Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
"""setup.check_shards: a metadata-only GGUF shard that ends at its header (no alignment padding, #2346) is whole;
a shard whose tensor data runs past its end is still refused."""
import struct
import sys
import tempfile
import unittest
from pathlib import Path
ROOT = Path(__file__).resolve().parent.parent
sys.path.insert(0, str(ROOT))
import setup # noqa: E402
def _str(s):
b = s.encode()
return struct.pack("<Q", len(b)) + b
def gguf(path, tensors=(), pad_data=0):
"""A GGUF v3 file with one string key and the given (name, n_floats) F32 tensors."""
kv = _str("general.name") + struct.pack("<I", 8) + _str("t")
tinfo = b""
off = 0
for name, n in tensors:
tinfo += _str(name) + struct.pack("<I", 1) + struct.pack("<Q", n) + struct.pack("<I", 0) + struct.pack("<Q", off)
off += ((n * 4 + 31) // 32) * 32
head = b"GGUF" + struct.pack("<IQQ", 3, len(tensors), 1) + kv + tinfo
body = head
if tensors:
body += b"\0" * ((32 - len(head) % 32) % 32)
body += b"\0" * pad_data
path.write_bytes(body)
return len(head)
class ShardCheck(unittest.TestCase):
def setUp(self):
self.d = Path(tempfile.mkdtemp())
def test_metadata_only_shard_without_padding_is_whole(self):
p = self.d / "m-00001-of-00006.gguf"
n = gguf(p)
self.assertNotEqual(n % 32, 0, "the test file must end unaligned, like the real shard")
setup.check_shards([p]) # must not stop setup
def test_truncated_tensor_data_is_still_refused(self):
p = self.d / "t-00002-of-00006.gguf"
gguf(p, tensors=[("w", 64)], pad_data=100) # needs 256 bytes of data, has 100
with self.assertRaises(SystemExit):
setup.check_shards([p])
if __name__ == "__main__":
unittest.main()