Pick an experiment.
Settings
System
Scheduled events
| at (s) | kind | target / values |
|---|
Run
YAML (edit directly; form changes apply on top)
Jobs
| run | experiment | status | progress | wall |
|---|
Runs
Select two runs to compare.
| run | experiment | seed | duration | completed | goodput req/s by class | J/token | PUE | trips |
|---|
Validate the simulator on real hardware
Simulates the experiment open in the Experiments tab under several seeds, replays the same workload against a live
vLLM server (identical arrivals and token counts) several times per seed, and reports accuracy (simulated vs measured, same seed),
the hardware's own run-to-run noise, and whether the measurement falls inside the simulator's seed band.
The server must be up and reachable from this machine (e.g. an SSH tunnel to the Spark). Each bench repeat takes the experiment's
duration plus the drain of its last requests.
Running
Validations
| name | experiment | hardware | pass | fail | no data | in seed band | bench | measured |
|---|
System
Live view: power node load (% of rating), rack draw against budget, loop temperatures. Click a box for its actions; what you trigger applies at the current simulated time.
Events
Chat with the live server
Talks to the vLLM server directly (through this tokenx server). Each reply is measured: time to first token,
decode speed, and, with telemetry, the GPU's power over the reply and the energy per output token (GPU only).