Merge pull request #5165 from klimaleksus/fix-sequential-vae

Make VAE step sequential to prevent VRAM spikes, will fix #3059, #2082, #2561, #3462

Author: AUTOMATIC1111

modules/processing.py

@@ -530,8 +530,8 @@ def infotext(iteration=0, position_in_batch=0):
             with devices.autocast():
                 samples_ddim = p.sample(conditioning=c, unconditional_conditioning=uc, seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength, prompts=prompts)
 
-            samples_ddim = samples_ddim.to(devices.dtype_vae)
-            x_samples_ddim = decode_first_stage(p.sd_model, samples_ddim)
+            x_samples_ddim = [decode_first_stage(p.sd_model, samples_ddim[i:i+1].to(dtype=devices.dtype_vae))[0].cpu() for i in range(samples_ddim.size(0))]
+            x_samples_ddim = torch.stack(x_samples_ddim).float()
             x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
 
             del samples_ddim