GPU

Install linescope[scalene] on Python 3.11–3.14, then enable GPU collection:

from linescope import profile

with profile(backend="scalene", gpu=True):
    train_model()

The CLI equivalent is linescope --backend scalene --gpu train.py. Cell magic accepts the same flags. See the GPU notebook for a CUDA workload.

Synchronize queued device work inside the session to include it in the measurement period. gpu=True collects metrics; it does not select a device for your calculations.

Collection support

Scalene supports NVIDIA and AWS Neuron metrics. Apple MPS metrics are unavailable. Trace warns and continues without GPU data; Tachyon rejects GPU requests.

NVIDIA utilization may include other processes when per-process accounting is unavailable. Device memory refers to the profiled process, but Windows WDDM drivers do not expose it through NVML. GPU time can still be collected.

Read the GPU view

The GPU view and source tables show estimated device time and sampled peak memory beside driver time. Collection failures appear as diagnostics.

GPU and CPU work can overlap. Do not add GPU time to driver wall time. Device memory is separate from process RAM and Spark metrics. Sampled peaks can miss brief spikes.

Measurement model

GPUStats stores line-level device measurements. A custom backend supplies them through RawLine.gpu with BackendCapabilities.gpu=True.