You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Runtime: Model L at 4K native falls back to HIP (D3D12 arena measured in 2560 MiB), which keeps ~1.4 GiB until exit #182
Found while measuring #85 on an RX 9070 XT, using the real D3D12 backend.
With Model L at 4K native output, the process keeps about 1,419 MiB of local GPU memory after the upscaler is destroyed. This happens without any output resize, so #85 doesn't cause it. The amount is the same on every create/destroy cycle, so memory doesn't grow without limit, but it isn't returned until the process exits.
In the same harness, Model M returns to within a few MiB of baseline after destroy.
To do:
find which allocation outlives the destroy for Model L;
fix it;
record a create/destroy loop showing memory back to baseline for Model L and Model M, at 4K native and 4K Quality.
changed the title [-]Runtime: Model L keeps about 1.4 GiB of GPU memory after its upscaler is destroyed (4K native)[/-][+]Runtime: Model L at 4K native falls back to HIP (D3D12 arena measured in 2560 MiB), which keeps ~1.4 GiB until exit[/+]on Nov 14, 2026
The cause turned out not to be a Model L allocation. Model L at 4K native needs 2738 MiB, but the D3D12 executor measured its arena in the default 2560 MiB. That pass ran out, so the context fell back to the HIP executor, and HIP keeps its instance and pooled arena until the process exits. That accounts for the constant +1419 MiB. Fix and create/destroy record: #195.
Activity
Issue #182 record: re-run on the merged tree (main with the Model L e…
[-]Runtime: Model L keeps about 1.4 GiB of GPU memory after its upscaler is destroyed (4K native)[/-][+]Runtime: Model L at 4K native falls back to HIP (D3D12 arena measured in 2560 MiB), which keeps ~1.4 GiB until exit[/+]phantoslayer21 commentedon Nov 14, 2026
The cause turned out not to be a Model L allocation. Model L at 4K native needs 2738 MiB, but the D3D12 executor measured its arena in the default 2560 MiB. That pass ran out, so the context fell back to the HIP executor, and HIP keeps its instance and pooled arena until the process exits. That accounts for the constant +1419 MiB. Fix and create/destroy record: #195.
D3D12 executor: measure a network larger than the default arena again…