Status of testing Apache Mahout Qumat 0.6.0rc2

[ryankert01]

Testing Apache Mahout Qumat 0.6.0rc2

RC2 supersedes RC1 (#1315) with four post-RC1 fixes. The full 130-PR changelog since 0.5.0 is unchanged — see #1315 for that list. Please re-test the four fixes below and re-validate anything you flagged on RC1.

Requirements

• Python 3.10, 3.11, or 3.12
• For QDP extension: Linux x86_64 with NVIDIA GPU and CUDA

Installation

# Core package
pip install --pre "qumat==0.6.0rc2"

# With QDP extension
pip install --pre "qumat[qdp]==0.6.0rc2"

Links

• PyPI qumat: https://pypi.org/project/qumat/0.6.0rc2/
• PyPI qumat-qdp: https://pypi.org/project/qumat-qdp/0.2.0rc2/
• Git tag: https://github.com/apache/mahout/releases/tag/mahout-qumat-0.6.0-RC2
• Milestone: https://github.com/apache/mahout/milestone/2
• Previous testing issue (RC1, closed): Status of testing Apache Mahout Qumat 0.6.0rc1 #1315

Fixes added in RC2 (please verify)

• fix(qdp): reject invalid non-amplitude benchmark combos (#1303)
• fix(qdp): fall back to CPU when GPU arch not in PyTorch's compiled list (#1323)
• fix(build): gate test_rust on nvcc presence, not just nvidia-smi (#1322)
• fix(qdp): guard batch size multiplication against overflow (#1324)

General regression check

• pip install --pre "qumat==0.6.0rc2" works on your platform
• If on Linux + NVIDIA GPU: pip install --pre "qumat[qdp]==0.6.0rc2" installs cleanly
• Anything you contributed to 0.6.0 still behaves correctly under RC2
• No regressions vs RC1 in your workflow

If nothing critical surfaces within 3–5 days, we'll move to the formal [VOTE] thread.

[andrewmusselman]

+1 (binding)

Tested RC2 on Linux + CUDA Toolkit 12.4 + GTX 1060 (Pascal, sm_61, same
setup as RC1) with CXX=g++-13 QDP_CUDA_ARCH_LIST=61+PTX.

PyPI installs:

• pip install --pre "qumat==0.6.0rc2": clean
• pip install --pre "qumat[qdp]==0.6.0rc2": clean (binary wheel)

Source build from mahout-qumat-0.6.0-RC2:

• maturin develop, cargo build, make test_rust: clean
• uv run pytest (CUDA visible): 785 pass, 80 fail
• uv run pytest (CUDA_VISIBLE_DEVICES=): 719 pass, 12 fail, 142 skipped

RC1 follow-ups verified:

• Makefile: detect the CUDA Toolkit, not just the driver, for test_rust #1322 (Makefile nvcc gate): test_rust runs cleanly on systems with the
  toolkit; the new HAS_NVCC gate fires as designed.
• qumat_qdp: intersect GPU compute capability with PyTorch's arch list #1323 (PyTorch arch fallback): the new UserWarning fires from both
  loader.py and api.py call sites on sm_61, and the 13 tests that route
  through _select_torch_device fall back to CPU and pass.

Three non-blocking follow-ups for 0.6.1:

1. PyTorch 2.9+cu128 wheel has a libcudart load-order quirk on systems
   without CUDA Toolkit 12.5+ in PATH: cudaGetDriverEntryPointByVersion
   is undefined at import. Workaround:

   LD_PRELOAD=$VIRTUAL_ENV/lib/python3.12/site-packages/nvidia/cuda_runtime/lib/libcudart.so.12

   Worth documenting in DEVELOPMENT.md or pinning the PyTorch wheel.

2. ~12 tests in testing/qdp/test_benchmark_api.py, test_bindings.py, and
   testing/qdp_python/test_quantum_data_loader.py instantiate QdpEngine(0)
   without an @skipif(no_gpu) guard. They error rather than skip when
   CUDA_VISIBLE_DEVICES= is set. Similarly, ~68 tests in test_bindings.py
   hardcode device="cuda" and fail on GPUs whose arch isn't in PyTorch's
   compiled list, despite qumat_qdp: intersect GPU compute capability with PyTorch's arch list #1323's fallback now being available for
   production code paths.

3. test_encode_from_gpu_ptr_f32_with_stream_non_default_success
   (qdp-core/tests/gpu_ptr_encoding.rs:795) fails during full
   make test_rust runs but passes 5/5 in isolation. Looks like
   test-isolation / GPU-state-pollution rather than a real defect in
   the f32+non-default-stream code path.

None of the above blocks the release; the shipped library behaves correctly
on supported hardware (Volta+) and all production code paths I exercised
are clean.