Implement multi-dimensional DistConv sharding#27

Merged

PatrickRMiles merged 26 commits intoLBANN:mainfrom

PatrickRMiles:miles30/multidim_distconv

Mar 26, 2026

Commits on Mar 10, 2026

update config with 3D num_shards and shard_dim
Patrick Miles
committed
update config util to expect 3D num_shards and shard_dim, add helper to deal with 1D inputs
Patrick Miles
committed
worker no longer needs to modify distconv params set in config -- just pass them as-is to the ParallelStrategy call
Patrick Miles
committed
implement multi-dimensional sharding for distconv
Patrick Miles
committed
update distconv param name scheme
Patrick Miles
committed
fix loss calc
Patrick Miles
committed

Commits on Mar 11, 2026

add sharded dice loss calculation to dice score util
Patrick Miles
committed
update evaluate to use sharded dice loss calc
Patrick Miles
committed
update trainer to use new evaluate; other small fixes/tweaks
Patrick Miles
committed
fix assert
Patrick Miles
committed
fix naming
Patrick Miles
committed
fix naming
Patrick Miles
committed
better default values
Patrick Miles
committed
missing import
Patrick Miles
committed
update distconv param names, default vals
Patrick Miles
committed
use np.prod instead of math.prod
Patrick Miles
committed
ruff
Patrick Miles
committed

Commits on Mar 17, 2026

Merge branch 'LBANN:main' into miles30/multidim_distconv
PatrickRMiles
authored

Commits on Mar 18, 2026

import math
Patrick Miles
committed
warmup logging and timing
Patrick Miles
committed
import math for prod
Patrick Miles
committed

Commits on Mar 26, 2026