Quantize model stack for ZeroGPU memory limits

AGENT.md

@@ -13,6 +13,10 @@ The base model, accelerated transformer, and Diffusers dependency are pinned to
 exact revisions. Update those pins deliberately and test the deployed Space
 before merging.
 
+The transformer and text encoder are dynamically quantized to FP8 before the
+pipeline is moved to its runtime device. Keep that ordering: the unquantized
+BF16 pipeline is too large for the ZeroGPU transfer path.
+
 ## Privacy boundary
 
 Image inference runs inside the Space. Do not add application code that sends

README.md

@@ -23,3 +23,5 @@ Pro Realism Edit Studio is a powerful image editor powered by [Qwen-Image-Edit-2
 ## Reproducibility
 
 The base model, accelerated transformer, and development version of Diffusers are pinned to exact revisions. If the accelerated four-step transformer cannot load, the Space stops with a clear error rather than silently switching to an incompatible model.
+
+The transformer and text encoder use dynamic FP8 quantization so the pipeline fits within Hugging Face ZeroGPU memory. This can introduce a small numerical difference from full BF16 inference, but avoids the accelerator failure caused by transferring an approximately 58 GB BF16 pipeline.

app.py

@@ -5,6 +5,8 @@ import torch
 import spaces
 
 from PIL import Image
+from torchao.quantization import Float8DynamicActivationFloat8WeightConfig
+from torchao.quantization import quantize_
 from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline
 from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel
 from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3
@@ -66,12 +68,28 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
     transformer=transformer,
     torch_dtype=dtype,
     token=hf_token,
-).to(device)
+)
 
 del transformer
 if torch.cuda.is_available():
     torch.cuda.empty_cache()
 
+# A full BF16 pipeline is roughly 58 GB and cannot be transferred into a
+# standard ZeroGPU allocation. Dynamic FP8 cuts the transformer/text-encoder
+# footprint enough to fit while preserving four-step Rapid-AIO behavior.
+fp8_config = Float8DynamicActivationFloat8WeightConfig()
+quantize_(pipe.transformer, fp8_config)
+try:
+    quantize_(pipe.text_encoder, fp8_config)
+except Exception as exc:
+    print(
+        f"Text-encoder FP8 quantization unavailable ({type(exc).__name__}); "
+        "continuing with the pinned text encoder.",
+        flush=True,
+    )
+
+pipe = pipe.to(device)
+
 # Apply the upstream FA3 optimization, but keep boot resilient if kernel support
 # differs on duplicated Spaces or the public fallback transformer.
 try: