Update app.py

app.py

@@ -186,7 +186,8 @@ def infer(prompt, negative_prompt, adapter_l_file, adapter_g_file, strength, noi
     
     # Apply CLIP-L adapter
     if adapter_l is not None:
-        anchor_l, delta_l, log_sigma_l, attn_l1, attn_l2, tau_l, g_pred_l, gate_l = adapter_l(t5_seq, clip_embeds["clip_l"])
+        clip_l_in = clip_embeds["clip_l"].to(torch.float32)
+        anchor_l, delta_l, log_sigma_l, attn_l1, attn_l2, tau_l, g_pred_l, gate_l = adapter_l(t5_seq, clip_l_in)
         gate_l_scaled = gate_l * gate_prob
         delta_l_final = delta_l * strength * gate_l_scaled
         clip_l_mod = clip_embeds["clip_l"] + delta_l_final
@@ -203,7 +204,10 @@ def infer(prompt, negative_prompt, adapter_l_file, adapter_g_file, strength, noi
     
     # Apply CLIP-G adapter  
     if adapter_g is not None:
-        anchor_g, delta_g, log_sigma_g, attn_g1, attn_g2, tau_g, g_pred_g, gate_g = adapter_g(t5_seq, clip_embeds["clip_g"])
+        # Float32 adapter input
+        clip_g_in = clip_embeds["clip_g"].to(torch.float32)
+
+        anchor_g, delta_g, log_sigma_g, attn_g1, attn_g2, tau_g, g_pred_g, gate_g = adapter_g(t5_seq, clip_g_in)
         gate_g_scaled = gate_g * gate_prob
         delta_g_final = delta_g * strength * gate_g_scaled
         clip_g_mod = clip_embeds["clip_g"] + delta_g_final