0, (width - height) // 2))
+        return result
+    else:
+        result = Image.new(pil_img.mode, (height, height), background_color)
+        result.paste(pil_img, ((height - width) // 2, 0))
+        return result
+
+
+class VLMImageProcessorConfig(PretrainedConfig):
+    model_type = "deepseek_vlm"
+    image_size: int
+    min_size: int
+    image_mean: Union[Tuple[float, float, float], List[float]]
+    image_std: Union[Tuple[float, float, float], List[float]]
+    rescale_factor: float
+    do_normalize: bool
+
+    def __init__(
+        self,
+        image_size: int,
+        min_size: int = 14,
+        image_mean: Union[Tuple[float, float, float], List[float]] = (
+            0.48145466,
+            0.4578275,
+            0.40821073,
+        ),
+        image_std: Union[Tuple[float, float, float], List[float]] = (
+            0.26862954,
+            0.26130258,
+            0.27577711,
+        ),
+        rescale_factor: float = 1.0 / 255.0,
+        do_normalize: bool = True,
+        **kwargs,
+    ):
+        self.image_size = image_size
+        self.min_size = min_size
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+
+        super().__init__(**kwargs)
+
+
+class VLMImageProcessor(BaseImageProcessor):
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        image_size: int,
+        min_size: int = 14,
+        image_mean: Union[Tuple[float, float, float], List[float]] = (
+            0.48145466,
+            0.4578275,
+            0.40821073,
+        ),
+        image_std: Union[Tuple[float, float, float], List[float]] = (
+            0.26862954,
+            0.26130258,
+            0.27577711,
+        ),
+        rescale_factor: float = 1.0 / 255.0,
+        do_normalize: bool = True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.image_size = image_size
+        self.rescale_factor = rescale_factor
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.min_size = min_size
+        self.do_normalize = do_normalize
+
+        if image_mean is None:
+            self.background_color = (127, 127, 127)
+        else:
+            self.background_color = tuple([int(x * 255) for x in image_mean])
+
+    def resize(self, pil_img: Image) -> np.ndarray:
+        """
+
+        Args:
+            pil_img (PIL.Image): [H, W, 3] in PIL.Image in RGB
+
+        Returns:
+            x (np.ndarray): [3, self.image_size, self.image_size]
+        """
+
+        width, height = pil_img.size
+        max_size = max(width, height)
+
+        size = [
+            max(int(height / max_size * self.image_size), self.min_size),
+            max(int(width / max_size * self.image_size), self.min_size),
+        ]
+
+        if width <= 0 or height <= 0 or size[0] <= 0 or size[1] <= 0:
+            print(f"orig size = {pil_img.size}, new size = {size}")
+            raise ValueError("Invalid size!")
+
+        pil_img = torchvision.transforms.functional.resize(
+            pil_img,
+            size,
+            interpolation=torchvision.transforms.functional.InterpolationMode.BICUBIC,
+            antialias=True,
+        )
+
+        pil_img = expand2square(pil_img, self.background_color)
+        x = to_numpy_array(pil_img)
+
+        # [H, W, 3] -> [3, H, W]
+        x = np.transpose(x, (2, 0, 1))
+
+        return x
+
+    def preprocess(self, images, return_tensors: str = "pt", **kwargs) -> BatchFeature:
+        # resize and pad to [self.image_size, self.image_size]
+        # then convert from [H, W, 3] to [3, H, W]
+        images: List[np.ndarray] = [self.resize(image) for image in images]
+
+        # resacle from [0, 255] -> [0, 1]
+        images = [
+            self.rescale(
+                image=image,
+                scale=self.rescale_factor,
+                input_data_format="channels_first",
+            )
+            for image in images
+        ]
+
+        # normalize
+        if self.do_normalize:
+            images = [
+                self.normalize(
+                    image=image,
+                    mean=self.image_mean,
+                    std=self.image_std,
+                    input_data_format="channels_first",
+                )
+                for image in images
+            ]
+
+        data = {"pixel_values": images}
+        return BatchFeature(data=data, tensor_type=return_tensors)
+
+    @property
+    def default_shape(self):
+        return [3, self.image_size, self.image_size]
+
+
+AutoImageProcessor.register(VLMImageProcessorConfig, VLMImageProcessor)
+
+
+if __name__ == "__main__":
+    image_processor = VLMImageProcessor(
+        image_size=1024,
+        image_mean=IMAGENET_INCEPTION_MEAN,
+        image_std=IMAGENET_INCEPTION_STD,
+        do_normalize=True,
+    )

deepseek_vl/models/modeling_vlm.py

@@ -0,0 +1,170 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+import torch
+from attrdict import AttrDict
+from einops import rearrange
+from transformers import (
+    AutoConfig,
+    AutoModelForCausalLM,
+    LlamaConfig,
+    LlamaForCausalLM,
+    PreTrainedModel,
+)
+from transformers.configuration_utils import PretrainedConfig
+
+from deepseek_vl.models.clip_encoder import CLIPVisionTower, HybridVisionTower
+from deepseek_vl.models.projector import MlpProjector
+
+
+def model_name_to_cls(cls_name):
+    if "MlpProjector" in cls_name:
+        cls = MlpProjector
+
+    elif "CLIPVisionTower" in cls_name:
+        cls = CLIPVisionTower
+
+    elif "HybridVisionTower" in cls_name:
+        cls = HybridVisionTower
+
+    else:
+        raise ValueError(f"class_name {cls_name} is invalid.")
+
+    return cls
+
+
+class VisionConfig(PretrainedConfig):
+    model_type = "vision"
+    cls: str = ""
+    params: AttrDict = {}
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+        self.cls = kwargs.get("cls", "")
+        if not isinstance(self.cls, str):
+            self.cls = self.cls.__name__
+
+        self.params = AttrDict(kwargs.get("params", {}))
+
+
+class AlignerConfig(PretrainedConfig):
+    model_type = "aligner"
+    cls: str = ""
+    params: AttrDict = {}
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+        self.cls = kwargs.get("cls", "")
+        if not isinstance(self.cls, str):
+            self.cls = self.cls.__name__
+
+        self.params = AttrDict(kwargs.get("params", {}))
+
+
+class MultiModalityConfig(PretrainedConfig):
+    model_type = "multi_modality"
+    vision_config: VisionConfig
+    aligner_config: AlignerConfig
+    language_config: LlamaConfig
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        vision_config = kwargs.get("vision_config", {})
+        self.vision_config = VisionConfig(**vision_config)
+
+        aligner_config = kwargs.get("aligner_config", {})
+        self.aligner_config = AlignerConfig(**aligner_config)
+
+        language_config = kwargs.get("language_config", {})
+        if isinstance(language_config, LlamaConfig):
+            self.language_config = language_config
+        else:
+            self.language_config = LlamaConfig(**language_config)
+
+
+class MultiModalityPreTrainedModel(PreTrainedModel):
+    config_class = MultiModalityConfig
+    base_model_prefix = "multi_modality"
+    _no_split_modules = []
+    _skip_keys_device_placement = "past_key_values"
+
+
+class MultiModalityCausalLM(MultiModalityPreTrainedModel):
+    def __init__(self, config: MultiModalityConfig):
+        super().__init__(config)
+
+        vision_config = config.vision_config
+        vision_cls = model_name_to_cls(vision_config.cls)
+        self.vision_model = vision_cls(**vision_config.params)
+
+        aligner_config = config.aligner_config
+        aligner_cls = model_name_to_cls(aligner_config.cls)
+        self.aligner = aligner_cls(aligner_config.params)
+
+        language_config = config.language_config
+        self.language_model = LlamaForCausalLM(language_config)
+
+    def prepare_inputs_embeds(
+        self,
+        input_ids: torch.LongTensor,
+        pixel_values: torch.FloatTensor,
+        images_seq_mask: torch.LongTensor,
+        images_emb_mask: torch.LongTensor,
+        **kwargs,
+    ):
+        """
+
+        Args:
+            input_ids (torch.LongTensor): [b, T]
+            pixel_values (torch.FloatTensor):   [b, n_images, 3, h, w]
+            images_seq_mask (torch.BoolTensor): [b, T]
+            images_emb_mask (torch.BoolTensor): [b, n_images, n_image_tokens]
+
+            assert torch.sum(images_seq_mask) == torch.sum(images_emb_mask)
+
+        Returns:
+            input_embeds (torch.Tensor): [b, T, D]
+        """
+
+        bs, n = pixel_values.shape[0:2]
+        images = rearrange(pixel_values, "b n c h w -> (b n) c h w")
+        # [b x n, T2, D]
+        images_embeds = self.aligner(self.vision_model(images))
+
+        # [b x n, T2, D] -> [b, n x T2, D]
+        images_embeds = rearrange(images_embeds, "(b n) t d -> b (n t) d", b=bs, n=n)
+        # [b, n, T2] -> [b, n x T2]
+        images_emb_mask = rearrange(images_emb_mask, "b n t -> b (n t)")
+
+        # [b, T, D]
+        input_ids[input_ids < 0] = 0  # ignore the image embeddings
+        inputs_embeds = self.language_model.get_input_embeddings()(input_ids)
+
+        # replace with the image embeddings
+        inputs_embeds[images_seq_mask] = images_embeds[images_emb_mask]
+
+        return inputs_embeds
+
+
+AutoConfig.register("vision", VisionConfig)
+AutoConfig.register("aligner", AlignerConfig)
+AutoConfig.register("multi_modality", MultiModalityConfig)
+AutoModelForCausalLM.register(MultiModalityConfig, MultiModalityCausalLM)

deepseek_vl/models/processing_vlm.py

@@ -0,0 +1,390 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from dataclasses import dataclass
+from typing import Dict, List
+
+import torch
+from PIL.Image import Image
+from transformers import LlamaTokenizerFast
+from transformers.processing_utils import ProcessorMixin
+
+from deepseek_vl.models.image_processing_vlm import VLMImageProcessor
+from deepseek_vl.utils.conversation import get_conv_template
+
+
+class DictOutput(object):
+    def keys(self):
+        return self.__dict__.keys()
+
+    def __getitem__(self, item):
+        return self.__dict__[item]
+
+    def __setitem__(self, key, value):
+        self.__dict__[key] = value
+
+
+@dataclass
+class VLChatProcessorOutput(DictOutput):
+    sft_format: str
+    input_ids: torch.Tensor
+    pixel_values: torch.Tensor
+    num_image_tokens: torch.IntTensor
+
+    def __len__(self):
+        return len(self.input_ids)
+
+
+@dataclass
+class BatchedVLChatProcessorOutput(DictOutput):
+    sft_format: List[str]
+    input_ids: torch.Tensor
+    pixel_values: torch.Tensor
+    attention_mask: torch.Tensor
+    images_seq_mask: torch.BoolTensor
+    images_emb_mask: torch.BoolTensor
+
+    def to(self, device, dtype=torch.bfloat16):
+        self.input_ids = self.input_ids.to(device)
+        self.attention_mask = self.attention_mask.to(device)
+        self.images_seq_mask = self.images_seq_mask.to(device)
+        self.images_emb_mask = self.images_emb_mask.to(device)
+        self.pixel_values = self.pixel_values.to(device=device, dtype=dtype)
+        return self
+
+
+class VLChatProcessor(ProcessorMixin):
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = ("LlamaTokenizer", "LlamaTokenizerFast")
+
+    attributes = ["image_processor", "tokenizer"]
+
+    system_prompt = (
+        "You are a helpful language and vision assistant. "
+        "You are able to understand the visual content that the user provides, "
+        "and assist the user with a variety of tasks using natural language."
+    )
+
+    def __init__(
+        self,
+        image_processor: VLMImageProcessor,
+        tokenizer: LlamaTokenizerFast,
+        image_tag: str = "<image_placeholder>",
+        num_image_tokens: int = 576,
+        add_special_token: bool = False,
+        sft_format: str = "deepseek",
+        mask_prompt: bool = True,
+        ignore_id: int = -100,
+        **kwargs,
+    ):
+        self.image_processor = image_processor
+        self.tokenizer = tokenizer
+
+        image_id = self.tokenizer.vocab.get(image_tag)
+        if image_id is None:
+            special_tokens = [image_tag]
+            special_tokens_dict = {"additional_special_tokens": special_tokens}
+            self.tokenizer.add_special_tokens(special_tokens_dict)
+            print(f"Add image tag = {image_tag} to the tokenizer")
+
+        self.image_tag = image_tag
+        self.num_image_tokens = num_image_tokens
+        self.add_special_token = add_special_token
+        self.sft_format = sft_format
+        self.mask_prompt = mask_prompt
+        self.ignore_id = ignore_id
+
+        super().__init__(
+            image_processor,
+            tokenizer,
+            image_tag,
+            num_image_tokens,
+            add_special_token,
+            sft_format,
+            mask_prompt,
+            ignore_id,
+            **kwargs,
+        )
+
+    def new_chat_template(self):
+        conv = get_conv_template(self.sft_format)
+        conv.set_system_message(self.system_prompt)
+        return conv
+
+    def apply_sft_template_for_multi_turn_prompts(
+        self,
+        conversations: List[Dict[str, str]],
+        sft_format: str = "deepseek",
+        system_prompt: str = "",
+    ):
+        """
+        Applies the SFT template to conversation.
+
+        An example of conversation:
+        conversation = [
+            {
+                "role": "User",
+                "content": "<image_placeholder> is Figure 1.\n<image_placeholder> is Figure 2.\nWhich image is brighter?",
+                "images": [
+                    "./multi-images/attribute_comparison_1.png",
+                    "./multi-images/attribute_comparison_2.png"
+                ]
+            },
+            {
+                "role": "Assistant",
+                "content": ""
+            }
+        ]
+
+        Args:
+            conversations (List[Dict]): A conversation with a List of Dict[str, str] text.
+            sft_format (str, optional): The format of the SFT template to use. Defaults to "deepseek".
+            system_prompt (str, optional): The system prompt to use in the SFT template. Defaults to "".
+
+        Returns:
+            sft_prompt (str): The formatted text.
+        """
+
+        conv = get_conv_template(sft_format)
+        conv.set_system_message(system_prompt)
+        for message in conversations:
+            conv.append_message(message["role"], message["content"].strip())
+        sft_prompt = conv.get_prompt().strip()
+
+        return sft_prompt
+
+    @property
+    def image_token(self):
+        return self.image_tag
+
+    @property
+    def image_id(self):
+        image_id = self.tokenizer.vocab.get(self.image_tag)
+        return image_id
+
+    @property
+    def pad_id(self):
+        pad_id = self.tokenizer.pad_token_id
+        if pad_id is None:
+            pad_id = self.tokenizer.eos_token_id
+
+        return pad_id
+
+    def add_image_token(
+        self,
+        image_indices: List[int],
+        input_ids: torch.LongTensor,
+    ):
+        """
+
+        Args:
+            image_indices (List[int]): [index_0, index_1, ..., index_j]
+            input_ids (torch.LongTensor): [N]
+
+        Returns:
+            input_ids (torch.LongTensor): [N + image tokens]
+            num_image_tokens (torch.IntTensor): [n_images]
+        """
+
+        input_slices = []
+
+        start = 0
+        for index in image_indices:
+            if self.add_special_token:
+                end = index + 1
+            else:
+                end = index
+
+            # original text tokens
+            input_slices.append(input_ids[start:end])
+
+            # add image tokens, and set the mask as False
+            input_slices.append(
+                self.image_id * torch.ones((self.num_image_tokens,), dtype=torch.long)
+            )
+            start = index + 1
+
+        # the left part
+        input_slices.append(input_ids[start:])
+
+        # concat all slices
+        input_ids = torch.cat(input_slices, dim=0)
+        num_image_tokens = torch.IntTensor([self.num_image_tokens] * len(image_indices))
+
+        return input_ids, num_image_tokens
+
+    def process_one(
+        self,
+        prompt: str = None,
+        conversations: List[Dict[str, str]] = None,
+        images: List[Image] = None,
+        **kwargs,
+    ):
+        """
+
+        Args:
+            prompt (str): the formatted prompt;
+            conversations (List[Dict]): conversations with a list of messages;
+            images (List[ImageType]): the list of images;
+            **kwargs:
+
+        Returns:
+            outputs (BaseProcessorOutput): the output of the processor,
+                - input_ids (torch.LongTensor): [N + image tokens]
+                - target_ids (torch.LongTensor): [N + image tokens]
+                - images (torch.FloatTensor): [n_images, 3, H, W]
+                - image_id (int): the id of the image token
+                - num_image_tokens (List[int]): the number of image tokens
+        """
+
+        assert (
+            prompt is None or conversations is None
+        ), "prompt and conversations cannot be used at the same time."
+
+        if prompt is None:
+            # apply sft format
+            sft_format = self.apply_sft_template_for_multi_turn_prompts(
+                conversations=conversations,
+                sft_format=self.sft_format,
+                system_prompt=self.system_prompt,
+            )
+        else:
+            sft_format = prompt
+
+        # tokenize
+        input_ids = self.tokenizer.encode(sft_format)
+        input_ids = torch.LongTensor(input_ids)
+
+        # add image tokens to the input_ids
+        image_token_mask: torch.BoolTensor = input_ids == self.image_id
+        image_indices = image_token_mask.nonzero()
+        input_ids, num_image_tokens = self.add_image_token(
+            image_indices=image_indices,
+            input_ids=input_ids,
+        )
+
+        # load images
+        images_outputs = self.image_processor(images, return_tensors="pt")
+
+        prepare = VLChatProcessorOutput(
+            sft_format=sft_format,
+            input_ids=input_ids,
+            pixel_values=images_outputs.pixel_values,
+            num_image_tokens=num_image_tokens,
+        )
+
+        return prepare
+
+    def __call__(
+        self,
+        *,
+        prompt: str = None,
+        conversations: List[Dict[str, str]] = None,
+        images: List[Image] = None,
+        force_batchify: bool = True,
+        **kwargs,
+    ):
+        """
+
+        Args:
+            prompt (str): the formatted prompt;
+            conversations (List[Dict]): conversations with a list of messages;
+            images (List[ImageType]): the list of images;
+            force_batchify (bool): force batchify the inputs;
+            **kwargs:
+
+        Returns:
+            outputs (BaseProcessorOutput): the output of the processor,
+                - input_ids (torch.LongTensor): [N + image tokens]
+                - images (torch.FloatTensor): [n_images, 3, H, W]
+                - image_id (int): the id of the image token
+                - num_image_tokens (List[int]): the number of image tokens
+        """
+
+        prepare = self.process_one(
+            prompt=prompt, conversations=conversations, images=images
+        )
+
+        if force_batchify:
+            prepare = self.batchify([prepare])
+
+        return prepare
+
+    def batchify(
+        self, prepare_list: List[VLChatProcessorOutput]
+    ) -> BatchedVLChatProcessorOutput:
+        """
+        Preprocesses the inputs for multimodal inference.
+
+        Args:
+            prepare_list (List[VLChatProcessorOutput]): A list of VLChatProcessorOutput.
+
+        Returns:
+            BatchedVLChatProcessorOutput: A dictionary of the inputs to use for multimodal inference.
+        """
+
+        batch_size = len(prepare_list)
+        sft_format = []
+        n_images = []
+        seq_lens = []
+        for prepare in prepare_list:
+            n_images.append(len(prepare.num_image_tokens))
+            seq_lens.append(len(prepare))
+
+        input_token_max_len = max(seq_lens)
+        max_n_images = max(1, max(n_images))
+
+        batched_input_ids = torch.full(
+            (batch_size, input_token_max_len), self.pad_id
+        ).long()  # FIXME
+        batched_attention_mask = torch.zeros((batch_size, input_token_max_len)).long()
+        batched_pixel_values = torch.zeros(
+            (batch_size, max_n_images, *self.image_processor.default_shape)
+        ).float()
+        batched_images_seq_mask = torch.zeros((batch_size, input_token_max_len)).bool()
+        batched_images_emb_mask = torch.zeros(
+            (batch_size, max_n_images, self.num_image_tokens)
+        ).bool()
+
+        for i, prepare in enumerate(prepare_list):
+            input_ids = prepare.input_ids
+            seq_len = len(prepare)
+            n_image = len(prepare.num_image_tokens)
+            # left-padding
+            batched_attention_mask[i, -seq_len:] = 1
+            batched_input_ids[i, -seq_len:] = torch.LongTensor(input_ids)
+            batched_images_seq_mask[i, -seq_len:] = input_ids == self.image_id
+
+            if n_image > 0:
+                batched_pixel_values[i, :n_image] = prepare.pixel_values
+                for j, n_image_tokens in enumerate(prepare.num_image_tokens):
+                    batched_images_emb_mask[i, j, :n_image_tokens] = True
+
+            sft_format.append(prepare.sft_format)
+
+        batched_prepares = BatchedVLChatProcessorOutput(
+            input_ids=batched_input_ids,
+            attention_mask=batched_attention_mask,
+            pixel_values=batched_pixel_values,
+            images_seq_mask=batched_images_seq_mask,
+            images_emb_mask=batched_images_emb_mask,
+            sft_format=sft_format,
+        )
+
+        return batched_prepares

deepseek_vl/models/projector.py

@@ -0,0 +1,100 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from typing import Tuple, Union
+
+import torch
+import torch.nn as nn
+from attrdict import AttrDict
+
+
+class MlpProjector(nn.Module):
+    def __init__(self, cfg):
+        super().__init__()
+
+        self.cfg = cfg
+
+        if cfg.projector_type == "identity":
+            modules = nn.Identity()
+
+        elif cfg.projector_type == "linear":
+            modules = nn.Linear(cfg.input_dim, cfg.n_embed)
+
+        elif cfg.projector_type == "mlp_gelu":
+            mlp_depth = cfg.get("depth", 1)
+            modules = [nn.Linear(cfg.input_dim, cfg.n_embed)]
+            for _ in range(1, mlp_depth):
+                modules.append(nn.GELU())
+                modules.append(nn.Linear(cfg.n_embed, cfg.n_embed))
+            modules = nn.Sequential(*modules)
+
+        elif cfg.projector_type == "low_high_hybrid_split_mlp_gelu":
+            mlp_depth = cfg.get("depth", 1)
+            self.high_up_proj = nn.Linear(cfg.input_dim, cfg.n_embed // 2)
+            self.low_up_proj = nn.Linear(cfg.input_dim, cfg.n_embed // 2)
+
+            modules = []
+            for _ in range(1, mlp_depth):
+                modules.append(nn.GELU())
+                modules.append(nn.Linear(cfg.n_embed, cfg.n_embed))
+            modules = nn.Sequential(*modules)
+
+        else:
+            raise ValueError(f"Unknown projector type: {cfg.projector_type}")
+
+        self.layers = modules
+
+    def forward(
+        self, x_or_tuple: Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor]
+    ):
+        """
+
+        Args:
+            x_or_tuple (Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor]:  if it is a tuple of torch.Tensor,
+                then it comes from the hybrid vision encoder, and x = high_res_x, low_res_x);
+                otherwise it is the feature from the single vision encoder.
+
+        Returns:
+            x (torch.Tensor): [b, s, c]
+        """
+
+        if isinstance(x_or_tuple, tuple):
+            # self.cfg.projector_type == "low_high_hybrid_split_mlp_gelu":
+            high_x, low_x = x_or_tuple
+            high_x = self.high_up_proj(high_x)
+            low_x = self.low_up_proj(low_x)
+            x = torch.concat([high_x, low_x], dim=-1)
+        else:
+            x = x_or_tuple
+
+        return self.layers(x)
+
+
+if __name__ == "__main__":
+    cfg = AttrDict(
+        input_dim=1024,
+        n_embed=2048,
+        depth=2,
+        projector_type="low_high_hybrid_split_mlp_gelu",
+    )
+    inputs = (torch.rand(4, 576, 1024), torch.rand(4, 576, 1024))
+
+    m = MlpProjector(cfg)
+    out = m(inputs)
+    print(out.shape)

deepseek_vl/models/sam.py

@@ -0,0 +1,593 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import copy
+from dataclasses import dataclass
+from functools import partial
+from typing import List, Optional, Tuple, Type, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class MLPBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        mlp_dim: int,
+        act: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        super().__init__()
+        self.lin1 = nn.Linear(embedding_dim, mlp_dim)
+        self.lin2 = nn.Linear(mlp_dim, embedding_dim)
+        self.act = act()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.lin2(self.act(self.lin1(x)))
+
+
+# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
+# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119  # noqa
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x
+
+
+# This class and its supporting functions below lightly adapted from the ViTDet backbone available at: https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/vit.py # noqa
+class ImageEncoderViT(nn.Module):
+    def __init__(
+        self,
+        img_size: int = 1024,
+        patch_size: int = 16,
+        in_chans: int = 3,
+        embed_dim: int = 768,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        out_chans: int = 256,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_abs_pos: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        global_attn_indexes: Tuple[int, ...] = (),
+        downsample_channels: Tuple[int, ...] = (512, 1024),
+    ) -> None:
+        """
+        Args:
+            img_size (int): Input image size.
+            patch_size (int): Patch size.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+            depth (int): Depth of ViT.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_abs_pos (bool): If True, use absolute positional embeddings.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks.
+            global_attn_indexes (list): Indexes for blocks using global attention.
+            downsample_channels (list): Channels for downsampling layers.
+        """
+        super().__init__()
+        self.img_size = img_size
+
+        self.patch_embed = PatchEmbed(
+            kernel_size=(patch_size, patch_size),
+            stride=(patch_size, patch_size),
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+        )
+
+        self.pos_embed: Optional[nn.Parameter] = None
+        if use_abs_pos:
+            # Initialize absolute positional embedding with pretrain image size.
+            self.pos_embed = nn.Parameter(
+                torch.zeros(
+                    1, img_size // patch_size, img_size // patch_size, embed_dim
+                )
+            )
+
+        self.blocks = nn.ModuleList()
+        for i in range(depth):
+            block = Block(
+                dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                norm_layer=norm_layer,
+                act_layer=act_layer,
+                use_rel_pos=use_rel_pos,
+                rel_pos_zero_init=rel_pos_zero_init,
+                window_size=window_size if i not in global_attn_indexes else 0,
+                input_size=(img_size // patch_size, img_size // patch_size),
+            )
+            self.blocks.append(block)
+
+        self.neck = nn.Sequential(
+            nn.Conv2d(
+                embed_dim,
+                out_chans,
+                kernel_size=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+            nn.Conv2d(
+                out_chans,
+                out_chans,
+                kernel_size=3,
+                padding=1,
+                bias=False,
+            ),
+            LayerNorm2d(out_chans),
+        )
+
+        in_channels = out_chans
+        downsamples = []
+        for i in range(len(downsample_channels)):
+            out_channels = downsample_channels[i]
+            downsamples.append(
+                nn.Conv2d(
+                    in_channels,
+                    out_channels,
+                    kernel_size=3,
+                    stride=2,
+                    padding=1,
+                    bias=False,
+                )
+            )
+            in_channels = out_channels
+        self.downsamples = nn.Sequential(*downsamples)
+
+        self.sam_hd = True
+        if self.sam_hd:
+            self.hd_alpha_downsamples = nn.Parameter(torch.zeros(1))
+            # self.neck_hd = nn.Linear(embed_dim, embed_dim)
+            self.neck_hd = copy.deepcopy(self.neck)
+            # self.downsamples_hd = copy.deepcopy(self.downsamples)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.patch_embed(x)
+        if self.pos_embed is not None:
+            x = x + self.pos_embed
+
+        global_features = []
+        for i, blk in enumerate(self.blocks):
+            x = blk(x)
+            if self.sam_hd and blk.window_size == 0:
+                global_features.append(x)
+
+        x = self.neck(x.permute(0, 3, 1, 2))
+        x_dtype = x.dtype
+        x = F.interpolate(
+            x.float(), size=(96, 96), mode="bilinear", align_corners=False
+        ).to(x_dtype)
+        x = self.downsamples(x)
+
+        if self.sam_hd:
+            first_global_feature = self.neck_hd(global_features[0].permute(0, 3, 1, 2))
+            x_dtype = first_global_feature.dtype
+            first_global_feature = F.interpolate(
+                first_global_feature.float(),
+                size=(96, 96),
+                mode="bilinear",
+                align_corners=False,
+            )
+            first_global_feature = self.downsamples(first_global_feature.to(x_dtype))
+            x = x + first_global_feature * self.hd_alpha_downsamples
+
+        return x
+
+
+class Block(nn.Module):
+    """Transformer blocks with support of window attention and residual propagation blocks"""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        norm_layer: Type[nn.Module] = nn.LayerNorm,
+        act_layer: Type[nn.Module] = nn.GELU,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        window_size: int = 0,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads in each ViT block.
+            mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+            qkv_bias (bool): If True, add a learnable bias to query, key, value.
+            norm_layer (nn.Module): Normalization layer.
+            act_layer (nn.Module): Activation layer.
+            use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            window_size (int): Window size for window attention blocks. If it equals 0, then
+                use global attention.
+            input_size (tuple(int, int) or None): Input resolution for calculating the relative
+                positional parameter size.
+        """
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            use_rel_pos=use_rel_pos,
+            rel_pos_zero_init=rel_pos_zero_init,
+            input_size=input_size if window_size == 0 else (window_size, window_size),
+        )
+
+        self.norm2 = norm_layer(dim)
+        self.mlp = MLPBlock(
+            embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer
+        )
+
+        self.window_size = window_size
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shortcut = x
+        x = self.norm1(x)
+        # Window partition
+        if self.window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, self.window_size)
+
+        x = self.attn(x)
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, self.window_size, pad_hw, (H, W))
+
+        x = shortcut + x
+        x = x + self.mlp(self.norm2(x))
+
+        return x
+
+
+class Attention(nn.Module):
+    """Multi-head Attention block with relative position embeddings."""
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = True,
+        use_rel_pos: bool = False,
+        rel_pos_zero_init: bool = True,
+        input_size: Optional[Tuple[int, int]] = None,
+    ) -> None:
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads.
+            qkv_bias (bool):  If True, add a learnable bias to query, key, value.
+            rel_pos (bool): If True, add relative positional embeddings to the attention map.
+            rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
+            input_size (tuple(int, int) or None): Input resolution for calculating the relative
+                positional parameter size.
+        """
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.proj = nn.Linear(dim, dim)
+
+        self.use_rel_pos = use_rel_pos
+        if self.use_rel_pos:
+            assert (
+                input_size is not None
+            ), "Input size must be provided if using relative positional encoding."
+            # initialize relative positional embeddings
+            self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
+            self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, H, W, _ = x.shape
+        # qkv with shape (3, B, nHead, H * W, C)
+        qkv = (
+            self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        )
+        # q, k, v with shape (B * nHead, H * W, C)
+        q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)
+
+        def do_attention(q, k, v):
+            attn = (q * self.scale) @ k.transpose(-2, -1)
+            if self.use_rel_pos:
+                attn = add_decomposed_rel_pos(
+                    attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W)
+                )
+
+            attn = attn.softmax(dim=-1)
+            x = (
+                (attn @ v)
+                .view(B, self.num_heads, H, W, -1)
+                .permute(0, 2, 3, 1, 4)
+                .reshape(B, H, W, -1)
+            )
+
+            return x
+
+        # from haiscale.utils import on_demand_checkpoint
+        # x = on_demand_checkpoint(do_attention, q, k, v)
+        x = do_attention(q, k, v)
+        x = self.proj(x)
+
+        return x
+
+
+def window_partition(
+    x: torch.Tensor, window_size: int
+) -> Tuple[torch.Tensor, Tuple[int, int]]:
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = (
+        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    )
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(
+    windows: torch.Tensor,
+    window_size: int,
+    pad_hw: Tuple[int, int],
+    hw: Tuple[int, int],
+) -> torch.Tensor:
+    """
+    Window unpartition into original sequences and removing padding.
+    Args:
+        windows (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(
+        B, Hp // window_size, Wp // window_size, window_size, window_size, -1
+    )
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
+    """
+    Get relative positional embeddings according to the relative positions of
+        query and key sizes.
+    Args:
+        q_size (int): size of query q.
+        k_size (int): size of key k.
+        rel_pos (Tensor): relative position embeddings (L, C).
+
+    Returns:
+        Extracted positional embeddings according to relative positions.
+    """
+    max_rel_dist = int(2 * max(q_size, k_size) - 1)
+    # Interpolate rel pos if needed.
+    if rel_pos.shape[0] != max_rel_dist:
+        # Interpolate rel pos.
+        rel_pos_resized = F.interpolate(
+            rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
+            size=max_rel_dist,
+            mode="linear",
+        )
+        rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
+    else:
+        rel_pos_resized = rel_pos
+
+    # Scale the coords with short length if shapes for q and k are different.
+    q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
+    k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
+    relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
+
+    return rel_pos_resized[relative_coords.long()]
+
+
+def add_decomposed_rel_pos(
+    attn: torch.Tensor,
+    q: torch.Tensor,
+    rel_pos_h: torch.Tensor,
+    rel_pos_w: torch.Tensor,
+    q_size: Tuple[int, int],
+    k_size: Tuple[int, int],
+) -> torch.Tensor:
+    """
+    Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
+    https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py   # noqa B950
+    Args:
+        attn (Tensor): attention map.
+        q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
+        rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
+        rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
+        q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
+        k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
+
+    Returns:
+        attn (Tensor): attention map with added relative positional embeddings.
+    """
+    q_h, q_w = q_size
+    k_h, k_w = k_size
+    Rh = get_rel_pos(q_h, k_h, rel_pos_h)
+    Rw = get_rel_pos(q_w, k_w, rel_pos_w)
+
+    B, _, dim = q.shape
+    r_q = q.reshape(B, q_h, q_w, dim)
+    rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
+    rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
+
+    attn = (
+        attn.view(B, q_h, q_w, k_h, k_w)
+        + rel_h[:, :, :, :, None]
+        + rel_w[:, :, :, None, :]
+    ).view(B, q_h * q_w, k_h * k_w)
+
+    return attn
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, int] = (16, 16),
+        stride: Tuple[int, int] = (16, 16),
+        padding: Tuple[int, int] = (0, 0),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ) -> None:
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x
+
+
+@dataclass
+class SAMViTCfg:
+    image_size: Union[Tuple[int, int], int] = 1024
+    width: int = 1024
+    layers: int = 23
+    heads: int = 16
+    patch_size: int = 16
+    window_size: int = 14
+    prompt_embed_dim: int = 256
+    global_attn_indexes: Union[List[int], Tuple[int]] = (5, 11, 17, 23)
+    downsample_channels: Union[List[int], Tuple[int]] = (512, 1024)
+
+
+SAM_MODEL_CONFIG = {
+    "sam_vit_b": {
+        "width": 768,
+        "layers": 12,
+        "heads": 12,
+        "global_attn_indexes": [2, 5, 8, 11],
+        "downsample_channels": (),
+    },
+    "sam_b_downsample": {
+        "width": 768,
+        "layers": 12,
+        "heads": 12,
+        "global_attn_indexes": [2, 5, 8, 11],
+        "downsample_channels": (512, 1024),
+    },
+    "sam_vit_l": {
+        "width": 1024,
+        "layers": 24,
+        "heads": 16,
+        "global_attn_indexes": [5, 11, 17, 23],
+        "downsample_channels": (),
+    },
+    "sam_vit_h": {
+        "width": 1280,
+        "layers": 32,
+        "heads": 16,
+        "global_attn_indexes": [7, 15, 23, 31],
+        "downsample_channels": (),
+    },
+}
+
+
+def create_sam_vit(
+    model_name: str = "sam_b_downsample",
+    image_size: int = 1024,
+    ckpt_path: str = "",
+    **kwargs,
+):
+    assert (
+        model_name in SAM_MODEL_CONFIG.keys()
+    ), f"model name: {model_name} should be in {SAM_MODEL_CONFIG.keys()}"
+
+    sam_cfg = SAMViTCfg(**SAM_MODEL_CONFIG[model_name])
+    image_encoder = ImageEncoderViT(
+        depth=sam_cfg.layers,
+        embed_dim=sam_cfg.width,
+        img_size=image_size,
+        mlp_ratio=4,
+        norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
+        num_heads=sam_cfg.heads,
+        patch_size=sam_cfg.patch_size,
+        qkv_bias=True,
+        use_rel_pos=True,
+        global_attn_indexes=sam_cfg.global_attn_indexes,
+        window_size=14,
+        out_chans=sam_cfg.prompt_embed_dim,
+        downsample_channels=sam_cfg.downsample_channels,
+    )
+
+    if ckpt_path:
+        state_dict = torch.load(ckpt_path)
+        image_encoder.load_state_dict(state_dict, strict=False)
+        print(f"SAM-ViT restores from {ckpt_path}")
+
+    return image_encoder
+
+
+if __name__ == "__main__":
+    x = torch.zeros(2, 3, 1024, 1024).bfloat16()
+    # x.permute(0, 3, 1, 2)
+    net = create_sam_vit().bfloat16()
+    out = net(x)
+    print(x.shape, out.shape)

deepseek_vl/models/siglip_vit.py

@@ -0,0 +1,681 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+# https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/vision_transformer.py
+import math
+import warnings
+from dataclasses import dataclass
+from functools import partial
+from typing import (
+    Callable,
+    Dict,
+    Final,
+    List,
+    Literal,
+    Optional,
+    Sequence,
+    Set,
+    Tuple,
+    Type,
+    Union,
+)
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from timm.layers import (
+    AttentionPoolLatent,
+    DropPath,
+    LayerType,
+    Mlp,
+    PatchDropout,
+    PatchEmbed,
+    resample_abs_pos_embed,
+)
+from timm.models._manipulate import checkpoint_seq, named_apply
+
+
+def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    with torch.no_grad():
+        # Values are generated by using a truncated uniform distribution and
+        # then using the inverse CDF for the normal distribution.
+        # Get upper and lower cdf values
+        l = norm_cdf((a - mean) / std)  # noqa: E741
+        u = norm_cdf((b - mean) / std)
+
+        # Uniformly fill tensor with values from [l, u], then translate to
+        # [2l-1, 2u-1].
+        tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+        # Use inverse cdf transform for normal distribution to get truncated
+        # standard normal
+        tensor.erfinv_()
+
+        # Transform to proper mean, std
+        tensor.mul_(std * math.sqrt(2.0))
+        tensor.add_(mean)
+
+        # Clamp to ensure it's in the proper range
+        tensor.clamp_(min=a, max=b)
+        return tensor
+
+
+def trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):
+    # type: (torch.Tensor, float, float, float, float) -> torch.Tensor
+    r"""The original timm.models.layers.weight_init.trunc_normal_ can not handle bfloat16 yet, here we first
+    convert the tensor to float32, apply the trunc_normal_() in float32, and then convert it back to its orignal dtype.
+    Fills the input Tensor with values drawn from a truncated normal distribution. The values are effectively drawn
+    from the normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \leq \text{mean} \leq b`.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    Examples:
+        >>> w = torch.empty(3, 5)
+        >>> nn.init.trunc_normal_(w)
+    """
+
+    with torch.no_grad():
+        dtype = tensor.dtype
+        tensor_fp32 = tensor.float()
+        tensor_fp32 = _no_grad_trunc_normal_(tensor_fp32, mean, std, a, b)
+        tensor_dtype = tensor_fp32.to(dtype=dtype)
+        tensor.copy_(tensor_dtype)
+
+
+def init_weights(self):
+    if self.pos_embed is not None:
+        trunc_normal_(self.pos_embed, std=self.pos_embed.shape[1] ** -0.5)
+    trunc_normal_(self.latent, std=self.latent_dim**-0.5)
+
+
+def init_weights_vit_timm(module: nn.Module, name: str = "") -> None:
+    """ViT weight initialization, original timm impl (for reproducibility)"""
+    if isinstance(module, nn.Linear):
+        trunc_normal_(module.weight, std=0.02)
+        if module.bias is not None:
+            nn.init.zeros_(module.bias)
+    elif hasattr(module, "init_weights"):
+        module.init_weights()
+
+
+class Attention(nn.Module):
+    fused_attn: Final[bool]
+
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int = 8,
+        qkv_bias: bool = False,
+        qk_norm: bool = False,
+        attn_drop: float = 0.0,
+        proj_drop: float = 0.0,
+        norm_layer: nn.Module = nn.LayerNorm,
+    ) -> None:
+        super().__init__()
+        assert dim % num_heads == 0, "dim should be divisible by num_heads"
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim**-0.5
+        # self.fused_attn = use_fused_attn()
+        self.fused_attn = True
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop) if proj_drop > 0.0 else nn.Identity()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, N, C = x.shape
+        qkv = (
+            self.qkv(x)
+            .reshape(B, N, 3, self.num_heads, self.head_dim)
+            .permute(2, 0, 3, 1, 4)
+        )
+        q, k, v = qkv.unbind(0)
+        q, k = self.q_norm(q), self.k_norm(k)
+
+        if self.fused_attn:
+            x = F.scaled_dot_product_attention(
+                q,
+                k,
+                v,
+                dropout_p=self.attn_drop.p if self.training else 0.0,
+            )
+        else:
+            q = q * self.scale
+            attn = q @ k.transpose(-2, -1)
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+            x = attn @ v
+
+        x = x.transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class LayerScale(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        init_values: float = 1e-5,
+        inplace: bool = False,
+    ) -> None:
+        super().__init__()
+        self.inplace = inplace
+        self.gamma = nn.Parameter(init_values * torch.ones(dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x.mul_(self.gamma) if self.inplace else x * self.gamma
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = False,
+        qk_norm: bool = False,
+        proj_drop: float = 0.0,
+        attn_drop: float = 0.0,
+        init_values: Optional[float] = None,
+        drop_path: float = 0.0,
+        act_layer: nn.Module = nn.GELU,
+        norm_layer: nn.Module = nn.LayerNorm,
+        mlp_layer: nn.Module = Mlp,
+    ) -> None:
+        super().__init__()
+        self.norm1 = norm_layer(dim)
+        self.attn = Attention(
+            dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            attn_drop=attn_drop,
+            proj_drop=proj_drop,
+            norm_layer=norm_layer,
+        )
+        self.ls1 = (
+            LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+        )
+        self.drop_path1 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        self.norm2 = norm_layer(dim)
+        self.mlp = mlp_layer(
+            in_features=dim,
+            hidden_features=int(dim * mlp_ratio),
+            act_layer=act_layer,
+            drop=proj_drop,
+        )
+        self.ls2 = (
+            LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+        )
+        self.drop_path2 = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x + self.drop_path1(self.ls1(self.attn(self.norm1(x))))
+        x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
+        return x
+
+
+class VisionTransformer(nn.Module):
+    """Vision Transformer
+
+    A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale`
+        - https://arxiv.org/abs/2010.11929
+    """
+
+    dynamic_img_size: Final[bool]
+
+    def __init__(
+        self,
+        img_size: Union[int, Tuple[int, int]] = 224,
+        patch_size: Union[int, Tuple[int, int]] = 16,
+        in_chans: int = 3,
+        num_classes: int = 1000,
+        global_pool: Literal["", "avg", "token", "map"] = "token",
+        embed_dim: int = 768,
+        depth: int = 12,
+        num_heads: int = 12,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        init_values: Optional[float] = None,
+        class_token: bool = True,
+        no_embed_class: bool = False,
+        reg_tokens: int = 0,
+        pre_norm: bool = False,
+        fc_norm: Optional[bool] = None,
+        dynamic_img_size: bool = False,
+        dynamic_img_pad: bool = False,
+        drop_rate: float = 0.0,
+        pos_drop_rate: float = 0.0,
+        patch_drop_rate: float = 0.0,
+        proj_drop_rate: float = 0.0,
+        attn_drop_rate: float = 0.0,
+        drop_path_rate: float = 0.0,
+        weight_init: Literal["skip", "jax", "jax_nlhb", "moco", ""] = "",
+        embed_layer: Callable = PatchEmbed,
+        norm_layer: Optional[LayerType] = None,
+        act_layer: Optional[LayerType] = None,
+        block_fn: Type[nn.Module] = Block,
+        mlp_layer: Type[nn.Module] = Mlp,
+        ignore_head: bool = False,
+    ) -> None:
+        """
+        Args:
+            img_size: Input image size.
+            patch_size: Patch size.
+            in_chans: Number of image input channels.
+            num_classes: Mumber of classes for classification head.
+            global_pool: Type of global pooling for final sequence (default: 'token').
+            embed_dim: Transformer embedding dimension.
+            depth: Depth of transformer.
+            num_heads: Number of attention heads.
+            mlp_ratio: Ratio of mlp hidden dim to embedding dim.
+            qkv_bias: Enable bias for qkv projections if True.
+            init_values: Layer-scale init values (layer-scale enabled if not None).
+            class_token: Use class token.
+            no_embed_class: Don't include position embeddings for class (or reg) tokens.
+            reg_tokens: Number of register tokens.
+            fc_norm: Pre head norm after pool (instead of before), if None, enabled when global_pool == 'avg'.
+            drop_rate: Head dropout rate.
+            pos_drop_rate: Position embedding dropout rate.
+            attn_drop_rate: Attention dropout rate.
+            drop_path_rate: Stochastic depth rate.
+            weight_init: Weight initialization scheme.
+            embed_layer: Patch embedding layer.
+            norm_layer: Normalization layer.
+            act_layer: MLP activation layer.
+            block_fn: Transformer block layer.
+        """
+        super().__init__()
+        assert global_pool in ("", "avg", "token", "map")
+        assert class_token or global_pool != "token"
+        use_fc_norm = global_pool == "avg" if fc_norm is None else fc_norm
+        # norm_layer = get_norm_layer(norm_layer) or partial(nn.LayerNorm, eps=1e-6)
+        # act_layer = get_act_layer(act_layer) or nn.GELU
+        norm_layer = partial(nn.LayerNorm, eps=1e-6)
+        act_layer = nn.GELU
+
+        self.num_classes = num_classes
+        self.global_pool = global_pool
+        self.num_features = self.embed_dim = (
+            embed_dim  # num_features for consistency with other models
+        )
+        self.num_prefix_tokens = 1 if class_token else 0
+        self.num_prefix_tokens += reg_tokens
+        self.num_reg_tokens = reg_tokens
+        self.has_class_token = class_token
+        self.no_embed_class = (
+            no_embed_class  # don't embed prefix positions (includes reg)
+        )
+        self.dynamic_img_size = dynamic_img_size
+        self.grad_checkpointing = False
+        self.ignore_head = ignore_head
+
+        embed_args = {}
+        if dynamic_img_size:
+            # flatten deferred until after pos embed
+            embed_args.update(dict(strict_img_size=False, output_fmt="NHWC"))
+        self.patch_embed = embed_layer(
+            img_size=img_size,
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=embed_dim,
+            bias=not pre_norm,  # disable bias if pre-norm is used (e.g. CLIP)
+            dynamic_img_pad=dynamic_img_pad,
+            **embed_args,
+        )
+        num_patches = self.patch_embed.num_patches
+
+        self.cls_token = (
+            nn.Parameter(torch.zeros(1, 1, embed_dim)) if class_token else None
+        )
+        self.reg_token = (
+            nn.Parameter(torch.zeros(1, reg_tokens, embed_dim)) if reg_tokens else None
+        )
+        embed_len = (
+            num_patches if no_embed_class else num_patches + self.num_prefix_tokens
+        )
+        self.pos_embed = nn.Parameter(torch.randn(1, embed_len, embed_dim) * 0.02)
+        self.pos_drop = nn.Dropout(p=pos_drop_rate)
+        if patch_drop_rate > 0:
+            self.patch_drop = PatchDropout(
+                patch_drop_rate,
+                num_prefix_tokens=self.num_prefix_tokens,
+            )
+        else:
+            self.patch_drop = nn.Identity()
+        self.norm_pre = norm_layer(embed_dim) if pre_norm else nn.Identity()
+
+        dpr = [
+            x.item() for x in torch.linspace(0, drop_path_rate, depth)
+        ]  # stochastic depth decay rule
+        self.blocks = nn.Sequential(
+            *[
+                block_fn(
+                    dim=embed_dim,
+                    num_heads=num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_norm=qk_norm,
+                    init_values=init_values,
+                    proj_drop=proj_drop_rate,
+                    attn_drop=attn_drop_rate,
+                    drop_path=dpr[i],
+                    norm_layer=norm_layer,
+                    act_layer=act_layer,
+                    mlp_layer=mlp_layer,
+                )
+                for i in range(depth)
+            ]
+        )
+        self.norm = norm_layer(embed_dim) if not use_fc_norm else nn.Identity()
+
+        # Classifier Head
+        if global_pool == "map":
+            AttentionPoolLatent.init_weights = init_weights
+            self.attn_pool = AttentionPoolLatent(
+                self.embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                norm_layer=norm_layer,
+            )
+        else:
+            self.attn_pool = None
+        self.fc_norm = norm_layer(embed_dim) if use_fc_norm else nn.Identity()
+        self.head_drop = nn.Dropout(drop_rate)
+        self.head = (
+            nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+        )
+
+        if weight_init != "skip":
+            self.init_weights(weight_init)
+
+    def init_weights(self, mode: Literal["jax", "jax_nlhb", "moco", ""] = "") -> None:
+        assert mode in ("jax", "jax_nlhb", "moco", "")
+        # head_bias = -math.log(self.num_classes) if "nlhb" in mode else 0.0
+        trunc_normal_(self.pos_embed, std=0.02)
+        if self.cls_token is not None:
+            nn.init.normal_(self.cls_token, std=1e-6)
+        named_apply(init_weights_vit_timm, self)
+
+    @torch.jit.ignore
+    def no_weight_decay(self) -> Set:
+        return {"pos_embed", "cls_token", "dist_token"}
+
+    @torch.jit.ignore
+    def group_matcher(self, coarse: bool = False) -> Dict:
+        return dict(
+            stem=r"^cls_token|pos_embed|patch_embed",  # stem and embed
+            blocks=[(r"^blocks\.(\d+)", None), (r"^norm", (99999,))],
+        )
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable: bool = True) -> None:
+        self.grad_checkpointing = enable
+
+    @torch.jit.ignore
+    def get_classifier(self) -> nn.Module:
+        return self.head
+
+    def reset_classifier(self, num_classes: int, global_pool=None) -> None:
+        self.num_classes = num_classes
+        if global_pool is not None:
+            assert global_pool in ("", "avg", "token", "map")
+            if global_pool == "map" and self.attn_pool is None:
+                assert (
+                    False
+                ), "Cannot currently add attention pooling in reset_classifier()."
+            elif global_pool != "map " and self.attn_pool is not None:
+                self.attn_pool = None  # remove attention pooling
+            self.global_pool = global_pool
+        self.head = (
+            nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()
+        )
+
+    def _pos_embed(self, x: torch.Tensor) -> torch.Tensor:
+        if self.dynamic_img_size:
+            B, H, W, C = x.shape
+            pos_embed = resample_abs_pos_embed(
+                self.pos_embed,
+                (H, W),
+                num_prefix_tokens=0 if self.no_embed_class else self.num_prefix_tokens,
+            )
+            x = x.view(B, -1, C)
+        else:
+            pos_embed = self.pos_embed
+
+        to_cat = []
+        if self.cls_token is not None:
+            to_cat.append(self.cls_token.expand(x.shape[0], -1, -1))
+        if self.reg_token is not None:
+            to_cat.append(self.reg_token.expand(x.shape[0], -1, -1))
+
+        if self.no_embed_class:
+            # deit-3, updated JAX (big vision)
+            # position embedding does not overlap with class token, add then concat
+            x = x + pos_embed
+            if to_cat:
+                x = torch.cat(to_cat + [x], dim=1)
+        else:
+            # original timm, JAX, and deit vit impl
+            # pos_embed has entry for class token, concat then add
+            if to_cat:
+                x = torch.cat(to_cat + [x], dim=1)
+            x = x + pos_embed
+
+        return self.pos_drop(x)
+
+    def _intermediate_layers(
+        self,
+        x: torch.Tensor,
+        n: Union[int, Sequence] = 1,
+    ) -> List[torch.Tensor]:
+        outputs, num_blocks = [], len(self.blocks)
+        take_indices = set(
+            range(num_blocks - n, num_blocks) if isinstance(n, int) else n
+        )
+
+        # forward pass
+        x = self.patch_embed(x)
+        x = self._pos_embed(x)
+        x = self.patch_drop(x)
+        x = self.norm_pre(x)
+        for i, blk in enumerate(self.blocks):
+            x = blk(x)
+            if i in take_indices:
+                outputs.append(x)
+
+        return outputs
+
+    def get_intermediate_layers(
+        self,
+        x: torch.Tensor,
+        n: Union[int, Sequence] = 1,
+        reshape: bool = False,
+        return_prefix_tokens: bool = False,
+        norm: bool = False,
+    ) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor]]]:
+        """Intermediate layer accessor (NOTE: This is a WIP experiment).
+        Inspired by DINO / DINOv2 interface
+        """
+        # take last n blocks if n is an int, if in is a sequence, select by matching indices
+        outputs = self._intermediate_layers(x, n)
+        if norm:
+            outputs = [self.norm(out) for out in outputs]
+        prefix_tokens = [out[:, 0 : self.num_prefix_tokens] for out in outputs]
+        outputs = [out[:, self.num_prefix_tokens :] for out in outputs]
+
+        if reshape:
+            grid_size = self.patch_embed.grid_size
+            outputs = [
+                out.reshape(x.shape[0], grid_size[0], grid_size[1], -1)
+                .permute(0, 3, 1, 2)
+                .contiguous()
+                for out in outputs
+            ]
+
+        if return_prefix_tokens:
+            return tuple(zip(outputs, prefix_tokens))
+        return tuple(outputs)
+
+    def forward_features(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.patch_embed(x)
+        x = self._pos_embed(x)
+        x = self.patch_drop(x)
+        x = self.norm_pre(x)
+        if self.grad_checkpointing and not torch.jit.is_scripting():
+            x = checkpoint_seq(self.blocks, x)
+        else:
+            x = self.blocks(x)
+        x = self.norm(x)
+        return x
+
+    def forward_head(self, x: torch.Tensor, pre_logits: bool = False) -> torch.Tensor:
+        if self.attn_pool is not None:
+            x = self.attn_pool(x)
+        elif self.global_pool == "avg":
+            x = x[:, self.num_prefix_tokens :].mean(dim=1)
+        elif self.global_pool:
+            x = x[:, 0]  # class token
+        x = self.fc_norm(x)
+        x = self.head_drop(x)
+        return x if pre_logits else self.head(x)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.forward_features(x)
+        if not self.ignore_head:
+            x = self.forward_head(x)
+        return x
+
+
+@dataclass
+class SigLIPVisionCfg:
+    width: int = 1152
+    layers: Union[Tuple[int, int, int, int], int] = 27
+    heads: int = 16
+    patch_size: int = 14
+    image_size: Union[Tuple[int, int], int] = 336
+    global_pool: str = "map"
+    mlp_ratio: float = 3.7362
+    class_token: bool = False
+    num_classes: int = 0
+    use_checkpoint: bool = False
+
+
+SigLIP_MODEL_CONFIG = {
+    "siglip_so400m_patch14_384": {
+        "image_size": 336,
+        "patch_size": 14,
+        "width": 1152,
+        "layers": 27,
+        "heads": 16,
+        "mlp_ratio": 3.7362,
+        "global_pool": "map",
+        "use_checkpoint": False,
+    },
+    "siglip_so400m_patch14_224": {
+        "image_size": 224,
+        "patch_size": 14,
+        "width": 1152,
+        "layers": 27,
+        "heads": 16,
+        "mlp_ratio": 3.7362,
+        "global_pool": "map",
+        "use_checkpoint": False,
+    },
+    "siglip_large_patch16_384": {
+        "image_size": 384,
+        "patch_size": 16,
+        "width": 1024,
+        "layers": 24,
+        "heads": 16,
+        "mlp_ratio": 4,
+        "global_pool": "map",
+        "use_checkpoint": False,
+    },
+}
+
+
+def create_siglip_vit(
+    model_name: str = "siglip_so400m_patch14_384",
+    image_size: int = 384,
+    select_layer: int = -1,
+    ckpt_path: str = "",
+    **kwargs,
+):
+    assert (
+        model_name in SigLIP_MODEL_CONFIG.keys()
+    ), f"model name should be in {SigLIP_MODEL_CONFIG.keys()}"
+
+    vision_cfg = SigLIPVisionCfg(**SigLIP_MODEL_CONFIG[model_name])
+
+    if select_layer <= 0:
+        layers = min(vision_cfg.layers, vision_cfg.layers + select_layer + 1)
+    else:
+        layers = min(vision_cfg.layers, select_layer)
+
+    model = VisionTransformer(
+        img_size=image_size,
+        patch_size=vision_cfg.patch_size,
+        embed_dim=vision_cfg.width,
+        depth=layers,
+        num_heads=vision_cfg.heads,
+        mlp_ratio=vision_cfg.mlp_ratio,
+        class_token=vision_cfg.class_token,
+        global_pool=vision_cfg.global_pool,
+        ignore_head=kwargs.get("ignore_head", True),
+        weight_init=kwargs.get("weight_init", "skip"),
+        num_classes=0,
+    )
+
+    if ckpt_path:
+        state_dict = torch.load(ckpt_path, map_location="cpu")
+
+        incompatible_keys = model.load_state_dict(state_dict, strict=False)
+        print(
+            f"SigLIP-ViT restores from {ckpt_path},\n"
+            f"\tincompatible_keys:', {incompatible_keys}."
+        )
+
+    return model

deepseek_vl/serve/app_deepseek.py

@@ -0,0 +1,514 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+# -*- coding:utf-8 -*-
+
+import base64
+from io import BytesIO
+
+import gradio as gr
+import torch
+from app_modules.gradio_utils import (
+    cancel_outputing,
+    delete_last_conversation,
+    reset_state,
+    reset_textbox,
+    transfer_input,
+    wrap_gen_fn,
+)
+from app_modules.overwrites import reload_javascript
+from app_modules.presets import CONCURRENT_COUNT, description, description_top, title
+from app_modules.utils import configure_logger, is_variable_assigned, strip_stop_words
+
+from deepseek_vl.serve.inference import (
+    convert_conversation_to_prompts,
+    deepseek_generate,
+    load_model,
+)
+from deepseek_vl.utils.conversation import SeparatorStyle
+
+
+def load_models():
+    models = {
+        "DeepSeek-VL 7B": "deepseek-ai/deepseek-vl-7b-chat",
+    }
+
+    for model_name in models:
+        models[model_name] = load_model(models[model_name])
+
+    return models
+
+
+logger = configure_logger()
+models = load_models()
+MODELS = sorted(list(models.keys()))
+
+
+def generate_prompt_with_history(
+    text, image, history, vl_chat_processor, tokenizer, max_length=2048
+):
+    """
+    Generate a prompt with history for the deepseek application.
+
+    Args:
+        text (str): The text prompt.
+        image (str): The image prompt.
+        history (list): List of previous conversation messages.
+        tokenizer: The tokenizer used for encoding the prompt.
+        max_length (int): The maximum length of the prompt.
+
+    Returns:
+        tuple: A tuple containing the generated prompt, image list, conversation, and conversation copy. If the prompt could not be generated within the max_length limit, returns None.
+    """
+
+    sft_format = "deepseek"
+    user_role_ind = 0
+    bot_role_ind = 1
+
+    # Initialize conversation
+    conversation = vl_chat_processor.new_chat_template()
+
+    if history:
+        conversation.messages = history
+
+    if image is not None:
+        if "<image_placeholder>" not in text:
+            text = (
+                "<image_placeholder>" + "\n" + text
+            )  # append the <image_placeholder> in a new line after the text prompt
+        text = (text, image)
+
+    conversation.append_message(conversation.roles[user_role_ind], text)
+    conversation.append_message(conversation.roles[bot_role_ind], "")
+
+    # Create a copy of the conversation to avoid history truncation in the UI
+    conversation_copy = conversation.copy()
+    logger.info("=" * 80)
+    logger.info(get_prompt(conversation))
+
+    rounds = len(conversation.messages) // 2
+
+    for _ in range(rounds):
+        current_prompt = get_prompt(conversation)
+        current_prompt = (
+            current_prompt.replace("</s>", "")
+            if sft_format == "deepseek"
+            else current_prompt
+        )
+
+        if torch.tensor(tokenizer.encode(current_prompt)).size(-1) <= max_length:
+            return conversation_copy
+
+        if len(conversation.messages) % 2 != 0:
+            gr.Error("The messages between user and assistant are not paired.")
+            return
+
+        try:
+            for _ in range(2):  # pop out two messages in a row
+                conversation.messages.pop(0)
+        except IndexError:
+            gr.Error("Input text processing failed, unable to respond in this round.")
+            return None
+
+    gr.Error("Prompt could not be generated within max_length limit.")
+    return None
+
+
+def to_gradio_chatbot(conv):
+    """Convert the conversation to gradio chatbot format."""
+    ret = []
+    for i, (role, msg) in enumerate(conv.messages[conv.offset :]):
+        if i % 2 == 0:
+            if type(msg) is tuple:
+                msg, image = msg
+                if isinstance(image, str):
+                    with open(image, "rb") as f:
+                        data = f.read()
+                    img_b64_str = base64.b64encode(data).decode()
+                    image_str = f'<video src="data:video/mp4;base64,{img_b64_str}" controls width="426" height="240"></video>'
+                    msg = msg.replace("\n".join(["<image_placeholder>"] * 4), image_str)
+                else:
+                    max_hw, min_hw = max(image.size), min(image.size)
+                    aspect_ratio = max_hw / min_hw
+                    max_len, min_len = 800, 400
+                    shortest_edge = int(min(max_len / aspect_ratio, min_len, min_hw))
+                    longest_edge = int(shortest_edge * aspect_ratio)
+                    W, H = image.size
+                    if H > W:
+                        H, W = longest_edge, shortest_edge
+                    else:
+                        H, W = shortest_edge, longest_edge
+                    image = image.resize((W, H))
+                    buffered = BytesIO()
+                    image.save(buffered, format="JPEG")
+                    img_b64_str = base64.b64encode(buffered.getvalue()).decode()
+                    img_str = f'<img src="data:image/png;base64,{img_b64_str}" alt="user upload image" />'
+                    msg = msg.replace("<image_placeholder>", img_str)
+            ret.append([msg, None])
+        else:
+            ret[-1][-1] = msg
+    return ret
+
+
+def to_gradio_history(conv):
+    """Convert the conversation to gradio history state."""
+    return conv.messages[conv.offset :]
+
+
+def get_prompt(conv) -> str:
+    """Get the prompt for generation."""
+    system_prompt = conv.system_template.format(system_message=conv.system_message)
+    if conv.sep_style == SeparatorStyle.DeepSeek:
+        seps = [conv.sep, conv.sep2]
+        if system_prompt == "" or system_prompt is None:
+            ret = ""
+        else:
+            ret = system_prompt + seps[0]
+        for i, (role, message) in enumerate(conv.messages):
+            if message:
+                if type(message) is tuple:  # multimodal message
+                    message, _ = message
+                ret += role + ": " + message + seps[i % 2]
+            else:
+                ret += role + ":"
+        return ret
+    else:
+        return conv.get_prompt
+
+
+@wrap_gen_fn
+def predict(
+    text,
+    image,
+    chatbot,
+    history,
+    top_p,
+    temperature,
+    repetition_penalty,
+    max_length_tokens,
+    max_context_length_tokens,
+    model_select_dropdown,
+):
+    """
+    Function to predict the response based on the user's input and selected model.
+
+    Parameters:
+    user_text (str): The input text from the user.
+    user_image (str): The input image from the user.
+    chatbot (str): The chatbot's name.
+    history (str): The history of the chat.
+    top_p (float): The top-p parameter for the model.
+    temperature (float): The temperature parameter for the model.
+    max_length_tokens (int): The maximum length of tokens for the model.
+    max_context_length_tokens (int): The maximum length of context tokens for the model.
+    model_select_dropdown (str): The selected model from the dropdown.
+
+    Returns:
+    generator: A generator that yields the chatbot outputs, history, and status.
+    """
+    print("running the prediction function")
+    try:
+        tokenizer, vl_gpt, vl_chat_processor = models[model_select_dropdown]
+
+        if text == "":
+            yield chatbot, history, "Empty context."
+            return
+    except KeyError:
+        yield [[text, "No Model Found"]], [], "No Model Found"
+        return
+
+    conversation = generate_prompt_with_history(
+        text,
+        image,
+        history,
+        vl_chat_processor,
+        tokenizer,
+        max_length=max_context_length_tokens,
+    )
+    prompts = convert_conversation_to_prompts(conversation)
+
+    stop_words = conversation.stop_str
+    gradio_chatbot_output = to_gradio_chatbot(conversation)
+
+    full_response = ""
+    with torch.no_grad():
+        for x in deepseek_generate(
+            prompts=prompts,
+            vl_gpt=vl_gpt,
+            vl_chat_processor=vl_chat_processor,
+            tokenizer=tokenizer,
+            stop_words=stop_words,
+            max_length=max_length_tokens,
+            temperature=temperature,
+            repetition_penalty=repetition_penalty,
+            top_p=top_p,
+        ):
+            full_response += x
+            response = strip_stop_words(full_response, stop_words)
+            conversation.update_last_message(response)
+            gradio_chatbot_output[-1][1] = response
+            yield gradio_chatbot_output, to_gradio_history(
+                conversation
+            ), "Generating..."
+
+    print("flushed result to gradio")
+    torch.cuda.empty_cache()
+
+    if is_variable_assigned("x"):
+        print(f"{model_select_dropdown}:\n{text}\n{'-' * 80}\n{x}\n{'=' * 80}")
+        print(
+            f"temperature: {temperature}, top_p: {top_p}, repetition_penalty: {repetition_penalty}, max_length_tokens: {max_length_tokens}"
+        )
+
+    yield gradio_chatbot_output, to_gradio_history(conversation), "Generate: Success"
+
+
+def retry(
+    text,
+    image,
+    chatbot,
+    history,
+    top_p,
+    temperature,
+    repetition_penalty,
+    max_length_tokens,
+    max_context_length_tokens,
+    model_select_dropdown,
+):
+    if len(history) == 0:
+        yield (chatbot, history, "Empty context")
+        return
+
+    chatbot.pop()
+    history.pop()
+    text = history.pop()[-1]
+    if type(text) is tuple:
+        text, image = text
+
+    yield from predict(
+        text,
+        image,
+        chatbot,
+        history,
+        top_p,
+        temperature,
+        repetition_penalty,
+        max_length_tokens,
+        max_context_length_tokens,
+        model_select_dropdown,
+    )
+
+
+def build_demo(MODELS):
+    with open("deepseek_vl/serve/assets/custom.css", "r", encoding="utf-8") as f:
+        customCSS = f.read()
+
+    with gr.Blocks(theme=gr.themes.Soft()) as demo:
+        history = gr.State([])
+        input_text = gr.State()
+        input_image = gr.State()
+
+        with gr.Row():
+            gr.HTML(title)
+            status_display = gr.Markdown("Success", elem_id="status_display")
+        gr.Markdown(description_top)
+
+        with gr.Row(equal_height=True):
+            with gr.Column(scale=4):
+                with gr.Row():
+                    chatbot = gr.Chatbot(
+                        elem_id="deepseek_chatbot",
+                        show_share_button=True,
+                        likeable=True,
+                        bubble_full_width=False,
+                        height=600,
+                    )
+                with gr.Row():
+                    with gr.Column(scale=4):
+                        text_box = gr.Textbox(
+                            show_label=False, placeholder="Enter text", container=False
+                        )
+                    with gr.Column(
+                        min_width=70,
+                    ):
+                        submitBtn = gr.Button("Send")
+                    with gr.Column(
+                        min_width=70,
+                    ):
+                        cancelBtn = gr.Button("Stop")
+                with gr.Row():
+                    emptyBtn = gr.Button(
+                        "🧹 New Conversation",
+                    )
+                    retryBtn = gr.Button("🔄 Regenerate")
+                    delLastBtn = gr.Button("🗑️ Remove Last Turn")
+
+            with gr.Column():
+                image_box = gr.Image(type="pil")
+
+                with gr.Tab(label="Parameter Setting") as parameter_row:
+                    top_p = gr.Slider(
+                        minimum=-0,
+                        maximum=1.0,
+                        value=0.95,
+                        step=0.05,
+                        interactive=True,
+                        label="Top-p",
+                    )
+                    temperature = gr.Slider(
+                        minimum=0,
+                        maximum=1.0,
+                        value=0.1,
+                        step=0.1,
+                        interactive=True,
+                        label="Temperature",
+                    )
+                    repetition_penalty = gr.Slider(
+                        minimum=0.0,
+                        maximum=2.0,
+                        value=1.1,
+                        step=0.1,
+                        interactive=True,
+                        label="Repetition penalty",
+                    )
+                    max_length_tokens = gr.Slider(
+                        minimum=0,
+                        maximum=4096,
+                        value=2048,
+                        step=8,
+                        interactive=True,
+                        label="Max Generation Tokens",
+                    )
+                    max_context_length_tokens = gr.Slider(
+                        minimum=0,
+                        maximum=4096,
+                        value=4096,
+                        step=128,
+                        interactive=True,
+                        label="Max History Tokens",
+                    )
+                    model_select_dropdown = gr.Dropdown(
+                        label="Select Models",
+                        choices=MODELS,
+                        multiselect=False,
+                        value=MODELS[0],
+                        interactive=True,
+                    )
+
+        examples_list = [
+            [
+                "deepseek_vl/serve/examples/rap.jpeg",
+                "Can you write me a master rap song that rhymes very well based on this image?",
+            ],
+            [
+                "deepseek_vl/serve/examples/app.png",
+                "What is this app about?",
+            ],
+            [
+                "deepseek_vl/serve/examples/pipeline.png",
+                "Help me write a python code based on the image.",
+            ],
+            [
+                "deepseek_vl/serve/examples/chart.png",
+                "Could you help me to re-draw this picture with python codes?",
+            ],
+            [
+                "deepseek_vl/serve/examples/mirror.png",
+                "How many people are there in the image. Why?",
+            ],
+            [
+                "deepseek_vl/serve/examples/puzzle.png",
+                "Can this 2 pieces combine together?",
+            ],
+        ]
+        gr.Examples(examples=examples_list, inputs=[image_box, text_box])
+        gr.Markdown(description)
+
+        input_widgets = [
+            input_text,
+            input_image,
+            chatbot,
+            history,
+            top_p,
+            temperature,
+            repetition_penalty,
+            max_length_tokens,
+            max_context_length_tokens,
+            model_select_dropdown,
+        ]
+        output_widgets = [chatbot, history, status_display]
+
+        transfer_input_args = dict(
+            fn=transfer_input,
+            inputs=[text_box, image_box],
+            outputs=[input_text, input_image, text_box, image_box, submitBtn],
+            show_progress=True,
+        )
+
+        predict_args = dict(
+            fn=predict,
+            inputs=input_widgets,
+            outputs=output_widgets,
+            show_progress=True,
+        )
+
+        retry_args = dict(
+            fn=retry,
+            inputs=input_widgets,
+            outputs=output_widgets,
+            show_progress=True,
+        )
+
+        reset_args = dict(
+            fn=reset_textbox, inputs=[], outputs=[text_box, status_display]
+        )
+
+        predict_events = [
+            text_box.submit(**transfer_input_args).then(**predict_args),
+            submitBtn.click(**transfer_input_args).then(**predict_args),
+        ]
+
+        emptyBtn.click(reset_state, outputs=output_widgets, show_progress=True)
+        emptyBtn.click(**reset_args)
+        retryBtn.click(**retry_args)
+
+        delLastBtn.click(
+            delete_last_conversation,
+            [chatbot, history],
+            output_widgets,
+            show_progress=True,
+        )
+
+        cancelBtn.click(cancel_outputing, [], [status_display], cancels=predict_events)
+
+    return demo
+
+
+if __name__ == "__main__":
+    demo = build_demo(MODELS)
+    demo.title = "DeepSeek-VL Chatbot"
+
+    reload_javascript()
+    demo.queue(concurrency_count=CONCURRENT_COUNT).launch(
+        share=False,
+        favicon_path="deepseek_vl/serve/assets/favicon.ico",
+        inbrowser=False,
+        server_name="0.0.0.0",
+        server_port=8122,
+    )

deepseek_vl/serve/app_modules/gradio_utils.py

@@ -0,0 +1,94 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from functools import wraps
+
+import gradio as gr
+
+
+def wrap_gen_fn(gen_fn):
+    @wraps(gen_fn)
+    def wrapped_gen_fn(prompt, *args, **kwargs):
+        try:
+            yield from gen_fn(prompt, *args, **kwargs)
+        except gr.Error as g_err:
+            raise g_err
+        except Exception as e:
+            raise gr.Error(f"Failed to generate text: {e}") from e
+
+    return wrapped_gen_fn
+
+
+def delete_last_conversation(chatbot, history):
+    if len(history) % 2 != 0:
+        gr.Error("history length is not even")
+        return (
+            chatbot,
+            history,
+            "Delete Done",
+        )
+
+    if len(chatbot) > 0:
+        chatbot.pop()
+
+    if len(history) > 0 and len(history) % 2 == 0:
+        history.pop()
+        history.pop()
+
+    return (
+        chatbot,
+        history,
+        "Delete Done",
+    )
+
+
+def reset_state():
+    return [], [], None, "Reset Done"
+
+
+def reset_textbox():
+    return gr.update(value=""), ""
+
+
+def cancel_outputing():
+    return "Stop Done"
+
+
+def transfer_input(input_text, input_image):
+    print("transferring input text and input image")
+    return (
+        input_text,
+        input_image,
+        gr.update(value=""),
+        gr.update(value=None),
+        gr.Button(visible=True),
+    )
+
+
+class State:
+    interrupted = False
+
+    def interrupt(self):
+        self.interrupted = True
+
+    def recover(self):
+        self.interrupted = False
+
+
+shared_state = State()

deepseek_vl/serve/app_modules/overwrites.py

@@ -0,0 +1,81 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from __future__ import annotations
+
+import logging
+from typing import List, Tuple
+
+from app_modules.presets import gr
+from app_modules.utils import convert_asis, convert_mdtext, detect_converted_mark
+
+
+def compact_text_chunks(self, prompt, text_chunks: List[str]) -> List[str]:
+    logging.debug("Compacting text chunks...🚀🚀🚀")
+    combined_str = [c.strip() for c in text_chunks if c.strip()]
+    combined_str = [f"[{index+1}] {c}" for index, c in enumerate(combined_str)]
+    combined_str = "\n\n".join(combined_str)
+    # resplit based on self.max_chunk_overlap
+    text_splitter = self.get_text_splitter_given_prompt(prompt, 1, padding=1)
+    return text_splitter.split_text(combined_str)
+
+
+def postprocess(
+    self, y: List[Tuple[str | None, str | None]]
+) -> List[Tuple[str | None, str | None]]:
+    """
+    Parameters:
+        y: List of tuples representing the message and response pairs. Each message and response should be a string, which may be in Markdown format.
+    Returns:
+        List of tuples representing the message and response. Each message and response will be a string of HTML.
+    """
+    if y is None or y == []:
+        return []
+    temp = []
+    for x in y:
+        user, bot = x
+        if not detect_converted_mark(user):
+            user = convert_asis(user)
+        if not detect_converted_mark(bot):
+            bot = convert_mdtext(bot)
+        temp.append((user, bot))
+    return temp
+
+
+with open("deepseek_vl/serve/assets/custom.js", "r", encoding="utf-8") as f, open(
+    "deepseek_vl/serve/assets/Kelpy-Codos.js", "r", encoding="utf-8"
+) as f2:
+    customJS = f.read()
+    kelpyCodos = f2.read()
+
+
+def reload_javascript():
+    print("Reloading javascript...")
+    js = f"<script>{customJS}</script><script>{kelpyCodos}</script>"
+
+    def template_response(*args, **kwargs):
+        res = GradioTemplateResponseOriginal(*args, **kwargs)
+        res.body = res.body.replace(b"</html>", f"{js}</html>".encode("utf8"))
+        res.init_headers()
+        return res
+
+    gr.routes.templates.TemplateResponse = template_response
+
+
+GradioTemplateResponseOriginal = gr.routes.templates.TemplateResponse

deepseek_vl/serve/app_modules/presets.py

@@ -0,0 +1,96 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+# -*- coding:utf-8 -*-
+import gradio as gr
+
+title = """<h1 align="left" style="min-width:200px; margin-top:0;">Chat with DeepSeek-VL </h1>"""
+description_top = """"""
+description = """"""
+CONCURRENT_COUNT = 10
+
+
+ALREADY_CONVERTED_MARK = "<!-- ALREADY CONVERTED BY PARSER. -->"
+
+small_and_beautiful_theme = gr.themes.Soft(
+    primary_hue=gr.themes.Color(
+        c50="#EBFAF2",
+        c100="#CFF3E1",
+        c200="#A8EAC8",
+        c300="#77DEA9",
+        c400="#3FD086",
+        c500="#02C160",
+        c600="#06AE56",
+        c700="#05974E",
+        c800="#057F45",
+        c900="#04673D",
+        c950="#2E5541",
+        name="small_and_beautiful",
+    ),
+    secondary_hue=gr.themes.Color(
+        c50="#576b95",
+        c100="#576b95",
+        c200="#576b95",
+        c300="#576b95",
+        c400="#576b95",
+        c500="#576b95",
+        c600="#576b95",
+        c700="#576b95",
+        c800="#576b95",
+        c900="#576b95",
+        c950="#576b95",
+    ),
+    neutral_hue=gr.themes.Color(
+        name="gray",
+        c50="#f6f7f8",
+        # c100="#f3f4f6",
+        c100="#F2F2F2",
+        c200="#e5e7eb",
+        c300="#d1d5db",
+        c400="#B2B2B2",
+        c500="#808080",
+        c600="#636363",
+        c700="#515151",
+        c800="#393939",
+        # c900="#272727",
+        c900="#2B2B2B",
+        c950="#171717",
+    ),
+    radius_size=gr.themes.sizes.radius_sm,
+).set(
+    # button_primary_background_fill="*primary_500",
+    button_primary_background_fill_dark="*primary_600",
+    # button_primary_background_fill_hover="*primary_400",
+    # button_primary_border_color="*primary_500",
+    button_primary_border_color_dark="*primary_600",
+    button_primary_text_color="white",
+    button_primary_text_color_dark="white",
+    button_secondary_background_fill="*neutral_100",
+    button_secondary_background_fill_hover="*neutral_50",
+    button_secondary_background_fill_dark="*neutral_900",
+    button_secondary_text_color="*neutral_800",
+    button_secondary_text_color_dark="white",
+    # background_fill_primary="#F7F7F7",
+    # background_fill_primary_dark="#1F1F1F",
+    # block_title_text_color="*primary_500",
+    block_title_background_fill_dark="*primary_900",
+    block_label_background_fill_dark="*primary_900",
+    input_background_fill="#F6F6F6",
+    # chatbot_code_background_color_dark="*neutral_950",
+)

deepseek_vl/serve/app_modules/utils.py

@@ -0,0 +1,228 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+# -*- coding:utf-8 -*-
+from __future__ import annotations
+
+import html
+import logging
+import os
+import re
+import time
+
+import mdtex2html
+from app_modules.presets import ALREADY_CONVERTED_MARK
+from markdown import markdown
+from pygments import highlight
+from pygments.formatters import HtmlFormatter
+from pygments.lexers import ClassNotFound, get_lexer_by_name, guess_lexer
+
+logger = logging.getLogger("gradio_logger")
+
+
+def configure_logger():
+    logger = logging.getLogger("gradio_logger")
+    logger.setLevel(logging.DEBUG)
+
+    timestr = time.strftime("%Y%m%d-%H%M%S")
+    os.makedirs("deepseek_vl/serve/logs", exist_ok=True)
+    file_handler = logging.FileHandler(
+        f"deepseek_vl/serve/logs/{timestr}_gradio_log.log"
+    )
+    console_handler = logging.StreamHandler()
+
+    formatter = logging.Formatter(
+        "%(asctime)s - %(name)s - %(levelname)s - %(message)s"
+    )
+    console_handler.setFormatter(formatter)
+    file_handler.setFormatter(formatter)
+
+    console_handler.setLevel(logging.INFO)
+    file_handler.setLevel(logging.INFO)
+
+    logger.addHandler(console_handler)
+    logger.addHandler(file_handler)
+
+    return logger
+
+
+def strip_stop_words(x, stop_words):
+    for w in stop_words:
+        if w in x:
+            return x[: x.index(w)].strip()
+    return x.strip()
+
+
+def format_output(history, text, x):
+    updated_history = history + [[text, x]]
+    a = [[y[0], convert_to_markdown(y[1])] for y in updated_history]
+    return a, updated_history
+
+
+def markdown_to_html_with_syntax_highlight(md_str):  # deprecated
+    def replacer(match):
+        lang = match.group(1) or "text"
+        code = match.group(2)
+
+        try:
+            lexer = get_lexer_by_name(lang, stripall=True)
+        except ValueError:
+            lexer = get_lexer_by_name("text", stripall=True)
+
+        formatter = HtmlFormatter()
+        highlighted_code = highlight(code, lexer, formatter)
+
+        return f'<pre><code class="{lang}">{highlighted_code}</code></pre>'
+
+    code_block_pattern = r"```(\w+)?\n([\s\S]+?)\n```"
+    md_str = re.sub(code_block_pattern, replacer, md_str, flags=re.MULTILINE)
+
+    html_str = markdown(md_str)
+    return html_str
+
+
+def normalize_markdown(md_text: str) -> str:  # deprecated
+    lines = md_text.split("\n")
+    normalized_lines = []
+    inside_list = False
+
+    for i, line in enumerate(lines):
+        if re.match(r"^(\d+\.|-|\*|\+)\s", line.strip()):
+            if not inside_list and i > 0 and lines[i - 1].strip() != "":
+                normalized_lines.append("")
+            inside_list = True
+            normalized_lines.append(line)
+        elif inside_list and line.strip() == "":
+            if i < len(lines) - 1 and not re.match(
+                r"^(\d+\.|-|\*|\+)\s", lines[i + 1].strip()
+            ):
+                normalized_lines.append(line)
+            continue
+        else:
+            inside_list = False
+            normalized_lines.append(line)
+
+    return "\n".join(normalized_lines)
+
+
+def convert_mdtext(md_text):
+    code_block_pattern = re.compile(r"```(.*?)(?:```|$)", re.DOTALL)
+    inline_code_pattern = re.compile(r"`(.*?)`", re.DOTALL)
+    code_blocks = code_block_pattern.findall(md_text)
+    non_code_parts = code_block_pattern.split(md_text)[::2]
+
+    result = []
+    for non_code, code in zip(non_code_parts, code_blocks + [""]):
+        if non_code.strip():
+            non_code = normalize_markdown(non_code)
+            if inline_code_pattern.search(non_code):
+                result.append(markdown(non_code, extensions=["tables"]))
+            else:
+                result.append(mdtex2html.convert(non_code, extensions=["tables"]))
+        if code.strip():
+            code = f"\n```{code}\n\n```"
+            code = markdown_to_html_with_syntax_highlight(code)
+            result.append(code)
+    result = "".join(result)
+    result += ALREADY_CONVERTED_MARK
+    return result
+
+
+def convert_asis(userinput):
+    return f'<p style="white-space:pre-wrap;">{html.escape(userinput)}</p>{ALREADY_CONVERTED_MARK}'
+
+
+def is_stop_word_or_prefix(s: str, stop_words: list) -> bool:
+    return any(s.endswith(stop_word) for stop_word in stop_words)
+
+
+def detect_converted_mark(userinput):
+    return bool(userinput.endswith(ALREADY_CONVERTED_MARK))
+
+
+def detect_language(code):
+    first_line = "" if code.startswith("\n") else code.strip().split("\n", 1)[0]
+    language = first_line.lower() if first_line else ""
+    code_without_language = code[len(first_line) :].lstrip() if first_line else code
+    return language, code_without_language
+
+
+def convert_to_markdown(text):
+    text = text.replace("$", "&#36;")
+    text = text.replace("\r\n", "\n")
+
+    def replace_leading_tabs_and_spaces(line):
+        new_line = []
+
+        for char in line:
+            if char == "\t":
+                new_line.append("&#9;")
+            elif char == " ":
+                new_line.append("&nbsp;")
+            else:
+                break
+        return "".join(new_line) + line[len(new_line) :]
+
+    markdown_text = ""
+    lines = text.split("\n")
+    in_code_block = False
+
+    for line in lines:
+        if in_code_block is False and line.startswith("```"):
+            in_code_block = True
+            markdown_text += f"{line}\n"
+        elif in_code_block is True and line.startswith("```"):
+            in_code_block = False
+            markdown_text += f"{line}\n"
+        elif in_code_block:
+            markdown_text += f"{line}\n"
+        else:
+            line = replace_leading_tabs_and_spaces(line)
+            line = re.sub(r"^(#)", r"\\\1", line)
+            markdown_text += f"{line}  \n"
+
+    return markdown_text
+
+
+def add_language_tag(text):
+    def detect_language(code_block):
+        try:
+            lexer = guess_lexer(code_block)
+            return lexer.name.lower()
+        except ClassNotFound:
+            return ""
+
+    code_block_pattern = re.compile(r"(```)(\w*\n[^`]+```)", re.MULTILINE)
+
+    def replacement(match):
+        code_block = match.group(2)
+        if match.group(2).startswith("\n"):
+            language = detect_language(code_block)
+            return (
+                f"```{language}{code_block}```" if language else f"```\n{code_block}```"
+            )
+        else:
+            return match.group(1) + code_block + "```"
+
+    text2 = code_block_pattern.sub(replacement, text)
+    return text2
+
+
+def is_variable_assigned(var_name: str) -> bool:
+    return var_name in locals()

deepseek_vl/serve/assets/Kelpy-Codos.js

@@ -0,0 +1,100 @@
+/**
+ * Copyright (c) 2023-2024 DeepSeek.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy of
+ * this software and associated documentation files (the "Software"), to deal in
+ * the Software without restriction, including without limitation the rights to
+ * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+ * the Software, and to permit persons to whom the Software is furnished to do so,
+ * subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+ * FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+ * COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+ * IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+// ==UserScript==
+// @name         Kelpy Codos
+// @namespace    https://github.com/Keldos-Li/Kelpy-Codos
+// @version      1.0.5
+// @author       Keldos; https://keldos.me/
+// @description  Add copy button to PRE tags before CODE tag, for Chuanhu ChatGPT especially.
+//               Based on Chuanhu ChatGPT version: ac04408 (2023-3-22)
+// @license      GPL-3.0
+// @grant        none
+// ==/UserScript==
+
+(function () {
+  "use strict";
+
+  function addCopyButton(pre) {
+    var code = pre.querySelector("code");
+    if (!code) {
+      return; // 如果没有找到 <code> 元素，则不添加按钮
+    }
+    var firstChild = code.firstChild;
+    if (!firstChild) {
+      return; // 如果 <code> 元素没有子节点，则不添加按钮
+    }
+    var button = document.createElement("button");
+    button.textContent = "\uD83D\uDCCE"; // 使用 📎 符号作为“复制”按钮的文本
+    button.style.position = "relative";
+    button.style.float = "right";
+    button.style.fontSize = "1em"; // 可选：调整按钮大小
+    button.style.background = "none"; // 可选：去掉背景颜色
+    button.style.border = "none"; // 可选：去掉边框
+    button.style.cursor = "pointer"; // 可选：显示指针样式
+    button.addEventListener("click", function () {
+      var range = document.createRange();
+      range.selectNodeContents(code);
+      range.setStartBefore(firstChild); // 将范围设置为第一个子节点之前
+      var selection = window.getSelection();
+      selection.removeAllRanges();
+      selection.addRange(range);
+
+      try {
+        var success = document.execCommand("copy");
+        if (success) {
+          button.textContent = "\u2714";
+          setTimeout(function () {
+            button.textContent = "\uD83D\uDCCE"; // 恢复按钮为“复制”
+          }, 2000);
+        } else {
+          button.textContent = "\u2716";
+        }
+      } catch (e) {
+        console.error(e);
+        button.textContent = "\u2716";
+      }
+
+      selection.removeAllRanges();
+    });
+    code.insertBefore(button, firstChild); // 将按钮插入到第一个子元素之前
+  }
+
+  function handleNewElements(mutationsList, observer) {
+    for (var mutation of mutationsList) {
+      if (mutation.type === "childList") {
+        for (var node of mutation.addedNodes) {
+          if (node.nodeName === "PRE") {
+            addCopyButton(node);
+          }
+        }
+      }
+    }
+  }
+
+  var observer = new MutationObserver(handleNewElements);
+  observer.observe(document.documentElement, {
+    childList: true,
+    subtree: true,
+  });
+
+  document.querySelectorAll("pre").forEach(addCopyButton);
+})();

deepseek_vl/serve/assets/custom.css

@@ -0,0 +1,355 @@
+/**
+ * Copyright (c) 2023-2024 DeepSeek.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy of
+ * this software and associated documentation files (the "Software"), to deal in
+ * the Software without restriction, including without limitation the rights to
+ * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+ * the Software, and to permit persons to whom the Software is furnished to do so,
+ * subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+ * FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+ * COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+ * IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+:root {
+  --chatbot-color-light: #f3f3f3;
+  --chatbot-color-dark: #121111;
+}
+
+/* status_display */
+#status_display {
+  display: flex;
+  min-height: 2.5em;
+  align-items: flex-end;
+  justify-content: flex-end;
+}
+#status_display p {
+  font-size: 0.85em;
+  font-family: monospace;
+  color: var(--body-text-color-subdued);
+}
+
+/* usage_display */
+#usage_display {
+  height: 1em;
+}
+#usage_display p {
+  padding: 0 1em;
+  font-size: 0.85em;
+  font-family: monospace;
+  color: var(--body-text-color-subdued);
+}
+/* list */
+ol:not(.options),
+ul:not(.options) {
+  padding-inline-start: 2em !important;
+}
+
+/* Thank @Keldos-Li for fixing it */
+/* Light mode (default) */
+#deepseek_chatbot {
+  background-color: var(--chatbot-color-light) !important;
+  color: #000000 !important;
+}
+[data-testid="bot"] {
+  background-color: #ffffff !important;
+}
+[data-testid="user"] {
+  background-color: #95ec69 !important;
+}
+
+/* Dark mode */
+.dark #deepseek_chatbot {
+  background-color: var(--chatbot-color-dark) !important;
+  color: #ffffff !important;
+}
+.dark [data-testid="bot"] {
+  background-color: #2c2c2c !important;
+}
+.dark [data-testid="user"] {
+  background-color: #26b561 !important;
+}
+
+#deepseek_chatbot {
+  height: 100%;
+  min-height: 800px;
+  flex-grow: 1;
+  overflow: auto;
+}
+
+[class*="message"] {
+  border-radius: var(--radius-xl) !important;
+  border: none;
+  padding: var(--spacing-xl) !important;
+  font-size: var(--text-md) !important;
+  line-height: var(--line-md) !important;
+  min-height: calc(var(--text-md) * var(--line-md) + 2 * var(--spacing-xl));
+  min-width: calc(var(--text-md) * var(--line-md) + 2 * var(--spacing-xl));
+}
+[data-testid="bot"] {
+  max-width: 85%;
+  border-bottom-left-radius: 0 !important;
+}
+[data-testid="user"] {
+  max-width: 85%;
+  width: auto !important;
+  border-bottom-right-radius: 0 !important;
+}
+/* Table */
+table {
+  margin: 1em 0;
+  border-collapse: collapse;
+  empty-cells: show;
+}
+td,
+th {
+  border: 1.2px solid var(--border-color-primary) !important;
+  padding: 0.2em;
+}
+thead {
+  background-color: rgba(175, 184, 193, 0.2);
+}
+thead th {
+  padding: 0.5em 0.2em;
+}
+/* Inline code */
+#deepseek_chatbot code {
+  display: inline;
+  white-space: break-spaces;
+  border-radius: 6px;
+  margin: 0 2px 0 2px;
+  padding: 0.2em 0.4em 0.1em 0.4em;
+  background-color: rgba(175, 184, 193, 0.2);
+}
+/* Code block */
+#deepseek_chatbot pre code {
+  display: block;
+  overflow: auto;
+  white-space: pre;
+  background-color: #1c1d1e !important;
+  border-radius: 10px;
+  padding: 1.4em 1.2em 0em 1.4em;
+  margin: 1.2em 2em 1.2em 0.5em;
+  color: #fdf8f8;
+  box-shadow: 6px 6px 16px hsla(0, 0%, 0%, 0.2);
+}
+/* Hightlight */
+#deepseek_chatbot .highlight {
+  background-color: transparent;
+}
+#deepseek_chatbot .highlight .hll {
+  background-color: #49483e;
+}
+#deepseek_chatbot .highlight .c {
+  color: #75715e;
+} /* Comment */
+#deepseek_chatbot .highlight .err {
+  color: #960050;
+  background-color: #1e0010;
+} /* Error */
+#deepseek_chatbot .highlight .k {
+  color: #66d9ef;
+} /* Keyword */
+#deepseek_chatbot .highlight .l {
+  color: #ae81ff;
+} /* Literal */
+#deepseek_chatbot .highlight .n {
+  color: #f8f8f2;
+} /* Name */
+#deepseek_chatbot .highlight .o {
+  color: #f92672;
+} /* Operator */
+#deepseek_chatbot .highlight .p {
+  color: #f8f8f2;
+} /* Punctuation */
+#deepseek_chatbot .highlight .ch {
+  color: #75715e;
+} /* Comment.Hashbang */
+#deepseek_chatbot .highlight .cm {
+  color: #75715e;
+} /* Comment.Multiline */
+#deepseek_chatbot .highlight .cp {
+  color: #75715e;
+} /* Comment.Preproc */
+#deepseek_chatbot .highlight .cpf {
+  color: #75715e;
+} /* Comment.PreprocFile */
+#deepseek_chatbot .highlight .c1 {
+  color: #75715e;
+} /* Comment.Single */
+#deepseek_chatbot .highlight .cs {
+  color: #75715e;
+} /* Comment.Special */
+#deepseek_chatbot .highlight .gd {
+  color: #f92672;
+} /* Generic.Deleted */
+#deepseek_chatbot .highlight .ge {
+  font-style: italic;
+} /* Generic.Emph */
+#deepseek_chatbot .highlight .gi {
+  color: #a6e22e;
+} /* Generic.Inserted */
+#deepseek_chatbot .highlight .gs {
+  font-weight: bold;
+} /* Generic.Strong */
+#deepseek_chatbot .highlight .gu {
+  color: #75715e;
+} /* Generic.Subheading */
+#deepseek_chatbot .highlight .kc {
+  color: #66d9ef;
+} /* Keyword.Constant */
+#deepseek_chatbot .highlight .kd {
+  color: #66d9ef;
+} /* Keyword.Declaration */
+#deepseek_chatbot .highlight .kn {
+  color: #f92672;
+} /* Keyword.Namespace */
+#deepseek_chatbot .highlight .kp {
+  color: #66d9ef;
+} /* Keyword.Pseudo */
+#deepseek_chatbot .highlight .kr {
+  color: #66d9ef;
+} /* Keyword.Reserved */
+#deepseek_chatbot .highlight .kt {
+  color: #66d9ef;
+} /* Keyword.Type */
+#deepseek_chatbot .highlight .ld {
+  color: #e6db74;
+} /* Literal.Date */
+#deepseek_chatbot .highlight .m {
+  color: #ae81ff;
+} /* Literal.Number */
+#deepseek_chatbot .highlight .s {
+  color: #e6db74;
+} /* Literal.String */
+#deepseek_chatbot .highlight .na {
+  color: #a6e22e;
+} /* Name.Attribute */
+#deepseek_chatbot .highlight .nb {
+  color: #f8f8f2;
+} /* Name.Builtin */
+#deepseek_chatbot .highlight .nc {
+  color: #a6e22e;
+} /* Name.Class */
+#deepseek_chatbot .highlight .no {
+  color: #66d9ef;
+} /* Name.Constant */
+#deepseek_chatbot .highlight .nd {
+  color: #a6e22e;
+} /* Name.Decorator */
+#deepseek_chatbot .highlight .ni {
+  color: #f8f8f2;
+} /* Name.Entity */
+#deepseek_chatbot .highlight .ne {
+  color: #a6e22e;
+} /* Name.Exception */
+#deepseek_chatbot .highlight .nf {
+  color: #a6e22e;
+} /* Name.Function */
+#deepseek_chatbot .highlight .nl {
+  color: #f8f8f2;
+} /* Name.Label */
+#deepseek_chatbot .highlight .nn {
+  color: #f8f8f2;
+} /* Name.Namespace */
+#deepseek_chatbot .highlight .nx {
+  color: #a6e22e;
+} /* Name.Other */
+#deepseek_chatbot .highlight .py {
+  color: #f8f8f2;
+} /* Name.Property */
+#deepseek_chatbot .highlight .nt {
+  color: #f92672;
+} /* Name.Tag */
+#deepseek_chatbot .highlight .nv {
+  color: #f8f8f2;
+} /* Name.Variable */
+#deepseek_chatbot .highlight .ow {
+  color: #f92672;
+} /* Operator.Word */
+#deepseek_chatbot .highlight .w {
+  color: #f8f8f2;
+} /* Text.Whitespace */
+#deepseek_chatbot .highlight .mb {
+  color: #ae81ff;
+} /* Literal.Number.Bin */
+#deepseek_chatbot .highlight .mf {
+  color: #ae81ff;
+} /* Literal.Number.Float */
+#deepseek_chatbot .highlight .mh {
+  color: #ae81ff;
+} /* Literal.Number.Hex */
+#deepseek_chatbot .highlight .mi {
+  color: #ae81ff;
+} /* Literal.Number.Integer */
+#deepseek_chatbot .highlight .mo {
+  color: #ae81ff;
+} /* Literal.Number.Oct */
+#deepseek_chatbot .highlight .sa {
+  color: #e6db74;
+} /* Literal.String.Affix */
+#deepseek_chatbot .highlight .sb {
+  color: #e6db74;
+} /* Literal.String.Backtick */
+#deepseek_chatbot .highlight .sc {
+  color: #e6db74;
+} /* Literal.String.Char */
+#deepseek_chatbot .highlight .dl {
+  color: #e6db74;
+} /* Literal.String.Delimiter */
+#deepseek_chatbot .highlight .sd {
+  color: #e6db74;
+} /* Literal.String.Doc */
+#deepseek_chatbot .highlight .s2 {
+  color: #e6db74;
+} /* Literal.String.Double */
+#deepseek_chatbot .highlight .se {
+  color: #ae81ff;
+} /* Literal.String.Escape */
+#deepseek_chatbot .highlight .sh {
+  color: #e6db74;
+} /* Literal.String.Heredoc */
+#deepseek_chatbot .highlight .si {
+  color: #e6db74;
+} /* Literal.String.Interpol */
+#deepseek_chatbot .highlight .sx {
+  color: #e6db74;
+} /* Literal.String.Other */
+#deepseek_chatbot .highlight .sr {
+  color: #e6db74;
+} /* Literal.String.Regex */
+#deepseek_chatbot .highlight .s1 {
+  color: #e6db74;
+} /* Literal.String.Single */
+#deepseek_chatbot .highlight .ss {
+  color: #e6db74;
+} /* Literal.String.Symbol */
+#deepseek_chatbot .highlight .bp {
+  color: #f8f8f2;
+} /* Name.Builtin.Pseudo */
+#deepseek_chatbot .highlight .fm {
+  color: #a6e22e;
+} /* Name.Function.Magic */
+#deepseek_chatbot .highlight .vc {
+  color: #f8f8f2;
+} /* Name.Variable.Class */
+#deepseek_chatbot .highlight .vg {
+  color: #f8f8f2;
+} /* Name.Variable.Global */
+#deepseek_chatbot .highlight .vi {
+  color: #f8f8f2;
+} /* Name.Variable.Instance */
+#deepseek_chatbot .highlight .vm {
+  color: #f8f8f2;
+} /* Name.Variable.Magic */
+#deepseek_chatbot .highlight .il {
+  color: #ae81ff;
+} /* Literal.Number.Integer.Long */

deepseek_vl/serve/assets/custom.js

@@ -0,0 +1,22 @@
+/**
+ * Copyright (c) 2023-2024 DeepSeek.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy of
+ * this software and associated documentation files (the "Software"), to deal in
+ * the Software without restriction, including without limitation the rights to
+ * use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+ * the Software, and to permit persons to whom the Software is furnished to do so,
+ * subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+ * FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+ * COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+ * IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+// custom javascript here

deepseek_vl/serve/inference.py

@@ -0,0 +1,170 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+from threading import Thread
+from typing import List
+
+import torch
+import transformers
+from transformers import (
+    AutoModelForCausalLM,
+    StoppingCriteria,
+    StoppingCriteriaList,
+    TextIteratorStreamer,
+)
+
+from deepseek_vl.models import MultiModalityCausalLM, VLChatProcessor
+from deepseek_vl.utils.conversation import Conversation
+
+
+def load_model(model_path):
+    vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(model_path)
+    tokenizer = vl_chat_processor.tokenizer
+    vl_gpt: MultiModalityCausalLM = AutoModelForCausalLM.from_pretrained(
+        model_path, trust_remote_code=True
+    )
+    vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()
+    return tokenizer, vl_gpt, vl_chat_processor
+
+
+def convert_conversation_to_prompts(conversation: Conversation):
+    prompts = []
+    messages = conversation.messages
+
+    for i in range(0, len(messages), 2):
+        prompt = {
+            "role": messages[i][0],
+            "content": (
+                messages[i][1][0]
+                if isinstance(messages[i][1], tuple)
+                else messages[i][1]
+            ),
+            "images": [messages[i][1][1]] if isinstance(messages[i][1], tuple) else [],
+        }
+        response = {"role": messages[i + 1][0], "content": messages[i + 1][1]}
+        prompts.extend([prompt, response])
+
+    return prompts
+
+
+class StoppingCriteriaSub(StoppingCriteria):
+    def __init__(self, stops=[], encounters=1):
+        super().__init__()
+        self.stops = [stop.to("cuda") for stop in stops]
+
+    def __call__(
+        self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs
+    ):
+        for stop in self.stops:
+            if input_ids.shape[-1] < len(stop):
+                continue
+            if torch.all((stop == input_ids[0][-len(stop) :])).item():
+                return True
+
+        return False
+
+
+@torch.inference_mode()
+def deepseek_generate(
+    prompts: list,
+    vl_gpt: torch.nn.Module,
+    vl_chat_processor,
+    tokenizer: transformers.PreTrainedTokenizer,
+    stop_words: list,
+    max_length: int = 256,
+    temperature: float = 1.0,
+    top_p: float = 1.0,
+    repetition_penalty=1.1,
+):
+    prompts = prompts
+    pil_images = list()
+    for message in prompts:
+        if "images" not in message:
+            continue
+        for pil_img in message["images"]:
+            pil_images.append(pil_img)
+
+    prepare_inputs = vl_chat_processor(
+        conversations=prompts, images=pil_images, force_batchify=True
+    ).to(vl_gpt.device)
+
+    return generate(
+        vl_gpt,
+        tokenizer,
+        prepare_inputs,
+        max_length,
+        temperature,
+        repetition_penalty,
+        top_p,
+        stop_words,
+    )
+
+
+@torch.inference_mode()
+def generate(
+    vl_gpt,
+    tokenizer,
+    prepare_inputs,
+    max_gen_len: int = 256,
+    temperature: float = 0,
+    repetition_penalty=1.1,
+    top_p: float = 0.95,
+    stop_words: List[str] = [],
+):
+    """Stream the text output from the multimodality model with prompt and image inputs."""
+    inputs_embeds = vl_gpt.prepare_inputs_embeds(**prepare_inputs)
+
+    streamer = TextIteratorStreamer(tokenizer)
+
+    stop_words_ids = [
+        torch.tensor(tokenizer.encode(stop_word)) for stop_word in stop_words
+    ]
+    stopping_criteria = StoppingCriteriaList(
+        [StoppingCriteriaSub(stops=stop_words_ids)]
+    )
+
+    generation_config = dict(
+        inputs_embeds=inputs_embeds,
+        attention_mask=prepare_inputs.attention_mask,
+        pad_token_id=tokenizer.eos_token_id,
+        bos_token_id=tokenizer.bos_token_id,
+        eos_token_id=tokenizer.eos_token_id,
+        max_new_tokens=max_gen_len,
+        do_sample=True,
+        use_cache=True,
+        streamer=streamer,
+        stopping_criteria=stopping_criteria,
+    )
+
+    if temperature > 0:
+        generation_config.update(
+            {
+                "do_sample": True,
+                "top_p": top_p,
+                "temperature": temperature,
+                "repetition_penalty": repetition_penalty,
+            }
+        )
+    else:
+        generation_config["do_sample"] = False
+
+    thread = Thread(target=vl_gpt.language_model.generate, kwargs=generation_config)
+    thread.start()
+
+    yield from streamer

deepseek_vl/utils/__init__.py

@@ -0,0 +1,18 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

deepseek_vl/utils/conversation.py

@@ -0,0 +1,348 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+"""
+From https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+"""
+
+import dataclasses
+from enum import IntEnum, auto
+from typing import Dict, List
+
+
+class SeparatorStyle(IntEnum):
+    """Separator styles."""
+
+    ADD_COLON_SINGLE = auto()
+    ADD_COLON_TWO = auto()
+    ADD_COLON_SPACE_SINGLE = auto()
+    NO_COLON_SINGLE = auto()
+    NO_COLON_TWO = auto()
+    ADD_NEW_LINE_SINGLE = auto()
+    LLAMA2 = auto()
+    CHATGLM = auto()
+    CHATML = auto()
+    CHATINTERN = auto()
+    DOLLY = auto()
+    RWKV = auto()
+    PHOENIX = auto()
+    ROBIN = auto()
+    DeepSeek = auto()
+    PLAIN = auto()
+    ALIGNMENT = auto()
+
+
+@dataclasses.dataclass
+class Conversation:
+    """A class that manages prompt templates and keeps all conversation history."""
+
+    # The name of this template
+    name: str
+    # The template of the system prompt
+    system_template: str = "{system_message}"
+    # The system message
+    system_message: str = ""
+    # The names of two roles
+    roles: List[str] = (("USER", "ASSISTANT"),)
+    # All messages. Each item is (role, message).
+    messages: List[List[str]] = ()
+    # The number of few shot examples
+    offset: int = 0
+    # The separator style and configurations
+    sep_style: SeparatorStyle = SeparatorStyle.ADD_COLON_SINGLE
+    sep: str = "\n"
+    sep2: str = None
+    # Stop criteria (the default one is EOS token)
+    stop_str: str = None
+    # Stops generation if meeting any token in this list
+    stop_token_ids: List[int] = None
+
+    def get_prompt(self) -> str:
+        """Get the prompt for generation."""
+        system_prompt = self.system_template.format(system_message=self.system_message)
+
+        if self.sep_style == SeparatorStyle.DeepSeek:
+            seps = [self.sep, self.sep2]
+            if system_prompt == "" or system_prompt is None:
+                ret = ""
+            else:
+                ret = system_prompt + seps[0]
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    ret += role + ": " + message + seps[i % 2]
+                else:
+                    ret += role + ":"
+            return ret
+        elif self.sep_style == SeparatorStyle.LLAMA2:
+            seps = [self.sep, self.sep2]
+            if self.system_message:
+                ret = system_prompt
+            else:
+                ret = "[INST] "
+            for i, (role, message) in enumerate(self.messages):
+                tag = self.roles[i % 2]
+                if message:
+                    if type(message) is tuple:  # multimodal message
+                        message, _ = message
+                    if i == 0:
+                        ret += message + " "
+                    else:
+                        ret += tag + " " + message + seps[i % 2]
+                else:
+                    ret += tag
+            return ret
+        elif self.sep_style == SeparatorStyle.PLAIN:
+            seps = [self.sep, self.sep2]
+            ret = ""
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    if i % 2 == 0:
+                        ret += message + seps[i % 2]
+                    else:
+                        ret += message + seps[i % 2]
+                else:
+                    ret += ""
+            return ret
+        elif self.sep_style == SeparatorStyle.ALIGNMENT:
+            seps = [self.sep, self.sep2]
+            ret = ""
+            for i, (role, message) in enumerate(self.messages):
+                if message:
+                    if type(message) is tuple:
+                        message, _, _ = message
+                    if i % 2 == 0:
+                        ret += "<image>\n" + seps[i % 2]
+                    else:
+                        ret += message + seps[i % 2]
+                else:
+                    ret += ""
+            return ret
+        else:
+            raise ValueError(f"Invalid style: {self.sep_style}")
+
+    def get_prompt_for_current_round(self, content=None):
+        """Get current round formatted question prompt during sft training"""
+        if self.sep_style == SeparatorStyle.PLAIN:
+            formatted_question = "<image>\n"
+        elif self.sep_style == SeparatorStyle.DeepSeek:
+            formatted_question = (
+                f"{self.roles[0]}: " + content.strip() + self.sep + f"{self.roles[1]}:"
+            )
+        else:
+            raise ValueError(f"Unsupported sep_style: {self.sep_style}")
+        return formatted_question
+
+    def set_system_message(self, system_message: str):
+        """Set the system message."""
+        self.system_message = system_message
+
+    def append_message(self, role: str, message: str):
+        """Append a new message."""
+        self.messages.append([role, message])
+
+    def reset_message(self):
+        """Reset a new message."""
+        self.messages = []
+
+    def update_last_message(self, message: str):
+        """Update the last output.
+
+        The last message is typically set to be None when constructing the prompt,
+        so we need to update it in-place after getting the response from a model.
+        """
+        self.messages[-1][1] = message
+
+    def to_gradio_chatbot(self):
+        """Convert the conversation to gradio chatbot format."""
+        ret = []
+        for i, (role, msg) in enumerate(self.messages[self.offset :]):
+            if i % 2 == 0:
+                ret.append([msg, None])
+            else:
+                ret[-1][-1] = msg
+        return ret
+
+    def to_openai_api_messages(self):
+        """Convert the conversation to OpenAI chat completion format."""
+        system_prompt = self.system_template.format(system_message=self.system_message)
+        ret = [{"role": "system", "content": system_prompt}]
+
+        for i, (_, msg) in enumerate(self.messages[self.offset :]):
+            if i % 2 == 0:
+                ret.append({"role": "user", "content": msg})
+            else:
+                if msg is not None:
+                    ret.append({"role": "assistant", "content": msg})
+        return ret
+
+    def copy(self):
+        return Conversation(
+            name=self.name,
+            system_template=self.system_template,
+            system_message=self.system_message,
+            roles=self.roles,
+            messages=[[x, y] for x, y in self.messages],
+            offset=self.offset,
+            sep_style=self.sep_style,
+            sep=self.sep,
+            sep2=self.sep2,
+            stop_str=self.stop_str,
+            stop_token_ids=self.stop_token_ids,
+        )
+
+    def dict(self):
+        return {
+            "template_name": self.name,
+            "system_message": self.system_message,
+            "roles": self.roles,
+            "messages": self.messages,
+            "offset": self.offset,
+        }
+
+
+# A global registry for all conversation templates
+conv_templates: Dict[str, Conversation] = {}
+
+
+def register_conv_template(template: Conversation, override: bool = False):
+    """Register a new conversation template."""
+    if not override:
+        assert (
+            template.name not in conv_templates
+        ), f"{template.name} has been registered."
+
+    conv_templates[template.name] = template
+
+
+def get_conv_template(name: str) -> Conversation:
+    """Get a conversation template."""
+    return conv_templates[name].copy()
+
+
+# llava_llama2 template
+register_conv_template(
+    Conversation(
+        name="llava_llama2",
+        system_message="You are a helpful language and vision assistant. "
+        "You are able to understand the visual content that the user provides, "
+        "and assist the user with a variety of tasks using natural language.",
+        system_template="[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n",
+        roles=("[INST]", "[/INST]"),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.LLAMA2,
+        sep=" ",
+        sep2=" </s><s>",
+        stop_token_ids=[2],
+    )
+)
+
+# llama2 template
+# reference: https://github.com/facebookresearch/llama/blob/cfc3fc8c1968d390eb830e65c63865e980873a06/llama/generation.py#L212
+register_conv_template(
+    Conversation(
+        name="llama-2",
+        system_template="[INST] <<SYS>>\n{system_message}\n<</SYS>>\n\n",
+        roles=("[INST]", "[/INST]"),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.LLAMA2,
+        sep=" ",
+        sep2=" </s><s>",
+        stop_token_ids=[2],
+    )
+)
+
+
+# deepseek template
+register_conv_template(
+    Conversation(
+        name="deepseek",
+        system_template="{system_message}",
+        # system_message="You are a helpful assistant. Please answer truthfully and write out your "
+        # "thinking step by step to be sure you get the right answer.",
+        system_message="",
+        roles=("User", "Assistant"),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.DeepSeek,
+        sep="\n\n",
+        sep2="<｜end▁of▁sentence｜>",
+        stop_token_ids=[100001],
+        stop_str=["User:", "<｜end▁of▁sentence｜>"],
+    )
+)
+
+register_conv_template(
+    Conversation(
+        name="plain",
+        system_template="",
+        system_message="",
+        roles=("", ""),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.PLAIN,
+        sep="",
+        sep2="",
+        stop_token_ids=[2],
+        stop_str=["</s>"],
+    )
+)
+
+
+register_conv_template(
+    Conversation(
+        name="alignment",
+        system_template="",
+        system_message="",
+        roles=("", ""),
+        messages=(),
+        offset=0,
+        sep_style=SeparatorStyle.ALIGNMENT,
+        sep="",
+        sep2="",
+        stop_token_ids=[2],
+        stop_str=["</s>"],
+    )
+)
+
+
+if __name__ == "__main__":
+    # print("Llama-2 template:")
+    # conv = get_conv_template("llama-2")
+    # conv.set_system_message("You are a helpful, respectful and honest assistant.")
+    # conv.append_message(conv.roles[0], "Hello!")
+    # conv.append_message(conv.roles[1], "Hi!")
+    # conv.append_message(conv.roles[0], "How are you?")
+    # conv.append_message(conv.roles[1], None)
+    # print(conv.get_prompt())
+
+    # print("\n")
+
+    print("deepseek template:")
+    conv = get_conv_template("deepseek")
+    conv.append_message(conv.roles[0], "Hello!")
+    conv.append_message(conv.roles[1], "Hi! This is Tony.")
+    conv.append_message(conv.roles[0], "Who are you?")
+    conv.append_message(conv.roles[1], "I am a helpful assistant.")
+    conv.append_message(conv.roles[0], "How are you?")
+    conv.append_message(conv.roles[1], None)
+    print(conv.get_prompt())

deepseek_vl/utils/io.py

@@ -0,0 +1,89 @@
+# Copyright (c) 2023-2024 DeepSeek.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy of
+# this software and associated documentation files (the "Software"), to deal in
+# the Software without restriction, including without limitation the rights to
+# use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+# the Software, and to permit persons to whom the Software is furnished to do so,
+# subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS
+# FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR
+# COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER
+# IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+# CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+import json
+from typing import Dict, List
+
+import PIL.Image
+import torch
+import base64
+import io
+from transformers import AutoModelForCausalLM
+
+from deepseek_vl.models import MultiModalityCausalLM, VLChatProcessor
+
+
+def load_pretrained_model(model_path: str):
+    vl_chat_processor: VLChatProcessor = VLChatProcessor.from_pretrained(model_path)
+    tokenizer = vl_chat_processor.tokenizer
+
+    vl_gpt: MultiModalityCausalLM = AutoModelForCausalLM.from_pretrained(
+        model_path, trust_remote_code=True
+    )
+    vl_gpt = vl_gpt.to(torch.bfloat16).cuda().eval()
+
+    return tokenizer, vl_chat_processor, vl_gpt
+
+
+def load_pil_images(conversations: List[Dict[str, str]]) -> List[PIL.Image.Image]:
+    """
+
+    Support file path or base64 images.
+
+    Args:
+        conversations (List[Dict[str, str]]): the conversations with a list of messages. An example is :
+            [
+                {
+                    "role": "User",
+                    "content": "<image_placeholder>\nExtract all information from this image and convert them into markdown format.",
+                    "images": ["./examples/table_datasets.png"]
+                },
+                {"role": "Assistant", "content": ""},
+            ]
+
+    Returns:
+        pil_images (List[PIL.Image.Image]): the list of PIL images.
+
+    """
+
+    pil_images = []
+
+    for message in conversations:
+        if "images" not in message:
+            continue
+
+        for image_data in message["images"]:
+            if image_data.startswith("data:image"):
+                # Image data is in base64 format
+                _, image_data = image_data.split(",", 1)
+                image_bytes = base64.b64decode(image_data)
+                pil_img = PIL.Image.open(io.BytesIO(image_bytes))
+            else:
+                # Image data is a file path
+                pil_img = PIL.Image.open(image_data)
+            pil_img = pil_img.convert("RGB")
+            pil_images.append(pil_img)
+
+    return pil_images
+
+
+def load_json(filepath):
+    with open(filepath, "r") as f:
+        data = json.load(f)
+        return data