Skip to content

perf: accelerate VAE direct 3D convolutions #1996

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
leejet merged 2 commits into from
Sep 19, 2026
Merged

perf: accelerate VAE direct 3D convolutions #1996

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0fc1ac9
feat: extend --vae-conv-direct to 3D convolutions
leejet Sep 18, 2026
e2d8004
perf: update ggml for faster 3d direct convolutions
leejet Sep 19, 2026
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
2 changes: 1 addition & 1 deletion examples/common/common.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -624,7 +624,7 @@ ArgOptions SDContextParams::get_options() {
true, &diffusion_conv_direct},
{"",
"--vae-conv-direct",
"use ggml_conv2d_direct in the vae model",
"use direct 2D and 3D convolutions in the vae model",
true, &vae_conv_direct},
};

Expand Down
2 changes: 1 addition & 1 deletion ggml
Open in desktop
Submodule ggml updated 4 files
+338 −0 src/ggml-cuda/conv3d.cu
+8 −0 src/ggml-cuda/conv3d.cuh
+8 −0 src/ggml-cuda/ggml-cuda.cu
+5 −0 tests/test-backend-ops.cpp
9 changes: 7 additions & 2 deletions src/core/ggml_extend.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -452,8 +452,13 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
int d0,
int d1,
int d2,
bool force_prec_f32) {
if (force_prec_f32) {
bool force_prec_f32,
bool direct) {
if (direct) {
int64_t OC = w->ne[3] / IC;
int64_t N = x->ne[3] / IC;
x = ggml_conv_3d_direct(ctx, w, x, s0, s1, s2, p0, p1, p2, d0, d1, d2, (int)IC, (int)N, (int)OC);
} else if (force_prec_f32) {
ggml_tensor* im2col = ggml_im2col_3d(ctx, w, x, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2, w->type);

int64_t OC = w->ne[3] / IC;
Expand Down
3 changes: 2 additions & 1 deletion src/core/ggml_extend.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,7 +153,8 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
int d0 = 1,
int d1 = 1,
int d2 = 1,
bool force_prec_f32 = false);
bool force_prec_f32 = false,
bool direct = false);

// w: [OC，IC, KD, 1 * 1]
// x: [N, IC, ID, IH*IW]
Expand Down
1 change: 1 addition & 0 deletions src/core/ggml_runner.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -530,6 +530,7 @@ GGMLRunnerContext GGMLRunner::get_context() {
runner_ctx.linear_scale = linear_scale;
runner_ctx.attn_scale = attn_scale;
runner_ctx.conv2d_direct_enabled = conv2d_direct_enabled;
runner_ctx.conv3d_direct_enabled = conv3d_direct_enabled;
runner_ctx.circular_x_enabled = circular_x_enabled;
runner_ctx.circular_y_enabled = circular_y_enabled;
runner_ctx.weight_adapter = weight_adapter;
Expand Down
6 changes: 6 additions & 0 deletions src/core/ggml_runner.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -71,6 +71,7 @@ struct GGMLRunnerContext {
float linear_scale = 0.f;
float attn_scale = 0.f;
bool conv2d_direct_enabled = false;
bool conv3d_direct_enabled = false;
bool circular_x_enabled = false;
bool circular_y_enabled = false;
ggml_tensor* ip_context = nullptr;
Expand Down Expand Up @@ -178,6 +179,7 @@ struct GGMLRunner {
float linear_scale = 0.f;
float attn_scale = 0.f;
bool conv2d_direct_enabled = false;
bool conv3d_direct_enabled = false;
bool circular_x_enabled = false;
bool circular_y_enabled = false;

Expand Down Expand Up @@ -346,6 +348,10 @@ struct GGMLRunner {
conv2d_direct_enabled = enabled;
}

void set_conv3d_direct_enabled(bool enabled) {
conv3d_direct_enabled = enabled;
}

void set_circular_axes(bool circular_x, bool circular_y) {
circular_x_enabled = circular_x;
circular_y_enabled = circular_y;
Expand Down
2 changes: 1 addition & 1 deletion src/model/common/ggml_block.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -728,7 +728,7 @@ class Conv3d : public UnaryBlock {
std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
std::get<2>(padding), std::get<1>(padding), std::get<0>(padding),
std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation),
force_prec_f32);
force_prec_f32, ctx->conv3d_direct_enabled);
}
};

Expand Down
3 changes: 2 additions & 1 deletion src/model/vae/wan_vae.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -78,7 +78,8 @@ namespace WAN {
return ggml_ext_conv_3d(ctx->ggml_ctx, ctx->backend, x, w, b, in_channels,
std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
0, 0, 0,
std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation));
std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation),
false, ctx->conv3d_direct_enabled);
}
};

Expand Down
4 changes: 3 additions & 1 deletion src/pipeline/model_builders.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -584,10 +584,12 @@ namespace sd::model_builders {
}

if (sd_ctx_params->vae_conv_direct) {
LOG_INFO("Using Conv2d direct in the vae model");
LOG_INFO("Using Conv2d/Conv3d direct in the vae model");
result.vae->set_conv2d_direct_enabled(true);
result.vae->set_conv3d_direct_enabled(true);
if (result.preview) {
result.preview->set_conv2d_direct_enabled(true);
result.preview->set_conv3d_direct_enabled(true);
}
}
if (result.vae) {
Expand Down
Loading