VRAM Purge

ComfyUI-VRAM-Purge

Adds a Purge VRAM button to the ComfyUI main menu. It clears ComfyUI out of VRAM without restarting the server — unlike the restart-menu pack, this one just frees memory in place.

Why two buttons

A normal "unload all models" frees the model weights and PyTorch's cached allocations, but the CUDA context (a few hundred MB) stays resident for the life of the process. That residual is the "VRAM squatting" you want gone.

| Button | What it does | What stays in VRAM | | --- | --- | --- | | Purge VRAM | Unload all models + gc.collect() + torch.cuda.empty_cache() + ipc_collect() | GPU context (~0.2–1 GB) | | Purge VRAM (deep) | Everything above plus a device reset to destroy the GPU context | ~nothing (re-created on next generation) |

Use deep when you want ComfyUI completely out of VRAM (e.g. to hand the GPU to another app). The context is lazily re-created the next time you generate, so there is no penalty beyond a one-time re-init.

How it works

The endpoint (POST /vram_purge/purge) runs, in order:

1. comfy.model_management.unload_all_models() — unloads every loaded model (checkpoint, CLIP, VAE, diffusion model, LoRA, controlnet, adapter, upscaler, ...).
2. gc.collect() — drops lingering Python references to CUDA tensors.
3. comfy.model_management.soft_empty_cache(force=True) — returns PyTorch's cached CUDA blocks to the driver (synchronize + empty_cache + ipc_collect).
4. (deep only) a device reset via ctypes — releases the GPU context. NVIDIA: cuDeviceReset / cudaDeviceReset via libcuda.so.1. AMD: hipDeviceReset via libhip.so. Other backends: skipped (no C-level reset), so deep == soft there.
5. Reports a before/after VRAM snapshot. In deep mode it also reports a driver-level free/total reading (cuMemGetInfo_v2 / hipMemGetInfo) so you can verify the context is actually gone — querying PyTorch after the reset would re-create the context and inflate the number.

Install

Drop the folder into your custom_nodes/ directory:

custom_nodes/ComfyUI-VRAM-Purge/
    __init__.py
    server_purge.py
    pyproject.toml
    web/
        purge_menu.js

Restart ComfyUI once (to load the extension), then the two buttons appear in the main menu bar.

Notes / caveats

• Soft purge works on every backend ComfyUI supports (NVIDIA, AMD/ROCm, Apple MPS, Intel XPU, Ascend NPU, MLU, DirectML, CPU) — it only uses ComfyUI's backend-agnostic unload/cache functions.
• Deep mode's context reset is NVIDIA + AMD. It calls the vendor driver directly (libcuda.so.1 / libhip.so). On other backends (MPS, XPU, ...) the reset is skipped and deep behaves like soft; the response says so.
• Deep mode is best-effort. The device reset blocks until outstanding GPU work finishes. Click it when ComfyUI is idle (not mid-generation) to avoid a hang. If it fails, the response says so and nothing is broken — you just keep the context.
• No restart. The server keeps running; your workflow, queue, and loaded settings are untouched. Only GPU memory is released.
• The soft purge is the safe default and matches what ComfyUI's own "unload all models" machinery does.