Commit b069e526d for llama.cpp
commit b069e526deabf7207906ddfae1eb61551b2f612c
Author: Jai Patel <61297975+jpatel2005@users.noreply.github.com>
Date: Wed Nov 18 09:28:47 2026 -0500
gguf-py : fix tensor offset overflow in GGUFReader (#32495)
* gguf-py : fix tensor offset overflow in GGUFReader
* gguf-py : cast alignment padding to int
diff --git a/gguf-py/gguf/gguf_reader.py b/gguf-py/gguf/gguf_reader.py
index bf3e08380..d8b07ef01 100644
--- a/gguf-py/gguf/gguf_reader.py
+++ b/gguf-py/gguf/gguf_reader.py
@@ -189,7 +189,7 @@ class GGUFReader:
raise ValueError('Invalid alignment: must be a non-zero power of two')
padding = offs % self.alignment
if padding != 0:
- offs += self.alignment - padding
+ offs += int(self.alignment - padding)
self.data_offset = offs
self._build_tensors(offs, tensors_fields)
@@ -350,7 +350,9 @@ class GGUFReader:
np_dims = tuple(reversed(dims.tolist()))
block_size, type_size = GGML_QUANT_SIZES[ggml_type]
n_bytes = n_elems * type_size // block_size
- data_offs = int(start_offs + offset_tensor[0])
+ data_offs = start_offs + int(offset_tensor[0])
+ if data_offs + n_bytes > self.data.nbytes:
+ raise ValueError(f'Tensor {tensor_name} data exceeds file size {self.data.nbytes}')
item_type: npt.DTypeLike
if ggml_type == GGMLQuantizationType.F16:
item_count = n_elems
diff --git a/gguf-py/tests/test_gguf_reader_validation.py b/gguf-py/tests/test_gguf_reader_validation.py
index 98f30a969..e6564ed7d 100644
--- a/gguf-py/tests/test_gguf_reader_validation.py
+++ b/gguf-py/tests/test_gguf_reader_validation.py
@@ -5,7 +5,7 @@ import pytest
from gguf.gguf_reader import GGUFReader
-def _write_gguf(path, n_dims_field, dims):
+def _write_gguf(path, n_dims_field, dims, offset=0):
buf = b'GGUF' + struct.pack('<IQQ', 3, 1, 0) # version 3, 1 tensor, 0 kv
name = b'bad_tensor'
buf += struct.pack('<Q', len(name)) + name
@@ -13,7 +13,7 @@ def _write_gguf(path, n_dims_field, dims):
for d in dims:
buf += struct.pack('<Q', d)
buf += struct.pack('<I', 0) # dtype F32
- buf += struct.pack('<Q', 0) # tensor offset
+ buf += struct.pack('<Q', offset) # tensor offset
buf += b'\x00' * 64
path.write_bytes(buf)
@@ -35,3 +35,11 @@ def test_dims_product_no_uint64_wraparound(tmp_path):
_write_gguf(p, len(dims), dims)
with pytest.raises(ValueError):
GGUFReader(p)
+
+
+def test_tensor_offset_no_uint64_wraparound(tmp_path):
+ # offset 2**64 - 1 must be rejected, not wrap around
+ p = tmp_path / 'evil_offset.gguf'
+ _write_gguf(p, 1, [4], offset=2**64 - 1)
+ with pytest.raises(ValueError, match='exceeds file size'):
+ GGUFReader(p)