vramtop

The htop for GPU memory. Beautiful. Zero-friction. NVIDIA-first.

A terminal UI for monitoring GPU memory usage across all your NVIDIA GPUs. See per-process VRAM consumption, memory phase detection, OOM prediction, and framework-aware enrichment -- all without instrumenting your code.

Install

pip install vramtop

Requires Python 3.10+ and NVIDIA drivers (make sure nvidia-smi works).

Optional extras

pip install vramtop[pelt]    # PELT changepoint detection (requires ruptures + numpy)

Quick start

vramtop

That's it. vramtop discovers all NVIDIA GPUs and shows live memory usage.

Features

• Per-process VRAM tracking -- see exactly which processes use how much GPU memory
• Memory phase detection -- classifies each process as stable, growing, shrinking, or volatile
• OOM prediction -- warns before your GPU runs out of memory, with time-to-OOM estimates
• Survival verdicts -- per-process OK / TIGHT / OOM risk assessment with spike detection
• Framework detection -- automatically identifies PyTorch, JAX, vLLM, Ollama, SGLang, llama.cpp, TGI
• Model file scanning -- detects which model files are loaded from /proc/fd
• Container awareness -- shows Docker/Podman container IDs per process
• PELT changepoint analysis -- optional segment-level memory phase analysis with labeled segments
• Deep mode -- get PyTorch-internal memory stats (allocated vs reserved vs active) via IPC
• HTTP scraping -- pulls KV cache usage from vLLM, Ollama, SGLang, llama.cpp metrics endpoints
• 6 themes -- dark (GitHub-dark), light, nord, catppuccin, dracula, solarized
• CSV export -- log GPU and process data to CSV for offline analysis
• SVG screenshots -- save terminal screenshots as SVG
• Accessibility -- respects NO_COLOR, --accessible mode with text labels

Usage

Monitor (default)

vramtop                          # Launch the TUI
vramtop --refresh-rate 2         # Poll every 2 seconds (default: 1)
vramtop --no-kill                # Disable process kill functionality
vramtop --accessible             # Text labels instead of Unicode symbols
vramtop --export-csv gpu.csv     # Log data to CSV while monitoring
vramtop --config path/to/config.toml

Wrap (deep mode)

Run any command with deep mode reporting enabled:

vramtop wrap -- python train.py

This sets VRAMTOP_REPORT=1 in the subprocess environment. If your training script calls vramtop.report(), it will send PyTorch memory internals to vramtop over a Unix socket.

Deep mode (in your code)

Add two lines to your PyTorch script to report allocator-level memory stats:

import vramtop
vramtop.report()

# ... your training code ...

vramtop's TUI will automatically discover the reporter socket and show allocated, reserved, and active memory breakdowns in the detail panel.

Keyboard shortcuts

Key	Action
q	Quit
d	Open detail panel for selected process
k	Kill selected process (SIGTERM, then SIGKILL)
t	Cycle through themes
s	Save SVG screenshot
?	Show help
Up/Down	Navigate process list

Configuration

Config file location: ~/.config/vramtop/config.toml

Reload without restarting: send SIGHUP (kill -HUP $(pgrep vramtop)).

[general]
refresh_rate = 1.0       # Poll interval in seconds (0.5-10)
history_length = 300     # Samples to keep for sparkline/analysis

[display]
theme = "dark"           # dark, light, nord, catppuccin, dracula, solarized
layout = "auto"          # auto, full, compact, mini
show_other_users = false
enable_kill = true
phase_indicator = true

[alerts]
oom_warning_seconds = 300
temp_warning_celsius = 85
oom_min_confidence = 0.3

[scraping]
enable = true
rate_limit_seconds = 5
timeout_ms = 500
allowed_ports = [8000, 8080, 11434, 3000]

[oom_prediction]
algorithm = "variance"   # variance or pelt
min_sustained_samples = 10
min_rate_mb_per_sec = 1.0

Layout modes

vramtop auto-detects terminal size and picks the best layout:

Mode	Min size	Shows
Full	120x30	GPU cards, memory bars, sparklines, process tables, phase badges
Compact	80x20	GPU cards, memory bars, process tables
Mini	40x10	Minimal GPU summary

Override with layout = "full" in config or let "auto" handle it.

How it works

vramtop reads GPU data through NVML (nvidia-ml-py) -- the same library that powers nvidia-smi. No root access needed, no kernel modules, no LD_PRELOAD hacks.

Data sources by tier:

1. NVML (always) -- total/used/free VRAM, per-process memory, temperature, power, clocks
2. /proc enrichment (always) -- framework detection from cmdline/maps, model files from fd, container from cgroup
3. HTTP scraping (opt-in) -- KV cache metrics from inference server endpoints (localhost only, port-owner verified)
4. Deep mode IPC (opt-in) -- PyTorch allocator stats via Unix domain socket

Security model:

• All /proc reads are same-UID only
• HTTP scraping enforces: localhost-only, port-owner verification, rate limiting, no redirects, 64KB size cap
• Kill requires confirmation (SIGTERM first, SIGKILL only after 5s timeout)
• All external strings sanitized (ANSI/control char stripping, 256-char truncation)

Requirements

• Python 3.10+
• NVIDIA GPU with drivers installed
• Linux (uses /proc for enrichment)

License

MIT