Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
"""Generate offline capabilities from the exact pinned llama.cpp converter AST.
Supply the unmodified convert_hf_to_gguf.py from the recorded commit. This script
never imports the converter or fetches source at runtime. See docs/serving-and-export.md.
"""
from __future__ import annotations
import argparse
import ast
import hashlib
import json
import re
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
OUTPUT = ROOT / "src/soup_cli/utils/llama_cpp_capabilities.json"
def extract_capabilities(source: bytes) -> dict[str, list[str]]:
"""Resolve the same class inheritance/model_arch rule as ModelBase.register."""
tree = ast.parse(source)
classes = {node.name: node for node in tree.body if isinstance(node, ast.ClassDef)}
def model_arch(name: str, visited: frozenset[str] = frozenset()) -> str | None:
if name in visited:
raise ValueError(f"cyclic converter class inheritance: {name}")
node = classes[name]
for statement in node.body:
if (isinstance(statement, ast.Assign)
and any(isinstance(target, ast.Name) and target.id == "model_arch"
for target in statement.targets)):
if not isinstance(statement.value, ast.Attribute):
raise ValueError(f"unrecognized model_arch on {name}")
return statement.value.attr
for base in node.bases:
if isinstance(base, ast.Name) and base.id in classes:
value = model_arch(base.id, visited | {name})
if value is not None:
return value
return None
kinds: dict[str, set[str]] = {"TEXT": set(), "VISION": set()}
for node in classes.values():
for decorator in node.decorator_list:
if (isinstance(decorator, ast.Call)
and isinstance(decorator.func, ast.Attribute)
and isinstance(decorator.func.value, ast.Name)
and decorator.func.value.id == "ModelBase"
and decorator.func.attr == "register"):
arch = model_arch(node.name)
if arch is None or not decorator.args or decorator.keywords:
raise ValueError(f"unrecognized registration on {node.name}")
kind = "VISION" if arch == "CLIP_VISION" else "TEXT"
for argument in decorator.args:
name = ast.literal_eval(argument)
if not isinstance(name, str) or not name:
raise ValueError(f"non-string architecture on {node.name}")
kinds[kind].add(name)
if not kinds["TEXT"]:
raise ValueError("converter has no recognized text registrations")
return {kind: sorted(names) for kind, names in kinds.items()}
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--converter", type=Path, required=True)
parser.add_argument("--commit", required=True)
parser.add_argument("--tag", required=True)
parser.add_argument("--output", type=Path, default=OUTPUT)
arguments = parser.parse_args()
export_tree = ast.parse((ROOT / "src/soup_cli/commands/export.py").read_text(encoding="utf-8"))
pin = next(ast.literal_eval(node.value) for node in export_tree.body
if isinstance(node, ast.Assign)
and any(isinstance(target, ast.Name) and target.id == "LLAMA_CPP_TAG"
for target in node.targets))
if arguments.tag != pin:
parser.error(f"--tag must match LLAMA_CPP_TAG={pin}")
if not re.fullmatch(r"[0-9a-f]{40}", arguments.commit):
parser.error("--commit must be a full immutable Git commit SHA")
source = arguments.converter.read_bytes()
# Record both the file hash and Git's blob identity, not a reformatted copy.
blob = f"blob {len(source)}\0".encode() + source
manifest = {
"schema_version": 1,
"tag": arguments.tag,
"commit": arguments.commit,
"source_url": (
f"https://github.com/ggml-org/llama.cpp/blob/{arguments.commit}/convert_hf_to_gguf.py"
),
"source_sha256": hashlib.sha256(source).hexdigest(),
"source_git_blob": hashlib.sha1(blob).hexdigest(),
"architectures": extract_capabilities(source),
}
arguments.output.write_text(json.dumps(manifest, indent=2) + "\n", encoding="utf-8")
if __name__ == "__main__":
main()