tt-toplike
official
A vibrant htop-style visualizer for Tenstorrent hardware written in Rust. Real-time process and utilization view for TT accelerators.
Releases
🐧
tt-toplike-app_0.13.4_amd64_jammy.deb
🐧
tt-toplike-app_0.13.4_amd64_noble.deb
📦
tt-toplike-tui-0.13.4-macos-universal.tar.gz
⬇
tt-toplike-tui-0.13.4-windows-x86_64.zip
+2 more
⚙ Requires PPA — setup instructions ↗
⚙ Requires PPA — setup instructions ↗
4 previous releases
📋 Changelog
# Changelog
The **canonical, complete release log lives in [`debian/changelog`](debian/changelog)** —
that's the file the `.deb` packages are built from and where every release is
recorded in full. This file is a friendly pointer plus a summary of the most
recent releases; it deliberately does not duplicate the whole history.
To see everything:
```bash
less debian/changelog # full history
git tag # released versions
```
## Recent releases
### 0.11.0
- **Add**: Training view (`t`) — a full-screen visualization of a live
tt-train run, drawn as the network it is: a character grid of transformer
blocks and attention heads fed by token particles, a loss "mountain range"
under a twinkling aurora nightscape that opens up as the model converges,
plus chip telemetry alongside it.
- **Auto-attaches with no command**: scans running processes for a tt-train
example binary (`nano_gpt`, `mnist_mlp`, `linear_regression`), then resolves
`/proc/<pid>/fd/1` to find and tail that process's log. Checkpoint saves
are detected by mtime on the run's rolling checkpoint file.
- **Honest limitation**: tt-train's per-step stream can only be tailed if its
stdout was redirected to a real file at launch (`> train.log`). If fd 1 is
a pipe or tty, retroactively reading it is an OS-level impossibility, not a
gap in this tool — the view says so and falls back to what it can still
see (process liveness, chip telemetry, and checkpoint mtime) rather than
drawing a fake loss curve. Gradient norms, MFU, and throughput counters
aren't emitted live either, so the view derives tokens/sec and ETA only
from what it can actually read.
- Nine independent color channels (loss magnitude, run-history timeline,
loss-delta direction, forward/backward sweep cadence, cache compile/steady
state, checkpoint bursts, plus chip temp/power) — see the new legend and
explain overlays.
### 0.10.3
- **Security fix**: a direct-vLLM host process's own `PATH` was forwarded
onto the `sh` the monitor spawns to probe it — a bare `Command::new("sh")`
lookup resolves via whatever `PATH` ends up on the builder, so an
attacker-controlled process (gated only by `MESH_DEVICE`/`TT_METAL_HOME`,
which the process itself controls) could have redirected the root-run
monitor into executing a planted binary. Fixed by invoking the interpreter
via an absolute path and never forwarding the target's own `PATH`.
- **Fix**: four GDDR-telemetry correctness gaps — a stale 8-channel cap
dropping real fault data for Blackhole's channels 8–11, Starfield's DDR
planets missing an `enabled` check present in every other memory
visualization, the Insights GDDR temp/ECC rows silently disappearing
instead of falling back on partially-usable real data, and a missing ECC
counter parsing as a genuine zero instead of "not reported".
- **Perf**: a host-keyed service's per-tick probe redundantly re-read
`/proc/<pid>/environ` up to four times and re-walked
Works on
wormhole
blackhole