Skip to content

CuPy device qualification

CuPy is an optional runtime array namespace. Normal compiled evaluators discover it from CuPy state arrays; there is no new compiler backend flag or provider setting. The compiler's required CPU NumPy/JAX/Torch matrix and the optional Numba CPU lowering remain separate contracts.

Install one CUDA family explicitly

From this checkout, uv sync --dev --extra cupy installs cupy-cuda12x>=14.2,<15. This extra deliberately selects the CUDA 12 wheel family. It leaves CUDA runtime/toolkit installation to the environment, keeping those large components optional. If the system does not supply them, explicitly install cupy-cuda12x[ctk] in the same environment. Use one CuPy distribution per environment; the CUDA 13 wheel family is a separate installation and qualification choice. See the official CuPy installation guide.

A GPU and compatible NVIDIA driver are required for execution. Package import or successful dependency installation alone does not qualify a device. Runtime state and floating parameter inputs in these tests are native CuPy arrays with one declared dtype. Use cupy.asarray to construct them; CuPy scalar dtype aliases are NumPy scalar types, so a dtype constructor alone does not establish GPU storage. See CuPy's documented differences.

Required device command and optional CPU skips

just qualify-cupy
# In an already prepared environment, retain the executed-case report:
python scripts/qualify_backend.py cupy --junitxml cupy-results.xml

The required command checks CUDA initialization and a real device operation before pytest, then requires executed cases with zero skips. Missing CuPy, missing or broken CUDA, missing devices, failed tests and skipped cases are failures. It cannot report CPU fallback as qualified CuPy execution.

The baseline suite instead uses an explicitly optional device fixture. Without CuPy or hardware it reports ten skips, with reasons. Standard hosted jobs do not install this optional GPU namespace or require CUDA devices; their green status is not CuPy coverage. Run the required command on suitable hardware and retain its versions/device metadata, source/wheel hashes, logs and JUnit records.

Bounded contract

Ten cases exercise both float32 and float64. They reuse a compiled artifact for NumPy and CuPy calls and check native GPU result storage, dtype, flat ordering, structured shapes and input immutability. The hierarchical balance fixture has independently written equations, shaped shared/local parameters, aliases, stage transport/reductions and both independent and coupled groups. A strided GPU state also follows the same flat layout. Additional cases check direct constants and native scalar time, layout errors, and bounded symbolic state/parameter and observable derivative values.

Host transfer is explicit in numerical assertions via cupy.asnumpy, after native device/dtype checks. This checks compiler values and layouts on the recorded device/runtime. It does not qualify every compiler operator, independent block vectorization, sparse downstream solvers, multiple devices, CuPy automatic differentiation, statistical distributions, SBI training or CUDA 13/ROCm. It makes no GPU performance or memory-reduction claim.

Device evidence

The final, independently reviewed workstation run on 23 October 2026 used an NVIDIA GeForce RTX 5070 Ti Laptop GPU (Blackwell, compute capability 12.0), driver 580.173.02 and Python 3.12.12. Its exact compiler head was 271fd43d59874c3c910fe180252bb7108c3d52ee; the pinned engine and inference heads were 76edf594f5d4462f9c359482378e662e58668234 and 713a2b76119e174dab75290eaf5bff9ffce63af6. All sources were acquired from their GitHub remotes, checked out detached and clean, built as wheels and imported only from their isolated environments.

CuPy 14.2.0 passed all ten cases with zero skips or failures in both the current and minimum dependency lanes. Each CuPy lane and the separate JAX 0.12.0 and Torch 2.14.1+cu130 lanes passed all four float32/float64 and retained-history core cases. JAX and Torch each passed 14 observation density/replay cases, ten prior-density cases and the masked/censored gradient check. Torch additionally preserved local/global RNG isolation and rejected a CPU generator for CUDA sampling. The recorded validator confirmed every required case and all 40 recorded commands exited successfully.

The workstation was compute-idle rather than headless. Only the display server, desktop shell and active editor held graphics contexts; there was no competing compute context. The before/after snapshots recorded 442/432 MiB of 12,227 MiB framebuffer use and 6%/5% instantaneous utilization. These functional checks have no timing threshold. Whole-process peak RSS was 451.2/452.2 MiB for the two CuPy core lanes, 984.7/1231.2 MiB for JAX core/statistics and 733.0/1354.7 MiB for Torch core/statistics. Torch statistics peaked at 0.752 MiB allocated and 2.0 MiB reserved device memory. Those process and allocator measures have different scopes and are not application or training-memory results.

All 40 recorded commands exited successfully; all 17 hosted checks were green, and independent review approved the exact source and receipt before PR #339 merged and issue #182 closed. Earlier receipts at d26a86d and 0c1306e are superseded. A separate JAX 0.12 CUDA run qualified the held-table gradient fix at exact commit abdc5eebfe1196ee06f27db5ace3409a5f368f15 before PR #345 merged. Raw records remain private because they contain machine paths and process metadata. A repeated, nonfatal JAX hwloc topology warning is retained with them. CPU skips cannot satisfy this device gate, and the bounded receipts do not qualify performance, estimator training, compiled Torch GPU loops, Diffrax GPU execution, multiple devices, CUDA 13 CuPy wheels or consumer adoption.