[droot]

…he Sweep tab

Live state

• The Fleet chips and the Timeline's current turn come from /api/v1/inspect/now, polled twice a second, instead of replaying the activity history about 8 s behind. A turn, op or batch of sampler requests that just ended is kept as a finished span until the history has it. On an API server without the endpoint (404), the page falls back to the replay, so it works with or without it.
• On a shared LoRA runtime, each job's chip is lit only while that job's op or sampler requests run.

Fleet

• A sweep selector entry, "no sweep": the pool with no job's counts or chips, any worker holding a GPU drawn as a neutral chip. ?sweep=none opens on it; starting a sweep switches back to the active sweeps.
• A time-sliced worker is drawn on the GPU only in its turns, even before its first one, so a worker waiting for the GPU no longer shows as holding it. The timeline's tick step follows its width, and its legend counts holders that do not fit.

Sweep

• Cards under the spec show the recipe every job runs: GSM8K, GRPO, 8 x 8 answers per step, 512 tokens, learning rates and LoRA rank, and the spec's totals as it is edited. The batch settings are constants in app.py that the Jobs are launched with, so the cards show what runs.
• The job list is collapsed until opened, so the page before a launch shows the spec and not earlier sweeps.
• "demo" is gone from the header and title, runs are named sweep-HHMMSS, the keynote-20 preset is mixed-20, and the qwen-8 preset is gone.

Run on open-rl-dra-use5 (8 x H100) with 16- and 20-job sweeps.