Qwenception

Pick a baseline inference run, add as many as four alternatives, and replay what every run wanted to emit from the identical recorded prefix.

Step 1 · establish the reference universe

Choose your baseline hardware

Every comparison starts from one canonical Triton, BF16-weight, BF16-key/value-cache run.

Prompt dataset Output range Show Tie rule Collapse agreement runs Keep active fork in view

Speed Loading catalog…

Recorded path playback Choose a prompt dataset, then press Play outputs.

The retained continuation will appear here one token at a time.

Available comparisons for this starting point

Find a run Hardware Change type

PositionRecorded outputMeasured next-token predictions