Skip to content
C-Kernel-Engine docs Antshiv Open Source

C-KERNEL-ENGINE · GUIDE

Explore a run in the IR Visualizer

Open a run you already know produced an answer. The intermediate representation (IR) is the compiler's record of the operations it selected for that run. You can follow one operation from its place in the sequence down to the buffers it reads and writes.

Generate the report from your run

Use the directory you passed to --run. The example below uses the Gemma directory from server setup. For the default Qwen2 walkthrough, use $HOME/.cache/ck-engine-v8/models/Qwen--Qwen2-0.5B-Instruct-GGUF instead. The list command shows the models in your cache.

Examplebash
.venv/bin/python version/v8/tools/open_ir_visualizer_v8.py --list
RUN="$HOME/cke-runs/gemma3-270m-1k"
.venv/bin/python version/v8/tools/open_ir_visualizer_v8.py \
  --run "$RUN" --html-only --generate --output "$RUN/ir_report.html"
xdg-open "$RUN/ir_report.html"

On a headless host, copy the generated HTML to your desktop and open it in a browser. The HTML-only option reads existing artifacts without collecting a new profile. For a new CLI run you can also add --generate-visualizer to the model command.

Follow one token

  1. Select Prefill Mode and open Kernel Flow. Find a projection in one transformer layer and note its inputs. Prefill processes the prompt tokens together.
  2. Switch to Decode Mode. Find the same layer for a single new token. Compare the listed calls and buffers. Memory Layout shows where the compiled runtime placed those buffers.
  3. Open Dataflow Graph to inspect dependencies exposed by this run. Use Parity Debug when you have a reference comparison. An empty comparison panel means the corresponding evidence has yet to be collected.
  4. Open Profile for measured timings if the run contains a profile. To collect a new one, see the launcher's --with-profile options and record your prompt, machine, thread count and context alongside the result.

Our local Qwen2 report loaded 16 artifacts. It contained inference layouts and kernel flows. Training and profiler artifacts were absent from that run. For live request activity while the server is running, open http://127.0.0.1:8080/viz.

Read your first run · CKE visualizer launcher