Keep a Float32 network Float32: training history, _norm, PoissonTensor defaults, the optimizer step size and two parameterlength counts - #295

Merged

michakraus merged 17 commits into

mainfrom

fix/audit-part-b

Sep 18, 2026

Commits on Sep 18, 2026

Type a Float32 network's training history as Float32 (B1)

michakraus
and
claude
committed
0c7e43e
Make all three arms of _norm keep their argument's element type (B2)

michakraus
and
claude
committed
738ef7f
Remove PoissonTensor's undocumented CPU/GPU element-type split (B3)

michakraus
and
claude
committed
3d79818
Type DataLoader(::EnsembleSolution)'s zero-filled buffer (B4)

michakraus
and
claude
committed
315340f
Scale the GO-native leaf optimizer step in the parameter's own type (B5)

michakraus
and
claude
committed
90f0e99
Make parameterlength exact by construction with integer division (B6)

michakraus
and
claude
committed
595f7ef
Record the Part B element-type fixes in CHANGELOG.md

michakraus
and
claude
committed
90da43e
Drop the loss-value element-type assert added for B1

michakraus
and
claude
committed
586b03f
Record the B8 decision on ClassificationLayer's element type in CHANGELOG.md

michakraus
and
claude
committed
c313e6d
Fix step_size_element_type.jl's regression claim (critic round 3)

michakraus
and
claude
committed
00d5658
Revert the no-op Float64(...) wrapper on _default_step_size's literals

michakraus
and
claude
committed
2021242
Rewrite the historical comment on PoissonTensor's backend constructors

michakraus
and
claude
committed
8ba8ae2
Record B10: DataLoader(::EnsembleSolution)'s zero-fix has no reachable test

michakraus
and
claude
committed
e1c68da
Rewrite the five remaining historical-narration comments in test/

michakraus
and
claude
committed
e30620d
Keep _norm general, and re-point the references this branch shifted

michakraus
and
claude
committed
4db79d3
Re-point the getindex citation the rebase left behind

michakraus
and
claude
committed
a07aad2
Answer the review of #295

michakraus
and
claude
committed
0f300a2

Navigation Menu

Keep a Float32 network Float32: training history, _norm, PoissonTensor defaults, the optimizer step size and two parameterlength counts - #295

Keep a Float32 network Float32: training history, _norm, PoissonTensor defaults, the optimizer step size and two parameterlength counts#295
michakraus merged 17 commits into
mainfrom
fix/audit-part-b

Commits on Sep 18, 2026

Type a Float32 network's training history as Float32 (B1)

Make all three arms of _norm keep their argument's element type (B2)

Remove PoissonTensor's undocumented CPU/GPU element-type split (B3)

Type DataLoader(::EnsembleSolution)'s zero-filled buffer (B4)

Scale the GO-native leaf optimizer step in the parameter's own type (B5)

Make parameterlength exact by construction with integer division (B6)

Record the Part B element-type fixes in CHANGELOG.md

Drop the loss-value element-type assert added for B1

Record the B8 decision on ClassificationLayer's element type in CHANGELOG.md

Fix step_size_element_type.jl's regression claim (critic round 3)

Revert the no-op Float64(...) wrapper on _default_step_size's literals

Rewrite the historical comment on PoissonTensor's backend constructors

Record B10: DataLoader(::EnsembleSolution)'s zero-fix has no reachable test

Rewrite the five remaining historical-narration comments in test/

Keep _norm general, and re-point the references this branch shifted

Re-point the getindex citation the rebase left behind

Answer the review of #295