Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
75 lines
3.3 KiB
Python
75 lines
3.3 KiB
Python
#!/usr/bin/env python3
|
|
"""Package an Irodori inference checkpoint and DACVAE as a standalone GGUF."""
|
|
|
|
from __future__ import annotations
|
|
|
|
import argparse
|
|
import json
|
|
import subprocess
|
|
import tempfile
|
|
from pathlib import Path
|
|
|
|
from safetensors import safe_open
|
|
|
|
K_QUANT_TYPES = ("q4_k", "q5_k", "q6_k")
|
|
K_QUANT_BLOCK = 256
|
|
|
|
|
|
def k_quant_fallbacks(source) -> list[str]:
|
|
"""2-D weights that a k-quant would leave in F32 because their input width
|
|
is not a multiple of 256 (v4 Large's 1152-wide T5Gemma 2 projections).
|
|
They are stored as Q8_0 instead. Embeddings keep the converter's F16 rule."""
|
|
names = []
|
|
for name in source.keys():
|
|
shape = source.get_slice(name).get_shape()
|
|
if (name.endswith(".weight") and len(shape) == 2 and "embed" not in name.lower()
|
|
and shape[-1] % K_QUANT_BLOCK != 0 and shape[-1] % 32 == 0):
|
|
names.append(name)
|
|
return names
|
|
|
|
|
|
def main() -> None:
|
|
parser = argparse.ArgumentParser(description=__doc__)
|
|
parser.add_argument("--model", type=Path, required=True, help="Irodori model.safetensors.")
|
|
parser.add_argument("--codec", type=Path, required=True, help="DACVAE weights.safetensors.")
|
|
parser.add_argument("--output", type=Path, required=True)
|
|
parser.add_argument("--type", choices=["orig", "f16", "bf16", "q8_0", *K_QUANT_TYPES], default="q8_0")
|
|
parser.add_argument("--converter", type=Path, default=Path("build/debug/bin/audiocpp_gguf"))
|
|
parser.add_argument("--overwrite", action="store_true")
|
|
parser.add_argument("--log", action="store_true")
|
|
args = parser.parse_args()
|
|
model = args.model.resolve(strict=True)
|
|
codec = args.codec.resolve(strict=True)
|
|
converter = args.converter.resolve(strict=True)
|
|
with safe_open(model, framework="pt", device="cpu") as source:
|
|
metadata = source.metadata() or {}
|
|
fallbacks = k_quant_fallbacks(source) if args.type in K_QUANT_TYPES else []
|
|
config = json.loads(metadata["config_json"])
|
|
sidecars = {"model_config.json": config}
|
|
if config.get("text_encoder_type") == "pretrained":
|
|
sidecars["text_encoder_config.json"] = json.loads(metadata["text_encoder_config_json"])
|
|
tokenizer = model.parent / "tokenizer" / "tokenizer.json"
|
|
if not tokenizer.is_file():
|
|
raise FileNotFoundError(f"Missing checkpoint tokenizer: {tokenizer}")
|
|
args.output.parent.mkdir(parents=True, exist_ok=True)
|
|
with tempfile.TemporaryDirectory(prefix="irodori_gguf_") as directory:
|
|
command = [str(converter), "--family", "irodori_tts",
|
|
"--input", f"model_weights={model}",
|
|
"--input", f"codec_weights={codec}",
|
|
"--root", str(model.parent), "--type", args.type,
|
|
"--output", str(args.output.resolve())]
|
|
for name, value in sidecars.items():
|
|
path = Path(directory) / name
|
|
path.write_text(json.dumps(value, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
|
|
command.extend(["--sidecar", f"{path}={name}"])
|
|
for name in fallbacks:
|
|
command.extend(["--keep-type", f"model_weights/{name}=q8_0"])
|
|
if args.overwrite:
|
|
command.append("--overwrite")
|
|
if args.log:
|
|
command.append("--log")
|
|
subprocess.run(command, check=True)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|