deepmd.kernels.triton.sezm.tile_config_data#
Built-in launch-configuration data for the shape-tuned SeZM Triton kernels.
This module is pure data: one nested mapping per GPU model, keyed by the exact device name reported by torch.cuda.get_device_name(). The query layer in tile_configs selects the sub-mapping of the running GPU and resolves individual keys; devices without an entry here fall back to the conservative defaults of every kernel family (correct on any CUDA device, merely not tuned).
Entry semantics#
Every per-family table maps an exact shape key to either a launch configuration tuple or None:
a tuple is the winning configuration measured by the sweep;
Nonerecords that the sweep ran and the tuned kernel did not beat its baseline for this key (win-list families) or that the default configuration itself won (default-keyed families) – the fallback is the measured optimum, not a guess;an absent key means the shape was never swept on this GPU. The freeze auto-tuner (
sweep_tile_configs.tune_missing_configs()) treats only absent keys as work.
Key conventions and value layouts are documented in tile_configs; regeneration is documented in sweep_tile_configs. All entries below were swept at production edge counts (3e5 to 6.5e5 edges) with the (C_wide, lmax)-keyed families measured at n_focus = 2.