Update app_quant_latent.py

app_quant_latent.py

@@ -444,8 +444,9 @@ try:
         transformer=transformer,
         text_encoder=text_encoder,
         torch_dtype=torch_dtype,
-        attn_implementation="kernels-community/vllm-flash-attn3",
+       
     )
+    pipe.transformer.set_attention_backend("_flash_3") 
    # pipe.load_lora_weights("bdsqlsz/qinglong_DetailedEyes_Z-Image", weight_name="qinglong_detailedeye_z-imageV2(comfy).safetensors", adapter_name="lora")
     pipe.load_lora_weights("rahul7star/ZImageLora",
                            weight_name="NSFW/doggystyle_pov.safetensors", adapter_name="lora")
@@ -467,22 +468,27 @@ log_system_stats("AFTER PIPELINE BUILD")
 # -----------------------------
 # Monkey-patch prepare_latents
 # -----------------------------
-original_prepare_latents = pipe.prepare_latents
+# -----------------------------
+# Monkey-patch prepare_latents
+# -----------------------------
+if pipe is not None:
+    original_prepare_latents = pipe.prepare_latents
 
-def logged_prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
-    # Call original method
-    result_latents = original_prepare_latents(
-        batch_size, num_channels_latents, height, width, dtype, device, generator, latents
-    )
+    def logged_prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        result_latents = original_prepare_latents(
+            batch_size, num_channels_latents, height, width, dtype, device, generator, latents
+        )
+        log_msg = f"🔹 prepare_latents called | shape={result_latents.shape}, dtype={result_latents.dtype}, device={result_latents.device}"
+        if hasattr(self, "_latents_log"):
+            self._latents_log.append(log_msg)
+        else:
+            self._latents_log = [log_msg]
+        return result_latents
 
-    # Save info for logging
-    log_msg = f"🔹 prepare_latents called | shape={result_latents.shape}, dtype={result_latents.dtype}, device={result_latents.device}"
-    if hasattr(self, "_latents_log"):
-        self._latents_log.append(log_msg)
-    else:
-        self._latents_log = [log_msg]
+    pipe.prepare_latents = logged_prepare_latents.__get__(pipe)
+else:
+    log("❌ WARNING: Pipe not initialized; skipping prepare_latents patch")
 
-    return result_latents
 
 # Apply patch
 pipe.prepare_latents = logged_prepare_latents.__get__(pipe)