Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-18 16:55:14 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	.github/workflows/build.yml
#	.github/workflows/close-issue.yml
#	.github/workflows/server.yml
#	AUTHORS
#	CMakeLists.txt
#	Makefile
#	README.md
#	cmake/llama.pc.in
#	common/CMakeLists.txt
#	docs/build.md
#	examples/batched.swift/Sources/main.swift
#	examples/llama.swiftui/llama.cpp.swift/LibLlama.swift
#	examples/llava/CMakeLists.txt
#	examples/llava/clip.h
#	examples/run/run.cpp
#	examples/server/README.md
#	ggml/CMakeLists.txt
#	ggml/src/ggml-cuda/CMakeLists.txt
#	ggml/src/ggml-hip/CMakeLists.txt
#	ggml/src/ggml-musa/CMakeLists.txt
#	scripts/sync-ggml.last
#	tests/CMakeLists.txt
#	tests/test-backend-ops.cpp
#	tests/test-chat-template.cpp
#	tests/test-grammar-integration.cpp
#	tests/test-json-schema-to-grammar.cpp
This commit is contained in:
Concedo
2025-02-07 00:52:31 +08:00
parent 8fef9f3fb5 2fb3c32a16
commit db6db9dff9
142 changed files with 74261 additions and 10014 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/fetch_server_test_models.py
-105
View File
@@ -1,105 +0,0 @@
#!/usr/bin/env python
'''
This script fetches all the models used in the server tests.
This is useful for slow tests that use larger models, to avoid them timing out on the model downloads.
It is meant to be run from the root of the repository.
Example:
python scripts/fetch_server_test_models.py
( cd examples/server/tests && ./tests.sh -v -x -m slow )
'''
import ast
import glob
import logging
import os
from typing import Generator
from pydantic import BaseModel
from typing import Optional
import subprocess
class HuggingFaceModel(BaseModel):
hf_repo: str
hf_file: Optional[str] = None
class Config:
frozen = True
def collect_hf_model_test_parameters(test_file) -> Generator[HuggingFaceModel, None, None]:
try:
with open(test_file) as f:
tree = ast.parse(f.read())
except Exception as e:
logging.error(f'collect_hf_model_test_parameters failed on {test_file}: {e}')
return
for node in ast.walk(tree):
if isinstance(node, ast.FunctionDef):
for dec in node.decorator_list:
if isinstance(dec, ast.Call) and isinstance(dec.func, ast.Attribute) and dec.func.attr == 'parametrize':
param_names = ast.literal_eval(dec.args[0]).split(",")
if "hf_repo" not in param_names:
continue
raw_param_values = dec.args[1]
if not isinstance(raw_param_values, ast.List):
logging.warning(f'Skipping non-list parametrize entry at {test_file}:{node.lineno}')
continue
hf_repo_idx = param_names.index("hf_repo")
hf_file_idx = param_names.index("hf_file") if "hf_file" in param_names else None
for t in raw_param_values.elts:
if not isinstance(t, ast.Tuple):
logging.warning(f'Skipping non-tuple parametrize entry at {test_file}:{node.lineno}')
continue
yield HuggingFaceModel(
hf_repo=ast.literal_eval(t.elts[hf_repo_idx]),
hf_file=ast.literal_eval(t.elts[hf_file_idx]) if hf_file_idx is not None else None)
if __name__ == '__main__':
logging.basicConfig(level=logging.INFO, format='%(levelname)s: %(message)s')
models = sorted(list(set([
model
for test_file in glob.glob('examples/server/tests/unit/test_*.py')
for model in collect_hf_model_test_parameters(test_file)
])), key=lambda m: (m.hf_repo, m.hf_file))
logging.info(f'Found {len(models)} models in parameterized tests:')
for m in models:
logging.info(f' - {m.hf_repo} / {m.hf_file}')
cli_path = os.environ.get(
'LLAMA_SERVER_BIN_PATH',
os.path.join(
os.path.dirname(__file__),
'../build/bin/Release/llama-cli.exe' if os.name == 'nt' else '../build/bin/llama-cli'))
for m in models:
if '<' in m.hf_repo or (m.hf_file is not None and '<' in m.hf_file):
continue
if m.hf_file is not None and '-of-' in m.hf_file:
logging.warning(f'Skipping model at {m.hf_repo} / {m.hf_file} because it is a split file')
continue
logging.info(f'Using llama-cli to ensure model {m.hf_repo}/{m.hf_file} was fetched')
cmd = [
cli_path,
'-hfr', m.hf_repo,
*([] if m.hf_file is None else ['-hff', m.hf_file]),
'-n', '1',
'-p', 'Hey',
'--no-warmup',
'--log-disable',
'-no-cnv']
if m.hf_file != 'tinyllamas/stories260K.gguf' and 'Mistral-Nemo' not in m.hf_repo:
cmd.append('-fa')
try:
subprocess.check_call(cmd)
except subprocess.CalledProcessError:
logging.error(f'Failed to fetch model at {m.hf_repo} / {m.hf_file} with command:\n {" ".join(cmd)}')
exit(1)
scripts/get_chat_template.py
-77
View File
@@ -1,77 +0,0 @@
#!/usr/bin/env python
'''
Fetches the Jinja chat template of a HuggingFace model.
If a model has multiple chat templates, you can specify the variant name.
Syntax:
./scripts/get_chat_template.py model_id [variant]
Examples:
./scripts/get_chat_template.py NousResearch/Meta-Llama-3-8B-Instruct
./scripts/get_chat_template.py NousResearch/Hermes-3-Llama-3.1-8B tool_use
./scripts/get_chat_template.py meta-llama/Llama-3.2-3B-Instruct
'''
import json
import re
import sys
def get_chat_template(model_id, variant=None):
try:
# Use huggingface_hub library if available.
# Allows access to gated models if the user has access and ran `huggingface-cli login`.
from huggingface_hub import hf_hub_download
with open(hf_hub_download(repo_id=model_id, filename="tokenizer_config.json")) as f:
config_str = f.read()
except ImportError:
import requests
assert re.match(r"^[\w.-]+/[\w.-]+$", model_id), f"Invalid model ID: {model_id}"
response = requests.get(f"https://huggingface.co/{model_id}/resolve/main/tokenizer_config.json")
if response.status_code == 401:
raise Exception('Access to this model is gated, please request access, authenticate with `huggingface-cli login` and make sure to run `pip install huggingface_hub`')
response.raise_for_status()
config_str = response.text
try:
config = json.loads(config_str)
except json.JSONDecodeError:
# Fix https://huggingface.co/NousResearch/Meta-Llama-3-8B-Instruct/blob/main/tokenizer_config.json
# (Remove extra '}' near the end of the file)
config = json.loads(re.sub(r'\}([\n\s]*\}[\n\s]*\],[\n\s]*"clean_up_tokenization_spaces")', r'\1', config_str))
chat_template = config['chat_template']
if isinstance(chat_template, str):
return chat_template
else:
variants = {
ct['name']: ct['template']
for ct in chat_template
}
def format_variants():
return ', '.join(f'"{v}"' for v in variants.keys())
if variant is None:
if 'default' not in variants:
raise Exception(f'Please specify a chat template variant (one of {format_variants()})')
variant = 'default'
sys.stderr.write(f'Note: picked "default" chat template variant (out of {format_variants()})\n')
elif variant not in variants:
raise Exception(f"Variant {variant} not found in chat template (found {format_variants()})")
return variants[variant]
def main(args):
if len(args) < 1:
raise ValueError("Please provide a model ID and an optional variant name")
model_id = args[0]
variant = None if len(args) < 2 else args[1]
template = get_chat_template(model_id, variant)
sys.stdout.write(template)
if __name__ == '__main__':
main(sys.argv[1:])
scripts/hf.sh
-112
View File
@@ -1,112 +0,0 @@
#!/bin/bash
#
# Shortcut for downloading HF models
#
# Usage:
# ./llama-cli -m $(./scripts/hf.sh https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
# ./llama-cli -m $(./scripts/hf.sh --url https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/blob/main/mixtral-8x7b-v0.1.Q4_K_M.gguf)
# ./llama-cli -m $(./scripts/hf.sh --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf)
#
# all logs go to stderr
function log {
echo "$@" 1>&2
}
function usage {
log "Usage: $0 [[--url] <url>] [--repo <repo>] [--file <file>] [--outdir <dir> [-h|--help]"
exit 1
}
# check for curl or wget
function has_cmd {
if ! [ -x "$(command -v $1)" ]; then
return 1
fi
}
if has_cmd wget; then
cmd="wget -q -c -O %s/%s %s"
elif has_cmd curl; then
cmd="curl -C - -f --output-dir %s -o %s -L %s"
else
log "[E] curl or wget not found"
exit 1
fi
url=""
repo=""
file=""
outdir="."
# parse args
while [[ $# -gt 0 ]]; do
case "$1" in
--url)
url="$2"
shift 2
;;
--repo)
repo="$2"
shift 2
;;
--file)
file="$2"
shift 2
;;
--outdir)
outdir="$2"
shift 2
;;
-h|--help)
usage
;;
*)
url="$1"
shift
;;
esac
done
if [ -n "$repo" ] && [ -n "$file" ]; then
url="https://huggingface.co/$repo/resolve/main/$file"
fi
if [ -z "$url" ]; then
log "[E] missing --url"
usage
fi
# check if the URL is a HuggingFace model, and if so, try to download it
is_url=false
if [[ ${#url} -gt 22 ]]; then
if [[ ${url:0:22} == "https://huggingface.co" ]]; then
is_url=true
fi
fi
if [ "$is_url" = false ]; then
log "[E] invalid URL, must start with https://huggingface.co"
exit 0
fi
# replace "blob/main" with "resolve/main"
url=${url/blob\/main/resolve\/main}
basename=$(basename $url)
log "[+] attempting to download $basename"
if [ -n "$cmd" ]; then
cmd=$(printf "$cmd" "$outdir" "$basename" "$url")
log "[+] $cmd"
if $cmd; then
echo $outdir/$basename
exit 0
fi
fi
log "[-] failed to download"
exit 1