Latest origin/main has increased VRAM usage

[Skelp]

Previously easy-to-do for my RTX 4070 Ti SUPER (16GB), I can't seem to generate any tracks anymore without hitting OOM.
Same settings used before and after updating and re-building.
1.7B LM, SFT DiT, compilation deselected, 215s of generation, default batch size (2), and everything else (except prompt and lyrics) kept at default.
Seems to occur right after generating the audio codes

2026-02-09 11:36:28.862 | INFO     | acestep.llm_inference:generate_with_stop_condition:1312 - Batch Phase 2 completed in 11.31s. Generated codes: [1075, 1075]
2026-02-09 11:36:28.862 | INFO     | acestep.handler:generate_music:3265 - [generate_music] Starting generation...
2026-02-09 11:36:28.862 | INFO     | acestep.handler:generate_music:3268 - [generate_music] Preparing inputs...
2026-02-09 11:36:28.947 | INFO     | acestep.handler:_load_model_context:902 - [_load_model_context] Loading vae to cuda
2026-02-09 11:36:29.197 | INFO     | acestep.handler:_load_model_context:915 - [_load_model_context] Loaded vae to cuda in 0.2496s
2026-02-09 11:36:29.197 | INFO     | acestep.handler:_prepare_batch:2117 - [generate_music] Decoding audio codes for item 0...
2026-02-09 11:36:29.197 | INFO     | acestep.handler:_load_model_context:902 - [_load_model_context] Loading model to cuda
2026-02-09 11:36:29.486 | INFO     | acestep.handler:_load_model_context:921 - [_load_model_context] Offloading vae to CPU
2026-02-09 11:36:29.783 | INFO     | acestep.handler:_load_model_context:935 - [_load_model_context] Offloaded vae to CPU in 0.2977s
2026-02-09 11:36:29.785 | ERROR    | acestep.handler:generate_music:3564 - [generate_music] Generation failed

[ChuxiJ]

We are working on a stable version to ensure that it tests correctly on every device. If it's convenient for you, please use the latest code to check the actual VRAM usage you need. You can adjust your tier in the GPU config section to get a better experience.

[Skelp]

Tier6 was and still is what is being reported. Even with batch size of 1 it now goes OOM. I don't know what you refer to by "GPU config section" though.

[ChuxiJ]

try: #417

[Skelp]

Okay, that did reduce me to Tier6a and fixed the OOM, but also locked me in a batch size of 1. Previously, before the OOM issues occured, I was easily able to run generation with a batch size of 2. This is still what I would consider a regression.

[ChuxiJ]

Could you tell me how you use it? I’m currently doing a simulation test with a 16GB card and found that batch size 2 works. I’d like to know how you launch it, so I can adjust some hyperparameters accordingly.

[Skelp]

uv run acestep
Everything else as described in the initial issue description. Those settings were used prior to updating acestep1.5 and after. Before it worked with batch size 2, now only with 1.