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.

Saved AdaLN tables are checked for NaN and infinity with an integer mask on the half-precision...

Bowen Xue07a85becc401
diff --git a/freevideo_engine/adaln.py b/freevideo_engine/adaln.py
index 28c2e8a..4fbbea8 100644
--- a/freevideo_engine/adaln.py
+++ b/freevideo_engine/adaln.py
@@ -12,6 +12,20 @@ from safetensors.torch import load_file, save_file
import torch
+def all_finite(tensor):
+ """Whether every element is finite; reads half-precision bits instead of calling isfinite.
+
+ A bfloat16 or float16 table is finite exactly when its exponent field is not
+ all ones, so one integer mask answers the question about ten times faster on
+ the CPU than isfinite() on those dtypes.
+ """
+ mask = {torch.bfloat16: 0x7F80, torch.float16: 0x7C00}.get(tensor.dtype)
+ if mask is None:
+ return bool(torch.isfinite(tensor).all())
+ bits = tensor.reshape(-1).view(torch.int16)
+ return not bool((bits & mask).eq(mask).any())
+
+
class ScheduleCursor:
def __init__(self, timesteps):
self.timesteps = timesteps
@@ -174,7 +188,7 @@ class TableCache:
# These are small immutable constants; their producer's GPU/runtime is
# deliberately irrelevant. Validate before transferring to this device.
state = load_file(path, device=self.device)
- if any(not bool(torch.isfinite(tensor).all()) for tensor in state.values()):
+ if any(not all_finite(tensor) for tensor in state.values()):
raise ValueError('Nonfinite AdaLN model asset')
return [state[f'step_{i}'].chunk(6, dim=-1) for i in range(steps)]
@@ -199,7 +213,7 @@ class TableCache:
path = self.root / f'{index:02d}.safetensors'
temporary = path.with_suffix('.partial')
state = {name: tensor.cpu() for name, tensor in module.state_dict().items()}
- if any(not bool(torch.isfinite(tensor).all()) for tensor in state.values()):
+ if any(not all_finite(tensor) for tensor in state.values()):
raise ValueError('Cannot save nonfinite AdaLN constants')
save_file(state, temporary)
row = {'bytes': temporary.stat().st_size, 'sha256': assets.file_hash(temporary)}