Load model on CPU at startup to bypass slow ZeroGPU disk packing

spaces 0.51.0 (released 2026-07-10) switched to disk-based tensor
packing (~19s pack + ~19s unpack per inference), breaking the ~7s
inference that worked with spaces 0.50.4's pinned-memory approach.

HuggingFace force-injects spaces==0.51.0 so pinning requirements.txt
doesn't help. Instead:
- Load transformer and pipeline on CPU at startup (ZeroGPU has nothing
to pack since no tensors are on GPU)
- pipe.to(device) inside @spaces.GPU: ~3-4s PCIe transfer vs ~19s disk
- pipe.to("cpu") in finally block so ZeroGPU has nothing to repack

Also reverts bad spaces==0.50.4 pin (build error) and raises GPU
duration defaults to accommodate the explicit model-move time.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -70,7 +70,7 @@ _hb = _start_heartbeat("transformer")
 _transformer = QwenImageTransformer2DModel.from_pretrained(
     "prithivMLmods/Qwen-Image-Edit-Rapid-AIO-V23",
     torch_dtype=dtype,
-    device_map="cuda",
+    device_map="cpu",
 )
 _hb.set()
 print(f"[startup] transformer loaded in {time.perf_counter()-_t0_load:.1f}s", flush=True)
@@ -82,7 +82,7 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
     "FireRedTeam/FireRed-Image-Edit-1.1",
     transformer=_transformer,
     torch_dtype=dtype,
-).to(device)
+)
 _hb.set()
 print(f"[startup] pipeline loaded in {time.perf_counter()-_t1_load:.1f}s", flush=True)
 
@@ -372,7 +372,9 @@ def _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height, m
         print(f"[infer] GPU: {p.name}, total={p.total_memory/1024**3:.1f}GB, cap={p.major}.{p.minor}")
         torch.cuda.reset_peak_memory_stats()
 
-    print(f"[infer] {_gpu_mem_str(_cuda_ok)} — t={time.perf_counter()-t0:.1f}s")
+    print(f"[infer] moving pipe to {device}... — t={time.perf_counter()-t0:.1f}s")
+    pipe.to(device)
+    print(f"[infer] pipe on {device} — {_gpu_mem_str(_cuda_ok)} — t={time.perf_counter()-t0:.1f}s")
 
     torch.cuda.empty_cache()
     print(f"[infer] cache cleared — {_gpu_mem_str(_cuda_ok)}")
@@ -418,6 +420,8 @@ def _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height, m
         timer.print_timings()
         raise
     finally:
+        print(f"[infer] moving pipe to CPU... t={time.perf_counter()-t0:.1f}s")
+        pipe.to("cpu")
         gc.collect()
         torch.cuda.empty_cache()
         print(f"[infer] ===== END t={time.perf_counter()-t0:.1f}s =====")
@@ -432,7 +436,7 @@ with gr.Blocks() as demo:
     guidance_scale = gr.Slider(minimum=1.0, maximum=10.0, step=0.1, value=1.0, elem_id="gradio-guidance", elem_classes="hidden-input", container=False)
     steps = gr.Slider(minimum=1, maximum=50, step=1, value=4, elem_id="gradio-steps", elem_classes="hidden-input", container=False)
     mode = gr.Textbox(value="fast", elem_id="gradio-mode", elem_classes="hidden-input", container=False)
-    gpu_duration = gr.Slider(minimum=10, maximum=120, step=5, value=20, elem_id="gradio-gpu-duration", elem_classes="hidden-input", container=False)
+    gpu_duration = gr.Slider(minimum=10, maximum=120, step=5, value=30, elem_id="gradio-gpu-duration", elem_classes="hidden-input", container=False)
     result = gr.Image(elem_id="gradio-result", elem_classes="hidden-input", container=False, format="png")
 
     example_idx = gr.Textbox(value="", elem_id="example-idx-input", elem_classes="hidden-input", container=False)

requirements.txt

@@ -7,7 +7,7 @@ pyarrow
 sentencepiece
 torchvision
 kernels
-spaces==0.50.4
+spaces
 hf_xet
 gradio
 pytest

static/mode_toggle.js

@@ -1,6 +1,6 @@
 () => {
 window.__selectedMode = 'fast';
-var MODE_GPU_DURATION = { fast: 20, high_detail: 40 };
+var MODE_GPU_DURATION = { fast: 30, high_detail: 60 };
 window.__setMode = function(m) {
     window.__selectedMode = m;
     var fast = document.getElementById('mode-btn-fast');

templates/app.html

@@ -120,8 +120,8 @@
                 <div class="settings-group-body">
                     <div class="slider-row">
                         <label>GPU Duration (s)</label>
-                        <input type="range" id="custom-gpu-duration" min="10" max="120" step="5" value="20">
-                        <span class="slider-val" id="custom-gpu-duration-val">20</span>
+                        <input type="range" id="custom-gpu-duration" min="10" max="120" step="5" value="30">
+                        <span class="slider-val" id="custom-gpu-duration-val">30</span>
                     </div>
                     <div class="slider-row">
                         <label>Seed</label>