v15 Phase 3: FP8 quantization with torchao

app.py

@@ -1,4 +1,4 @@
-"""Z-Image-Turbo v14.4 - Optimized with SDPA + VAE optimizations"""
+"""Z-Image-Turbo v15 - Phase 3: FP8 Quantization + SDPA + VAE optimizations"""
 
 import random
 import torch
@@ -30,6 +30,15 @@ pipe_i2i = ZImageImg2ImgPipeline.from_pretrained(
 )
 pipe_i2i.to("cuda")
 
+# Phase 3: FP8 Quantization (30-50% speedup, 50% memory reduction)
+try:
+    from torchao.quantization import quantize_, float8_dynamic_activation_float8_weight
+    quantize_(pipe_t2i.transformer, float8_dynamic_activation_float8_weight())
+    quantize_(pipe_i2i.transformer, float8_dynamic_activation_float8_weight())
+    print("FP8 quantization enabled on transformers!")
+except Exception as e:
+    print(f"FP8 quantization not available: {e}")
+
 # Phase 1+2 Optimizations: VAE optimizations for large images
 try:
     pipe_t2i.vae.enable_tiling()
@@ -45,7 +54,7 @@ try:
 except Exception as e:
     print(f"VAE slicing not available: {e}")
 
-print("Pipelines ready with optimizations!")
+print("Pipelines ready with all optimizations!")
 
 STYLES = ["None", "Photorealistic", "Cinematic", "Anime", "Digital Art", 
           "Oil Painting", "Watercolor", "3D Render", "Fantasy", "Sci-Fi"]