baalajimaestro/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Merge branch 'main' into setup

Browse source
This commit is contained in:
Andrei Betlen 2023-05-26 17:53:58 -04:00
commit 447a3d249e
22 changed files with 1095 additions and 254 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
.gitmodules vendored
View file

@ -1,3 +1,3 @@
[submodule "vendor/llama.cpp"] [submodule "vendor/llama.cpp"]
path = vendor/llama.cpp path = vendor/llama.cpp
url = git@github.com:ggerganov/llama.cpp.git url = https://github.com/ggerganov/llama.cpp.git

12
CHANGELOG.md Normal file
View file

@ -0,0 +1,12 @@
# Changelog
All notable changes to this project will be documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [Unreleased]
### Added
- Added first version of the changelog

2
CMakeLists.txt
View file

@ -28,4 +28,4 @@ else()
LIBRARY DESTINATION llama_cpp LIBRARY DESTINATION llama_cpp
RUNTIME DESTINATION llama_cpp RUNTIME DESTINATION llama_cpp
) )
endif(UNIX) endif()

24
README.md
View file

@ -15,6 +15,8 @@ This package provides:
- OpenAI-like API - OpenAI-like API
- LangChain compatibility - LangChain compatibility
Documentation is available at [https://abetlen.github.io/llama-cpp-python](https://abetlen.github.io/llama-cpp-python).
## Installation from PyPI (recommended) ## Installation from PyPI (recommended)
Install from PyPI (requires a c compiler): Install from PyPI (requires a c compiler):
@ -26,6 +28,18 @@ pip install llama-cpp-python
The above command will attempt to install the package and build build `llama.cpp` from source. The above command will attempt to install the package and build build `llama.cpp` from source.
This is the recommended installation method as it ensures that `llama.cpp` is built with the available optimizations for your system. This is the recommended installation method as it ensures that `llama.cpp` is built with the available optimizations for your system.
If you have previously installed `llama-cpp-python` through pip and want to upgrade your version or rebuild the package with different compiler options, please add the following flags to ensure that the package is rebuilt correctly:
```bash
pip install llama-cpp-python --force-reinstall --upgrade --no-cache-dir
```
Note: If you are using Apple Silicon (M1) Mac, make sure you have installed a version of Python that supports arm64 architecture. For example:
```
wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-MacOSX-arm64.sh
bash Miniforge3-MacOSX-arm64.sh
```
Otherwise, while installing it will build the llama.ccp x86 version which will be 10x slower on Apple Silicon (M1) Mac.
### Installation with OpenBLAS / cuBLAS / CLBlast ### Installation with OpenBLAS / cuBLAS / CLBlast
@ -35,19 +49,19 @@ Use the `FORCE_CMAKE=1` environment variable to force the use of `cmake` and ins
To install with OpenBLAS, set the `LLAMA_OPENBLAS=1` environment variable before installing: To install with OpenBLAS, set the `LLAMA_OPENBLAS=1` environment variable before installing:
```bash ```bash
LLAMA_OPENBLAS=1 FORCE_CMAKE=1 pip install llama-cpp-python CMAKE_ARGS="-DLLAMA_OPENBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python
``` ```
To install with cuBLAS, set the `LLAMA_CUBLAS=1` environment variable before installing: To install with cuBLAS, set the `LLAMA_CUBLAS=1` environment variable before installing:
```bash ```bash
LLAMA_CUBLAS=1 FORCE_CMAKE=1 pip install llama-cpp-python CMAKE_ARGS="-DLLAMA_CUBLAS=on" FORCE_CMAKE=1 pip install llama-cpp-python
``` ```
To install with CLBlast, set the `LLAMA_CLBLAST=1` environment variable before installing: To install with CLBlast, set the `LLAMA_CLBLAST=1` environment variable before installing:
```bash ```bash
LLAMA_CLBLAST=1 FORCE_CMAKE=1 pip install llama-cpp-python CMAKE_ARGS="-DLLAMA_CLBLAST=on" FORCE_CMAKE=1 pip install llama-cpp-python
``` ```
@ -102,7 +116,7 @@ Navigate to [http://localhost:8000/docs](http://localhost:8000/docs) to see the
A Docker image is available on [GHCR](https://ghcr.io/abetlen/llama-cpp-python). To run the server: A Docker image is available on [GHCR](https://ghcr.io/abetlen/llama-cpp-python). To run the server:
```bash ```bash
docker run --rm -it -p8000:8000 -v /path/to/models:/models -eMODEL=/models/ggml-model-name.bin ghcr.io/abetlen/llama-cpp-python:latest docker run --rm -it -p 8000:8000 -v /path/to/models:/models -e MODEL=/models/ggml-model-name.bin ghcr.io/abetlen/llama-cpp-python:latest
``` ```
## Low-level API ## Low-level API
@ -120,7 +134,7 @@ Below is a short example demonstrating how to use the low-level API to tokenize
>>> ctx = llama_cpp.llama_init_from_file(b"./models/7b/ggml-model.bin", params) >>> ctx = llama_cpp.llama_init_from_file(b"./models/7b/ggml-model.bin", params)
>>> max_tokens = params.n_ctx >>> max_tokens = params.n_ctx
# use ctypes arrays for array params # use ctypes arrays for array params
>>> tokens = (llama_cppp.llama_token * int(max_tokens))() >>> tokens = (llama_cpp.llama_token * int(max_tokens))()
>>> n_tokens = llama_cpp.llama_tokenize(ctx, b"Q: Name the planets in the solar system? A: ", tokens, max_tokens, add_bos=llama_cpp.c_bool(True)) >>> n_tokens = llama_cpp.llama_tokenize(ctx, b"Q: Name the planets in the solar system? A: ", tokens, max_tokens, add_bos=llama_cpp.c_bool(True))
>>> llama_cpp.llama_free(ctx) >>> llama_cpp.llama_free(ctx)
``` ```

51
docker/Dockerfile Normal file
View file

@ -0,0 +1,51 @@
# Define the image argument and provide a default value
ARG IMAGE=python:3-slim-bullseye
# Use the image as specified
FROM ${IMAGE}
# Re-declare the ARG after FROM
ARG IMAGE
# Update and upgrade the existing packages
RUN apt-get update && apt-get upgrade -y && apt-get install -y --no-install-recommends \
python3 \
python3-pip \
ninja-build \
build-essential
RUN python3 -m pip install --upgrade pip pytest cmake scikit-build setuptools fastapi uvicorn sse-starlette
# Perform the conditional installations based on the image
RUN echo "Image: ${IMAGE}" && \
if [ "${IMAGE}" = "python:3-slim-bullseye" ] ; then \
echo "OpenBLAS install:" && \
apt-get install -y --no-install-recommends libopenblas-dev && \
LLAMA_OPENBLAS=1 pip install llama-cpp-python --verbose; \
else \
echo "CuBLAS install:" && \
LLAMA_CUBLAS=1 pip install llama-cpp-python --verbose; \
fi
# Clean up apt cache
RUN rm -rf /var/lib/apt/lists/*
# Set a working directory for better clarity
WORKDIR /app
# Copy files to the app directory
RUN echo "Installing model...this can take some time..."
COPY ./model.bin /app/model.bin
COPY ./start_server.sh /app/start_server.sh
# Make the server start script executable
RUN chmod +x /app/start_server.sh
# Set environment variable for the host
ENV HOST=0.0.0.0
# Expose a port for the server
EXPOSE 8000
# Run the server start script
CMD ["/bin/sh", "/app/start_server.sh"]

5
Dockerfile.cuda → docker/Dockerfile.cuda_simple
View file

@ -1,4 +1,5 @@
FROM nvidia/cuda:12.1.1-devel-ubuntu20.04 ARG CUDA_IMAGE="12.1.1-devel-ubuntu22.04"
FROM ${CUDA_IMAGE}
# We need to set the host to 0.0.0.0 to allow outside access # We need to set the host to 0.0.0.0 to allow outside access
ENV HOST 0.0.0.0 ENV HOST 0.0.0.0
@ -12,4 +13,4 @@ RUN python3 -m pip install --upgrade pip pytest cmake scikit-build setuptools fa
RUN LLAMA_CUBLAS=1 python3 setup.py develop RUN LLAMA_CUBLAS=1 python3 setup.py develop
# Run the server # Run the server
CMD python3 -m llama_cpp.server CMD python3 -m llama_cpp.server

0
Dockerfile → docker/Dockerfile.openblas_simple
View file

46
docker/README.md Normal file
View file

@ -0,0 +1,46 @@
# Dockerfiles for building the llama-cpp-python server
- `Dockerfile.openblas_simple` - a simple Dockerfile for non-GPU OpenBLAS
- `Dockerfile.cuda_simple` - a simple Dockerfile for CUDA accelerated CuBLAS
- `hug_model.py` - a Python utility for interactively choosing and downloading the latest `5_1` quantized models from [huggingface.co/TheBloke]( https://huggingface.co/TheBloke)
- `Dockerfile` - a single OpenBLAS and CuBLAS combined Dockerfile that automatically installs a previously downloaded model `model.bin`
# Get model from Hugging Face
`python3 ./hug_model.py`
You should now have a model in the current directory and `model.bin` symlinked to it for the subsequent Docker build and copy step. e.g.
```
docker $ ls -lh *.bin
-rw-rw-r-- 1 user user 4.8G May 23 18:30 <downloaded-model-file>.q5_1.bin
lrwxrwxrwx 1 user user 24 May 23 18:30 model.bin -> <downloaded-model-file>.q5_1.bin
```
**Note #1:** Make sure you have enough disk space to download the model. As the model is then copied into the image you will need at least
**TWICE** as much disk space as the size of the model:
| Model | Quantized size |
|------:|----------------:|
| 7B | 5 GB |
| 13B | 10 GB |
| 30B | 25 GB |
| 65B | 50 GB |
**Note #2:** If you want to pass or tune additional parameters, customise `./start_server.sh` before running `docker build ...`
# Install Docker Server
**Note #3:** This was tested with Docker running on Linux. If you can get it working on Windows or MacOS, please update this `README.md` with a PR!
[Install Docker Engine](https://docs.docker.com/engine/install)
# Use OpenBLAS
Use if you don't have a NVidia GPU. Defaults to `python:3-slim-bullseye` Docker base image and OpenBLAS:
## Build:
`docker build --build-arg -t openblas .`
## Run:
`docker run --cap-add SYS_RESOURCE -t openblas`
# Use CuBLAS
Requires a NVidia GPU with sufficient VRAM (approximately as much as the size above) and Docker NVidia support (see [container-toolkit/install-guide](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html))
## Build:
`docker build --build-arg IMAGE=nvidia/cuda:12.1.1-devel-ubuntu22.04 -t cublas .`
## Run:
`docker run --cap-add SYS_RESOURCE -t cublas`

116
docker/hug_model.py Normal file
View file

@ -0,0 +1,116 @@
import requests
import json
import os
import struct
def make_request(url, params=None):
print(f"Making request to {url}...")
response = requests.get(url, params=params)
if response.status_code == 200:
return json.loads(response.text)
else:
print(f"Request failed with status code {response.status_code}")
return None
def check_magic_and_version(filename):
with open(filename, 'rb') as f:
# Read the first 6 bytes from the file
data = f.read(6)
# Unpack the binary data, interpreting the first 4 bytes as a little-endian unsigned int
# and the next 2 bytes as a little-endian unsigned short
magic, version = struct.unpack('<I H', data)
print(f"magic: 0x{magic:08x}, version: 0x{version:04x}, file: {filename}")
return magic, version
def download_file(url, destination):
print(f"Downloading {url} to {destination}...")
response = requests.get(url, stream=True)
if response.status_code == 200:
with open(destination, 'wb') as f:
total_downloaded = 0
for chunk in response.iter_content(chunk_size=1024):
if chunk: # filter out keep-alive new chunks
f.write(chunk)
total_downloaded += len(chunk)
if total_downloaded >= 10485760: # 10 MB
print('.', end='', flush=True)
total_downloaded = 0
print("\nDownload complete.")
# Creating a symbolic link from destination to "model.bin"
if os.path.isfile("model.bin"):
os.remove("model.bin") # remove the existing link if any
os.symlink(destination, "model.bin")
else:
print(f"Download failed with status code {response.status_code}")
def get_user_choice(model_list):
# Print the enumerated list
print("\n")
for i, (model_id, rfilename) in enumerate(model_list):
print(f"{i+1}: Model ID: {model_id}, RFilename: {rfilename}")
# Get user's choice
choice = input("Choose a model to download by entering the corresponding number: ")
try:
index = int(choice) - 1
if 0 <= index < len(model_list):
# Return the chosen model
return model_list[index]
else:
print("Invalid choice.")
except ValueError:
print("Invalid input. Please enter a number corresponding to a model.")
except IndexError:
print("Invalid choice. Index out of range.")
return None
import argparse
def main():
# Create an argument parser
parser = argparse.ArgumentParser(description='Process the model version.')
parser.add_argument('-v', '--version', type=int, default=0x0003,
help='an integer for the version to be used')
# Parse the arguments
args = parser.parse_args()
# Define the parameters
params = {
"author": "TheBloke", # Filter by author
"tags": "llama"
}
models = make_request('https://huggingface.co/api/models', params=params)
if models is None:
return
model_list = []
# Iterate over the models
for model in models:
model_id = model['id']
model_info = make_request(f'https://huggingface.co/api/models/{model_id}')
if model_info is None:
continue
for sibling in model_info.get('siblings', []):
rfilename = sibling.get('rfilename')
if rfilename and 'q5_1' in rfilename:
model_list.append((model_id, rfilename))
model_choice = get_user_choice(model_list)
if model_choice is not None:
model_id, rfilename = model_choice
url = f"https://huggingface.co/{model_id}/resolve/main/{rfilename}"
download_file(url, rfilename)
_, version = check_magic_and_version(rfilename)
if version != args.version:
print(f"Warning: Expected version {args.version}, but found different version in the file.")
if __name__ == '__main__':
main()

11
docker/start_server.sh Executable file
View file

@ -0,0 +1,11 @@
#!/bin/sh
# For mmap support
ulimit -l unlimited
if [ "$IMAGE" = "python:3-slim-bullseye" ]; then
python3 -B -m llama_cpp.server --model /app/model.bin
else
# You may have to reduce --n_gpu_layers=1000 to 20 or less if you don't have enough VRAM
python3 -B -m llama_cpp.server --model /app/model.bin --n_gpu_layers=1000
fi

4
docs/index.md
View file

@ -112,8 +112,12 @@ python3 setup.py develop
show_root_heading: true show_root_heading: true
::: llama_cpp.LlamaCache ::: llama_cpp.LlamaCache
options:
show_root_heading: true
::: llama_cpp.LlamaState ::: llama_cpp.LlamaState
options:
show_root_heading: true
::: llama_cpp.llama_cpp ::: llama_cpp.llama_cpp
options: options:

19
examples/low_level_api/low_level_api_chat_cpp.py
View file

@ -368,10 +368,10 @@ n_keep = {self.params.n_keep}
id = llama_cpp.llama_sample_token_mirostat_v2(self.ctx, candidates_p, llama_cpp.c_float(self.params.mirostat_tau), llama_cpp.c_float(self.params.mirostat_eta), llama_cpp.c_float(mirostat_mu)) id = llama_cpp.llama_sample_token_mirostat_v2(self.ctx, candidates_p, llama_cpp.c_float(self.params.mirostat_tau), llama_cpp.c_float(self.params.mirostat_eta), llama_cpp.c_float(mirostat_mu))
else: else:
# Temperature sampling # Temperature sampling
llama_cpp.llama_sample_top_k(self.ctx, candidates_p, top_k) llama_cpp.llama_sample_top_k(self.ctx, candidates_p, top_k, min_keep=llama_cpp.c_size_t(1))
llama_cpp.llama_sample_tail_free(self.ctx, candidates_p, llama_cpp.c_float(self.params.tfs_z)) llama_cpp.llama_sample_tail_free(self.ctx, candidates_p, llama_cpp.c_float(self.params.tfs_z), min_keep=llama_cpp.c_size_t(1))
llama_cpp.llama_sample_typical(self.ctx, candidates_p, llama_cpp.c_float(self.params.typical_p)) llama_cpp.llama_sample_typical(self.ctx, candidates_p, llama_cpp.c_float(self.params.typical_p), min_keep=llama_cpp.c_size_t(1))
llama_cpp.llama_sample_top_p(self.ctx, candidates_p, llama_cpp.c_float(self.params.top_p)) llama_cpp.llama_sample_top_p(self.ctx, candidates_p, llama_cpp.c_float(self.params.top_p), min_keep=llama_cpp.c_size_t(1))
llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp)) llama_cpp.llama_sample_temperature(self.ctx, candidates_p, llama_cpp.c_float(self.params.temp))
id = llama_cpp.llama_sample_token(self.ctx, candidates_p) id = llama_cpp.llama_sample_token(self.ctx, candidates_p)
# print("`{}`".format(candidates_p.size)) # print("`{}`".format(candidates_p.size))
@ -382,12 +382,15 @@ n_keep = {self.params.n_keep}
# replace end of text token with newline token when in interactive mode # replace end of text token with newline token when in interactive mode
if (id == llama_cpp.llama_token_eos() and self.params.interactive and not self.params.instruct): if (id == llama_cpp.llama_token_eos() and self.params.interactive and not self.params.instruct):
id = self.llama_token_newline[0] id = self.llama_token_newline[0]
self.embd.append(id)
if (self.use_antiprompt()): if (self.use_antiprompt()):
# tokenize and inject first reverse prompt # tokenize and inject first reverse prompt
self.embd_inp += self.first_antiprompt[0] self.embd_inp += self.first_antiprompt[0]
for id in self.first_antiprompt[0]:
# add it to the context self.embd.append(id)
self.embd.append(id) else:
# add it to the context
self.embd.append(id)
# echo this to console # echo this to console
self.output_echo = True self.output_echo = True
@ -493,7 +496,7 @@ n_keep = {self.params.n_keep}
# Contains multi-byte UTF8 # Contains multi-byte UTF8
for num, pattern in [(2, 192), (3, 224), (4, 240)]: for num, pattern in [(2, 192), (3, 224), (4, 240)]:
# Bitwise AND check # Bitwise AND check
if pattern & int.from_bytes(cur_char) == pattern: if pattern & int.from_bytes(cur_char, 'little') == pattern:
self.multibyte_fix = [cur_char] + ([None] * (num-1)) self.multibyte_fix = [cur_char] + ([None] * (num-1))
# Stop incomplete bytes from passing # Stop incomplete bytes from passing

89
examples/notebooks/Guidance.ipynb Normal file
View file

File diff suppressed because one or more lines are too long

597
llama_cpp/llama.py
View file

@ -4,7 +4,17 @@ import uuid
import time import time
import math import math
import multiprocessing import multiprocessing
from typing import List, Optional, Union, Generator, Sequence, Iterator, Deque, Tuple from typing import (
List,
Optional,
Union,
Generator,
Sequence,
Iterator,
Deque,
Tuple,
Callable,
)
from collections import deque, OrderedDict from collections import deque, OrderedDict
from . import llama_cpp from . import llama_cpp
@ -15,9 +25,7 @@ class LlamaCache:
"""Cache for a llama.cpp model.""" """Cache for a llama.cpp model."""
def __init__(self, capacity_bytes: int = (2 << 30)): def __init__(self, capacity_bytes: int = (2 << 30)):
self.cache_state: OrderedDict[ self.cache_state: OrderedDict[Tuple[int, ...], "LlamaState"] = OrderedDict()
Tuple[llama_cpp.llama_token, ...], "LlamaState"
] = OrderedDict()
self.capacity_bytes = capacity_bytes self.capacity_bytes = capacity_bytes
@property @property
@ -26,8 +34,8 @@ class LlamaCache:
def _find_longest_prefix_key( def _find_longest_prefix_key(
self, self,
key: Tuple[llama_cpp.llama_token, ...], key: Tuple[int, ...],
) -> Optional[Tuple[llama_cpp.llama_token, ...]]: ) -> Optional[Tuple[int, ...]]:
min_len = 0 min_len = 0
min_key = None min_key = None
keys = ( keys = (
@ -39,7 +47,7 @@ class LlamaCache:
min_key = k min_key = k
return min_key return min_key
def __getitem__(self, key: Sequence[llama_cpp.llama_token]) -> "LlamaState": def __getitem__(self, key: Sequence[int]) -> "LlamaState":
key = tuple(key) key = tuple(key)
_key = self._find_longest_prefix_key(key) _key = self._find_longest_prefix_key(key)
if _key is None: if _key is None:
@ -48,10 +56,10 @@ class LlamaCache:
self.cache_state.move_to_end(_key) self.cache_state.move_to_end(_key)
return value return value
def __contains__(self, key: Sequence[llama_cpp.llama_token]) -> bool: def __contains__(self, key: Sequence[int]) -> bool:
return self._find_longest_prefix_key(tuple(key)) is not None return self._find_longest_prefix_key(tuple(key)) is not None
def __setitem__(self, key: Sequence[llama_cpp.llama_token], value: "LlamaState"): def __setitem__(self, key: Sequence[int], value: "LlamaState"):
key = tuple(key) key = tuple(key)
if key in self.cache_state: if key in self.cache_state:
del self.cache_state[key] del self.cache_state[key]
@ -63,7 +71,7 @@ class LlamaCache:
class LlamaState: class LlamaState:
def __init__( def __init__(
self, self,
eval_tokens: Deque[llama_cpp.llama_token], eval_tokens: Deque[int],
eval_logits: Deque[List[float]], eval_logits: Deque[List[float]],
llama_state, # type: llama_cpp.Array[llama_cpp.c_uint8] llama_state, # type: llama_cpp.Array[llama_cpp.c_uint8]
llama_state_size: int, llama_state_size: int,
@ -74,6 +82,24 @@ class LlamaState:
self.llama_state_size = llama_state_size self.llama_state_size = llama_state_size
LogitsProcessor = Callable[[List[int], List[float]], List[float]]
class LogitsProcessorList(List[LogitsProcessor]):
def __call__(self, input_ids: List[int], scores: List[float]) -> List[float]:
for processor in self:
scores = processor(input_ids, scores)
return scores
StoppingCriteria = Callable[[List[int], List[float]], bool]
class StoppingCriteriaList(List[StoppingCriteria]):
def __call__(self, input_ids: List[int], logits: List[float]) -> bool:
return any([stopping_criteria(input_ids, logits) for stopping_criteria in self])
class Llama: class Llama:
"""High-level Python wrapper for a llama.cpp model.""" """High-level Python wrapper for a llama.cpp model."""
@ -83,6 +109,7 @@ class Llama:
# NOTE: These parameters are likely to change in the future. # NOTE: These parameters are likely to change in the future.
n_ctx: int = 512, n_ctx: int = 512,
n_parts: int = -1, n_parts: int = -1,
n_gpu_layers: int = 0,
seed: int = 1337, seed: int = 1337,
f16_kv: bool = True, f16_kv: bool = True,
logits_all: bool = False, logits_all: bool = False,
@ -128,7 +155,7 @@ class Llama:
self.params = llama_cpp.llama_context_default_params() self.params = llama_cpp.llama_context_default_params()
self.params.n_ctx = n_ctx self.params.n_ctx = n_ctx
self.params.n_parts = n_parts self.params.n_gpu_layers = n_gpu_layers
self.params.seed = seed self.params.seed = seed
self.params.f16_kv = f16_kv self.params.f16_kv = f16_kv
self.params.logits_all = logits_all self.params.logits_all = logits_all
@ -139,7 +166,7 @@ class Llama:
self.last_n_tokens_size = last_n_tokens_size self.last_n_tokens_size = last_n_tokens_size
self.n_batch = min(n_ctx, n_batch) self.n_batch = min(n_ctx, n_batch)
self.eval_tokens: Deque[llama_cpp.llama_token] = deque(maxlen=n_ctx) self.eval_tokens: Deque[int] = deque(maxlen=n_ctx)
self.eval_logits: Deque[List[float]] = deque(maxlen=n_ctx if logits_all else 1) self.eval_logits: Deque[List[float]] = deque(maxlen=n_ctx if logits_all else 1)
self.cache: Optional[LlamaCache] = None self.cache: Optional[LlamaCache] = None
@ -149,6 +176,10 @@ class Llama:
self.lora_base = lora_base self.lora_base = lora_base
self.lora_path = lora_path self.lora_path = lora_path
### DEPRECATED ###
self.n_parts = n_parts
### DEPRECATED ###
if not os.path.exists(model_path): if not os.path.exists(model_path):
raise ValueError(f"Model path does not exist: {model_path}") raise ValueError(f"Model path does not exist: {model_path}")
@ -174,7 +205,30 @@ class Llama:
if self.verbose: if self.verbose:
print(llama_cpp.llama_print_system_info().decode("utf-8"), file=sys.stderr) print(llama_cpp.llama_print_system_info().decode("utf-8"), file=sys.stderr)
def tokenize(self, text: bytes) -> List[llama_cpp.llama_token]: self._n_vocab = self.n_vocab()
self._n_ctx = self.n_ctx()
data = (llama_cpp.llama_token_data * self._n_vocab)(
*[
llama_cpp.llama_token_data(
id=llama_cpp.llama_token(i),
logit=llama_cpp.c_float(0.0),
p=llama_cpp.c_float(0.0),
)
for i in range(self._n_vocab)
]
)
size = llama_cpp.c_size_t(self._n_vocab)
sorted = False
candidates = llama_cpp.llama_token_data_array(
data=data,
size=size,
sorted=sorted,
)
self._candidates = candidates
self._token_nl = Llama.token_nl()
self._token_eos = Llama.token_eos()
def tokenize(self, text: bytes, add_bos: bool = True) -> List[int]:
"""Tokenize a string. """Tokenize a string.
Args: Args:
@ -187,20 +241,32 @@ class Llama:
A list of tokens. A list of tokens.
""" """
assert self.ctx is not None assert self.ctx is not None
n_ctx = llama_cpp.llama_n_ctx(self.ctx) n_ctx = self._n_ctx
tokens = (llama_cpp.llama_token * int(n_ctx))() tokens = (llama_cpp.llama_token * n_ctx)()
n_tokens = llama_cpp.llama_tokenize( n_tokens = llama_cpp.llama_tokenize(
self.ctx, self.ctx,
text, text,
tokens, tokens,
n_ctx, llama_cpp.c_int(n_ctx),
llama_cpp.c_bool(True), llama_cpp.c_bool(add_bos),
) )
if int(n_tokens) < 0: if n_tokens < 0:
raise RuntimeError(f'Failed to tokenize: text="{text}" n_tokens={n_tokens}') n_tokens = abs(n_tokens)
tokens = (llama_cpp.llama_token * n_tokens)()
n_tokens = llama_cpp.llama_tokenize(
self.ctx,
text,
tokens,
llama_cpp.c_int(n_tokens),
llama_cpp.c_bool(add_bos),
)
if n_tokens < 0:
raise RuntimeError(
f'Failed to tokenize: text="{text}" n_tokens={n_tokens}'
)
return list(tokens[:n_tokens]) return list(tokens[:n_tokens])
def detokenize(self, tokens: List[llama_cpp.llama_token]) -> bytes: def detokenize(self, tokens: List[int]) -> bytes:
"""Detokenize a list of tokens. """Detokenize a list of tokens.
Args: Args:
@ -212,7 +278,9 @@ class Llama:
assert self.ctx is not None assert self.ctx is not None
output = b"" output = b""
for token in tokens: for token in tokens:
output += llama_cpp.llama_token_to_str(self.ctx, token) output += llama_cpp.llama_token_to_str(
self.ctx, llama_cpp.llama_token(token)
)
return output return output
def set_cache(self, cache: Optional[LlamaCache]): def set_cache(self, cache: Optional[LlamaCache]):
@ -228,14 +296,14 @@ class Llama:
self.eval_tokens.clear() self.eval_tokens.clear()
self.eval_logits.clear() self.eval_logits.clear()
def eval(self, tokens: Sequence[llama_cpp.llama_token]): def eval(self, tokens: Sequence[int]):
"""Evaluate a list of tokens. """Evaluate a list of tokens.
Args: Args:
tokens: The list of tokens to evaluate. tokens: The list of tokens to evaluate.
""" """
assert self.ctx is not None assert self.ctx is not None
n_ctx = int(llama_cpp.llama_n_ctx(self.ctx)) n_ctx = self._n_ctx
for i in range(0, len(tokens), self.n_batch): for i in range(0, len(tokens), self.n_batch):
batch = tokens[i : min(len(tokens), i + self.n_batch)] batch = tokens[i : min(len(tokens), i + self.n_batch)]
n_past = min(n_ctx - len(batch), len(self.eval_tokens)) n_past = min(n_ctx - len(batch), len(self.eval_tokens))
@ -247,18 +315,16 @@ class Llama:
n_past=llama_cpp.c_int(n_past), n_past=llama_cpp.c_int(n_past),
n_threads=llama_cpp.c_int(self.n_threads), n_threads=llama_cpp.c_int(self.n_threads),
) )
if int(return_code) != 0: if return_code != 0:
raise RuntimeError(f"llama_eval returned {return_code}") raise RuntimeError(f"llama_eval returned {return_code}")
# Save tokens # Save tokens
self.eval_tokens.extend(batch) self.eval_tokens.extend(batch)
# Save logits # Save logits
rows = n_tokens if self.params.logits_all else 1 rows = n_tokens if self.params.logits_all else 1
n_vocab = llama_cpp.llama_n_vocab(self.ctx) n_vocab = self._n_vocab
cols = int(n_vocab) cols = n_vocab
logits_view = llama_cpp.llama_get_logits(self.ctx) logits_view = llama_cpp.llama_get_logits(self.ctx)
logits: List[List[float]] = [ logits = [logits_view[i * cols : (i + 1) * cols] for i in range(rows)]
[logits_view[i * cols + j] for j in range(cols)] for i in range(rows)
]
self.eval_logits.extend(logits) self.eval_logits.extend(logits)
def _sample( def _sample(
@ -275,28 +341,33 @@ class Llama:
mirostat_mode: llama_cpp.c_int, mirostat_mode: llama_cpp.c_int,
mirostat_tau: llama_cpp.c_float, mirostat_tau: llama_cpp.c_float,
mirostat_eta: llama_cpp.c_float, mirostat_eta: llama_cpp.c_float,
penalize_nl: bool = True,
logits_processor: Optional[LogitsProcessorList] = None,
): ):
assert self.ctx is not None assert self.ctx is not None
assert len(self.eval_logits) > 0 assert len(self.eval_logits) > 0
n_vocab = int(llama_cpp.llama_n_vocab(self.ctx)) n_vocab = self._n_vocab
n_ctx = self._n_ctx
top_k = llama_cpp.c_int(n_vocab) if top_k.value <= 0 else top_k
last_n_tokens_size = (
llama_cpp.c_int(n_ctx)
if last_n_tokens_size.value < 0
else last_n_tokens_size
)
logits = self.eval_logits[-1] logits = self.eval_logits[-1]
data = (llama_cpp.llama_token_data * n_vocab)(
*[ if logits_processor is not None:
llama_cpp.llama_token_data( logits = logits_processor(list(self.eval_tokens), logits)
id=llama_cpp.llama_token(i), self.eval_logits[-1] = logits
logit=logits[i],
p=llama_cpp.c_float(0.0), nl_logit = logits[self._token_nl]
) candidates = self._candidates
for i in range(n_vocab) for i, logit in enumerate(logits):
] candidates.data[i].id = llama_cpp.llama_token(i)
) candidates.data[i].logit = llama_cpp.c_float(logit)
size = llama_cpp.c_size_t(n_vocab) candidates.data[i].p = llama_cpp.c_float(0.0)
sorted = False candidates.sorted = llama_cpp.c_bool(False)
candidates = llama_cpp.llama_token_data_array( candidates.size = llama_cpp.c_size_t(n_vocab)
data=data,
size=size,
sorted=sorted,
)
llama_cpp.llama_sample_repetition_penalty( llama_cpp.llama_sample_repetition_penalty(
ctx=self.ctx, ctx=self.ctx,
last_tokens_data=last_n_tokens_data, last_tokens_data=last_n_tokens_data,
@ -312,6 +383,8 @@ class Llama:
alpha_frequency=frequency_penalty, alpha_frequency=frequency_penalty,
alpha_presence=presence_penalty, alpha_presence=presence_penalty,
) )
if not penalize_nl:
candidates.data[self._token_nl].logit = llama_cpp.c_float(nl_logit)
if temp.value == 0.0: if temp.value == 0.0:
return llama_cpp.llama_sample_token_greedy( return llama_cpp.llama_sample_token_greedy(
ctx=self.ctx, ctx=self.ctx,
@ -394,6 +467,8 @@ class Llama:
mirostat_mode: int = 0, mirostat_mode: int = 0,
mirostat_eta: float = 0.1, mirostat_eta: float = 0.1,
mirostat_tau: float = 5.0, mirostat_tau: float = 5.0,
penalize_nl: bool = True,
logits_processor: Optional[LogitsProcessorList] = None,
): ):
"""Sample a token from the model. """Sample a token from the model.
@ -425,24 +500,27 @@ class Llama:
mirostat_mode=llama_cpp.c_int(mirostat_mode), mirostat_mode=llama_cpp.c_int(mirostat_mode),
mirostat_tau=llama_cpp.c_float(mirostat_tau), mirostat_tau=llama_cpp.c_float(mirostat_tau),
mirostat_eta=llama_cpp.c_float(mirostat_eta), mirostat_eta=llama_cpp.c_float(mirostat_eta),
penalize_nl=penalize_nl,
logits_processor=logits_processor,
) )
def generate( def generate(
self, self,
tokens: Sequence[llama_cpp.llama_token], tokens: Sequence[int],
top_k: int, top_k: int = 40,
top_p: float, top_p: float = 0.95,
temp: float, temp: float = 0.80,
repeat_penalty: float, repeat_penalty: float = 1.1,
reset: bool = True, reset: bool = True,
frequency_penalty: float = 0.0, frequency_penalty: float = 0.0,
presence_penalty: float = 0.0, presence_penalty: float = 0.0,
tfs_z: float = 1.0,
mirostat_mode: int = 0, mirostat_mode: int = 0,
mirostat_tau: float = 5.0, mirostat_tau: float = 5.0,
mirostat_eta: float = 0.1, mirostat_eta: float = 0.1,
) -> Generator[ logits_processor: Optional[LogitsProcessorList] = None,
llama_cpp.llama_token, Optional[Sequence[llama_cpp.llama_token]], None stopping_criteria: Optional[StoppingCriteriaList] = None,
]: ) -> Generator[int, Optional[Sequence[int]], None]:
"""Create a generator of tokens from a prompt. """Create a generator of tokens from a prompt.
Examples: Examples:
@ -495,16 +573,24 @@ class Llama:
repeat_penalty=repeat_penalty, repeat_penalty=repeat_penalty,
frequency_penalty=frequency_penalty, frequency_penalty=frequency_penalty,
presence_penalty=presence_penalty, presence_penalty=presence_penalty,
tfs_z=tfs_z,
mirostat_mode=mirostat_mode, mirostat_mode=mirostat_mode,
mirostat_tau=mirostat_tau, mirostat_tau=mirostat_tau,
mirostat_eta=mirostat_eta, mirostat_eta=mirostat_eta,
logits_processor=logits_processor,
) )
if stopping_criteria is not None and stopping_criteria(
list(self.eval_tokens), self.eval_logits[-1]
):
return
tokens_or_none = yield token tokens_or_none = yield token
tokens = [token] tokens = [token]
if tokens_or_none is not None: if tokens_or_none is not None:
tokens.extend(tokens_or_none) tokens.extend(tokens_or_none)
def create_embedding(self, input: str) -> Embedding: def create_embedding(
self, input: Union[str, List[str]], model: Optional[str] = None
) -> Embedding:
"""Embed a string. """Embed a string.
Args: Args:
@ -514,6 +600,7 @@ class Llama:
An embedding object. An embedding object.
""" """
assert self.ctx is not None assert self.ctx is not None
model_name: str = model if model is not None else self.model_path
if self.params.embedding == False: if self.params.embedding == False:
raise RuntimeError( raise RuntimeError(
@ -523,30 +610,40 @@ class Llama:
if self.verbose: if self.verbose:
llama_cpp.llama_reset_timings(self.ctx) llama_cpp.llama_reset_timings(self.ctx)
tokens = self.tokenize(input.encode("utf-8")) if isinstance(input, str):
self.reset() inputs = [input]
self.eval(tokens) else:
n_tokens = len(tokens) inputs = input
embedding = llama_cpp.llama_get_embeddings(self.ctx)[
: llama_cpp.llama_n_embd(self.ctx)
]
data: List[EmbeddingData] = []
total_tokens = 0
for index, input in enumerate(inputs):
tokens = self.tokenize(input.encode("utf-8"))
self.reset()
self.eval(tokens)
n_tokens = len(tokens)
total_tokens += n_tokens
embedding = llama_cpp.llama_get_embeddings(self.ctx)[
: llama_cpp.llama_n_embd(self.ctx)
]
data.append(
{
"object": "embedding",
"embedding": embedding,
"index": index,
}
)
if self.verbose: if self.verbose:
llama_cpp.llama_print_timings(self.ctx) llama_cpp.llama_print_timings(self.ctx)
return { return {
"object": "list", "object": "list",
"data": [ "data": data,
{ "model": model_name,
"object": "embedding",
"embedding": embedding,
"index": 0,
}
],
"model": self.model_path,
"usage": { "usage": {
"prompt_tokens": n_tokens, "prompt_tokens": total_tokens,
"total_tokens": n_tokens, "total_tokens": total_tokens,
}, },
} }
@ -570,35 +667,39 @@ class Llama:
top_p: float = 0.95, top_p: float = 0.95,
logprobs: Optional[int] = None, logprobs: Optional[int] = None,
echo: bool = False, echo: bool = False,
stop: Optional[List[str]] = [], stop: Optional[Union[str, List[str]]] = [],
frequency_penalty: float = 0.0, frequency_penalty: float = 0.0,
presence_penalty: float = 0.0, presence_penalty: float = 0.0,
repeat_penalty: float = 1.1, repeat_penalty: float = 1.1,
top_k: int = 40, top_k: int = 40,
stream: bool = False, stream: bool = False,
tfs_z: float = 1.0,
mirostat_mode: int = 0, mirostat_mode: int = 0,
mirostat_tau: float = 5.0, mirostat_tau: float = 5.0,
mirostat_eta: float = 0.1, mirostat_eta: float = 0.1,
model: Optional[str] = None,
stopping_criteria: Optional[StoppingCriteriaList] = None,
logits_processor: Optional[LogitsProcessorList] = None,
) -> Union[Iterator[Completion], Iterator[CompletionChunk]]: ) -> Union[Iterator[Completion], Iterator[CompletionChunk]]:
assert self.ctx is not None assert self.ctx is not None
completion_id: str = f"cmpl-{str(uuid.uuid4())}" completion_id: str = f"cmpl-{str(uuid.uuid4())}"
created: int = int(time.time()) created: int = int(time.time())
completion_tokens: List[llama_cpp.llama_token] = [] completion_tokens: List[int] = []
# Add blank space to start of prompt to match OG llama tokenizer # Add blank space to start of prompt to match OG llama tokenizer
prompt_tokens: List[llama_cpp.llama_token] = self.tokenize( prompt_tokens: List[int] = self.tokenize(b" " + prompt.encode("utf-8"))
b" " + prompt.encode("utf-8")
)
text: bytes = b"" text: bytes = b""
returned_characters: int = 0 returned_tokens: int = 0
stop = stop if stop is not None else [] stop = (
stop if isinstance(stop, list) else [stop] if isinstance(stop, str) else []
)
model_name: str = model if model is not None else self.model_path
if self.verbose: if self.verbose:
llama_cpp.llama_reset_timings(self.ctx) llama_cpp.llama_reset_timings(self.ctx)
if len(prompt_tokens) + max_tokens > int(llama_cpp.llama_n_ctx(self.ctx)): if len(prompt_tokens) + max_tokens > self._n_ctx:
raise ValueError( raise ValueError(f"Requested tokens exceed context window of {self._n_ctx}")
f"Requested tokens exceed context window of {llama_cpp.llama_n_ctx(self.ctx)}"
)
if stop != []: if stop != []:
stop_sequences = [s.encode("utf-8") for s in stop] stop_sequences = [s.encode("utf-8") for s in stop]
@ -634,14 +735,17 @@ class Llama:
top_k=top_k, top_k=top_k,
top_p=top_p, top_p=top_p,
temp=temperature, temp=temperature,
tfs_z=tfs_z,
mirostat_mode=mirostat_mode, mirostat_mode=mirostat_mode,
mirostat_tau=mirostat_tau, mirostat_tau=mirostat_tau,
mirostat_eta=mirostat_eta, mirostat_eta=mirostat_eta,
frequency_penalty=frequency_penalty, frequency_penalty=frequency_penalty,
presence_penalty=presence_penalty, presence_penalty=presence_penalty,
repeat_penalty=repeat_penalty, repeat_penalty=repeat_penalty,
stopping_criteria=stopping_criteria,
logits_processor=logits_processor,
): ):
if token == llama_cpp.llama_token_eos(): if token == self._token_eos:
text = self.detokenize(completion_tokens) text = self.detokenize(completion_tokens)
finish_reason = "stop" finish_reason = "stop"
break break
@ -671,63 +775,189 @@ class Llama:
break break
if stream: if stream:
start = returned_characters
longest = 0
# We want to avoid yielding any characters from # We want to avoid yielding any characters from
# the generated text if they are part of a stop # the generated text if they are part of a stop
# sequence. # sequence.
first_stop_position = 0
for s in stop_sequences: for s in stop_sequences:
for i in range(len(s), 0, -1): for i in range(len(s), 0, -1):
if all_text.endswith(s[:i]): if all_text.endswith(s[:i]):
if i > longest: if i > first_stop_position:
longest = i first_stop_position = i
break break
text = all_text[: len(all_text) - longest]
returned_characters += len(text[start:]) token_end_position = 0
yield { remaining_tokens = completion_tokens[returned_tokens:]
"id": completion_id, remaining_length = len(self.detokenize(remaining_tokens))
"object": "text_completion", for token in remaining_tokens:
"created": created, token_end_position += len(self.detokenize([token]))
"model": self.model_path, # Check if stop sequence is in the token
"choices": [ if token_end_position >= (
{ remaining_length - first_stop_position - 1
"text": text[start:].decode("utf-8", errors="ignore"), ):
"index": 0, break
"logprobs": None, logprobs_or_none: Optional[CompletionLogprobs] = None
"finish_reason": None, if logprobs is not None:
token_str = self.detokenize([token]).decode(
"utf-8", errors="ignore"
)
text_offset = len(prompt) + len(
self.detokenize(completion_tokens[:returned_tokens])
)
token_offset = len(prompt_tokens) + returned_tokens
logits = self.eval_logits[token_offset - 1]
current_logprobs = Llama.logits_to_logprobs(logits)
sorted_logprobs = list(
sorted(
zip(current_logprobs, range(len(current_logprobs))),
reverse=True,
)
)
top_logprob = {
self.detokenize([i]).decode(
"utf-8", errors="ignore"
): logprob
for logprob, i in sorted_logprobs[:logprobs]
} }
], top_logprob.update({token_str: current_logprobs[int(token)]})
} logprobs_or_none = {
"tokens": [
self.detokenize([token]).decode(
"utf-8", errors="ignore"
)
],
"text_offset": [text_offset],
"token_logprobs": [sorted_logprobs[int(token)][0]],
"top_logprobs": [top_logprob],
}
returned_tokens += 1
yield {
"id": completion_id,
"object": "text_completion",
"created": created,
"model": model_name,
"choices": [
{
"text": self.detokenize([token]).decode(
"utf-8", errors="ignore"
),
"index": 0,
"logprobs": logprobs_or_none,
"finish_reason": None,
}
],
}
if len(completion_tokens) >= max_tokens: if len(completion_tokens) >= max_tokens:
text = self.detokenize(completion_tokens) text = self.detokenize(completion_tokens)
finish_reason = "length" finish_reason = "length"
break break
if stopping_criteria is not None and stopping_criteria(
list(self.eval_tokens), self.eval_logits[-1]
):
text = self.detokenize(completion_tokens)
finish_reason = "stop"
if self.verbose:
llama_cpp.llama_print_timings(self.ctx)
if stream:
remaining_tokens = completion_tokens[returned_tokens:]
all_text = self.detokenize(remaining_tokens)
any_stop = [s for s in stop_sequences if s in all_text]
if len(any_stop) > 0:
end = min(all_text.index(stop) for stop in any_stop)
else:
end = len(all_text)
token_end_position = 0
for token in remaining_tokens:
token_end_position += len(self.detokenize([token]))
logprobs_or_none: Optional[CompletionLogprobs] = None
if logprobs is not None:
token_str = self.detokenize([token]).decode(
"utf-8", errors="ignore"
)
text_offset = len(prompt) + len(
self.detokenize(completion_tokens[:returned_tokens])
)
token_offset = len(prompt_tokens) + returned_tokens - 1
logits = self.eval_logits[token_offset]
current_logprobs = Llama.logits_to_logprobs(logits)
sorted_logprobs = list(
sorted(
zip(current_logprobs, range(len(current_logprobs))),
reverse=True,
)
)
top_logprob = {
self.detokenize([i]).decode("utf-8", errors="ignore"): logprob
for logprob, i in sorted_logprobs[:logprobs]
}
top_logprob.update({token_str: current_logprobs[int(token)]})
logprobs_or_none = {
"tokens": [
self.detokenize([token]).decode("utf-8", errors="ignore")
],
"text_offset": [text_offset],
"token_logprobs": [sorted_logprobs[int(token)][0]],
"top_logprobs": [top_logprob],
}
if token_end_position >= end:
last_text = self.detokenize([token])
if token_end_position == end - 1:
break
returned_tokens += 1
yield {
"id": completion_id,
"object": "text_completion",
"created": created,
"model": model_name,
"choices": [
{
"text": last_text[
: len(last_text) - (token_end_position - end)
].decode("utf-8", errors="ignore"),
"index": 0,
"logprobs": logprobs_or_none,
"finish_reason": finish_reason,
}
],
}
break
returned_tokens += 1
yield {
"id": completion_id,
"object": "text_completion",
"created": created,
"model": model_name,
"choices": [
{
"text": self.detokenize([token]).decode(
"utf-8", errors="ignore"
),
"index": 0,
"logprobs": logprobs_or_none,
"finish_reason": finish_reason
if returned_tokens == len(completion_tokens)
else None,
}
],
}
if self.cache:
if self.verbose:
print("Llama._create_completion: cache save", file=sys.stderr)
self.cache[prompt_tokens + completion_tokens] = self.save_state()
return
if self.cache: if self.cache:
if self.verbose: if self.verbose:
print("Llama._create_completion: cache save", file=sys.stderr) print("Llama._create_completion: cache save", file=sys.stderr)
self.cache[prompt_tokens + completion_tokens] = self.save_state() self.cache[prompt_tokens + completion_tokens] = self.save_state()
if stream:
yield {
"id": completion_id,
"object": "text_completion",
"created": created,
"model": self.model_path,
"choices": [
{
"text": text[returned_characters:].decode(
"utf-8", errors="ignore"
),
"index": 0,
"logprobs": None,
"finish_reason": finish_reason,
}
],
}
return
text_str = text.decode("utf-8", errors="ignore") text_str = text.decode("utf-8", errors="ignore")
if echo: if echo:
@ -738,13 +968,19 @@ class Llama:
logprobs_or_none: Optional[CompletionLogprobs] = None logprobs_or_none: Optional[CompletionLogprobs] = None
if logprobs is not None: if logprobs is not None:
text_offset = 0 text_offset = 0 if echo else len(prompt)
token_offset = 0 if echo else len(prompt_tokens[1:])
text_offsets: List[int] = [] text_offsets: List[int] = []
token_logprobs: List[float] = [] token_logprobs: List[Optional[float]] = []
tokens: List[str] = [] tokens: List[str] = []
top_logprobs: List[Dict[str, float]] = [] top_logprobs: List[Optional[Dict[str, float]]] = []
if echo:
# Remove leading BOS token
all_tokens = prompt_tokens[1:] + completion_tokens
else:
all_tokens = completion_tokens
all_tokens = prompt_tokens + completion_tokens
all_token_strs = [ all_token_strs = [
self.detokenize([token]).decode("utf-8", errors="ignore") self.detokenize([token]).decode("utf-8", errors="ignore")
for token in all_tokens for token in all_tokens
@ -752,7 +988,7 @@ class Llama:
all_logprobs = [ all_logprobs = [
Llama.logits_to_logprobs(list(map(float, row))) Llama.logits_to_logprobs(list(map(float, row)))
for row in self.eval_logits for row in self.eval_logits
] ][token_offset:]
for token, token_str, logprobs_token in zip( for token, token_str, logprobs_token in zip(
all_tokens, all_token_strs, all_logprobs all_tokens, all_token_strs, all_logprobs
): ):
@ -765,14 +1001,18 @@ class Llama:
) )
) )
token_logprobs.append(sorted_logprobs[int(token)][0]) token_logprobs.append(sorted_logprobs[int(token)][0])
top_logprob = { top_logprob: Optional[Dict[str, float]] = {
self.detokenize([llama_cpp.llama_token(i)]).decode( self.detokenize([i]).decode("utf-8", errors="ignore"): logprob
"utf-8", errors="ignore"
): logprob
for logprob, i in sorted_logprobs[:logprobs] for logprob, i in sorted_logprobs[:logprobs]
} }
top_logprob.update({token_str: sorted_logprobs[int(token)][0]}) top_logprob.update({token_str: logprobs_token[int(token)]})
top_logprobs.append(top_logprob) top_logprobs.append(top_logprob)
# Weird idosincracy of the OpenAI API where
# token_logprobs and top_logprobs are null for
# the first token.
if echo and len(all_tokens) > 0:
token_logprobs[0] = None
top_logprobs[0] = None
logprobs_or_none = { logprobs_or_none = {
"tokens": tokens, "tokens": tokens,
"text_offset": text_offsets, "text_offset": text_offsets,
@ -780,14 +1020,11 @@ class Llama:
"top_logprobs": top_logprobs, "top_logprobs": top_logprobs,
} }
if self.verbose:
llama_cpp.llama_print_timings(self.ctx)
yield { yield {
"id": completion_id, "id": completion_id,
"object": "text_completion", "object": "text_completion",
"created": created, "created": created,
"model": self.model_path, "model": model_name,
"choices": [ "choices": [
{ {
"text": text_str, "text": text_str,
@ -812,15 +1049,19 @@ class Llama:
top_p: float = 0.95, top_p: float = 0.95,
logprobs: Optional[int] = None, logprobs: Optional[int] = None,
echo: bool = False, echo: bool = False,
stop: Optional[List[str]] = [], stop: Optional[Union[str, List[str]]] = [],
frequency_penalty: float = 0.0, frequency_penalty: float = 0.0,
presence_penalty: float = 0.0, presence_penalty: float = 0.0,
repeat_penalty: float = 1.1, repeat_penalty: float = 1.1,
top_k: int = 40, top_k: int = 40,
stream: bool = False, stream: bool = False,
tfs_z: float = 1.0,
mirostat_mode: int = 0, mirostat_mode: int = 0,
mirostat_tau: float = 5.0, mirostat_tau: float = 5.0,
mirostat_eta: float = 0.1, mirostat_eta: float = 0.1,
model: Optional[str] = None,
stopping_criteria: Optional[StoppingCriteriaList] = None,
logits_processor: Optional[LogitsProcessorList] = None,
) -> Union[Completion, Iterator[CompletionChunk]]: ) -> Union[Completion, Iterator[CompletionChunk]]:
"""Generate text from a prompt. """Generate text from a prompt.
@ -858,9 +1099,13 @@ class Llama:
repeat_penalty=repeat_penalty, repeat_penalty=repeat_penalty,
top_k=top_k, top_k=top_k,
stream=stream, stream=stream,
tfs_z=tfs_z,
mirostat_mode=mirostat_mode, mirostat_mode=mirostat_mode,
mirostat_tau=mirostat_tau, mirostat_tau=mirostat_tau,
mirostat_eta=mirostat_eta, mirostat_eta=mirostat_eta,
model=model,
stopping_criteria=stopping_criteria,
logits_processor=logits_processor,
) )
if stream: if stream:
chunks: Iterator[CompletionChunk] = completion_or_chunks chunks: Iterator[CompletionChunk] = completion_or_chunks
@ -877,15 +1122,19 @@ class Llama:
top_p: float = 0.95, top_p: float = 0.95,
logprobs: Optional[int] = None, logprobs: Optional[int] = None,
echo: bool = False, echo: bool = False,
stop: Optional[List[str]] = [], stop: Optional[Union[str, List[str]]] = [],
frequency_penalty: float = 0.0, frequency_penalty: float = 0.0,
presence_penalty: float = 0.0, presence_penalty: float = 0.0,
repeat_penalty: float = 1.1, repeat_penalty: float = 1.1,
top_k: int = 40, top_k: int = 40,
stream: bool = False, stream: bool = False,
tfs_z: float = 1.0,
mirostat_mode: int = 0, mirostat_mode: int = 0,
mirostat_tau: float = 5.0, mirostat_tau: float = 5.0,
mirostat_eta: float = 0.1, mirostat_eta: float = 0.1,
model: Optional[str] = None,
stopping_criteria: Optional[StoppingCriteriaList] = None,
logits_processor: Optional[LogitsProcessorList] = None,
) -> Union[Completion, Iterator[CompletionChunk]]: ) -> Union[Completion, Iterator[CompletionChunk]]:
"""Generate text from a prompt. """Generate text from a prompt.
@ -923,9 +1172,13 @@ class Llama:
repeat_penalty=repeat_penalty, repeat_penalty=repeat_penalty,
top_k=top_k, top_k=top_k,
stream=stream, stream=stream,
tfs_z=tfs_z,
mirostat_mode=mirostat_mode, mirostat_mode=mirostat_mode,
mirostat_tau=mirostat_tau, mirostat_tau=mirostat_tau,
mirostat_eta=mirostat_eta, mirostat_eta=mirostat_eta,
model=model,
stopping_criteria=stopping_criteria,
logits_processor=logits_processor,
) )
def _convert_text_completion_to_chat( def _convert_text_completion_to_chat(
@ -993,14 +1246,16 @@ class Llama:
top_p: float = 0.95, top_p: float = 0.95,
top_k: int = 40, top_k: int = 40,
stream: bool = False, stream: bool = False,
stop: Optional[List[str]] = [], stop: Optional[Union[str, List[str]]] = [],
max_tokens: int = 256, max_tokens: int = 256,
presence_penalty: float = 0.0, presence_penalty: float = 0.0,
frequency_penalty: float = 0.0, frequency_penalty: float = 0.0,
repeat_penalty: float = 1.1, repeat_penalty: float = 1.1,
tfs_z: float = 1.0,
mirostat_mode: int = 0, mirostat_mode: int = 0,
mirostat_tau: float = 5.0, mirostat_tau: float = 5.0,
mirostat_eta: float = 0.1, mirostat_eta: float = 0.1,
model: Optional[str] = None,
) -> Union[ChatCompletion, Iterator[ChatCompletionChunk]]: ) -> Union[ChatCompletion, Iterator[ChatCompletionChunk]]:
"""Generate a chat completion from a list of messages. """Generate a chat completion from a list of messages.
@ -1017,7 +1272,9 @@ class Llama:
Returns: Returns:
Generated chat completion or a stream of chat completion chunks. Generated chat completion or a stream of chat completion chunks.
""" """
stop = stop if stop is not None else [] stop = (
stop if isinstance(stop, list) else [stop] if isinstance(stop, str) else []
)
chat_history = "".join( chat_history = "".join(
f'### {"Human" if message["role"] == "user" else "Assistant"}:{message["content"]}' f'### {"Human" if message["role"] == "user" else "Assistant"}:{message["content"]}'
for message in messages for message in messages
@ -1035,9 +1292,11 @@ class Llama:
repeat_penalty=repeat_penalty, repeat_penalty=repeat_penalty,
presence_penalty=presence_penalty, presence_penalty=presence_penalty,
frequency_penalty=frequency_penalty, frequency_penalty=frequency_penalty,
tfs_z=tfs_z,
mirostat_mode=mirostat_mode, mirostat_mode=mirostat_mode,
mirostat_tau=mirostat_tau, mirostat_tau=mirostat_tau,
mirostat_eta=mirostat_eta, mirostat_eta=mirostat_eta,
model=model,
) )
if stream: if stream:
chunks: Iterator[CompletionChunk] = completion_or_chunks # type: ignore chunks: Iterator[CompletionChunk] = completion_or_chunks # type: ignore
@ -1056,7 +1315,7 @@ class Llama:
verbose=self.verbose, verbose=self.verbose,
model_path=self.model_path, model_path=self.model_path,
n_ctx=self.params.n_ctx, n_ctx=self.params.n_ctx,
n_parts=self.params.n_parts, n_gpu_layers=self.params.n_gpu_layers,
seed=self.params.seed, seed=self.params.seed,
f16_kv=self.params.f16_kv, f16_kv=self.params.f16_kv,
logits_all=self.params.logits_all, logits_all=self.params.logits_all,
@ -1069,6 +1328,9 @@ class Llama:
n_threads=self.n_threads, n_threads=self.n_threads,
lora_base=self.lora_base, lora_base=self.lora_base,
lora_path=self.lora_path, lora_path=self.lora_path,
### DEPRECATED ###
n_parts=self.n_parts,
### DEPRECATED ###
) )
def __setstate__(self, state): def __setstate__(self, state):
@ -1076,6 +1338,7 @@ class Llama:
model_path=state["model_path"], model_path=state["model_path"],
n_ctx=state["n_ctx"], n_ctx=state["n_ctx"],
n_parts=state["n_parts"], n_parts=state["n_parts"],
n_gpu_layers=state["n_gpu_layers"],
seed=state["seed"], seed=state["seed"],
f16_kv=state["f16_kv"], f16_kv=state["f16_kv"],
logits_all=state["logits_all"], logits_all=state["logits_all"],
@ -1120,16 +1383,41 @@ class Llama:
if llama_cpp.llama_set_state_data(self.ctx, state.llama_state) != state_size: if llama_cpp.llama_set_state_data(self.ctx, state.llama_state) != state_size:
raise RuntimeError("Failed to set llama state data") raise RuntimeError("Failed to set llama state data")
def n_ctx(self) -> int:
"""Return the context window size."""
assert self.ctx is not None
return llama_cpp.llama_n_ctx(self.ctx)
def n_embd(self) -> int:
"""Return the embedding size."""
assert self.ctx is not None
return llama_cpp.llama_n_embd(self.ctx)
def n_vocab(self) -> int:
"""Return the vocabulary size."""
assert self.ctx is not None
return llama_cpp.llama_n_vocab(self.ctx)
def tokenizer(self) -> "LlamaTokenizer":
"""Return the tokenizer for this model."""
assert self.ctx is not None
return LlamaTokenizer(self)
@staticmethod @staticmethod
def token_eos() -> llama_cpp.llama_token: def token_eos() -> int:
"""Return the end-of-sequence token.""" """Return the end-of-sequence token."""
return llama_cpp.llama_token_eos() return llama_cpp.llama_token_eos()
@staticmethod @staticmethod
def token_bos() -> llama_cpp.llama_token: def token_bos() -> int:
"""Return the beginning-of-sequence token.""" """Return the beginning-of-sequence token."""
return llama_cpp.llama_token_bos() return llama_cpp.llama_token_bos()
@staticmethod
def token_nl() -> int:
"""Return the newline token."""
return llama_cpp.llama_token_nl()
@staticmethod @staticmethod
def logits_to_logprobs(logits: List[float]) -> List[float]: def logits_to_logprobs(logits: List[float]) -> List[float]:
exps = [math.exp(float(x)) for x in logits] exps = [math.exp(float(x)) for x in logits]
@ -1137,9 +1425,7 @@ class Llama:
return [math.log(x / sum_exps) for x in exps] return [math.log(x / sum_exps) for x in exps]
@staticmethod @staticmethod
def longest_token_prefix( def longest_token_prefix(a: Sequence[int], b: Sequence[int]):
a: Sequence[llama_cpp.llama_token], b: Sequence[llama_cpp.llama_token]
):
longest_prefix = 0 longest_prefix = 0
for _a, _b in zip(a, b): for _a, _b in zip(a, b):
if _a == _b: if _a == _b:
@ -1147,3 +1433,20 @@ class Llama:
else: else:
break break
return longest_prefix return longest_prefix
class LlamaTokenizer:
def __init__(self, llama: Llama):
self.llama = llama
def encode(self, text: str, add_bos: bool = True) -> List[int]:
return self.llama.tokenize(
text.encode("utf-8", errors="ignore"), add_bos=add_bos
)
def decode(self, tokens: List[int]) -> str:
return self.llama.detokenize(tokens).decode("utf-8", errors="ignore")
@classmethod
def from_ggml_file(cls, path: str) -> "LlamaTokenizer":
return cls(Llama(model_path=path, vocab_only=True))

267
llama_cpp/llama_cpp.py
View file

@ -44,15 +44,20 @@ def _load_shared_library(lib_base_name: str):
_base_path = _lib.parent.resolve() _base_path = _lib.parent.resolve()
_lib_paths = [_lib.resolve()] _lib_paths = [_lib.resolve()]
cdll_args = dict() # type: ignore
# Add the library directory to the DLL search path on Windows (if needed) # Add the library directory to the DLL search path on Windows (if needed)
if sys.platform == "win32" and sys.version_info >= (3, 8): if sys.platform == "win32" and sys.version_info >= (3, 8):
os.add_dll_directory(str(_base_path)) os.add_dll_directory(str(_base_path))
if "CUDA_PATH" in os.environ:
os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "bin"))
os.add_dll_directory(os.path.join(os.environ["CUDA_PATH"], "lib"))
cdll_args["winmode"] = 0
# Try to load the shared library, handling potential errors # Try to load the shared library, handling potential errors
for _lib_path in _lib_paths: for _lib_path in _lib_paths:
if _lib_path.exists(): if _lib_path.exists():
try: try:
return ctypes.CDLL(str(_lib_path)) return ctypes.CDLL(str(_lib_path), **cdll_args)
except Exception as e: except Exception as e:
raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}") raise RuntimeError(f"Failed to load shared library '{_lib_path}': {e}")
@ -67,31 +72,61 @@ _lib_base_name = "llama"
# Load the library # Load the library
_lib = _load_shared_library(_lib_base_name) _lib = _load_shared_library(_lib_base_name)
# C types # Misc
LLAMA_FILE_VERSION = c_int(1) c_float_p = POINTER(c_float)
LLAMA_FILE_MAGIC = b"ggjt" c_uint8_p = POINTER(c_uint8)
LLAMA_FILE_MAGIC_UNVERSIONED = b"ggml" c_size_t_p = POINTER(c_size_t)
LLAMA_SESSION_MAGIC = b"ggsn"
# llama.h bindings
# #define LLAMA_FILE_MAGIC_GGJT 0x67676a74u // 'ggjt'
LLAMA_FILE_MAGIC_GGJT = ctypes.c_uint(0x67676A74)
# #define LLAMA_FILE_MAGIC_GGLA 0x67676c61u // 'ggla'
LLAMA_FILE_MAGIC_GGLA = ctypes.c_uint(0x67676C61)
# #define LLAMA_FILE_MAGIC_GGMF 0x67676d66u // 'ggmf'
LLAMA_FILE_MAGIC_GGMF = ctypes.c_uint(0x67676D66)
# #define LLAMA_FILE_MAGIC_GGML 0x67676d6cu // 'ggml'
LLAMA_FILE_MAGIC_GGML = ctypes.c_uint(0x67676D6C)
# #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
LLAMA_FILE_MAGIC_GGSN = ctypes.c_uint(0x6767736E)
# #define LLAMA_FILE_VERSION 3
LLAMA_FILE_VERSION = c_int(3)
LLAMA_FILE_MAGIC = LLAMA_FILE_MAGIC_GGJT
LLAMA_FILE_MAGIC_UNVERSIONED = LLAMA_FILE_MAGIC_GGML
LLAMA_SESSION_MAGIC = LLAMA_FILE_MAGIC_GGSN
LLAMA_SESSION_VERSION = c_int(1) LLAMA_SESSION_VERSION = c_int(1)
# struct llama_context;
llama_context_p = c_void_p llama_context_p = c_void_p
# typedef int llama_token;
llama_token = c_int llama_token = c_int
llama_token_p = POINTER(llama_token) llama_token_p = POINTER(llama_token)
# typedef struct llama_token_data {
# llama_token id; // token id
# float logit; // log-odds of the token
# float p; // probability of the token
# } llama_token_data;
class llama_token_data(Structure): class llama_token_data(Structure):
_fields_ = [ _fields_ = [
("id", llama_token), # token id ("id", llama_token),
("logit", c_float), # log-odds of the token ("logit", c_float),
("p", c_float), # probability of the token ("p", c_float),
] ]
llama_token_data_p = POINTER(llama_token_data) llama_token_data_p = POINTER(llama_token_data)
# typedef struct llama_token_data_array {
# llama_token_data * data;
# size_t size;
# bool sorted;
# } llama_token_data_array;
class llama_token_data_array(Structure): class llama_token_data_array(Structure):
_fields_ = [ _fields_ = [
("data", llama_token_data_p), ("data", llama_token_data_p),
@ -102,53 +137,72 @@ class llama_token_data_array(Structure):
llama_token_data_array_p = POINTER(llama_token_data_array) llama_token_data_array_p = POINTER(llama_token_data_array)
# typedef void (*llama_progress_callback)(float progress, void *ctx);
llama_progress_callback = ctypes.CFUNCTYPE(None, c_float, c_void_p) llama_progress_callback = ctypes.CFUNCTYPE(None, c_float, c_void_p)
# struct llama_context_params {
# int n_ctx; // text context
# int n_gpu_layers; // number of layers to store in VRAM
# int seed; // RNG seed, -1 for random
# bool f16_kv; // use fp16 for KV cache
# bool logits_all; // the llama_eval() call computes all logits, not just the last one
# bool vocab_only; // only load the vocabulary, no weights
# bool use_mmap; // use mmap if possible
# bool use_mlock; // force system to keep model in RAM
# bool embedding; // embedding mode only
# // called with a progress value between 0 and 1, pass NULL to disable
# llama_progress_callback progress_callback;
# // context pointer passed to the progress callback
# void * progress_callback_user_data;
# };
class llama_context_params(Structure): class llama_context_params(Structure):
_fields_ = [ _fields_ = [
("n_ctx", c_int), # text context ("n_ctx", c_int),
("n_parts", c_int), # -1 for default ("n_gpu_layers", c_int),
("seed", c_int), # RNG seed, 0 for random ("seed", c_int),
("f16_kv", c_bool), # use fp16 for KV cache ("f16_kv", c_bool),
( (
"logits_all", "logits_all",
c_bool, c_bool,
), # the llama_eval() call computes all logits, not just the last one ),
("vocab_only", c_bool), # only load the vocabulary, no weights ("vocab_only", c_bool),
("use_mmap", c_bool), # use mmap if possible ("use_mmap", c_bool),
("use_mlock", c_bool), # force system to keep model in RAM ("use_mlock", c_bool),
("embedding", c_bool), # embedding mode only ("embedding", c_bool),
# called with a progress value between 0 and 1, pass NULL to disable
("progress_callback", llama_progress_callback), ("progress_callback", llama_progress_callback),
# context pointer passed to the progress callback
("progress_callback_user_data", c_void_p), ("progress_callback_user_data", c_void_p),
] ]
llama_context_params_p = POINTER(llama_context_params) llama_context_params_p = POINTER(llama_context_params)
# enum llama_ftype {
# LLAMA_FTYPE_ALL_F32 = 0,
# LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
# // LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // support has been removed
# // LLAMA_FTYPE_MOSTLY_Q4_3 = 6, // support has been removed
# LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors
# LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors
# };
LLAMA_FTYPE_ALL_F32 = c_int(0) LLAMA_FTYPE_ALL_F32 = c_int(0)
LLAMA_FTYPE_MOSTLY_F16 = c_int(1) # except 1d tensors LLAMA_FTYPE_MOSTLY_F16 = c_int(1)
LLAMA_FTYPE_MOSTLY_Q4_0 = c_int(2) # except 1d tensors LLAMA_FTYPE_MOSTLY_Q4_0 = c_int(2)
LLAMA_FTYPE_MOSTLY_Q4_1 = c_int(3) # except 1d tensors LLAMA_FTYPE_MOSTLY_Q4_1 = c_int(3)
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = c_int( LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = c_int(4)
4 LLAMA_FTYPE_MOSTLY_Q8_0 = c_int(7)
) # tok_embeddings.weight and output.weight are F16 LLAMA_FTYPE_MOSTLY_Q5_0 = c_int(8)
LLAMA_FTYPE_MOSTLY_Q4_2 = c_int(5) # except 1d tensors LLAMA_FTYPE_MOSTLY_Q5_1 = c_int(9)
# LLAMA_FTYPE_MOSTYL_Q4_3 = c_int(6) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q8_0 = c_int(7) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_0 = c_int(8) # except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_1 = c_int(9) # except 1d tensors
# Misc
c_float_p = POINTER(c_float)
c_uint8_p = POINTER(c_uint8)
c_size_t_p = POINTER(c_size_t)
# Functions
# LLAMA_API struct llama_context_params llama_context_default_params();
def llama_context_default_params() -> llama_context_params: def llama_context_default_params() -> llama_context_params:
return _lib.llama_context_default_params() return _lib.llama_context_default_params()
@ -157,6 +211,7 @@ _lib.llama_context_default_params.argtypes = []
_lib.llama_context_default_params.restype = llama_context_params _lib.llama_context_default_params.restype = llama_context_params
# LLAMA_API bool llama_mmap_supported();
def llama_mmap_supported() -> bool: def llama_mmap_supported() -> bool:
return _lib.llama_mmap_supported() return _lib.llama_mmap_supported()
@ -165,6 +220,7 @@ _lib.llama_mmap_supported.argtypes = []
_lib.llama_mmap_supported.restype = c_bool _lib.llama_mmap_supported.restype = c_bool
# LLAMA_API bool llama_mlock_supported();
def llama_mlock_supported() -> bool: def llama_mlock_supported() -> bool:
return _lib.llama_mlock_supported() return _lib.llama_mlock_supported()
@ -173,9 +229,33 @@ _lib.llama_mlock_supported.argtypes = []
_lib.llama_mlock_supported.restype = c_bool _lib.llama_mlock_supported.restype = c_bool
# Various functions for loading a ggml llama model. # // TODO: not great API - very likely to change
# Allocate (almost) all memory needed for the model. # // Initialize the llama + ggml backend
# Return NULL on failure # // Call once at the start of the program
# LLAMA_API void llama_init_backend();
def llama_init_backend():
return _lib.llama_init_backend()
_lib.llama_init_backend.argtypes = []
_lib.llama_init_backend.restype = None
# LLAMA_API int64_t llama_time_us();
def llama_time_us() -> int:
return _lib.llama_time_us()
_lib.llama_time_us.argtypes = []
_lib.llama_time_us.restype = ctypes.c_int64
# // Various functions for loading a ggml llama model.
# // Allocate (almost) all memory needed for the model.
# // Return NULL on failure
# LLAMA_API struct llama_context * llama_init_from_file(
# const char * path_model,
# struct llama_context_params params);
def llama_init_from_file( def llama_init_from_file(
path_model: bytes, params: llama_context_params path_model: bytes, params: llama_context_params
) -> llama_context_p: ) -> llama_context_p:
@ -187,8 +267,9 @@ _lib.llama_init_from_file.restype = llama_context_p
# Frees all allocated memory # Frees all allocated memory
# LLAMA_API void llama_free(struct llama_context * ctx);
def llama_free(ctx: llama_context_p): def llama_free(ctx: llama_context_p):
_lib.llama_free(ctx) return _lib.llama_free(ctx)
_lib.llama_free.argtypes = [llama_context_p] _lib.llama_free.argtypes = [llama_context_p]
@ -198,9 +279,14 @@ _lib.llama_free.restype = None
# TODO: not great API - very likely to change # TODO: not great API - very likely to change
# Returns 0 on success # Returns 0 on success
# nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given # nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given
# LLAMA_API int llama_model_quantize(
# const char * fname_inp,
# const char * fname_out,
# enum llama_ftype ftype,
# int nthread);
def llama_model_quantize( def llama_model_quantize(
fname_inp: bytes, fname_out: bytes, ftype: c_int, nthread: c_int fname_inp: bytes, fname_out: bytes, ftype: c_int, nthread: c_int
) -> c_int: ) -> int:
return _lib.llama_model_quantize(fname_inp, fname_out, ftype, nthread) return _lib.llama_model_quantize(fname_inp, fname_out, ftype, nthread)
@ -214,12 +300,17 @@ _lib.llama_model_quantize.restype = c_int
# The model needs to be reloaded before applying a new adapter, otherwise the adapter # The model needs to be reloaded before applying a new adapter, otherwise the adapter
# will be applied on top of the previous one # will be applied on top of the previous one
# Returns 0 on success # Returns 0 on success
# LLAMA_API int llama_apply_lora_from_file(
# struct llama_context * ctx,
# const char * path_lora,
# const char * path_base_model,
# int n_threads);
def llama_apply_lora_from_file( def llama_apply_lora_from_file(
ctx: llama_context_p, ctx: llama_context_p,
path_lora: c_char_p, path_lora: c_char_p,
path_base_model: c_char_p, path_base_model: c_char_p,
n_threads: c_int, n_threads: c_int,
) -> c_int: ) -> int:
return _lib.llama_apply_lora_from_file(ctx, path_lora, path_base_model, n_threads) return _lib.llama_apply_lora_from_file(ctx, path_lora, path_base_model, n_threads)
@ -228,7 +319,8 @@ _lib.llama_apply_lora_from_file.restype = c_int
# Returns the number of tokens in the KV cache # Returns the number of tokens in the KV cache
def llama_get_kv_cache_token_count(ctx: llama_context_p) -> c_int: # LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
def llama_get_kv_cache_token_count(ctx: llama_context_p) -> int:
return _lib.llama_get_kv_cache_token_count(ctx) return _lib.llama_get_kv_cache_token_count(ctx)
@ -237,6 +329,7 @@ _lib.llama_get_kv_cache_token_count.restype = c_int
# Sets the current rng seed. # Sets the current rng seed.
# LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed);
def llama_set_rng_seed(ctx: llama_context_p, seed: c_int): def llama_set_rng_seed(ctx: llama_context_p, seed: c_int):
return _lib.llama_set_rng_seed(ctx, seed) return _lib.llama_set_rng_seed(ctx, seed)
@ -247,7 +340,8 @@ _lib.llama_set_rng_seed.restype = None
# Returns the maximum size in bytes of the state (rng, logits, embedding # Returns the maximum size in bytes of the state (rng, logits, embedding
# and kv_cache) - will often be smaller after compacting tokens # and kv_cache) - will often be smaller after compacting tokens
def llama_get_state_size(ctx: llama_context_p) -> c_size_t: # LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);
def llama_get_state_size(ctx: llama_context_p) -> int:
return _lib.llama_get_state_size(ctx) return _lib.llama_get_state_size(ctx)
@ -258,10 +352,11 @@ _lib.llama_get_state_size.restype = c_size_t
# Copies the state to the specified destination address. # Copies the state to the specified destination address.
# Destination needs to have allocated enough memory. # Destination needs to have allocated enough memory.
# Returns the number of bytes copied # Returns the number of bytes copied
# LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst);
def llama_copy_state_data( def llama_copy_state_data(
ctx: llama_context_p, dest # type: Array[c_uint8] ctx: llama_context_p, dst # type: Array[c_uint8]
) -> int: ) -> int:
return _lib.llama_copy_state_data(ctx, dest) return _lib.llama_copy_state_data(ctx, dst)
_lib.llama_copy_state_data.argtypes = [llama_context_p, c_uint8_p] _lib.llama_copy_state_data.argtypes = [llama_context_p, c_uint8_p]
@ -270,6 +365,7 @@ _lib.llama_copy_state_data.restype = c_size_t
# Set the state reading from the specified address # Set the state reading from the specified address
# Returns the number of bytes read # Returns the number of bytes read
# LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
def llama_set_state_data( def llama_set_state_data(
ctx: llama_context_p, src # type: Array[c_uint8] ctx: llama_context_p, src # type: Array[c_uint8]
) -> int: ) -> int:
@ -281,13 +377,14 @@ _lib.llama_set_state_data.restype = c_size_t
# Save/load session file # Save/load session file
# LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
def llama_load_session_file( def llama_load_session_file(
ctx: llama_context_p, ctx: llama_context_p,
path_session: bytes, path_session: bytes,
tokens_out, # type: Array[llama_token] tokens_out, # type: Array[llama_token]
n_token_capacity: c_size_t, n_token_capacity: c_size_t,
n_token_count_out, # type: _Pointer[c_size_t] n_token_count_out, # type: _Pointer[c_size_t]
) -> c_size_t: ) -> int:
return _lib.llama_load_session_file( return _lib.llama_load_session_file(
ctx, path_session, tokens_out, n_token_capacity, n_token_count_out ctx, path_session, tokens_out, n_token_capacity, n_token_count_out
) )
@ -303,12 +400,13 @@ _lib.llama_load_session_file.argtypes = [
_lib.llama_load_session_file.restype = c_size_t _lib.llama_load_session_file.restype = c_size_t
# LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
def llama_save_session_file( def llama_save_session_file(
ctx: llama_context_p, ctx: llama_context_p,
path_session: bytes, path_session: bytes,
tokens, # type: Array[llama_token] tokens, # type: Array[llama_token]
n_token_count: c_size_t, n_token_count: c_size_t,
) -> c_size_t: ) -> int:
return _lib.llama_save_session_file(ctx, path_session, tokens, n_token_count) return _lib.llama_save_session_file(ctx, path_session, tokens, n_token_count)
@ -325,13 +423,19 @@ _lib.llama_save_session_file.restype = c_size_t
# tokens + n_tokens is the provided batch of new tokens to process # 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 # n_past is the number of tokens to use from previous eval calls
# Returns 0 on success # Returns 0 on success
# LLAMA_API int llama_eval(
# struct llama_context * ctx,
# const llama_token * tokens,
# int n_tokens,
# int n_past,
# int n_threads);
def llama_eval( def llama_eval(
ctx: llama_context_p, ctx: llama_context_p,
tokens, # type: Array[llama_token] tokens, # type: Array[llama_token]
n_tokens: c_int, n_tokens: c_int,
n_past: c_int, n_past: c_int,
n_threads: c_int, n_threads: c_int,
) -> c_int: ) -> int:
return _lib.llama_eval(ctx, tokens, n_tokens, n_past, n_threads) return _lib.llama_eval(ctx, tokens, n_tokens, n_past, n_threads)
@ -344,13 +448,19 @@ _lib.llama_eval.restype = c_int
# Returns the number of tokens on success, no more than n_max_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 # Returns a negative number on failure - the number of tokens that would have been returned
# TODO: not sure if correct # TODO: not sure if correct
# LLAMA_API int llama_tokenize(
# struct llama_context * ctx,
# const char * text,
# llama_token * tokens,
# int n_max_tokens,
# bool add_bos);
def llama_tokenize( def llama_tokenize(
ctx: llama_context_p, ctx: llama_context_p,
text: bytes, text: bytes,
tokens, # type: Array[llama_token] tokens, # type: Array[llama_token]
n_max_tokens: c_int, n_max_tokens: c_int,
add_bos: c_bool, add_bos: c_bool,
) -> c_int: ) -> int:
return _lib.llama_tokenize(ctx, text, tokens, n_max_tokens, add_bos) return _lib.llama_tokenize(ctx, text, tokens, n_max_tokens, add_bos)
@ -358,7 +468,8 @@ _lib.llama_tokenize.argtypes = [llama_context_p, c_char_p, llama_token_p, c_int,
_lib.llama_tokenize.restype = c_int _lib.llama_tokenize.restype = c_int
def llama_n_vocab(ctx: llama_context_p) -> c_int: # LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
def llama_n_vocab(ctx: llama_context_p) -> int:
return _lib.llama_n_vocab(ctx) return _lib.llama_n_vocab(ctx)
@ -366,7 +477,8 @@ _lib.llama_n_vocab.argtypes = [llama_context_p]
_lib.llama_n_vocab.restype = c_int _lib.llama_n_vocab.restype = c_int
def llama_n_ctx(ctx: llama_context_p) -> c_int: # LLAMA_API int llama_n_ctx (const struct llama_context * ctx);
def llama_n_ctx(ctx: llama_context_p) -> int:
return _lib.llama_n_ctx(ctx) return _lib.llama_n_ctx(ctx)
@ -374,7 +486,8 @@ _lib.llama_n_ctx.argtypes = [llama_context_p]
_lib.llama_n_ctx.restype = c_int _lib.llama_n_ctx.restype = c_int
def llama_n_embd(ctx: llama_context_p) -> c_int: # LLAMA_API int llama_n_embd (const struct llama_context * ctx);
def llama_n_embd(ctx: llama_context_p) -> int:
return _lib.llama_n_embd(ctx) return _lib.llama_n_embd(ctx)
@ -387,6 +500,7 @@ _lib.llama_n_embd.restype = c_int
# Can be mutated in order to change the probabilities of the next token # Can be mutated in order to change the probabilities of the next token
# Rows: n_tokens # Rows: n_tokens
# Cols: n_vocab # Cols: n_vocab
# LLAMA_API float * llama_get_logits(struct llama_context * ctx);
def llama_get_logits( def llama_get_logits(
ctx: llama_context_p, ctx: llama_context_p,
): # type: (...) -> Array[float] # type: ignore ): # type: (...) -> Array[float] # type: ignore
@ -399,6 +513,7 @@ _lib.llama_get_logits.restype = c_float_p
# Get the embeddings for the input # Get the embeddings for the input
# shape: [n_embd] (1-dimensional) # shape: [n_embd] (1-dimensional)
# LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
def llama_get_embeddings( def llama_get_embeddings(
ctx: llama_context_p, ctx: llama_context_p,
): # type: (...) -> Array[float] # type: ignore ): # type: (...) -> Array[float] # type: ignore
@ -410,6 +525,7 @@ _lib.llama_get_embeddings.restype = c_float_p
# Token Id -> String. Uses the vocabulary in the provided context # Token Id -> String. Uses the vocabulary in the provided context
# LLAMA_API const char * llama_token_to_str(const struct llama_context * ctx, llama_token token);
def llama_token_to_str(ctx: llama_context_p, token: llama_token) -> bytes: def llama_token_to_str(ctx: llama_context_p, token: llama_token) -> bytes:
return _lib.llama_token_to_str(ctx, token) return _lib.llama_token_to_str(ctx, token)
@ -420,7 +536,8 @@ _lib.llama_token_to_str.restype = c_char_p
# Special tokens # Special tokens
def llama_token_bos() -> llama_token: # LLAMA_API llama_token llama_token_bos();
def llama_token_bos() -> int:
return _lib.llama_token_bos() return _lib.llama_token_bos()
@ -428,7 +545,8 @@ _lib.llama_token_bos.argtypes = []
_lib.llama_token_bos.restype = llama_token _lib.llama_token_bos.restype = llama_token
def llama_token_eos() -> llama_token: # LLAMA_API llama_token llama_token_eos();
def llama_token_eos() -> int:
return _lib.llama_token_eos() return _lib.llama_token_eos()
@ -436,7 +554,8 @@ _lib.llama_token_eos.argtypes = []
_lib.llama_token_eos.restype = llama_token _lib.llama_token_eos.restype = llama_token
def llama_token_nl() -> llama_token: # LLAMA_API llama_token llama_token_nl();
def llama_token_nl() -> int:
return _lib.llama_token_nl() return _lib.llama_token_nl()
@ -448,6 +567,7 @@ _lib.llama_token_nl.restype = llama_token
# @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix. # @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
# LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float penalty);
def llama_sample_repetition_penalty( def llama_sample_repetition_penalty(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
@ -471,6 +591,7 @@ _lib.llama_sample_repetition_penalty.restype = None
# @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details. # @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
# LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, const llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
def llama_sample_frequency_and_presence_penalties( def llama_sample_frequency_and_presence_penalties(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
@ -501,6 +622,7 @@ _lib.llama_sample_frequency_and_presence_penalties.restype = None
# @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits. # @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
# LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
def llama_sample_softmax( def llama_sample_softmax(
ctx: llama_context_p, candidates # type: _Pointer[llama_token_data] ctx: llama_context_p, candidates # type: _Pointer[llama_token_data]
): ):
@ -515,6 +637,7 @@ _lib.llama_sample_softmax.restype = None
# @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 # @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
# LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep);
def llama_sample_top_k( def llama_sample_top_k(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
@ -534,6 +657,7 @@ _lib.llama_sample_top_k.restype = None
# @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751 # @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
# LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
def llama_sample_top_p( def llama_sample_top_p(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
@ -553,6 +677,7 @@ _lib.llama_sample_top_p.restype = None
# @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/. # @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
# LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep);
def llama_sample_tail_free( def llama_sample_tail_free(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
@ -572,6 +697,7 @@ _lib.llama_sample_tail_free.restype = None
# @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666. # @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
# LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep);
def llama_sample_typical( def llama_sample_typical(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
@ -590,6 +716,7 @@ _lib.llama_sample_typical.argtypes = [
_lib.llama_sample_typical.restype = None _lib.llama_sample_typical.restype = None
# LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
def llama_sample_temperature( def llama_sample_temperature(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
@ -612,6 +739,7 @@ _lib.llama_sample_temperature.restype = None
# @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. # @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
# @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm. # @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
# @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. # @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
# LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
def llama_sample_token_mirostat( def llama_sample_token_mirostat(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
@ -619,7 +747,7 @@ def llama_sample_token_mirostat(
eta: c_float, eta: c_float,
m: c_int, m: c_int,
mu, # type: _Pointer[c_float] mu, # type: _Pointer[c_float]
) -> llama_token: ) -> int:
return _lib.llama_sample_token_mirostat(ctx, candidates, tau, eta, m, mu) return _lib.llama_sample_token_mirostat(ctx, candidates, tau, eta, m, mu)
@ -639,13 +767,14 @@ _lib.llama_sample_token_mirostat.restype = llama_token
# @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text. # @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
# @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates. # @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
# @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal. # @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
# LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
def llama_sample_token_mirostat_v2( def llama_sample_token_mirostat_v2(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
tau: c_float, tau: c_float,
eta: c_float, eta: c_float,
mu, # type: _Pointer[c_float] mu, # type: _Pointer[c_float]
) -> llama_token: ) -> int:
return _lib.llama_sample_token_mirostat_v2(ctx, candidates, tau, eta, mu) return _lib.llama_sample_token_mirostat_v2(ctx, candidates, tau, eta, mu)
@ -660,10 +789,11 @@ _lib.llama_sample_token_mirostat_v2.restype = llama_token
# @details Selects the token with the highest probability. # @details Selects the token with the highest probability.
# LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
def llama_sample_token_greedy( def llama_sample_token_greedy(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
) -> llama_token: ) -> int:
return _lib.llama_sample_token_greedy(ctx, candidates) return _lib.llama_sample_token_greedy(ctx, candidates)
@ -675,10 +805,11 @@ _lib.llama_sample_token_greedy.restype = llama_token
# @details Randomly selects a token from the candidates based on their probabilities. # @details Randomly selects a token from the candidates based on their probabilities.
# LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
def llama_sample_token( def llama_sample_token(
ctx: llama_context_p, ctx: llama_context_p,
candidates, # type: _Pointer[llama_token_data_array] candidates, # type: _Pointer[llama_token_data_array]
) -> llama_token: ) -> int:
return _lib.llama_sample_token(ctx, candidates) return _lib.llama_sample_token(ctx, candidates)
@ -692,6 +823,7 @@ _lib.llama_sample_token.restype = llama_token
# Performance information # Performance information
# LLAMA_API void llama_print_timings(struct llama_context * ctx);
def llama_print_timings(ctx: llama_context_p): def llama_print_timings(ctx: llama_context_p):
_lib.llama_print_timings(ctx) _lib.llama_print_timings(ctx)
@ -700,6 +832,7 @@ _lib.llama_print_timings.argtypes = [llama_context_p]
_lib.llama_print_timings.restype = None _lib.llama_print_timings.restype = None
# LLAMA_API void llama_reset_timings(struct llama_context * ctx);
def llama_reset_timings(ctx: llama_context_p): def llama_reset_timings(ctx: llama_context_p):
_lib.llama_reset_timings(ctx) _lib.llama_reset_timings(ctx)
@ -709,9 +842,19 @@ _lib.llama_reset_timings.restype = None
# Print system information # Print system information
# LLAMA_API const char * llama_print_system_info(void);
def llama_print_system_info() -> bytes: def llama_print_system_info() -> bytes:
return _lib.llama_print_system_info() return _lib.llama_print_system_info()
_lib.llama_print_system_info.argtypes = [] _lib.llama_print_system_info.argtypes = []
_lib.llama_print_system_info.restype = c_char_p _lib.llama_print_system_info.restype = c_char_p
###################################################################################################
_llama_initialized = False
if not _llama_initialized:
llama_init_backend()
_llama_initialized = True

6
llama_cpp/llama_types.py
View file

@ -1,4 +1,4 @@
from typing import List, Optional, Dict, Union from typing import List, Optional, Dict
from typing_extensions import TypedDict, NotRequired, Literal from typing_extensions import TypedDict, NotRequired, Literal
@ -22,9 +22,9 @@ class Embedding(TypedDict):
class CompletionLogprobs(TypedDict): class CompletionLogprobs(TypedDict):
text_offset: List[int] text_offset: List[int]
token_logprobs: List[float] token_logprobs: List[Optional[float]]
tokens: List[str] tokens: List[str]
top_logprobs: List[Dict[str, float]] top_logprobs: List[Optional[Dict[str, float]]]
class CompletionChoice(TypedDict): class CompletionChoice(TypedDict):

72
llama_cpp/server/app.py
View file

@ -1,4 +1,5 @@
import json import json
import logging
import multiprocessing import multiprocessing
from threading import Lock from threading import Lock
from typing import List, Optional, Union, Iterator, Dict from typing import List, Optional, Union, Iterator, Dict
@ -16,7 +17,16 @@ class Settings(BaseSettings):
model: str = Field( model: str = Field(
description="The path to the model to use for generating completions." description="The path to the model to use for generating completions."
) )
model_alias: Optional[str] = Field(
default=None,
description="The alias of the model to use for generating completions.",
)
n_ctx: int = Field(default=2048, ge=1, description="The context size.") n_ctx: int = Field(default=2048, ge=1, description="The context size.")
n_gpu_layers: int = Field(
default=0,
ge=0,
description="The number of layers to put on the GPU. The rest will be on the CPU.",
)
n_batch: int = Field( n_batch: int = Field(
default=512, ge=1, description="The batch size to use per eval." default=512, ge=1, description="The batch size to use per eval."
) )
@ -59,6 +69,7 @@ class Settings(BaseSettings):
router = APIRouter() router = APIRouter()
settings: Optional[Settings] = None
llama: Optional[llama_cpp.Llama] = None llama: Optional[llama_cpp.Llama] = None
@ -80,6 +91,7 @@ def create_app(settings: Optional[Settings] = None):
global llama global llama
llama = llama_cpp.Llama( llama = llama_cpp.Llama(
model_path=settings.model, model_path=settings.model,
n_gpu_layers=settings.n_gpu_layers,
f16_kv=settings.f16_kv, f16_kv=settings.f16_kv,
use_mlock=settings.use_mlock, use_mlock=settings.use_mlock,
use_mmap=settings.use_mmap, use_mmap=settings.use_mmap,
@ -95,6 +107,12 @@ def create_app(settings: Optional[Settings] = None):
if settings.cache: if settings.cache:
cache = llama_cpp.LlamaCache(capacity_bytes=settings.cache_size) cache = llama_cpp.LlamaCache(capacity_bytes=settings.cache_size)
llama.set_cache(cache) llama.set_cache(cache)
def set_settings(_settings: Settings):
global settings
settings = _settings
set_settings(settings)
return app return app
@ -106,6 +124,10 @@ def get_llama():
yield llama yield llama
def get_settings():
yield settings
model_field = Field(description="The model to use for generating completions.") model_field = Field(description="The model to use for generating completions.")
max_tokens_field = Field( max_tokens_field = Field(
@ -152,9 +174,23 @@ repeat_penalty_field = Field(
+ "Repeat penalty is a hyperparameter used to penalize the repetition of token sequences during text generation. It helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient.", + "Repeat penalty is a hyperparameter used to penalize the repetition of token sequences during text generation. It helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient.",
) )
presence_penalty_field = Field(
default=0.0,
ge=-2.0,
le=2.0,
description="Positive values penalize new tokens based on whether they appear in the text so far, increasing the model's likelihood to talk about new topics.",
)
frequency_penalty_field = Field(
default=0.0,
ge=-2.0,
le=2.0,
description="Positive values penalize new tokens based on their existing frequency in the text so far, decreasing the model's likelihood to repeat the same line verbatim.",
)
class CreateCompletionRequest(BaseModel): class CreateCompletionRequest(BaseModel):
prompt: Optional[str] = Field( prompt: Union[str, List[str]] = Field(
default="", description="The prompt to generate completions for." default="", description="The prompt to generate completions for."
) )
suffix: Optional[str] = Field( suffix: Optional[str] = Field(
@ -168,20 +204,20 @@ class CreateCompletionRequest(BaseModel):
default=False, default=False,
description="Whether to echo the prompt in the generated text. Useful for chatbots.", description="Whether to echo the prompt in the generated text. Useful for chatbots.",
) )
stop: Optional[List[str]] = stop_field stop: Optional[Union[str, List[str]]] = stop_field
stream: bool = stream_field stream: bool = stream_field
logprobs: Optional[int] = Field( logprobs: Optional[int] = Field(
default=None, default=None,
ge=0, ge=0,
description="The number of logprobs to generate. If None, no logprobs are generated.", description="The number of logprobs to generate. If None, no logprobs are generated.",
) )
presence_penalty: Optional[float] = presence_penalty_field
frequency_penalty: Optional[float] = frequency_penalty_field
# ignored or currently unsupported # ignored or currently unsupported
model: Optional[str] = model_field model: Optional[str] = model_field
n: Optional[int] = 1 n: Optional[int] = 1
logprobs: Optional[int] = Field(None) logprobs: Optional[int] = Field(None)
presence_penalty: Optional[float] = 0
frequency_penalty: Optional[float] = 0
best_of: Optional[int] = 1 best_of: Optional[int] = 1
logit_bias: Optional[Dict[str, float]] = Field(None) logit_bias: Optional[Dict[str, float]] = Field(None)
user: Optional[str] = Field(None) user: Optional[str] = Field(None)
@ -209,10 +245,13 @@ CreateCompletionResponse = create_model_from_typeddict(llama_cpp.Completion)
def create_completion( def create_completion(
request: CreateCompletionRequest, llama: llama_cpp.Llama = Depends(get_llama) request: CreateCompletionRequest, llama: llama_cpp.Llama = Depends(get_llama)
): ):
if isinstance(request.prompt, list):
assert len(request.prompt) <= 1
request.prompt = request.prompt[0] if len(request.prompt) > 0 else ""
completion_or_chunks = llama( completion_or_chunks = llama(
**request.dict( **request.dict(
exclude={ exclude={
"model",
"n", "n",
"best_of", "best_of",
"logit_bias", "logit_bias",
@ -221,15 +260,22 @@ def create_completion(
) )
) )
if request.stream: if request.stream:
async def server_sent_events(
chunks: Iterator[llama_cpp.CompletionChunk],
):
for chunk in chunks:
yield dict(data=json.dumps(chunk))
chunks: Iterator[llama_cpp.CompletionChunk] = completion_or_chunks # type: ignore chunks: Iterator[llama_cpp.CompletionChunk] = completion_or_chunks # type: ignore
return EventSourceResponse(dict(data=json.dumps(chunk)) for chunk in chunks) return EventSourceResponse(server_sent_events(chunks))
completion: llama_cpp.Completion = completion_or_chunks # type: ignore completion: llama_cpp.Completion = completion_or_chunks # type: ignore
return completion return completion
class CreateEmbeddingRequest(BaseModel): class CreateEmbeddingRequest(BaseModel):
model: Optional[str] = model_field model: Optional[str] = model_field
input: str = Field(description="The input to embed.") input: Union[str, List[str]] = Field(description="The input to embed.")
user: Optional[str] user: Optional[str]
class Config: class Config:
@ -250,7 +296,7 @@ CreateEmbeddingResponse = create_model_from_typeddict(llama_cpp.Embedding)
def create_embedding( def create_embedding(
request: CreateEmbeddingRequest, llama: llama_cpp.Llama = Depends(get_llama) request: CreateEmbeddingRequest, llama: llama_cpp.Llama = Depends(get_llama)
): ):
return llama.create_embedding(**request.dict(exclude={"model", "user"})) return llama.create_embedding(**request.dict(exclude={"user"}))
class ChatCompletionRequestMessage(BaseModel): class ChatCompletionRequestMessage(BaseModel):
@ -269,12 +315,12 @@ class CreateChatCompletionRequest(BaseModel):
top_p: float = top_p_field top_p: float = top_p_field
stop: Optional[List[str]] = stop_field stop: Optional[List[str]] = stop_field
stream: bool = stream_field stream: bool = stream_field
presence_penalty: Optional[float] = presence_penalty_field
frequency_penalty: Optional[float] = frequency_penalty_field
# ignored or currently unsupported # ignored or currently unsupported
model: Optional[str] = model_field model: Optional[str] = model_field
n: Optional[int] = 1 n: Optional[int] = 1
presence_penalty: Optional[float] = 0
frequency_penalty: Optional[float] = 0
logit_bias: Optional[Dict[str, float]] = Field(None) logit_bias: Optional[Dict[str, float]] = Field(None)
user: Optional[str] = Field(None) user: Optional[str] = Field(None)
@ -311,7 +357,6 @@ def create_chat_completion(
completion_or_chunks = llama.create_chat_completion( completion_or_chunks = llama.create_chat_completion(
**request.dict( **request.dict(
exclude={ exclude={
"model",
"n", "n",
"logit_bias", "logit_bias",
"user", "user",
@ -354,13 +399,16 @@ GetModelResponse = create_model_from_typeddict(ModelList)
@router.get("/v1/models", response_model=GetModelResponse) @router.get("/v1/models", response_model=GetModelResponse)
def get_models( def get_models(
settings: Settings = Depends(get_settings),
llama: llama_cpp.Llama = Depends(get_llama), llama: llama_cpp.Llama = Depends(get_llama),
) -> ModelList: ) -> ModelList:
return { return {
"object": "list", "object": "list",
"data": [ "data": [
{ {
"id": llama.model_path, "id": settings.model_alias
if settings.model_alias is not None
else llama.model_path,
"object": "model", "object": "model",
"owned_by": "me", "owned_by": "me",
"permissions": [], "permissions": [],

12
poetry.lock generated
View file

@ -463,14 +463,14 @@ socks = ["socksio (>=1.0.0,<2.0.0)"]
[[package]] [[package]]
name = "httpx" name = "httpx"
version = "0.24.0" version = "0.24.1"
description = "The next generation HTTP client." description = "The next generation HTTP client."
category = "dev" category = "dev"
optional = false optional = false
python-versions = ">=3.7" python-versions = ">=3.7"
files = [ files = [
{file = "httpx-0.24.0-py3-none-any.whl", hash = "sha256:447556b50c1921c351ea54b4fe79d91b724ed2b027462ab9a329465d147d5a4e"}, {file = "httpx-0.24.1-py3-none-any.whl", hash = "sha256:06781eb9ac53cde990577af654bd990a4949de37a28bdb4a230d434f3a30b9bd"},
{file = "httpx-0.24.0.tar.gz", hash = "sha256:507d676fc3e26110d41df7d35ebd8b3b8585052450f4097401c9be59d928c63e"}, {file = "httpx-0.24.1.tar.gz", hash = "sha256:5853a43053df830c20f8110c5e69fe44d035d850b2dfe795e196f00fdb774bdd"},
] ]
[package.dependencies] [package.dependencies]
@ -800,14 +800,14 @@ mkdocs = ">=1.1"
[[package]] [[package]]
name = "mkdocs-material" name = "mkdocs-material"
version = "9.1.11" version = "9.1.14"
description = "Documentation that simply works" description = "Documentation that simply works"
category = "dev" category = "dev"
optional = false optional = false
python-versions = ">=3.7" python-versions = ">=3.7"
files = [ files = [
{file = "mkdocs_material-9.1.11-py3-none-any.whl", hash = "sha256:fbc86d50ec2cf34d40d5c4365780f290ceedde23f1a0704323b34e7f16b0c0dd"}, {file = "mkdocs_material-9.1.14-py3-none-any.whl", hash = "sha256:b56a9f955ed32d38333715cbbf68ce38f683bf38610c65094fa4ef2db9f08bcd"},
{file = "mkdocs_material-9.1.11.tar.gz", hash = "sha256:f5d473eb79d6640a5e668d4b2ab5b9de5e76ae0a0e2d864112df0cfe9016dc1d"}, {file = "mkdocs_material-9.1.14.tar.gz", hash = "sha256:1ae74cc5464ef2f64574d4884512efed7f4db386fb9bc6af20fd427d7a702f49"},
] ]
[package.dependencies] [package.dependencies]

6
pyproject.toml
View file

@ -1,6 +1,6 @@
[tool.poetry] [tool.poetry]
name = "llama_cpp_python" name = "llama_cpp_python"
version = "0.1.48" version = "0.1.55"
description = "Python bindings for the llama.cpp library" description = "Python bindings for the llama.cpp library"
authors = ["Andrei Betlen <abetlen@gmail.com>"] authors = ["Andrei Betlen <abetlen@gmail.com>"]
license = "MIT" license = "MIT"
@ -24,9 +24,9 @@ black = "^23.3.0"
twine = "^4.0.2" twine = "^4.0.2"
mkdocs = "^1.4.3" mkdocs = "^1.4.3"
mkdocstrings = {extras = ["python"], version = "^0.21.2"} mkdocstrings = {extras = ["python"], version = "^0.21.2"}
mkdocs-material = "^9.1.11" mkdocs-material = "^9.1.14"
pytest = "^7.3.1" pytest = "^7.3.1"
httpx = "^0.24.0" httpx = "^0.24.1"
scikit-build = "0.13" scikit-build = "0.13"
[tool.poetry.extras] [tool.poetry.extras]

2
setup.py
View file

@ -10,7 +10,7 @@ setup(
description="A Python wrapper for llama.cpp", description="A Python wrapper for llama.cpp",
long_description=long_description, long_description=long_description,
long_description_content_type="text/markdown", long_description_content_type="text/markdown",
version="0.1.48", version="0.1.55",
author="Andrei Betlen", author="Andrei Betlen",
author_email="abetlen@gmail.com", author_email="abetlen@gmail.com",
license="MIT", license="MIT",

4
tests/test_llama.py
View file

@ -17,7 +17,7 @@ def test_llama():
# @pytest.mark.skip(reason="need to update sample mocking") # @pytest.mark.skip(reason="need to update sample mocking")
def test_llama_patch(monkeypatch): def test_llama_patch(monkeypatch):
llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True) llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True)
n_vocab = int(llama_cpp.llama_n_vocab(llama.ctx)) n_vocab = llama_cpp.llama_n_vocab(llama.ctx)
## Set up mock function ## Set up mock function
def mock_eval(*args, **kwargs): def mock_eval(*args, **kwargs):
@ -107,7 +107,7 @@ def test_llama_pickle():
def test_utf8(monkeypatch): def test_utf8(monkeypatch):
llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True) llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True)
n_vocab = int(llama_cpp.llama_n_vocab(llama.ctx)) n_vocab = llama_cpp.llama_n_vocab(llama.ctx)
## Set up mock function ## Set up mock function
def mock_eval(*args, **kwargs): def mock_eval(*args, **kwargs):

2
vendor/llama.cpp vendored

@ -1 +1 @@
Subproject commit 1b0fd454650ef4d68a980e3225488b79e6e9af25 Subproject commit 66874d4fbcc7866377246efbcee938e8cc9c7d76