Many entrances, one session · nsys-ai 0.3.0

nsys-ai

Open Nsight Systems profiles in the browser, terminal, or MCP and keep one evidence-first diagnosis moving through a shared session.

GitHub PyPI Python MIT
nsys-ai timeline viewer with the guided optimization workflow open
Carry one profile from diagnose through propose, re-profile, diff, and decide.

Try It Now

Get started in 60 seconds with a real GPU training profile

Quick Start — Example 20: Megatron-LM DistCA
# 1. Install $ pip install nsys-ai # 2. Clone & download a real profile (~700 MB) $ git clone https://github.com/GindaChen/nsys-ai && cd nsys-ai $ cd examples/example-20-megatron-distca $ python download_data.py # 3. Open the web viewer (default command) $ nsys-ai output/megatron_distca.nsys-rep # 4. Check profile health before analysis $ nsys-ai doctor output/megatron_distca.nsys-rep # 5. Publish deterministic findings to a handoff session $ nsys-ai diagnose output/megatron_distca.nsys-rep \ --session /tmp/nsys-ai/demo

See all examples → examples/

One session, five steps

Run the same handoff from CLI, Web, TUI, or MCP: deterministic evidence stays in one session while each surface does the job it is best at.

  1. 01

    Diagnose

    Run the default analysis pack and publish findings.

  2. 02

    Propose

    Turn a finding into a concrete, inspectable change.

  3. 03

    Re-profile

    Run the same workload and record a new RunSpec.

  4. 04

    Diff

    Compare before and after with canonical metric direction.

  5. 05

    Decide

    Accept or reject the change and keep the reason in session.

WEB

Web Timeline

Default experience: multi-GPU browser viewer with progressive rendering, NVTX hierarchy bars, pinch-to-zoom, trackpad pan, and AI chat. No --trim required.

TUI

Timeline TUI

Perfetto-style horizontal timeline with per-stream kernel visualization, NVTX hierarchy bars, and time-cursor navigation. Zoom, pan, and bookmark positions.

TREE

Tree TUI

Interactive NVTX hierarchy browser. Expand/collapse nodes, see kernel durations, filter by name or threshold, and navigate with vim-style keybindings.

HTML

HTML Exports

Generate interactive HTML viewers, Chrome Trace Event JSON files for Perfetto, and flat CSV/JSON exports for scripting and sharing.

SEARCH

Smart Search

Search kernels and NVTX annotations by name, with hierarchical search support. Filter by GPU, time window, and parent NVTX scope.

OVERLAP

Overlap Analysis

Compute vs NCCL overlap, collective breakdowns by type, and automatic training iteration detection for distributed workloads.

AI

AI Analysis

Optional planner and synthesizer over registered evidence: ask a question, inspect the skills that grounded it, and resume the answer from a shared session.

SESSION

Session Handoff

Move findings, proposals, RunSpecs, diffs, and decisions between CLI, Web, TUI, and MCP with one inspectable directory.

Keyboard-Driven Navigation

Navigation

← →Pan through time
↑ ↓Select stream
TabSnap to next kernel
+ -Zoom in / out
aToggle time mode

Analysis

/Filter by name
mMin duration filter
dDemangled names
CConfig panel
hHelp overlay

Bookmarks

BSave bookmark
'Bookmark list
, .Cycle bookmarks
`Jump back
[ ]Range bookmark

Documentation

Comprehensive guides for Nsight Systems profiling

→What to hand nsys-ai →Known limits & trust →Migrating to 0.3.0 →Time windows →Reading a diff →Troubleshooting →Environment variables →Analysis skills →Choosing a Web viewer →User guide →doctor →Support matrix →CI diff gate devIngest policy devSkill contract devSurface adapters devArtifact layout devSkill memo identity devDiff findings devTesting & coverage devArchitecture & release guide devRelease process 01CLI Reference 02SQLite Schema 03NVTX Annotations 04CUDA Trace 05NCCL Tracing 06Python / PyTorch 07Container Profiling 08Focused Profiling

Interactive viewers

Choose the surface that matches the question: hierarchy, timeline, or before-and-after change.

nsys-ai NVTX tree viewer

NVTX tree

Browse nested ranges and the kernels attributed to each node.

 nsys-ai progressive multi-GPU timeline

Progressive timeline

Follow streams, kernels, NVTX lanes, search, and session-backed workflow state.

 nsys-ai before-and-after diff viewer

Pair comparison

Inspect a real before-and-after result with verdict, comparability, regressions, and improvements.

More tools

SQLite Schema Explorer

Interactive browser for the Nsight SQLite schema — tables, foreign keys, example queries, and cross-highlighted concepts. Open right in your browser.

NVTX Stack Visualizations

Pre-built interactive HTML exports showing NVTX call stacks — full trace and single-iteration views with expandable nodes.