From 1b4851536d9a4b2f4a5e2f052e36f206b4a89846 Mon Sep 17 00:00:00 2001
From: anrp <anrp@anrp.net>
Date: Sat, 9 May 2026 09:46:26 +0000
Subject: [PATCH] fix: Reset model after saving merged model (#321)

* fix: Reset model after saving merged model

The adapter is lost and writes 0-byte adapters if you save an adapter after saving the merged model.

* Revert "Revert "Revert "fix: disable LoRA export for now" (#308)" (#319)"

This reverts commit 216c0899744388cbd8dadbf700f1619214836197.

* Add comment as to why resetting model is needed
---
 src/heretic/main.py | 71 +++++++++++++++++++++++----------------------
 1 file changed, 37 insertions(+), 34 deletions(-)

diff --git a/src/heretic/main.py b/src/heretic/main.py
index c64f9d3..48eece7 100644
--- a/src/heretic/main.py
+++ b/src/heretic/main.py
@@ -132,31 +132,26 @@ def obtain_merge_strategy(settings: Settings, model: Model) -> str | None:
             )
         print()
 
-        strategy = prompt_select(
-            "How do you want to proceed?",
-            choices=[
-                Choice(
-                    title="Merge LoRA into full model"
-                    + (
-                        ""
-                        if settings.quantization == QuantizationMethod.NONE
-                        else " (requires sufficient RAM)"
-                    ),
-                    value="merge",
+    strategy = prompt_select(
+        "How do you want to proceed?",
+        choices=[
+            Choice(
+                title="Merge LoRA into full model"
+                + (
+                    ""
+                    if settings.quantization == QuantizationMethod.NONE
+                    else " (requires sufficient RAM)"
                 ),
-                Choice(
-                    title="Cancel",
-                    value="cancel",
-                ),
-            ],
-        )
+                value="merge",
+            ),
+            Choice(
+                title="Save LoRA adapter only (can be merged later)",
+                value="adapter",
+            ),
+        ],
+    )
 
-        if strategy == "cancel":
-            return None
-
-        return strategy
-    else:
-        return "merge"
+    return strategy
 
 
 def run():
@@ -754,17 +749,23 @@ def run():
             print("* Parameters:")
             for name, value in get_trial_parameters(trial).items():
                 print(f"  * {name} = [bold]{value}[/]")
-            print("* Resetting model...")
-            model.reset_model()
-            print("* Abliterating...")
-            model.abliterate(
-                refusal_directions,
-                trial.user_attrs["direction_index"],
-                {
-                    k: AbliterationParameters(**v)
-                    for k, v in trial.user_attrs["parameters"].items()
-                },
-            )
+
+            # Per https://github.com/huggingface/peft/issues/868#issuecomment-1820642893 once a LoRA is merged it's
+            # expected to be empty. Provide a utility function to restore the previous LoRA-ified state.
+            def reset_trial_model() -> None:
+                print("* Resetting model...")
+                model.reset_model()
+                print("* Abliterating...")
+                model.abliterate(
+                    refusal_directions,
+                    trial.user_attrs["direction_index"],
+                    {
+                        k: AbliterationParameters(**v)
+                        for k, v in trial.user_attrs["parameters"].items()
+                    },
+                )
+
+            reset_trial_model()
 
             while True:
                 print()
@@ -812,6 +813,7 @@ def run():
                                 del merged_model
                                 empty_cache()
                                 model.tokenizer.save_pretrained(save_directory)
+                                reset_trial_model()
 
                             print(f"Model saved to [bold]{save_directory}[/].")
 
@@ -921,6 +923,7 @@ def run():
                                     private=private,
                                     token=token,
                                 )
+                                reset_trial_model()
 
                             if is_hf_path(settings.model):
                                 card = ModelCard.load(settings.model)