Skip to content

Make videos on the computer you already own. FreeVideo runs MiniMax H3 in as little as 8 GB of VRAM and 16 GB of RAM, and adapts its acceleration path to your hardware.

Image generation: a failed fall-back to streaming keeps the card able to retry it

Bowen Xue3a84876f7ffd
diff --git a/freevideo_engine/image_gen/image_runner.py b/freevideo_engine/image_gen/image_runner.py
index c90e914..b73d021 100644
--- a/freevideo_engine/image_gen/image_runner.py
+++ b/freevideo_engine/image_gen/image_runner.py
@@ -235,13 +235,14 @@ class ImageRunner:
"""A request no longer fits the whole model on the card: put the host copy back and stream."""
try:
with torch.no_grad():
- self.dit_stack = self.dit_stack.to_streamed()
- self.te_stack = self.te_stack.to_streamed()
+ if hasattr(self.dit_stack, 'to_streamed'):
+ self.dit_stack = self.dit_stack.to_streamed()
+ if hasattr(self.te_stack, 'to_streamed'):
+ self.te_stack = self.te_stack.to_streamed()
finally:
- # Each stack is usable on its own whichever one stopped, so the next request plans from stacks as they are.
- self.full_card = False
torch.cuda.empty_cache()
-
+ # Only now: a stack that stopped is still whole-model, so the next request tries again.
+ self.full_card = False
def budget(self):
if self.gpu_budget_bytes is not None:
return int(self.gpu_budget_bytes)