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.

Linux drops an idle text encoder without first copying its weights to memory

Bowen Xuefe022558067c
diff --git a/freevideo_engine/resident_models.py b/freevideo_engine/resident_models.py
index a85c9d4..c6fb185 100644
--- a/freevideo_engine/resident_models.py
+++ b/freevideo_engine/resident_models.py
@@ -2,6 +2,7 @@
import gc
import json
from pathlib import Path
+import sys
import time
import torch
@@ -95,19 +96,21 @@ def parked_bytes(model):
def release_encoder(clip, manager):
- """Discard our Windows mapped encoder without a full GPU-to-CPU copy.
+ """Discard our mapped encoder without a full GPU-to-CPU copy.
Native unpatching still releases pins, hooks and loaded-model registrations.
Once the bank drops its references, unused tensors can die where they are;
any other owner keeps valid tensors. Future encoding uses the mmap loader.
"""
from .encoder_checkpoint import load_clip_model_patcher
- from .system import windows
def owned(patcher):
factory = getattr(patcher, 'cached_patcher_init', None)
return isinstance(factory, (tuple, list)) and bool(factory) and factory[0] is load_clip_model_patcher
patcher = getattr(clip, 'patcher', None)
- targets = [patcher] if windows() and owned(patcher) else []
+ # A native unpatch copies every weight to the CPU first: 13 GiB for the video text encoder,
+ # enough to exceed a machine's memory in the moment it was meant to make room. The Mac
+ # releases its encoder its own way (macos_encoder) and keeps the native path here.
+ targets = [patcher] if owned(patcher) and sys.platform != 'darwin' else []
if targets:
for loaded in manager.current_loaded_models:
candidate = loaded.model