Did our run establish a usable installation?
No. Our run on October 1, 2026, at commit bfa4507 recorded zero exit codes for the install and build steps, alongside installation errors in the captured output. The installation log records build-backend failures caused by missing torch. A successful wrapper command is insufficient evidence that the intended package installed.
Our harness stores step status, exit code, elapsed time, parsed test counts and captured log tails separately. Those fields must be read together here: the install status says success, while the tail shows failed editable and wheel build-requirement discovery. The build status also says success, but its tail shows Python warnings rather than native compilation evidence. The stored record does not explain how those command outcomes arose.
The MrKeyoor testing methodology explains these kinds of records and their limits. That page provides provenance context; its later description does not establish which procedure version produced this October 1 run. This correction uses the existing record and includes no new execution.
What stopped test collection?
The stored parser summary records 0 passed, 0 failed, 0 skipped and 11 errors. It does not establish 11 executed tests or exhaustive coverage of DeepSelect. The raw footer ends with 11 collection errors in 0.70 seconds; the harness records 6 seconds for the overall test step.
The captured tail identifies several distinct problems in bundled CUTLASS material: a missing torch import in CuTeDSL collection, missing pycute imports in Python tests, and a sharding test that cannot import running_on_ci from pytest's assertion utilities. The evidence does not support assigning every collection error to one missing dependency.
These collection errors provide no CUDA or Ascend correctness result. Our sandbox record does not establish kernel execution, workload throughput, multi-device behavior or deployment readiness.
Does the kernel contract fit your workload?
DeepSelect supports NVIDIA CUDA and Huawei Ascend. The README at the recorded commit describes specialized TopK paths for sparse attention and sampling, with topk capped at 4096 and tuning focused on 512. Float32 input is supported only on CUDA; Ascend supports bfloat16 input and int32 output indices.
Check layout before integrating it. Input rows require byte-aligned strides, and the last dimension must be contiguous. Returned tensors may be non-contiguous. Caller-owned index buffers also have alignment requirements. The interface marks begin and hint as unsupported, despite their presence in the signature.
The checked-in CUDA build targets sm_100a and sm_103a, skips SM80 and SM90 in its target list, and requires NVCC 12.9 or newer. The Ascend build imports torch_npu and requires an Ascend toolkit installation. These source requirements are separate from the harness's recorded build status.
NaN behavior also needs an integration check. The README says the default abort_when_nan_found=True invokes trap() and aborts the kernel. CUDA rows whose length is at most topk skip the NaN check. Validate the behavior your application needs before routing model outputs through it.
The README's speedup figures are upstream benchmark claims. Our run did not reproduce them, and installation timings do not measure TopK performance.
Which alternatives answer a different need?
PyTorch provides tensors that can live on the CPU or GPU. Keep its built-in torch.topk as your comparison baseline when evaluating whether a specialized extension is worth integrating. PyTorch is also the framework DeepSelect extends, so retaining PyTorch and adopting DeepSelect are compatible choices.
CUTLASS provides reusable components for building custom CUDA kernels. It is a development toolkit to consider when you intend to implement or modify kernels, rather than a demonstrated replacement for DeepSelect's Python TopK interface. DeepSelect's source build itself includes CUTLASS headers.
How to read the ratings
The assessed dimensions use a five-point rubric: 1 means a basic requirement is blocked or undocumented; 2 means substantial qualification remains; 3 means the dimension is adequate within a documented scope; 4 means detailed evidence supports it with stated limits; 5 requires demonstrated coverage across the intended use. Each rating applies only to its named dimension. Unknown evidence receives no numerical rating.
Setup ease is 2 because missing imports blocked installation attempts and test collection despite the stored success statuses. Documentation is 4 because the supplied material explains installation, supported cases, layout requirements and NaN behavior, although documentation cannot establish successful execution. Maturity is 2 because the narrow hardware and input contract still requires qualification that our run did not provide. Community is unassessed: the supplied primary extracts do not establish contributor activity, response times or support quality.
What to do before adoption
First check the documented dtype, topk, stride and compiler requirements against your actual workload and hardware. Then establish that the native extension installs and imports in that environment, run correctness checks against your baseline, and measure the workload you intend to deploy. Include short rows, NaNs and caller-owned output buffers if your application uses them. This stored sandbox run cannot substitute for those checks.

