refactor(lora): make the fused chunk slicer public

The fidelity CLI slices fused saves the way try_load_lora does.

modules/lora/native_adapter.py

@@ -525,7 +525,7 @@ def slice_chunk_rows(t, chunk: ChunkSpec):
     return t.contiguous()
 
 
-def _slice_lora_chunk(w, chunk: ChunkSpec):
+def slice_lora_chunk(w, chunk: ChunkSpec):
     """Return a shallow copy of ``w`` with ``lora_up.weight`` sliced per ``chunk``; a dense bias follows a pure reorder."""
     out = dict(w)
     out["lora_up.weight"] = slice_chunk_rows(w["lora_up.weight"], chunk)
@@ -640,7 +640,7 @@ def try_load_lora(name, network_on_disk, lora_scale, *,
                     skipped += 1
                     continue
                 fused_out = w["lora_up.weight"].shape[0]
-                target_w = _slice_lora_chunk(w, chunk)
+                target_w = slice_lora_chunk(w, chunk)
                 target_w = slice_dora_scale(target_w, chunk, fused_out)
                 if target_w is None:
                     log.warning(f'Network load: type=LoRA name="{name}" arch={arch_name} key={network_key} per-input DoRA on fused target skipped (unsupported)')

test/test-minimax-native-adapters.py

@@ -370,7 +370,7 @@ def native_mapping(state_dict, network_alpha=None):
         if 'lora_down.weight' not in w or 'lora_up.weight' not in w:
             continue
         for path, chunk in native_adapter.resolve_group_targets(M.resolve_targets, prefix, base):
-            target = native_adapter._slice_lora_chunk(w, chunk) if chunk is not None else w # pylint: disable=protected-access
+            target = native_adapter.slice_lora_chunk(w, chunk) if chunk is not None else w
             alpha = network_alpha if 'alpha' not in target else float(target['alpha'])
             scale = 1.0 if alpha is None else alpha / target['lora_down.weight'].shape[0]
             out[path] = (target['lora_down.weight'], target['lora_up.weight'], scale)