Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-18 16:55:14 +02:00
Code Issues Packages Projects Releases Wiki Activity

sd: sync to master-504-636d3cb (#1969)

Browse Source 
* sd: sync to master-504-636d3cb

* sd: fix and simplify limit calculation

- restore the "arbitrarily high" 8192 limit, since it's used to turn
off the img_hard_limit (and if each side was always limited by 2048,
we wouldn't need hard_megapixel_res_limit)
- avoid changing the config cfg_square_limit during a generation
- apply the hard_megapixel_res_limit only in the configuration-changed
path, since the default path uses constants
- clean up comments

The calculation itself remains the same:
- default area limit: 832² for SD1.5/SD2, 1024² otherwise
- configured limit always between 64 and 2048
This commit is contained in:
Wagner Bruna
2026-02-13 21:12:08 -03:00
committed by GitHub
parent 7cccbb68b9
commit ae5183be10
20 changed files with 860 additions and 538 deletions
Download Patch File Download Diff File Expand all files Collapse all files
otherarch/sdcpp/stable-diffusion.cpp
+17 -8
View File
@@ -1702,7 +1702,7 @@ public:
if (vae_tiling_params.enabled) {
// split latent in 32x32 tiles and compute in several steps
auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
first_stage_model->compute(n_threads, in, true, &out, nullptr);
return first_stage_model->compute(n_threads, in, true, &out, nullptr);
};
silent_tiling(latents, result, get_vae_scale_factor(), 32, 0.5f, on_tiling);
@@ -1721,7 +1721,7 @@ public:
if (vae_tiling_params.enabled) {
// split latent in 64x64 tiles and compute in several steps
auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
tae_first_stage->compute(n_threads, in, true, &out, nullptr);
return tae_first_stage->compute(n_threads, in, true, &out, nullptr);
};
silent_tiling(latents, result, get_vae_scale_factor(), 64, 0.5f, on_tiling);
} else {
@@ -2690,7 +2690,7 @@ public:
LOG_DEBUG("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
first_stage_model->compute(n_threads, in, false, &out, work_ctx);
return first_stage_model->compute(n_threads, in, false, &out, work_ctx);
};
sd_tiling_non_square(x, result, vae_scale_factor, tile_size_x, tile_size_y, tile_overlap, on_tiling);
} else {
@@ -2701,7 +2701,7 @@ public:
if (vae_tiling_params.enabled && !encode_video) {
// split latent in 32x32 tiles and compute in several steps
auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
tae_first_stage->compute(n_threads, in, false, &out, nullptr);
return tae_first_stage->compute(n_threads, in, false, &out, nullptr);
};
sd_tiling(x, result, vae_scale_factor, 64, 0.5f, on_tiling);
} else {
@@ -2819,11 +2819,15 @@ public:
// split latent in 32x32 tiles and compute in several steps
auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
first_stage_model->compute(n_threads, in, true, &out, nullptr);
return first_stage_model->compute(n_threads, in, true, &out, nullptr);
};
sd_tiling_non_square(x, result, vae_scale_factor, tile_size_x, tile_size_y, tile_overlap, on_tiling);
} else {
first_stage_model->compute(n_threads, x, true, &result, work_ctx);
if (!first_stage_model->compute(n_threads, x, true, &result, work_ctx)) {
LOG_ERROR("Failed to decode latetnts");
first_stage_model->free_compute_buffer();
return nullptr;
}
}
first_stage_model->free_compute_buffer();
process_vae_output_tensor(result);
@@ -2831,11 +2835,15 @@ public:
if (vae_tiling_params.enabled) {
// split latent in 64x64 tiles and compute in several steps
auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
tae_first_stage->compute(n_threads, in, true, &out);
return tae_first_stage->compute(n_threads, in, true, &out);
};
sd_tiling(x, result, vae_scale_factor, 64, 0.5f, on_tiling);
} else {
tae_first_stage->compute(n_threads, x, true, &result);
if (!tae_first_stage->compute(n_threads, x, true, &result)) {
LOG_ERROR("Failed to decode latetnts");
tae_first_stage->free_compute_buffer();
return nullptr;
}
}
tae_first_stage->free_compute_buffer();
}
@@ -3625,6 +3633,7 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
ggml_free(work_ctx);
return nullptr;
}
memset(result_images, 0, batch_count * sizeof(sd_image_t));
for (size_t i = 0; i < decoded_images.size(); i++) {
result_images[i].width = width;