Edit photos by describing the change, on your own GPU. A free, open-source alternative to Nano Banana and ChatGPT image editing: photos stay on your computer, no daily limit. One installer for Windows, Mac and Linux sets up the best open model (Qwen-Image 2.1). MCP server for AI agents.
TypeScript ai-photo-editorcomfyui-alternativegenerative-fillimage-editingimage-generationinpaintinglinuxlocal-aimacosmcpmcp-servernano-banana-alternativeofflinephoto-editorprivacyqwen-imageqwen-image-2-1stable-diffusion-cppvulkanwindows
"""Writes Tests/KokonEngineTests/Fixtures/quant.json: random blocks of each GGUF type the engine reads, and
the values gguf-py (ggml's Python reference) unpacks them to. The Swift tests check the CPU and GPU
unpacking against these values, so the oracle is not the engine's own transcription of ggml.
python3 -m venv /tmp/gguf-venv && /tmp/gguf-venv/bin/pip install gguf numpy
/tmp/gguf-venv/bin/python -I engine-mac/tools/make_quant_fixtures.py engine-mac/Tests/KokonEngineTests/Fixtures/quant.json
"""
import json
import sys
import numpy as np
from gguf import GGMLQuantizationType as Q
from gguf import quants
TYPES = {"Q4_0": Q.Q4_0, "Q8_0": Q.Q8_0, "Q4_K": Q.Q4_K, "Q5_K": Q.Q5_K, "Q6_K": Q.Q6_K,
"F16": Q.F16, "BF16": Q.BF16}
BLOCK = {"Q4_0": (32, 18), "Q8_0": (32, 34), "Q4_K": (256, 144), "Q5_K": (256, 176), "Q6_K": (256, 210),
"F16": (1, 2), "BF16": (1, 2)}
# Where each block keeps its fp16 scales: random bytes there would often be NaN or huge.
HALVES = {"Q4_0": [0], "Q8_0": [0], "Q4_K": [0, 2], "Q5_K": [0, 2], "Q6_K": [208]}
def blocks(name, count, rng):
values, size = BLOCK[name]
raw = rng.integers(0, 256, size=(count, size), dtype=np.uint8)
if name in ("F16", "BF16"):
f = rng.standard_normal(count * 64).astype(np.float32) * 0.05
if name == "F16":
return f.astype(np.float16).view(np.uint8).reshape(-1)
return (f.view(np.uint32) >> 16).astype(np.uint16).view(np.uint8).reshape(-1)
scales = [0.0, -0.0, 1.0, -0.5, 6.1e-5, 3e-8, 0.0123, -0.002, 65504.0]
for b in range(count):
for at in HALVES[name]:
s = scales[b] if b < len(scales) else rng.standard_normal() * 0.01
raw[b, at:at + 2] = np.frombuffer(np.float16(s).tobytes(), dtype=np.uint8)
# Extreme quant bits in the first blocks: all zero, all ones.
if count > 2:
raw[0, max(HALVES[name]) + 2:] = 0
raw[1, max(HALVES[name]) + 2:] = 255
return raw.reshape(-1)
def main(path):
rng = np.random.default_rng(20261025)
out = {}
for name, qtype in TYPES.items():
count = 12 if BLOCK[name][0] > 1 else 1
data = blocks(name, count, rng)
values = quants.dequantize(data, qtype).astype(np.float32).reshape(-1)
out[name] = {"bytes": data.tobytes().hex(), "values": [float(v) for v in values]}
with open(path, "w") as f:
json.dump(out, f)
if __name__ == "__main__":
main(sys.argv[1])