From 0ff74be52f83c8a739de4a64bae2fcb11faad17b Mon Sep 17 00:00:00 2001
From: Ethanfel <ethan.fel@ts-pc.fr>
Date: Tue, 7 Jul 2026 19:49:26 +0200
Subject: [PATCH] fix(inline): thread model-class arch hint into auto_tune so
 captured chains score under the right preset
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Completes FIX #12: the hint reached optimize_merge but not auto_tune's
_normalize_stack, so inline autotuner-mode candidate scoring + the
community arch bucket used key-based detection (mis-tagging e.g.
attention-only Qwen as ACE-Step). File-based stacks have no virtual
items → hint unreachable → unchanged.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 lora_optimizer.py              |  8 +++++++-
 tests/test_inline_optimizer.py | 25 +++++++++++++++++++++++++
 2 files changed, 32 insertions(+), 1 deletion(-)

diff --git a/lora_optimizer.py b/lora_optimizer.py
index e1e910c..2373cbc 100644
--- a/lora_optimizer.py
+++ b/lora_optimizer.py
@@ -11572,7 +11572,13 @@ class LoRAAutoTuner(LoRAOptimizer):
             lora_stack = clean_stack
 
         # --- Normalize & validate stack ---
-        normalized_stack = self._normalize_stack(lora_stack, normalize_keys=normalize_keys)
+        # Pass the model-class arch hint so captured (inline) chains score
+        # candidates under the correct preset (else key-based detection
+        # mis-tags e.g. attention-only Qwen as ACE-Step). File-based stacks
+        # have no virtual items, so the hint is unreachable — identical.
+        normalized_stack = self._normalize_stack(
+            lora_stack, normalize_keys=normalize_keys,
+            _arch_hint=self._model_class_arch(model))
         active_loras = [item for item in normalized_stack if item["strength"] != 0]
         if not active_loras:
             return (model, clip, "No active LoRAs in stack.", "", None, None)
diff --git a/tests/test_inline_optimizer.py b/tests/test_inline_optimizer.py
index 8650eb1..c1263e8 100644
--- a/tests/test_inline_optimizer.py
+++ b/tests/test_inline_optimizer.py
@@ -1622,6 +1622,31 @@ class TestInlineAutoTunerDelegation(unittest.TestCase):
                     "analysis report", {"td": 1}, {"ld": 2})
         node._autotuner_delegate = types.SimpleNamespace(auto_tune=fake_auto_tune)
 
+    def test_autotune_threads_model_class_arch_hint(self):
+        """auto_tune must pass the model-class arch hint to _normalize_stack so
+        captured chains score candidates under the right preset (was hintless →
+        attention-only Qwen scored under the ACE-Step preset)."""
+        class QwenImage:  # mirrors comfy.model_base.QwenImage
+            pass
+        model = _FakePatcher({})
+        model.model = QwenImage()
+        delegate = lora_optimizer.LoRAAutoTuner()
+        seen = {}
+
+        class _Stop(Exception):
+            pass
+
+        def rec_normalize(stack, normalize_keys="disabled", _arch_hint=None):
+            seen["hint"] = _arch_hint
+            raise _Stop
+
+        delegate._normalize_stack = rec_normalize
+        stack = [{"name": "chain lora #1 [x]", "lora": {"k": _adapter()},
+                  "strength": 1.0, "_precomputed_diffs": True}]
+        with self.assertRaises(_Stop):
+            delegate.auto_tune(model, stack, 1.0)
+        self.assertEqual(seen.get("hint"), "qwen_image")
+
     def test_delegates_with_virtual_stack_and_stripped_model(self):
         model = self._model(_chain_patches(
             (0.8, {"a": _adapter()}), (0.5, {"a": _adapter()})))