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.

Keep the model warm after a successful video until something needs the memory, with no idle limit

Bowen Xue3ec8d4c0c80c
diff --git a/freevideo_engine/agent/guide.py b/freevideo_engine/agent/guide.py
index 863c650..a2d26c8 100644
--- a/freevideo_engine/agent/guide.py
+++ b/freevideo_engine/agent/guide.py
@@ -128,7 +128,7 @@ GUIDE = {
'without studio prompts. POST /jobs/{id}/cancel cancels only this job.',
'POST /jobs/{id}/resume continues failed, interrupted or cancelled generation; successful blocks stay. '
'refine_from is the separate preview second pass. Active or succeeded jobs cannot resume.',
- 'After a video the model stays on the graphics card and in memory for up to ten minutes, so the next '
+ 'After a video the model stays on the graphics card and in memory, so the next '
'video starts quickly; it goes at once when another program needs the memory. An agent that runs its own '
'model on the same card between videos (for example to check a result) calls POST /gpu/release (MCP freevideo_release_gpu) first '
'(it answers released=false while a video is running), or sets release_gpu_after=true on the generate job.',
diff --git a/freevideo_engine/resident_process.py b/freevideo_engine/resident_process.py
index d125b8d..6003da8 100644
--- a/freevideo_engine/resident_process.py
+++ b/freevideo_engine/resident_process.py
@@ -24,7 +24,7 @@ from .monitoring import save
ENV = 'FREEVIDEO_RESIDENT_SESSION'
GiB = 2**30
# A finished request keeps the worker for the next one while host memory is ample
-# (KEEP_IDLE_SECONDS idle at most), and it is given up at once when anything else
+# (for as long as nothing needs the memory, unless KEEP_IDLE_SECONDS sets a limit), and it is given up at once when anything else
# needs the memory: another program's GPU use grows, host memory runs low, an
# image or prompt model starts (those release it themselves), or the agent asks
# for the card back. A request that is not followed by a queued part and fails,
@@ -32,7 +32,7 @@ GiB = 2**30
# its way is held for HOLD_SECONDS.
HOLD_SECONDS = 90.
AGENT_HOLD_SECONDS = 30.
-KEEP_IDLE_SECONDS = 600.
+KEEP_IDLE_SECONDS = None
OTHER_GPU_GROWTH = 512 * 2**20
HOST_FLOOR = 4 * GiB
# Below this much host memory (free plus the idle worker's own), a kept worker's
@@ -396,7 +396,7 @@ class SessionOwner:
hold, next_part = KEEP_IDLE_SECONDS, 'warm'
else:
warm = False
- if not next_part or hold <= 0:
+ if not next_part or (hold is not None and hold <= 0):
return self.release('finished' if not next_part else 'no_hold')
reason = pressure(gpu, None) or self._tight()
if reason:
@@ -439,7 +439,7 @@ class SessionOwner:
def _watch(self, generation, since, hold, gpu, baseline):
while True:
time.sleep(.5)
- reason = ('idle' if time.monotonic() - since >= hold
+ reason = ('idle' if hold is not None and time.monotonic() - since >= hold
else pressure(gpu, baseline))
with self.mutex:
if generation != self.generation or self.active or self.process is None: