Skip to content

gpuwm-hex 0.3.1

Choose a tag to compare

@FahrenheitResearch FahrenheitResearch released this 15 Sep 03:44
· 21 commits to main since this release

gpuwm-hex 0.3.1

New:

  • A fine hex core at any CONUS point, from HRRR. gpuwm-hex mesh-plan --point LAT,LON
    plans and prices the 937.5 m ladder and, with --generate, builds, admits, registers
    and culls it; gpuwm-hex intermediate --source hrrr and gpuwm-hex lbc write the
    intermediates and the boundary files.
  • The surface/PBL cadence is a selectable knob: gpuwm-hex forecast --pbl-cadence SECONDS.
    The default stays the weld; a cadence that is not a whole number of steps is refused by name.
  • Local time stepping reaches a limited-area cull (--local-timestep), selectable and off
    by default.
  • The engine-verdict instrument ships in the tree (tools/measure_engine_verdicts.py); a
    re-pin is one command. An engine that is not on PyPI is admitted by a measured row of its
    sixteen seam bytes, and the door, doctor and the driver's checkout guard read that one row.
  • A seam-level engine A/B (tools/measure_engine_seam_ab.py) and a per-kernel
    capture-and-replay instrument (tools/kernel_element_ab.py).

Fixed:

  • Requires gpuwm 2.7.4. The sixteen seam pins are re-measured against the published wheel;
    the two files that moved since 2.7.3 do not reach the port's path, measured byte-neutral
    through the seam and on a one-hour point-cull forecast.
  • The forecast step is faster on the 937.5 m point cull, 1.22x per step at the landed tree:
    the CUDA admission is measured once, every recovery kernel launches once, the health gate
    and the regional validations run on the device, and the level-independent dycore kernels
    launch one thread per element. Every history frame is byte-identical and the contract deck
    is 8 of 8 bitwise, dual-run identical.
  • gpuwm-hex init stamps the run declarations a mesh-time capsule leaves absent, so the
    forecast driver no longer refuses a first point cull.
  • The boundary producer resolves through the engine ladder, so a bundle staged by
    gpuwm fetch-bridges opens the lbc leg the way it opens init.
  • The HRRR intermediate's map source fits the 32-character WPS field.