Gradient accumulation, autocast fix, new latent sampling method, etc

modules/sd_hijack_checkpoint.py

@@ -0,0 +1,10 @@
+from torch.utils.checkpoint import checkpoint
+
+def BasicTransformerBlock_forward(self, x, context=None):
+    return checkpoint(self._forward, x, context)
+
+def AttentionBlock_forward(self, x):
+    return checkpoint(self._forward, x)
+
+def ResBlock_forward(self, x, emb):
+    return checkpoint(self._forward, x, emb)