streamlit-ncu-rep-viewer 0.1.1

Unofficial Streamlit viewer for NVIDIA Nsight Compute (.ncu-rep) profiler reports (not affiliated with NVIDIA)

streamlit-ncu-rep-viewer

A Streamlit viewer that reproduces NVIDIA Nsight Compute's kernel-profiling visualizations from .ncu-rep report files — built for embedding in a Python Panel on comet.com.

Unofficial project. This is an independent, from-scratch viewer. It is not affiliated with, endorsed by, or distributed by NVIDIA. "Nsight" and "Nsight Compute" are trademarks of NVIDIA Corporation. See Third-party licenses for the ncu-report dependency.

Install

pip install streamlit-ncu-rep-viewer

This pulls in Streamlit, Plotly, pandas, and NVIDIA's standalone ncu-report report-reader wheel (no full CUDA / Nsight Compute toolkit required).

Usage

streamlit-ncu-rep-viewer                 # scan the current directory for *.ncu-rep
streamlit-ncu-rep-viewer path/to/dir     # scan a directory
streamlit-ncu-rep-viewer report.ncu-rep  # open a single report
streamlit-ncu-rep-viewer --samples       # open the bundled sample reports

Any extra arguments are forwarded to streamlit run, e.g.:

streamlit-ncu-rep-viewer --samples --server.port 8600 --server.headless true

Pick a report and kernel in the sidebar. Three tabs: Summary (all kernels), Details (per-kernel sections), Raw Metrics (searchable).

How it works

.ncu-rep  ──►  streamlit_ncu_rep_viewer.extract  ──►  JSON  ──►  app.py (Streamlit UI)
              (uses the ncu_report wheel)          (cached)   (pure Python + plotly)

• streamlit_ncu_rep_viewer/extract.py — dumps every metric + profiler rule from a .ncu-rep to JSON. Runnable directly: python -m streamlit_ncu_rep_viewer.extract report.ncu-rep.
• streamlit_ncu_rep_viewer/app.py — the viewer. Shells out to the extractor (cached) so report parsing is decoupled from rendering.
• streamlit_ncu_rep_viewer/cli.py — the streamlit-ncu-rep-viewer console entry point.
• streamlit_ncu_rep_viewer/sample_reports/ — bundled sample .ncu-rep files.

Details sections (per kernel)

• NVTX Ranges — the NVTX ranges active on the call stack when the kernel launched (shown only when the report contains NVTX data), so kernels map back to your code's phases. Also surfaced as a breadcrumb under the kernel header and an NVTX column in the Summary tab.
• GPU Speed Of Light Throughput — headline Compute (SM) vs Memory %, duration, elapsed cycles, SM/DRAM frequency, and a compute-vs-memory-bound verdict.
• Compute Workload Analysis — pipeline utilization (ALU/FMA/LSU/…), busiest pipe highlighted.
• Memory Workload Analysis — L1/TEX, L2, DRAM throughput + top memory-unit contributors (from the report's own breakdown definitions).
• Scheduler / Warp State Statistics — issue active, active warps/cycle, occupancy, and PC-sampled warp stall reasons.
• Instruction Statistics — executed IPC and per-pipe instruction mix.
• Launch Statistics, Occupancy (with occupancy-limiter chart), Analysis & Recommendations.

Development

pip install -e .
streamlit run streamlit_ncu_rep_viewer/app.py   # or: streamlit-ncu-rep-viewer --samples

Not implemented (data not present in the sample reports)

• Roofline chart — needs FLOP-count metrics (--set roofline / --set full profiling).
• Source page (SASS/PTX per-line counters) — needs source metrics enabled at profile time.

Scope is intentionally single-file/single-directory for now (no cross-report baselines).

License

MIT — see LICENSE. (Author/copyright placeholder; edit pyproject.toml and LICENSE before publishing.)

Third-party licenses

This package's own code is MIT-licensed, but it depends on ncu-report, which is proprietary NVIDIA software — not open source. It is distributed under the NVIDIA Software License Agreement (classifier License :: Other/Proprietary License; see the LICENSE.rst bundled inside the installed ncu_report package, and https://docs.nvidia.com/nsight-compute/CopyrightAndLicenses/index.html).

Notable terms of that agreement: it is non-transferable and non-sublicensable, it is intended for developing applications for systems with NVIDIA GPUs, and it restricts redistribution of the software as well as the publication of benchmark/competitive comparisons against non-NVIDIA platforms without NVIDIA's prior written permission.

streamlit-ncu-rep-viewer does not bundle or redistribute ncu-report; it is declared as an ordinary dependency and installed by pip from NVIDIA's own distribution, under NVIDIA's terms. Review that agreement before use or redistribution.