fix(regional): preserve shared model patch ancestry

Keep compatible parallel regional paths on one inherited model lineage, including NegPip interoperability, while retaining bounded fused and optional Triton execution paths.

Expand graph-shape, lifecycle, memory-safety, and runtime regressions across the supported attention families.
This commit is contained in:
Artificial Sweetener
2026-09-10 23:11:24 -04:00
parent c298efd8e8
commit 017e3fc7fe
44 changed files with 1855 additions and 231 deletions
+1 -1
View File
@@ -100,7 +100,7 @@ def test_regional_patch_stack_preserves_state_lineage_and_runtime_nesting() -> N
assert stack.user_model is source
assert stack.attention_model.parent is source
assert stack.sampling_model.parent is stack.attention_model
assert stack.sampling_model.parent is source
assert stack.sampling_model.patches["weight"] == [global_lora_patch]
assert (
source.get_wrappers("diffusion_model", "simple_syrup.attention_coupling") == []