DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-11 19:36:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

refactor: unify ggml operation helpers around runner context (#2314)

Browse Source
This commit is contained in:
leejet authored and GitHub committed 2026-11-10 01:58:04 +08:00
1 parent 73fef08224
commit a3c49b13be
76 files changed
+1983 -1831
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/model/upscaler/esrgan.hpp
+2 -2
View File
@@ -127,7 +127,7 @@ public:
x_cat = ggml_concat(ctx->ggml_ctx, x_cat, x4, 2);
auto x5 = conv5->forward(ctx, x_cat);
x5 = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx->ggml_ctx, x5, 0.2f), x);
x5 = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx, x5, 0.2f), x);
return x5;
}
};
@@ -152,7 +152,7 @@ public:
out = rdb2->forward(ctx, out);
out = rdb3->forward(ctx, out);
out = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx->ggml_ctx, out, 0.2f), x);
out = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx, out, 0.2f), x);
return out;
}
};
src/model/upscaler/flash_vsr.hpp
+1 -1
View File
@@ -146,7 +146,7 @@ namespace RelayVSR {
auto y = attention->forward(ctx, attention_norm->forward(ctx, x), qpe, kpe, local, name, condition_slot, start_slot, end_slot, history);
x = ggml_add(c, x, y);
y = feed_forward_gate_up_proj->forward(ctx, feed_forward_norm->forward(ctx, x));
auto parts = ggml_ext_chunk(c, y, 2, 0);
auto parts = ggml_ext_chunk(ctx, y, 2, 0);
y = ggml_mul(c, ggml_silu(c, parts[0]), parts[1]);
return ggml_add(c, x, feed_forward_down->forward(ctx, y));
}
src/model/upscaler/h3_latent_upscaler.hpp
+7 -7
View File
@@ -195,8 +195,8 @@ namespace H3VUpsampler {
ggml_tensor* emb_silu = ggml_silu(ctx->ggml_ctx, emb);
ggml_tensor* emb_out = emb_layers->forward(ctx, emb_silu);
ggml_tensor* emb_1d = ggml_reshape_1d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, emb_out), 2 * channels);
ggml_tensor* scale = ggml_reshape_4d(ctx->ggml_ctx, ggml_ext_slice(ctx->ggml_ctx, emb_1d, 0, 0, channels), 1, 1, 1, channels);
ggml_tensor* shift = ggml_reshape_4d(ctx->ggml_ctx, ggml_ext_slice(ctx->ggml_ctx, emb_1d, 0, channels, 2 * channels), 1, 1, 1, channels);
ggml_tensor* scale = ggml_reshape_4d(ctx->ggml_ctx, ggml_ext_slice(ctx, emb_1d, 0, 0, channels), 1, 1, 1, channels);
ggml_tensor* shift = ggml_reshape_4d(ctx->ggml_ctx, ggml_ext_slice(ctx, emb_1d, 0, channels, 2 * channels), 1, 1, 1, channels);
ggml_tensor* gain = ggml_scale_bias(ctx->ggml_ctx, scale, 1.f, 1.f);
h = out_norm->forward(ctx, h);
@@ -236,7 +236,7 @@ namespace H3VUpsampler {
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
ggml_tensor* w = params["weight"];
if (ctx->weight_adapter) {
w = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, w, prefix + "weight");
w = ctx->weight_adapter->patch_weight(ctx, w, prefix + "weight");
}
ggml_tensor* w32 = ggml_cast(ctx->ggml_ctx, w, GGML_TYPE_F32);
ggml_tensor* xa = x;
@@ -250,7 +250,7 @@ namespace H3VUpsampler {
// ggml_pad only pads the end; concatenate zeros for symmetric padding.
ggml_tensor* xp = xa;
if (pad > 0) {
ggml_tensor* zero = ggml_scale(ctx->ggml_ctx, ggml_ext_slice(ctx->ggml_ctx, xa, 2, 0, 1), 0.0f);
ggml_tensor* zero = ggml_scale(ctx->ggml_ctx, ggml_ext_slice(ctx, xa, 2, 0, 1), 0.0f);
ggml_tensor* lead = zero;
for (int i = 1; i < pad; ++i) {
lead = ggml_concat(ctx->ggml_ctx, lead, zero, 2);
@@ -261,15 +261,15 @@ namespace H3VUpsampler {
ggml_tensor* acc = nullptr;
for (int j = 0; j < kernel; ++j) {
ggml_tensor* xs = ggml_ext_slice(ctx->ggml_ctx, xp, 2, j, j + T);
ggml_tensor* wj = ggml_ext_slice(ctx->ggml_ctx, w32, 2, j, j + 1); // [1, 1, 1, C]
ggml_tensor* xs = ggml_ext_slice(ctx, xp, 2, j, j + T);
ggml_tensor* wj = ggml_ext_slice(ctx, w32, 2, j, j + 1); // [1, 1, 1, C]
ggml_tensor* term = ggml_mul(ctx->ggml_ctx, xs, wj);
acc = acc == nullptr ? term : ggml_add(ctx->ggml_ctx, acc, term);
}
if (has_bias) {
ggml_tensor* b = params["bias"];
if (ctx->weight_adapter) {
b = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, b, prefix + "bias");
b = ctx->weight_adapter->patch_weight(ctx, b, prefix + "bias");
}
acc = ggml_add(ctx->ggml_ctx, acc, ggml_reshape_4d(ctx->ggml_ctx, b, 1, 1, 1, channels));
}
src/model/upscaler/ltx_latent_upscaler.hpp
+17 -17
View File
@@ -158,15 +158,15 @@ namespace LTXVUpsampler {
const int64_t W = x->ne[0];
const int64_t H = x->ne[1];
const int64_t T = x->ne[2];
x = ggml_ext_cont(ctx->ggml_ctx, x);
x = ggml_ext_cont(ctx, x);
x = ggml_reshape_4d(ctx->ggml_ctx, x, W, H * T, num_channels, 1);
x = ggml_group_norm(ctx->ggml_ctx, x, num_groups, eps);
ggml_tensor* weight = params["weight"];
ggml_tensor* bias = params["bias"];
if (ctx->weight_adapter) {
weight = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, weight, prefix + "weight");
bias = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, bias, prefix + "bias");
weight = ctx->weight_adapter->patch_weight(ctx, weight, prefix + "weight");
bias = ctx->weight_adapter->patch_weight(ctx, bias, prefix + "bias");
}
weight = ggml_reshape_4d(ctx->ggml_ctx, weight, 1, 1, num_channels, 1);
bias = ggml_reshape_4d(ctx->ggml_ctx, bias, 1, 1, num_channels, 1);
@@ -218,9 +218,9 @@ namespace LTXVUpsampler {
int64_t w = x->ne[0];
GGML_ASSERT(x->ne[2] % (upscale_factor * upscale_factor) == 0);
// x: [b*f, c*p1*p2, h, w] -> [b*f, c, h*p1, w*p2]
x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 2, 0, 1, 3)); // [b*f, h, w, c*p1*p2]
x = ggml_reshape_3d(ctx->ggml_ctx, x, x->ne[0], x->ne[1] * x->ne[2], x->ne[3]); // [b*f, h*w, c*p1*p2]
return DiT::unpatchify(ctx->ggml_ctx, x, h, w, upscale_factor, upscale_factor, true);
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 2, 0, 1, 3)); // [b*f, h, w, c*p1*p2]
x = ggml_reshape_3d(ctx->ggml_ctx, x, x->ne[0], x->ne[1] * x->ne[2], x->ne[3]); // [b*f, h*w, c*p1*p2]
return DiT::unpatchify(ctx, x, h, w, upscale_factor, upscale_factor, true);
}
};
@@ -241,9 +241,9 @@ namespace LTXVUpsampler {
const int64_t C = x->ne[3] / upscale_factor;
// x: [b, c*p, f, h, w] -> [b, c, f*p, h, w]
x = ggml_ext_cont(ctx->ggml_ctx, x);
x = ggml_ext_cont(ctx, x);
x = ggml_reshape_4d(ctx->ggml_ctx, x, W * H, F, upscale_factor, C);
x = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));
x = ggml_ext_cont(ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));
return ggml_reshape_4d(ctx->ggml_ctx, x, W, H, F * upscale_factor, C);
}
};
@@ -324,11 +324,11 @@ namespace LTXVUpsampler {
auto blur_down = std::dynamic_pointer_cast<BlurDownsample>(blocks["blur_down"]);
// rearrange(x, "b c f h w -> (b f) c h w")
x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 1, 3, 2));
x = conv->forward(ctx, x);
x = pixel_shuffle->forward(ctx, x);
x = blur_down->forward(ctx, x);
return ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
return ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 1, 3, 2));
}
};
@@ -403,18 +403,18 @@ namespace LTXVUpsampler {
} else if (config.temporal_upsample) {
auto upsample_conv = std::dynamic_pointer_cast<Conv3d>(blocks["upsampler.0"]);
auto pixel_shuffle = std::dynamic_pointer_cast<TemporalPixelShuffleND>(blocks["upsampler.1"]);
x = upsample_conv->forward(ctx, x); // [b, c*2, f, h, w]
x = pixel_shuffle->forward(ctx, x); // [b, c, f*2, h, w]
x = ggml_ext_slice(ctx->ggml_ctx, x, 2, 1, x->ne[2]); // x[:, :, 1:, :, :]
x = upsample_conv->forward(ctx, x); // [b, c*2, f, h, w]
x = pixel_shuffle->forward(ctx, x); // [b, c, f*2, h, w]
x = ggml_ext_slice(ctx, x, 2, 1, x->ne[2]); // x[:, :, 1:, :, :]
} else {
auto upsample_conv = std::dynamic_pointer_cast<Conv2d>(blocks["upsampler.0"]);
auto pixel_shuffle = std::dynamic_pointer_cast<PixelShuffleND>(blocks["upsampler.1"]);
// rearrange(x, "b c f h w -> (b f) c h w"),
x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2)); // [b*f, c, h, w]
x = upsample_conv->forward(ctx, x); // [b*f, c*4, h, w]
x = pixel_shuffle->forward(ctx, x); // [b*f, c, h*2, w*2]
x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2)); // [b*c, f, h, w]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 1, 3, 2)); // [b*f, c, h, w]
x = upsample_conv->forward(ctx, x); // [b*f, c*4, h, w]
x = pixel_shuffle->forward(ctx, x); // [b*f, c, h*2, w*2]
x = ggml_ext_cont(ctx, ggml_ext_torch_permute(ctx, x, 0, 1, 3, 2)); // [b*c, f, h, w]
}
sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.spatial_up", "x");
src/model/upscaler/mosr.hpp
+7 -8
View File
@@ -118,9 +118,9 @@ public:
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
// [N, C, H, W] -> [N, H, W, C] so the norm reduces over channels.
x = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 2, 0, 3));
x = ggml_ext_layer_norm(ctx->ggml_ctx, x, params["weight"], params["bias"], 1e-6f);
return ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 2, 0, 1, 3));
x = ggml_ext_cont(ctx, ggml_permute(ctx->ggml_ctx, x, 1, 2, 0, 3));
x = ggml_ext_layer_norm(ctx, x, params["weight"], params["bias"], 1e-6f);
return ggml_ext_cont(ctx, ggml_permute(ctx->ggml_ctx, x, 2, 0, 1, 3));
}
};
@@ -141,9 +141,8 @@ public:
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
const int pad = kernel_size / 2;
x = ggml_conv_2d_dw_direct(ctx->ggml_ctx, params["weight"], ggml_ext_cont(ctx->ggml_ctx, x), 1, 1, pad, pad, 1, 1);
auto b = ggml_reshape_4d(ctx->ggml_ctx, params["bias"], 1, 1, channels, 1);
return ggml_add_inplace(ctx->ggml_ctx, x, b);
return ggml_ext_conv_2d_grouped(ctx, ggml_ext_cont(ctx, x), params["weight"], params["bias"], channels,
1, 1, pad, pad, 1, 1, true);
}
};
@@ -179,10 +178,10 @@ public:
auto i = channel_view(hidden, hidden - conv_ch);
auto c = conv->forward(ctx, channel_view(2 * hidden - conv_ch, conv_ch));
if (hidden > conv_ch) {
c = ggml_concat(ctx->ggml_ctx, ggml_ext_cont(ctx->ggml_ctx, i), c, 2);
c = ggml_concat(ctx->ggml_ctx, ggml_ext_cont(ctx, i), c, 2);
}
h = ggml_mul(ctx->ggml_ctx, c, mosr_mish(ctx->ggml_ctx, ggml_ext_cont(ctx->ggml_ctx, g)));
h = ggml_mul(ctx->ggml_ctx, c, mosr_mish(ctx->ggml_ctx, ggml_ext_cont(ctx, g)));
h = mosr_mish(ctx->ggml_ctx, fc2->forward(ctx, h));
return ggml_add(ctx->ggml_ctx, h, ggml_scale_bias(ctx->ggml_ctx, x, 1.f, -0.5f));
}