Revert "Revert "fix: disable LoRA export for now" (#308)"

This reverts commit da92f745de.

src/heretic/main.py

@@ -17,11 +17,15 @@ def _is_help_invocation() -> bool:
 if _is_help_invocation():
     Settings()  # ty:ignore[missing-argument]
 
+# FIXME: Rich progress bars are currently disabled because of rendering issues
+#        when used from multiple threads in parallel (e.g. by huggingface_hub).
+"""
 from .progress import patch_tqdm
 
 # This patches tqdm class definitions, which must happen
 # before any other module imports tqdm.
 patch_tqdm()
+"""
 
 import logging
 import math
@@ -425,9 +429,6 @@ def run():
 
     needs_full_residuals = settings.print_residual_geometry or settings.plot_residuals
 
-    good_residuals = None
-    bad_residuals = None
-
     if needs_full_residuals:
         print("* Obtaining residuals for good prompts...")
         good_residuals = model.get_residuals_batched(good_prompts)
@@ -465,8 +466,12 @@ def run():
             refusal_directions - projection_vector.unsqueeze(1) * good_directions
         )
         refusal_directions = F.normalize(refusal_directions, p=2, dim=1)
+        del good_directions, projection_vector
+
+    del good_means, bad_means
 
     # Clear cache before starting the optimization study.
+    # This should free up memory from the objects released with the del statements above.
     empty_cache()
 
     trial_index = 0

src/heretic/model.py

@@ -154,13 +154,15 @@ class Model:
         # so we don't need to do anything manually.
 
         print(f"* Transformer model with [bold]{len(self.get_layers())}[/] layers")
-        print("* Abliterable components:")
+
         all_components = {}
         for layer_index in range(len(self.get_layers())):
             for component, modules in self.get_layer_modules(layer_index).items():
                 if component not in all_components:
                     all_components[component] = 0
                 all_components[component] += len(modules)
+
+        print("* Abliterable components:")
         for component, count in all_components.items():
             print(f"  * [bold]{component}[/]: [bold]{count}[/] modules total")
 
@@ -368,8 +370,8 @@ class Model:
         with suppress(Exception):
             try_add("attn.o_proj", layer.self_attn.o_proj)  # ty:ignore[possibly-missing-attribute]
 
-        # Qwen3.5 MoE hybrid layers use GatedDeltaNet (linear attention) instead
+        # Qwen3.5 MoE hybrid layers use GatedDeltaNet (linear attention) instead of
-        # of standard self-attention, so self_attn.o_proj doesn't exist on those layers.
+        # standard self-attention, so self_attn.o_proj doesn't exist on those layers.
         with suppress(Exception):
             try_add("attn.o_proj", layer.linear_attn.out_proj)  # ty:ignore[possibly-missing-attribute]
 
@@ -403,11 +405,13 @@ class Model:
         return modules
 
     def get_abliterable_components(self) -> list[str]:
+        components: set[str] = set()
+
         # Scan all layers because hybrid models (e.g. Qwen3.5 MoE) have different
         # components on different layers (some have self_attn, others linear_attn).
-        components: set[str] = set()
         for layer_index in range(len(self.get_layers())):
             components.update(self.get_layer_modules(layer_index).keys())
+
         return sorted(components)
 
     def abliterate(
@@ -744,9 +748,8 @@ class Model:
         # The returned tensor has shape (prompt, token).
         logprobs = F.log_softmax(logits, dim=-1)
 
-        del outputs
-
         if self.settings.offload_outputs_to_cpu:
+            del outputs, logits
             logprobs = logprobs.cpu()
             empty_cache()
 

uv.lock

@@ -8,7 +8,7 @@ resolution-markers = [
 ]
 
 [options]
-exclude-newer = "2026-04-14T22:48:57.86057843Z"
+exclude-newer = "2026-04-16T07:30:09.771407348Z"
 exclude-newer-span = "P7D"
 
 [[package]]
@@ -1508,14 +1508,14 @@ wheels = [
 
 [[package]]
 name = "mako"
-version = "1.3.10"
+version = "1.3.11"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "markupsafe" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/9e/38/bd5b78a920a64d708fe6bc8e0a2c075e1389d53bef8413725c63ba041535/mako-1.3.10.tar.gz", hash = "sha256:99579a6f39583fa7e5630a28c3c1f440e4e97a414b80372649c0ce338da2ea28", size = 392474, upload-time = "2025-04-10T12:44:31.16Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/59/8a/805404d0c0b9f3d7a326475ca008db57aea9c5c9f2e1e39ed0faa335571c/mako-1.3.11.tar.gz", hash = "sha256:071eb4ab4c5010443152255d77db7faa6ce5916f35226eb02dc34479b6858069", size = 399811, upload-time = "2026-04-14T20:19:51.493Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/87/fb/99f81ac72ae23375f22b7afdb7642aba97c00a713c217124420147681a2f/mako-1.3.10-py3-none-any.whl", hash = "sha256:baef24a52fc4fc514a0887ac600f9f1cff3d82c61d4d700a1fa84d597b88db59", size = 78509, upload-time = "2025-04-10T12:50:53.297Z" },
+    { url = "https://files.pythonhosted.org/packages/68/a5/19d7aaa7e433713ffe881df33705925a196afb9532efc8475d26593921a6/mako-1.3.11-py3-none-any.whl", hash = "sha256:e372c6e333cf004aa736a15f425087ec977e1fcbd2966aae7f17c8dc1da27a77", size = 78503, upload-time = "2026-04-14T20:19:53.233Z" },
 ]
 
 [[package]]