Diagnostics

Inspect your installed runtime with:

tensorcx-doctor
tensorcx-doctor --json
tensorcx-doctor --json --check-tools

The standalone command reports installation problems even when the native runtime cannot be imported. Use python -m tensorcx_doctor --json if the console script is outside your PATH. Reports stay local; nothing is uploaded.

For a working runtime:

import tensorcx as cx

report = cx.diagnostics()
operations = cx.capabilities("cpu")
memory = cx.memory_info("cpu")

The report separates backends compiled into your package from devices currently available. Capability lists describe accepted data input dtypes; shapes, masks, embedding indices and other arguments retain their operation-specific rules. They are declared support, rather than an execution test.

CPU memory describes system physical memory, not your process or container’s allocation limit. CUDA memory describes device 0’s instantaneous free/total memory, including the effects of other users of that GPU. Missing measurements are null; Metal memory reporting is not implemented yet.

Cached CUDA pool blocks still occupy device memory. pool.stats() reports the payload owned by that pool, rather than process-wide or total GPU memory. Use host profiling and memory reuse to examine allocation, copy and wait costs, or try the runtime tools example. Native Metal profiling and device-event timings remain open.

LLVM checks do not execute tools by default. --check-tools runs bounded version checks for the optional LLVM 21.1.8 toolchain. Ordinary tensor operations do not need LLVM, and compiler readiness does not establish support for every kernel.

See the detailed contract and limitations.