Hi everyone. I'm still getting my head around how NemoClaw handles the underlying hardware, and this particular point has been bugging me. I've been reading about GPU memory isolation between tenants, and I keep seeing mentions of "VRAM residue" and that weights can't be purged "quickly."

Could someone help break down why that is, in simpler terms? I think I understand that when one tenant's model finishes, that memory should be freed for the next user. But from what I gather, it's not as simple as just deleting a file from RAM.

My basic understanding is that the GPU memory (VRAM) is managed by the driver and the CUDA runtime. When a model loads its weights, that memory is allocated and filled. To truly purge it, that specific memory space needs to be overwritten with something else—like zeros or another tenant's data. Until that happens, the old data just sits there. Is the delay because the system is waiting for a new allocation to overwrite it, or is there an active "clearing" step that's surprisingly slow?

Also, I'm a bit unclear on what role NVIDIA's hardware features (like MIG or the GPU's memory management units) actually play here. Do they prevent one tenant from *actively* reading another's leftover weights, but don't necessarily force a speedy wipe?

Thanks so much for any insight. This forum has been incredibly helpful as I try to learn.