qwenGPU / recorded execution

Full model · Metal

Implications

Security. The model’s weights carried out the AES rounds. A monitor watching only for calls to an encryption library would miss where that work happened. This modified model could encrypt data when given a program and a key. That capability matters when deciding what data and programs a model should receive.

Researcher view

Result for this saved trial

Inside the GPU

Warp scheduler

Program ROM · 32 block slots

Active instruction stages

Registers

Global memory

Hover over a word for its address and value.

How to read this replay

The model generated the tokens in this recording. The player decodes them into the warp, program block, instruction stages, four register lanes, and global-memory writes using the qwenGPU monitor's record format. A dot means the stream has not reported a register value. The left pane explains the experiment. The result card is an end-of-trial summary. Recorded compute time covers the launches and their checks; loading and compilation are separate. Guided playback speeds up long trials and advances between launches automatically.

Launch times include trace checking and recording; loading and compilation are excluded. Token spacing is interpolated. Untimed releases advance immediately at numeric speeds.