[Bug] `GGML_ASSERT` crash in SafeTensors loader on inconsistent tensor metadata

[professor-moody]

Git commit

$git rev-parse HEAD
8afbeb6

Operating System & Version

Fedora 43

GGML backends

CPU, HIP

Command-line arguments used

sd -m crash-029.safetensors --mode txt2img

Steps to reproduce

Steps to reproduce

1. Create a minimal 91-byte SafeTensors file with mismatched shape/dtype vs data_offsets:

import struct, json

header = json.dumps({
    "weight": {"dtype": "F64", "shape": [4], "data_offsets": [0, 16]}
})
buf = struct.pack('<Q', len(header)) + header.encode() + b'\x00' * 16

with open('crash-029.safetensors', 'wb') as f:
    f.write(buf)

2. Run: sd -m crash-029.safetensors --mode txt2img

What you expected to happen

What you expected to happen

The process should reject the file with an error message when tensor metadata is internally inconsistent (shape/dtype imply a different byte count than data_offsets).

What actually happened

What actually happened

The process crashes with GGML_ASSERT → ggml_abort() → abort().

init_from_safetensors_file() (src/model.cpp:588-595) uses GGML_ASSERT to check tensor size consistency. When dtype is "F64" with shape [4], the expected size is 32 bytes (4 × 8), but data_offsets says 16. The assertion tensor_storage.nbytes() * 2 == tensor_data_size fails → unconditional abort (not affected by -DNDEBUG).

Logs / error messages / stack trace

Logs / error messages / stack trace

GGML_ASSERT(tensor_storage.nbytes() * 2 == tensor_data_size) failed
  at src/model.cpp:589 in init_from_safetensors_file()
Aborted (core dumped)

Additional context / environment details

Additional context / environment details

Root cause — src/model.cpp:588-595:

} else if (dtype == "F64") {
    tensor_storage.is_f64 = true;
    GGML_ASSERT(tensor_storage.nbytes() * 2 == tensor_data_size);  // line 589 — crashes here
} else {
    GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size);      // line 595
}

Suggested fix — replace GGML_ASSERT with error returns:

 } else if (dtype == "F64") {
     tensor_storage.is_f64 = true;
-    GGML_ASSERT(tensor_storage.nbytes() * 2 == tensor_data_size);
+    if (tensor_storage.nbytes() * 2 != tensor_data_size) {
+        LOG_ERROR("SafeTensors: size mismatch for tensor '%s' (F64)\n", name.c_str());
+        return false;
+    }
 } else {
-    GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size);
+    if (tensor_storage.nbytes() != tensor_data_size) {
+        LOG_ERROR("SafeTensors: size mismatch for tensor '%s'\n", name.c_str());
+        return false;
+    }
 }

• x86_64, clang 19
• Reproduces on both sanitizer and release builds
• Found via fuzz testing with crafted SafeTensors files