Skip to content

fix VAE tiling for Qwen Image #873

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

Merged
merged 20 commits into from
Oct 12, 2025
Merged

fix VAE tiling for Qwen Image #873

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
20 commits
Select commit Hold shift + click to select a range
f88daa5
add qwen tokenizer
leejet Sep 20, 2025
fe4e731
add qwen2.5 vl support
leejet Sep 20, 2025
d8d4c26
mv qwen.hpp -> qwenvl.hpp
leejet Sep 21, 2025
d232509
add qwen image model
leejet Sep 21, 2025
cf19c6e
add qwen image t2i pipeline
leejet Sep 22, 2025
477911f
fix qwen image flash attn
leejet Sep 22, 2025
feb0279
add qwen image i2i pipline
leejet Sep 22, 2025
5af0bb0
change encoding of vocab_qwen.hpp to utf8
leejet Sep 22, 2025
a8d3aa0
Merge branch 'master' into qwen_image
leejet Sep 24, 2025
a3a2b2d
fix get_first_stage_encoding
leejet Sep 24, 2025
178a415
Merge branch 'master' into qwen_image
leejet Sep 25, 2025
94f4f29
Merge branch 'master' into qwen_image
leejet Sep 25, 2025
d19d4a5
Merge branch 'master' into qwen_image
leejet Oct 10, 2025
6ea2a75
apply jeffbolz f32 patch
leejet Oct 10, 2025
91dca1a
fix VAE tiling for Qwen Image
wbruna Oct 4, 2025
f1db617
fix encoding VAE tiling for Qwen Image
wbruna Oct 9, 2025
acbeb05
Merge branch 'master' into fix_vae_tiling_qwen
leejet Oct 12, 2025
9719dee
format code
leejet Oct 12, 2025
929f7ec
Merge branch 'master' into fix_vae_tiling_qwen
leejet Oct 12, 2025
3ad6d6f
fix vae_encode
leejet Oct 12, 2025
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
54 changes: 30 additions & 24 deletions ggml_extend.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -483,12 +483,15 @@ __STATIC_INLINE__ void ggml_split_tensor_2d(struct ggml_tensor* input,
int64_t width = output->ne[0];
int64_t height = output->ne[1];
int64_t channels = output->ne[2];
int64_t ne3 = output->ne[3];
GGML_ASSERT(input->type == GGML_TYPE_F32 && output->type == GGML_TYPE_F32);
for (int iy = 0; iy < height; iy++) {
for (int ix = 0; ix < width; ix++) {
for (int k = 0; k < channels; k++) {
float value = ggml_tensor_get_f32(input, ix + x, iy + y, k);
ggml_tensor_set_f32(output, value, ix, iy, k);
for (int l = 0; l < ne3; l++) {
float value = ggml_tensor_get_f32(input, ix + x, iy + y, k, l);
ggml_tensor_set_f32(output, value, ix, iy, k, l);
}
}
}
}
Expand All @@ -511,6 +514,7 @@ __STATIC_INLINE__ void ggml_merge_tensor_2d(struct ggml_tensor* input,
int64_t width = input->ne[0];
int64_t height = input->ne[1];
int64_t channels = input->ne[2];
int64_t ne3 = input->ne[3];

int64_t img_width = output->ne[0];
int64_t img_height = output->ne[1];
Expand All @@ -519,24 +523,26 @@ __STATIC_INLINE__ void ggml_merge_tensor_2d(struct ggml_tensor* input,
for (int iy = y_skip; iy < height; iy++) {
for (int ix = x_skip; ix < width; ix++) {
for (int k = 0; k < channels; k++) {
float new_value = ggml_tensor_get_f32(input, ix, iy, k);
if (overlap_x > 0 || overlap_y > 0) { // blend colors in overlapped area
float old_value = ggml_tensor_get_f32(output, x + ix, y + iy, k);

const float x_f_0 = (overlap_x > 0 && x > 0) ? (ix - x_skip) / float(overlap_x) : 1;
const float x_f_1 = (overlap_x > 0 && x < (img_width - width)) ? (width - ix) / float(overlap_x) : 1;
const float y_f_0 = (overlap_y > 0 && y > 0) ? (iy - y_skip) / float(overlap_y) : 1;
const float y_f_1 = (overlap_y > 0 && y < (img_height - height)) ? (height - iy) / float(overlap_y) : 1;

const float x_f = std::min(std::min(x_f_0, x_f_1), 1.f);
const float y_f = std::min(std::min(y_f_0, y_f_1), 1.f);

ggml_tensor_set_f32(
output,
old_value + new_value * ggml_smootherstep_f32(y_f) * ggml_smootherstep_f32(x_f),
x + ix, y + iy, k);
} else {
ggml_tensor_set_f32(output, new_value, x + ix, y + iy, k);
for (int l = 0; l < ne3; l++) {
float new_value = ggml_tensor_get_f32(input, ix, iy, k, l);
if (overlap_x > 0 || overlap_y > 0) { // blend colors in overlapped area
float old_value = ggml_tensor_get_f32(output, x + ix, y + iy, k, l);

const float x_f_0 = (overlap_x > 0 && x > 0) ? (ix - x_skip) / float(overlap_x) : 1;
const float x_f_1 = (overlap_x > 0 && x < (img_width - width)) ? (width - ix) / float(overlap_x) : 1;
const float y_f_0 = (overlap_y > 0 && y > 0) ? (iy - y_skip) / float(overlap_y) : 1;
const float y_f_1 = (overlap_y > 0 && y < (img_height - height)) ? (height - iy) / float(overlap_y) : 1;

const float x_f = std::min(std::min(x_f_0, x_f_1), 1.f);
const float y_f = std::min(std::min(y_f_0, y_f_1), 1.f);

ggml_tensor_set_f32(
output,
old_value + new_value * ggml_smootherstep_f32(y_f) * ggml_smootherstep_f32(x_f),
x + ix, y + iy, k, l);
} else {
ggml_tensor_set_f32(output, new_value, x + ix, y + iy, k, l);
}
}
}
}
Expand Down Expand Up @@ -852,8 +858,8 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
}

struct ggml_init_params params = {};
params.mem_size += input_tile_size_x * input_tile_size_y * input->ne[2] * sizeof(float); // input chunk
params.mem_size += output_tile_size_x * output_tile_size_y * output->ne[2] * sizeof(float); // output chunk
params.mem_size += input_tile_size_x * input_tile_size_y * input->ne[2] * input->ne[3] * sizeof(float); // input chunk
params.mem_size += output_tile_size_x * output_tile_size_y * output->ne[2] * output->ne[3] * sizeof(float); // output chunk
params.mem_size += 3 * ggml_tensor_overhead();
params.mem_buffer = NULL;
params.no_alloc = false;
Expand All @@ -868,8 +874,8 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
}

// tiling
ggml_tensor* input_tile = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, input_tile_size_x, input_tile_size_y, input->ne[2], 1);
ggml_tensor* output_tile = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, output_tile_size_x, output_tile_size_y, output->ne[2], 1);
ggml_tensor* input_tile = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, input_tile_size_x, input_tile_size_y, input->ne[2], input->ne[3]);
ggml_tensor* output_tile = ggml_new_tensor_4d(tiles_ctx, GGML_TYPE_F32, output_tile_size_x, output_tile_size_y, output->ne[2], output->ne[3]);
int num_tiles = num_tiles_x * num_tiles_y;
LOG_INFO("processing %i tiles", num_tiles);
pretty_progress(0, num_tiles, 0.0f);
Expand Down
15 changes: 12 additions & 3 deletions stable-diffusion.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1440,10 +1440,19 @@ class StableDiffusionGGML {
if (vae_tiling_params.enabled && !encode_video) {
// TODO wan2.2 vae support?
int C = sd_version_is_dit(version) ? 16 : 4;
if (!use_tiny_autoencoder) {
C *= 2;
int ne2;
int ne3;
if (sd_version_is_qwen_image(version)) {
ne2 = 1;
ne3 = C*x->ne[3];
} else {
if (!use_tiny_autoencoder) {
C *= 2;
}
ne2 = C;
ne3 = x->ne[3];
}
result = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, C, x->ne[3]);
result = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, ne2, ne3);
}

if (sd_version_is_qwen_image(version)) {
Expand Down
Loading