ComfyUI Base 3.21.0: pick_lane, the cheapest test GPU that holds a measured peak

[Soul-Craft]

What

py/gpu_fit.py learns which free GPU a smoke should go to.

• LANES, cheapest first: Kaggle T4 (16 GB) and L4 (24 GB), HF ZeroGPU large (48 GB) and xlarge (96 GB), Modal's $30 of monthly credits from a T4 to a B200, then RunPod (RTX PRO 6000, B200, B300), marked paid.
• pick_lane(peak_gib, minutes, needs) returns the first lane whose card holds the measured peak plus 10 % and whose own limit holds the run:
  • Kaggle: 12 h session
  • ZeroGPU: our 30-minute cap, 15 minutes on xlarge since it costs the quota twice
  • Modal: the minutes $30 buys at the GPU's list price, within Modal's 24 h per call
• needs takes fast dtypes (fp16 bf16 fp8 fp4, from the compute capability by fast_dtypes()), server (ZeroGPU runs per-call functions, so it never qualifies) and GPU classes (B200, B300). A class is judged by the existing fits(), so GPU_PREFER and pick_lane share one rule.
• CLI: python3 py/gpu_fit.py --peak-gib 35.7 --minutes 10 --needs bf16 prints lane=zerogpu-large ... paid=no. lane=none exits 1, and an unknown need exits 2.

Modal prices and the 24 h limit come from modal.com/pricing and Modal's timeouts guide (read 2026-10-17). They match ci/modal-smoke/modal_guard.py in the workspace.

Tests

suite/test_unit.py::test_unit_pick_lane_takes_the_cheapest_lane_that_holds_the_peak_and_the_run covers:

• the lane order and the headroom
• the ZeroGPU and Modal limits
• dtypes, server, B200 through fits()
• the paid fallback, None, and the CLI exit codes

The peaks are the measured ones: Photo Auditor 17.72 GB, Krea 2 bf16 35.7 GB, Flux 2 Dev fp8 about 71 GB.

Nothing ran on the Mac except py_compile, a CLI spot check and package.py --base --check. This PR's CI runs the suite. No pod used.

After merge: tag base-v3.21.0, then the workspace moves the base line of repos.txt in its own commit.

🤖 Generated with Claude Code