Update app.py

app.py

@@ -634,9 +634,18 @@ def _generate(
     # to call on every request. Filtered to `LOADED_LORAS`: a slider for a LoRA that failed at startup has no
     # adapter behind it, and `set_adapters` would raise if asked to activate a name that was never attached.
     if LOADED_LORAS:
-        active = {name: strength for name, strength in lora_strengths.items() if name in LOADED_LORAS}
-        if active:
-            PIPE.transformer.set_adapters(list(active), weights=list(active.values()))
+        # Filtered to strength > 0, not just "loaded": PEFT computes every adapter in the active list on every
+        # forward regardless of its weight (no early-exit for scale 0), so an adapter left active at 0.0 still
+        # costs a real lora_A/lora_B matmul per targeted Linear, every block, every step — overhead that scales
+        # with how many LoRAs are loaded, not how many are actually in use for a given request. Called
+        # unconditionally, even with an empty list, rather than only `if active:` — skipping the call when every
+        # slider is 0 would leave whichever adapters the *previous* request activated still live.
+        active = {
+            name: strength
+            for name, strength in lora_strengths.items()
+            if name in LOADED_LORAS and strength > 0
+        }
+        PIPE.transformer.set_adapters(list(active), weights=list(active.values()))
 
     if PLACEMENT == "lazy":
         PIPE.to("cuda")