Install

Not released

The measurement tool isn't released yet. The 0.0.1 packages on PyPI and crates.io hold the name and don't measure anything.

Ways to install, once released

Python, without installing	`uvx diescope run -- python train.py`
Python	`pip install diescope`
Rust	`cargo install diescope`
Linux binary	A static binary on GitHub Releases
macOS	`brew install diescope/tap/diescope`

One binary. Vendor libraries load at runtime, so the same Linux binary runs on machines with or without CUDA installed.

What it needs on NVIDIA

Linux.
The NVIDIA driver. Diescope loads its CUDA and NVML libraries at runtime.
CUPTI, from the CUDA toolkit. Diescope looks for libcupti.so on the library path and in the usual CUDA install folders.
Permission to read GPU performance counters. NVIDIA restricts them to root by default. DIE-Util needs them; GPU-Util doesn't.

Counter permission

Either run Diescope with sudo, or let normal users read the counters. To do that, add this line to a file in /etc/modprobe.d/:

/etc/modprobe.d/nvidia-profiling.conf

options nvidia NVreg_RestrictProfilingToAdminUsers=0

Then rebuild your initramfs if the nvidia module loads from it, and reboot. Check it took effect:

grep RmProfilingAdminOnly /proc/driver/nvidia/params

Expect RmProfilingAdminOnly: 0. NVIDIA's own guide: developer.nvidia.com/ERR_NVGPUCTRPERM.

Tested on

One GPU so far: an RTX 5070 Ti Laptop GPU (Blackwell). Other NVIDIA GPUs are untested.

Other chips

AMD GPUs are in testing. Apple silicon, Intel GPUs, Qualcomm chips, and CPUs are planned.

What it sends

Nothing. No account and no telemetry. Diescope makes no outbound network calls. Reports are files on your machine. The always-on agent can serve Prometheus metrics, only on an address you choose.