fix: remove CPU offload, use pipe.to(cuda) on xlarge for 22B model, adjust duration estimate

app.py

@@ -189,7 +189,6 @@ def _load_pipeline():
         adapter_name="3dreal-strong",
         weight_name="3DREAL-strong.safetensors",
     )
-    _pipe.enable_model_cpu_offload()
     _pipe_lora_variant = None
 
 
@@ -278,9 +277,9 @@ def _estimate_duration(*args, **kwargs):
         num_inference_steps = int(num_inference_steps)
     except (TypeError, ValueError):
         return 240
-    # 22B model with enable_model_cpu_offload(): first step ~28s (cold),
-    # subsequent ~5s each, plus VAE decode + video encoding ~20s.
-    base = 60 + num_inference_steps * 8
+    # 22B model on xlarge (96GB): pipe.to("cuda") ~20s cold start,
+    # ~3s per denoising step, VAE decode + video encoding ~15s.
+    base = 40 + num_inference_steps * 4
     frame_factor = max(1.0, num_frames / 41)
     return min(240, int(base * frame_factor))
 
@@ -318,6 +317,7 @@ def generate_video(
     prompt = _normalize_prompt(prompt)
 
     pipe = _select_lora(intensity)
+    pipe.to("cuda")
 
     width, height = RESOLUTION_MAP.get(resolution, RESOLUTION_MAP["720p"])
     num_frames = _round_num_frames(num_frames)