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.
Planner: fund the decoder from the request budget, not beside sampling-only workspace
Bowen Xueddf886dce61d
diff --git a/freevideo_engine/policy.py b/freevideo_engine/policy.pyindex db9fe9c..4607807 100644--- a/freevideo_engine/policy.py+++ b/freevideo_engine/policy.py@@ -913,7 +913,14 @@ def _choose(hardware: Hardware, *, vram_gib=None, ram_gib=None, attention='auto'# Scale the existing activation estimate with output and reference rows.# Keep the decoder's fixed workspace funded; neither weights nor its# temporal tile shrink in proportion to the number of output frames.- fixed_workspace = decoder_workspace(canvas)+ # The decoder runs after the transformer is released, and with it the+ # schedule tables, LoRA workspace and step cache: it is funded from the+ # request budget, as its resident blocks are below. gpu_budget has those+ # sampling allocations taken out already, so charge the decoder only what+ # they do not cover. Sampling still has to fit its own working set beside+ # them. Charging both refused full-size requests whose sampling and decode+ # each fit the card.+ fixed_workspace = max(0, decoder_workspace(canvas) - placement_workspace)# The staging estimate was measured on per-tensor FP8, whose FF keeps every# SwiGLU row for one scale (1.94 GiB at 72576 tokens). Int8 rows carry their# own scales and stash nothing. At 1344x768x243 on int8, with a 4.70 GiB@@ -948,7 +955,11 @@ def _choose(hardware: Hardware, *, vram_gib=None, ram_gib=None, attention='auto'# WDDM, and the two bounds above: the head group may not exceed its# measured requirement, and the no-weights path stays on the platform# whose margin was measured.- if not desktop and gpu_reserve_gib is None and free_gpu < 10 * GiB and gpu_budget < residual_need:+ # The trigger keeps the decoder's whole workspace. The credit above is for+ # admission; the cards this reduction already covered spend the headroom+ # keeping attention outputs on the GPU, which is worth more than the margin.+ if (not desktop and gpu_reserve_gib is None and free_gpu < 10 * GiB+ and gpu_budget < max(residual_need, decoder_workspace(canvas))):reserve_gpu = MIN_GPU_RESERVE_GIBrequest_gpu_budget = free_gpu - int(reserve_gpu * GiB)gpu_budget = request_gpu_budget - placement_workspace