Summary

The prebuilt libbitsandbytes_xpu.so shipped in the wheel is linked against
libsycl.so.8 (oneAPI 2025.x). On oneAPI 2026 — which ships only
libsycl.so.9, and which the current XPU PyTorch nightlies are built against —
the native library fails to dlopen. bitsandbytes then silently falls back to
the mock library, so any native-only call (e.g. lib.cget_managed_ptr, used by
paged optimizers) raises RuntimeError.

Environment

• bitsandbytes: 0.50.0.dev0 (prebuilt XPU wheel)
• PyTorch: 2.13.0.dev20260526+xpu
• oneAPI: 2026.0 (provides libsycl.so.9 only)
• GPU: Intel Data Center GPU Max 1550 (PVC)
• OS: Ubuntu 24.04, Python 3.12

Root cause

The wheel's XPU binary has a hard NEEDED libsycl.so.8:

$ ldd /opt/venv/lib/python3.12/site-packages/bitsandbytes/libbitsandbytes_xpu.so | grep sycl
        libsycl.so.8 => not found

But the environment (oneAPI 2026 + torch 2.13+xpu) only provides libsycl.so.9:

$ find /opt/intel /opt/venv -name 'libsycl.so*'
/opt/intel/oneapi/2026.0/lib/libsycl.so.9
/opt/venv/lib/libsycl.so.9
...

$ ldd .../torch/lib/libtorch_xpu.so | grep sycl
        libsycl.so.9 => .../libsycl.so.9

Reproduction

repro_bnb_xpu_sycl.py:

import bitsandbytes as bnb
import bitsandbytes.cextension as ce

print(f"bitsandbytes : {bnb.__version__} ({bnb.__file__})")
print(f"backend      : {ce.BNB_BACKEND}")
print(f"loaded lib   : {type(ce.lib).__name__}")

# bitsandbytes catches the real load error for XPU/CPU and replaces `lib`
# with a mock (ErrorHandlerMockBNBNativeLibrary). Re-run its own loader to
# surface the actual reason the native library could not be loaded.
print("\n--- bitsandbytes.cextension.get_native_library() ---")
try:
    ce.get_native_library()
    print("native library loaded OK")
except Exception as e:
    print(f"{type(e).__name__}: {e}")

Output:

$ python repro_bnb_xpu_sycl.py
bitsandbytes : 0.50.0.dev0 (/opt/venv/lib/python3.12/site-packages/bitsandbytes/__init__.py)
backend      : XPU
loaded lib   : ErrorHandlerMockBNBNativeLibrary

--- bitsandbytes.cextension.get_native_library() ---
OSError: libsycl.so.8: cannot open shared object file: No such file or directory

Impact

Because the native library never loads, lib is the mock, and native-only
functionality fails. For example, running a paged optimizer test:

RuntimeError: Attempted to use bitsandbytes native library functionality but it's not available.
...
Native code method attempted to call: lib.cget_managed_ptr()

Request

Please rebuild/ship the XPU wheel against SYCL 9 / oneAPI 2026, so
libbitsandbytes_xpu.so loads with current XPU PyTorch builds. Building from
source with the 2026 toolchain produces a .so linked to libsycl.so.9 and
resolves the issue locally, but the published wheel is still SYCL 8.