[Purfview]

When a file-like object is passed, CTranslate2 crashes with Windows error 0xC0000005 (access violation). [bytes object works]
Loading model files from memory (a file-like or bytes objects) was implemented here: #1030 [files arg documentation ]
I tested the fix on Windows with both file-like and bytes objects - works fine.

Code to reproduce a file-like object crash:

import os
import ctranslate2

model_dir = r"D:\faster-whisper-tiny"

files = {
    filename: open(os.path.join(model_dir, filename), "rb")
    for filename in os.listdir(model_dir)
    if os.path.isfile(os.path.join(model_dir, filename))
}

model = ctranslate2.models.Whisper("test-tiny", files=files)

print("No crash!", flush=True)

AI generated explanation:

Problem

When loading model files from file-like objects, create_model_reader() stores the dictionary value in auto content, which deduces to a non-owning py::handle.

When content = read() assigns the result of a file-like object's read() method, only the underlying Python object pointer is copied. The temporary owning py::object returned by read() is destroyed at the end of the statement, potentially leaving content pointing to a deallocated object.

For ordinary binary file handles, the returned bytes object is not otherwise retained. The subsequent conversion to std::string therefore accesses a dangling Python object, causing a use-after-free that can crash the process.

Fix

Store the dictionary value in an owning py::object, created with py::reinterpret_borrow<py::object>(). This ensures the contents returned by read() remain alive while being converted and registered with the in-memory model reader.

[jordimas]

Thanks so much!