Skip to content

fix: 512px VAE decode OOM - effective half-size tiling + unconditional retry #1932

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Hmission wants to merge 1 commit into
base: master
Choose a base branch
from
Open

fix: 512px VAE decode OOM - effective half-size tiling + unconditional retry #1932

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions src/core/backend_fit.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -370,6 +370,13 @@ namespace sd::backend_fit {
retry_mode = tiling_params.enabled ? "spatial+temporal" : "temporal";
} else if (!tiling_params.enabled) {
tiling_params.enabled = true;
// 512px VAE decode memory fix: get_tile_sizes() defaults to
// rel_size=1.0 (factor branch wins), which yields a full-size tile
// (tiling effectively disabled) and can exceed device buffer limits
// (Adreno 740: 512px decode graph ~1.94GB). Half-relative tiles make
// tiling effective (~416MB), no quality impact (VAE tiling overlaps).
tiling_params.rel_size_x = 0.5f;
tiling_params.rel_size_y = 0.5f;
if (tiling_params.tile_size_x <= 0) {
tiling_params.tile_size_x = 256;
}
Expand Down
5 changes: 4 additions & 1 deletion src/stable-diffusion.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3085,8 +3085,11 @@ class StableDiffusionGGML {
auto latents = first_stage_model->diffusion_to_vae_latents(x);
auto decoded = first_stage_model->decode(n_threads, latents, vae_tiling_params, decode_video, circular_x, circular_y);
const bool prefer_temporal_tiling = decode_video && first_stage_model->can_temporal_tile_decode();
// 512px VAE decode memory fix: retry with tiling unconditionally on
// decode failure (OOM) instead of gating on --auto-fit (default off),
// so mobile GPUs with small buffers recover automatically. Only fires
// when decode actually failed, no side effects on the happy path.
while (decoded.empty() &&
auto_fit_enabled &&
sd::backend_fit::prepare_vae_decode_retry_tiling(vae_tiling_params, prefer_temporal_tiling)) {
first_stage_model->free_compute_buffer();
decoded = first_stage_model->decode(n_threads, latents, vae_tiling_params, decode_video, circular_x, circular_y);
Expand Down