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.
Try cheaper Windows plans before host staging and flag idle VRAM (#32)
Bowen Xue6eb1b884f7dc
diff --git a/freevideo_engine/policy.py b/freevideo_engine/policy.pyindex 1dd4e43..050ecf5 100644--- a/freevideo_engine/policy.py+++ b/freevideo_engine/policy.py@@ -614,6 +614,19 @@ def choose(hardware: Hardware, *, vram_gib=None, ram_gib=None, attention='auto',# 10.00 GiB budget for a group measuring 10.53.while head > 4 and activation_bytes(head, effective_tokens) > gpu_budget:head //= 2+ # The GPU-output workspace below adds the wider group's surcharge to+ # its head-8 estimate, and a miss there moves attention outputs and+ # residuals to host memory. Narrow the group first; the second+ # transfer slot is given up below before anything is staged. A 32 GB+ # RTX 4080 SUPER at 1344x768x736 fit sixteen heads by 0.10 GiB, missed+ # their workspace by 4.77 and staged both, holding 14.5 of its 32 GB at+ # 360-595 s/step; eight heads needed 25.07 of its 30.40 GiB budget.+ # Without this, 26 to 30 s at 1344x768 staged and 31 to 38 s did not.+ if not small and not ampere:+ while head > 8 and (windows_gpu_output_workspace(effective_tokens, prefetch=False)+ + activation_bytes(head, effective_tokens)+ - activation_bytes(8, effective_tokens)) > gpu_budget:+ head //= 2else:# The widest group the request can afford at its own token count.# A fixed budget threshold cannot express this: 41472 tokens take@@ -835,6 +848,11 @@ def choose(hardware: Hardware, *, vram_gib=None, ram_gib=None, attention='auto',if host_need > ram_budget >= (50 - without_slot) * block_bytes + int(2.5 * GiB):prefetch = Falsewindows_workspace -= BLOCK_BYTES+ # The second transfer slot is worth about 2.5%; host attention outputs+ # and a host residual cost 15% to 2x. Give up the slot first.+ if prefetch and windows_workspace > gpu_budget >= windows_workspace - BLOCK_BYTES:+ prefetch = False+ windows_workspace -= BLOCK_BYTESresident = min(resident, max(0, int((gpu_budget - windows_workspace) / block_bytes)))# Do not enable FF recomputation merely to retain all host mappings.# Native 5060 Ti measurements show that doing so makes the same layerdiff --git a/freevideo_engine/support_report.py b/freevideo_engine/support_report.pyindex e3a5f4b..d1e920a 100644--- a/freevideo_engine/support_report.py+++ b/freevideo_engine/support_report.py@@ -88,6 +88,22 @@ def summarize(report, engine, encoding):hints.append(dict(kind='configuration', message='CPU attention/residual staging is active. Check live VRAM and ''policy decisions before attributing slow sampling to GPU compute.',message_zh='已启用 attention 输出或残差的 CPU 暂存;需结合当时剩余显存及策略理由判断采样变慢原因。'))+ passes = engine.get('sampling_passes')+ for index, row in enumerate(passes if isinstance(passes, list) else [], 1):+ # Staging is a capacity route. Several GiB of the budget left unused+ # means the plan was more conservative than the request needed, as in+ # issue #30: 13.5 of 30.3 GiB on a 32 GB card.+ staged = mapping(mapping(row).get('compute_configuration'))+ admission = mapping(mapping(row).get('pass_cache_admission'))+ budget, peak = number(admission.get('gpu_budget_bytes')), number(admission.get('measured_peak_reserved_bytes'))+ if (staged.get('attention_cpu_outputs') or staged.get('residual_offload')) and budget and peak \+ and budget - peak >= 4 * 2**30:+ hints.append(dict(kind='needs_investigation',+ message='Sampling pass %d staged in host memory while its peak used %.1f of a %.1f GiB VRAM budget. '+ 'The plan may be more conservative than this request needs; please report it with this file.'+ % (index, peak / 2**30, budget / 2**30),+ message_zh='第 %d 遍采样启用了内存暂存,但峰值只用了 %.1f/%.1f GiB 显存预算,策略可能过于保守,请附上本文件反馈。'+ % (index, peak / 2**30, budget / 2**30)))reserved = number(engine.get('torch_peak_reserved_bytes'))capacity = number(hardware.get('vram_total'))if hardware.get('system') == 'Windows' and reserved and capacity and reserved > capacity: