Initial commit

Author: abetlen

.gitignore

@@ -0,0 +1,164 @@
+.envrc
+
+models/
+
+# 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/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# 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
+
+# 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/

LICENSE.md

@@ -0,0 +1,9 @@
+MIT License
+
+Copyright (c) 2023 Andrei Betlen
+
+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.

README.md

@@ -0,0 +1,40 @@
+# `llama.cpp` Python Bindings
+
+Simple Python bindings for @ggerganov's [`llama.cpp`](https://github.com/ggerganov/llama.cpp) library.
+
+These bindings expose the low-level `llama.cpp` C API through a complete `ctypes` interface.
+This module also exposes a high-level Python API that is more convenient to use and follows a familiar format.
+
+# Install
+
+```bash
+pip install llama_cpp
+```
+
+# Usage
+
+```python
+>>> from llama_cpp import Llama
+>>> llm = Llama(model_path="models/7B/...")
+>>> output = llm("Q: Name the planets in the solar system? A: ", max_tokens=32, stop=["Q:", "\n"], echo=True)
+>>> print(output)
+{
+  "id": "cmpl-xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx",
+  "object": "text_completion",
+  "created": 1679561337,
+  "model": "models/7B/...",
+  "choices": [
+    {
+      "text": "Q: Name the planets in the solar system? A: Mercury, Venus, Earth, Mars, Jupiter, Saturn, Uranus, Neptune and Pluto.",
+      "index": 0,
+      "logprobs": None,
+      "finish_reason": "stop"
+    }
+  ],
+  "usage": {
+    "prompt_tokens": 14,
+    "completion_tokens": 28,
+    "total_tokens": 42
+  }
+}
+```

examples/basic.py

@@ -0,0 +1,8 @@
+import json
+from llama_cpp import Llama
+
+llm = Llama(model_path="models/...")
+
+output = llm("Q: Name the planets in the solar system? A: ", max_tokens=32, stop=["Q:", "\n"], echo=True)
+
+print(json.dumps(output, indent=2))

llama_cpp/__init__.py

@@ -0,0 +1,2 @@
+from .llama_cpp import *
+from .llama import *

llama_cpp/llama.py

@@ -0,0 +1,131 @@
+import uuid
+import time
+import multiprocessing
+from typing import List, Optional
+
+from . import llama_cpp
+
+class Llama:
+    def __init__(
+        self,
+        model_path: str,
+        n_ctx: int = 512,
+        n_parts: int = -1,
+        seed: int = 1337,
+        f16_kv: bool = False,
+        logits_all: bool = False,
+        vocab_only: bool = False,
+        n_threads: Optional[int] = None,
+        model_name: Optional[str]=None,
+    ):
+        self.model_path = model_path
+        self.model = model_name or model_path
+
+        self.params = llama_cpp.llama_context_default_params()
+        self.params.n_ctx = n_ctx
+        self.params.n_parts = n_parts
+        self.params.seed = seed
+        self.params.f16_kv = f16_kv
+        self.params.logits_all = logits_all
+        self.params.vocab_only = vocab_only
+
+        self.n_threads = n_threads or multiprocessing.cpu_count()
+
+        self.tokens = (llama_cpp.llama_token * self.params.n_ctx)()
+
+        self.ctx = llama_cpp.llama_init_from_file(
+            self.model_path.encode("utf-8"), self.params
+        )
+
+    def __call__(
+        self,
+        prompt: str,
+        suffix: Optional[str] = None,
+        max_tokens: int = 16,
+        temperature: float = 0.8,
+        top_p: float = 0.95,
+        echo: bool = False,
+        stop: List[str] = [],
+        repeat_penalty: float = 1.1,
+        top_k: int = 40,
+    ):
+        text = ""
+        finish_reason = "length"
+        completion_tokens = 0
+
+        prompt_tokens = llama_cpp.llama_tokenize(
+            self.ctx, prompt.encode("utf-8"), self.tokens, self.params.n_ctx, True
+        )
+
+        if prompt_tokens + max_tokens > self.params.n_ctx:
+            raise ValueError(
+                f"Requested tokens exceed context window of {self.params.n_ctx}"
+            )
+
+        for i in range(prompt_tokens):
+            llama_cpp.llama_eval(
+                self.ctx, (llama_cpp.c_int * 1)(self.tokens[i]), 1, i, self.n_threads
+            )
+
+        for i in range(max_tokens):
+            token = llama_cpp.llama_sample_top_p_top_k(
+                self.ctx,
+                self.tokens,
+                prompt_tokens + completion_tokens,
+                top_k=top_k,
+                top_p=top_p,
+                temp=temperature,
+                repeat_penalty=repeat_penalty,
+            )
+            if token == llama_cpp.llama_token_eos():
+                finish_reason = "stop"
+                break
+            text += llama_cpp.llama_token_to_str(self.ctx, token).decode("utf-8")
+            self.tokens[prompt_tokens + i] = token
+            completion_tokens += 1
+
+            any_stop = [s for s in stop if s in text]
+            if len(any_stop) > 0:
+                first_stop = any_stop[0]
+                text = text[: text.index(first_stop)]
+                finish_reason = "stop"
+                break
+
+            llama_cpp.llama_eval(
+                self.ctx,
+                (llama_cpp.c_int * 1)(self.tokens[prompt_tokens + i]),
+                1,
+                prompt_tokens + completion_tokens,
+                self.n_threads,
+            )
+
+        if echo:
+            text = prompt + text
+
+        if suffix is not None:
+            text = text + suffix
+
+        return {
+            "id": f"cmpl-{str(uuid.uuid4())}", # Likely to change
+            "object": "text_completion",
+            "created": int(time.time()),
+            "model": self.model, # Likely to change
+            "choices": [
+                {
+                    "text": text,
+                    "index": 0,
+                    "logprobs": None,
+                    "finish_reason": finish_reason,
+                }
+            ],
+            "usage": {
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "total_tokens": prompt_tokens + completion_tokens,
+            },
+        }
+
+    def __del__(self):
+        llama_cpp.llama_free(self.ctx)
+
+

llama_cpp/llama_cpp.py

@@ -0,0 +1,157 @@
+import ctypes
+
+from ctypes import c_int, c_float, c_double, c_char_p, c_void_p, c_bool, POINTER, Structure
+
+import pathlib
+
+# Load the library
+libfile = pathlib.Path(__file__).parent.parent / "libllama.so"
+lib = ctypes.CDLL(str(libfile))
+
+
+# C types
+llama_token = c_int
+llama_token_p = POINTER(llama_token)
+
+class llama_token_data(Structure):
+    _fields_ = [
+        ('id', llama_token), # token id
+        ('p', c_float), # probability of the token
+        ('plog', c_float), # log probability of the token
+    ]
+
+llama_token_data_p = POINTER(llama_token_data)
+
+class llama_context_params(Structure):
+    _fields_ = [
+        ('n_ctx', c_int), # text context
+        ('n_parts', c_int), # -1 for default
+        ('seed', c_int), # RNG seed, 0 for random
+        ('f16_kv', c_bool), # use fp16 for KV cache
+        ('logits_all', c_bool), # the llama_eval() call computes all logits, not just the last one
+
+        ('vocab_only', c_bool), # only load the vocabulary, no weights
+    ]
+
+llama_context_params_p = POINTER(llama_context_params)
+
+llama_context_p = c_void_p
+
+# C functions
+lib.llama_context_default_params.argtypes = []
+lib.llama_context_default_params.restype = llama_context_params
+
+lib.llama_init_from_file.argtypes = [c_char_p, llama_context_params]
+lib.llama_init_from_file.restype = llama_context_p
+
+lib.llama_free.argtypes = [llama_context_p]
+lib.llama_free.restype = None
+
+lib.llama_model_quantize.argtypes = [c_char_p, c_char_p, c_int, c_int]
+lib.llama_model_quantize.restype = c_int
+
+lib.llama_eval.argtypes = [llama_context_p, llama_token_p, c_int, c_int, c_int]
+lib.llama_eval.restype = c_int
+
+lib.llama_tokenize.argtypes = [llama_context_p, c_char_p, llama_token_p, c_int, c_bool]
+lib.llama_tokenize.restype = c_int
+
+lib.llama_n_vocab.argtypes = [llama_context_p]
+lib.llama_n_vocab.restype = c_int
+
+lib.llama_n_ctx.argtypes = [llama_context_p]
+lib.llama_n_ctx.restype = c_int
+
+lib.llama_get_logits.argtypes = [llama_context_p]
+lib.llama_get_logits.restype = POINTER(c_float)
+
+lib.llama_token_to_str.argtypes = [llama_context_p, llama_token]
+lib.llama_token_to_str.restype = c_char_p
+
+lib.llama_token_bos.argtypes = []
+lib.llama_token_bos.restype = llama_token
+
+lib.llama_token_eos.argtypes = []
+lib.llama_token_eos.restype = llama_token
+
+lib.llama_sample_top_p_top_k.argtypes = [llama_context_p, llama_token_p, c_int, c_int, c_double, c_double, c_double]
+lib.llama_sample_top_p_top_k.restype = llama_token
+
+lib.llama_print_timings.argtypes = [llama_context_p]
+lib.llama_print_timings.restype = None
+
+lib.llama_reset_timings.argtypes = [llama_context_p]
+lib.llama_reset_timings.restype = None
+
+lib.llama_print_system_info.argtypes = []
+lib.llama_print_system_info.restype = c_char_p
+
+# Python functions
+def llama_context_default_params() -> llama_context_params:
+    params = lib.llama_context_default_params()
+    return params
+
+def llama_init_from_file(path_model: bytes, params: llama_context_params) -> llama_context_p:
+    """Various functions for loading a ggml llama model.
+    Allocate (almost) all memory needed for the model.
+    Return NULL on failure """
+    return lib.llama_init_from_file(path_model, params)
+
+def llama_free(ctx: llama_context_p):
+    """Free all allocated memory"""
+    lib.llama_free(ctx)
+
+def llama_model_quantize(fname_inp: bytes, fname_out: bytes, itype: c_int, qk: c_int) -> c_int:
+    """Returns 0 on success"""
+    return lib.llama_model_quantize(fname_inp, fname_out, itype, qk)
+
+def llama_eval(ctx: llama_context_p, tokens: llama_token_p, n_tokens: c_int, n_past: c_int, n_threads: c_int) -> c_int:
+    """Run the llama inference to obtain the logits and probabilities for the next token.
+    tokens + n_tokens is the provided batch of new tokens to process
+    n_past is the number of tokens to use from previous eval calls
+    Returns 0 on success"""
+    return lib.llama_eval(ctx, tokens, n_tokens, n_past, n_threads)
+
+def llama_tokenize(ctx: llama_context_p, text: bytes, tokens: llama_token_p, n_max_tokens: c_int, add_bos: c_bool) -> c_int:
+    """Convert the provided text into tokens.
+    The tokens pointer must be large enough to hold the resulting tokens.
+    Returns the number of tokens on success, no more than n_max_tokens
+    Returns a negative number on failure - the number of tokens that would have been returned"""
+    return lib.llama_tokenize(ctx, text, tokens, n_max_tokens, add_bos)
+
+def llama_n_vocab(ctx: llama_context_p) -> c_int:
+    return lib.llama_n_vocab(ctx)
+
+def llama_n_ctx(ctx: llama_context_p) -> c_int:
+    return lib.llama_n_ctx(ctx)
+
+def llama_get_logits(ctx: llama_context_p):
+    """Token logits obtained from the last call to llama_eval()
+    The logits for the last token are stored in the last row
+    Can be mutated in order to change the probabilities of the next token
+    Rows: n_tokens
+    Cols: n_vocab"""
+    return lib.llama_get_logits(ctx)
+
+def llama_token_to_str(ctx: llama_context_p, token: int) -> bytes:
+    """Token Id -> String. Uses the vocabulary in the provided context"""
+    return lib.llama_token_to_str(ctx, token)
+
+def llama_token_bos() -> llama_token:
+    return lib.llama_token_bos()
+
+def llama_token_eos() -> llama_token:
+    return lib.llama_token_eos()
+
+def llama_sample_top_p_top_k(ctx: llama_context_p, last_n_tokens_data: llama_token_p, last_n_tokens_size: c_int, top_k: c_int, top_p: c_double, temp: c_double, repeat_penalty: c_double) -> llama_token:
+    return lib.llama_sample_top_p_top_k(ctx, last_n_tokens_data, last_n_tokens_size, top_k, top_p, temp, repeat_penalty)
+
+def llama_print_timings(ctx: llama_context_p):
+    lib.llama_print_timings(ctx)
+
+def llama_reset_timings(ctx: llama_context_p):
+    lib.llama_reset_timings(ctx)
+
+def llama_print_system_info() -> bytes:
+    """Print system informaiton"""
+    return lib.llama_print_system_info()

poetry.lock

@@ -0,0 +1,24 @@
+[tool.poetry]
+name = "llama_cpp"
+version = "0.1.0"
+description = "Python bindings for the llama.cpp library"
+authors = ["Andrei Betlen <abetlen@gmail.com>"]
+license = "MIT"
+readme = "README.md"
+homepage = "https://github.com/abetlen/llama_cpp_python"
+repository = "https://github.com/abetlen/llama_cpp_python"
+packages = [{include = "llama_cpp"}]
+include = [
+    "LICENSE.md",
+]
+
+[tool.poetry.dependencies]
+python = "^3.8.1"
+
+
+[tool.poetry.group.dev.dependencies]
+black = "^23.1.0"
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"

pyproject.toml

@@ -0,0 +1,42 @@
+import os
+import subprocess
+from setuptools import setup, Extension
+
+from distutils.command.build_ext import build_ext
+
+
+class build_ext_custom(build_ext):
+    def run(self):
+        build_dir = os.path.join(os.getcwd(), "build")
+        src_dir = os.path.join(os.getcwd(), "vendor", "llama.cpp")
+
+        os.makedirs(build_dir, exist_ok=True)
+
+        cmake_flags = [
+            "-DLLAMA_STATIC=Off",
+            "-DBUILD_SHARED_LIBS=On",
+            "-DCMAKE_CXX_FLAGS=-fPIC",
+            "-DCMAKE_C_FLAGS=-fPIC",
+        ]
+        subprocess.check_call(["cmake", src_dir, *cmake_flags], cwd=build_dir)
+        subprocess.check_call(["cmake", "--build", "."], cwd=build_dir)
+
+        # Move the shared library to the root directory
+        lib_path = os.path.join(build_dir, "libllama.so")
+        target_path = os.path.join(os.getcwd(), "libllama.so")
+        os.rename(lib_path, target_path)
+
+
+setup(
+    name="llama_cpp",
+    description="A Python wrapper for llama.cpp",
+    version="0.0.1",
+    author="Andrei Betlen",
+    author_email="abetlen@gmail.com",
+    license="MIT",
+    py_modules=["llama_cpp"],
+    ext_modules=[
+        Extension("libllama", ["vendor/llama.cpp"]),
+    ],
+    cmdclass={"build_ext": build_ext_custom},
+)