GPU & VRAM Management
Inspect GPU use, unload eligible models and plan a workload that fits
Last updated: 2026-10-27
Open the header's GPU status control for GPU information and memory tools. Settings > Performance separately controls the VRAM budget used when loading a supported pipeline. Settings > Vision controls Describe-model lifetime.
Loaded Models Panel
Loaded models lists known model/resource holders and their available state. Read the model name, active/busy status and footprint together. Some entries show a measured load-time footprint; others show offloaded, untracked or ncnn — untracked. An unavailable or zero attribution is not proof that a model consumes no GPU memory. Models/other work can own memory not represented by a simple sum of rows.
Total VRAM Bar
The device gauge shows overall card use, including other applications when the system can report it. It is different from the app's load-time model attribution. Numbers can update at different times; do not add overlapping readings into a new total or interpret a missing process reading as zero use.
Allocator pools
The Allocator pools section shows retained working-memory pools, separately from model rows. The visible Torch unused and CuPy pool readings describe different pools; they are not extra model entries.
Use Empty both pools when available to release reusable working memory without unloading models. Busy work can disable or refuse it. Inspect the returned result; Windows/device usage can remain high after unused blocks were released. A later task can grow the pools again. Repeated clicks do not guarantee more free memory.
Unload a Model
- Let work settle, then find the intended model in Loaded models.
- Use its Unload action when offered. For an active diffusion render, read the stop-and-unload confirmation; do not assume that unloading is harmless to running work.
- Wait for the operation's confirmed result and refreshed state before starting a conflicting job. Busy/loading/unknown owners can be ineligible for unload.
- Inspect the device reading as well as the row. A removed model entry does not guarantee that the driver has returned a particular number of bytes.
Unloading releases a loaded model; it does not uninstall its downloaded weights. Use Model Management only when you intend installation/repair/removal of those files.
Bulk Eviction
Unload All acts on eligible holders. Active work can require confirmation, unsafe/busy owners can be excluded and individual releases can fail. Read the combined result rather than assuming every resource was released. Other apps and driver allocations can remain, so this does not promise an empty GPU.
Preview Caches
Preview caches is separate from Loaded models and Allocator pools. Layers preview memory can include system RAM and estimated active texture memory. Parked work may keep reusable CPU previews while its active viewer resources have been released. Use its Clear only when the app offers it for parked entries; an active tool's draft must not be cleared as if it were disposable.
This temporary preview-memory action does not unload models or delete saved Layers/history. Disk cleanup belongs to Storage and Backup.
Performance Level Slider
In Settings > Performance, VRAM budget offers detected budget stops from Floor to Auto, or no slider when only one stop is available. Auto plans against the detected card. A lower stop limits the planner's budget and can change the selected precision/placement the next time a pipeline loads.
It is a planning cap, not a hard memory-allocation limit or a guarantee of fit. Lowering it is not a universal speed/quality ranking. Actual choices depend on the selected model, precision, hardware capability and offload support. Use Model Management and the selected entry's requirements instead of a fixed product-wide precision/VRAM table.
Automatic Eviction
The app can reclaim eligible idle models under memory pressure; it avoids assuming a busy owner is safe to release. Switching diffusion pipelines and loading another kind of task can have different outcomes. Keep resident for Vision still yields to pressure; see Vision and Describe.
If admission refuses a job, inspect the message, settle conflicting work and reduce size/add-ons or choose supported placement. Do not bypass ownership by clearing locks or killing unknown workers. A stop request is not proof of exit. Rendering and Recovery explains what to inspect before retrying.
VRAM Requirements by Pipeline
Requirements vary by model/variant and operation. Check System Requirements, the model's actual fit estimate and the source-specific guide. Keep enough system RAM for supported CPU offload too; offload does not make an unsupported model fit every GPU. For persistent errors, retain the displayed device/model/action and reviewed diagnostic report.