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

perf: accelerate INT8 diffusion inference with CUDA fusions (#2235)

Browse Source
This commit is contained in:
leejet authored and GitHub committed 2026-10-28 22:02:46 +08:00
1 parent 8cbe905fdd
commit 77b72a2efe
12 files changed
+113 -128
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/model/diffusion/qwen_image_2_1.hpp
+14 -15
View File
@@ -266,16 +266,18 @@ namespace Qwen {
blocks["img_mlp.out"] = std::make_shared<Linear>(config.intermediate_size, config.hidden_size, false);
}
static ggml_tensor* modulate(ggml_context* ctx, ggml_tensor* x, ggml_tensor* params, int64_t prefix_length, bool gate = false) {
static ggml_tensor* modulate(GGMLRunnerContext* runner_ctx, ggml_tensor* x, ggml_tensor* params, int64_t prefix_length, bool gate = false) {
auto ctx = runner_ctx->ggml_ctx;
auto rows = ggml_ext_chunk(ctx, params, 2, 1);
auto apply = [&](ggml_tensor* part, ggml_tensor* row) {
row = gate ? ggml_tanh(ctx, row) : ggml_scale_bias(ctx, row, 1.f, 1.f);
runner_ctx->expand_graph(row);
return ggml_mul(ctx, part, row);
};
auto target = apply(ggml_ext_slice(ctx, x, 1, prefix_length, x->ne[1]), rows[0]);
if (prefix_length == 0) {
return target;
return apply(x, rows[0]);
}
auto target = apply(ggml_ext_slice(ctx, x, 1, prefix_length, x->ne[1]), rows[0]);
auto prefix = apply(ggml_ext_slice(ctx, x, 1, 0, prefix_length), rows[1]);
return ggml_concat(ctx, prefix, target, 1);
}
@@ -283,25 +285,22 @@ namespace Qwen {
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, const std::vector<ggml_tensor*>& modulation, ggml_tensor* pe, const QwenImage21Layout& layout, const std::vector<ggml_tensor*>& masks, const QwenImage21PrefixCache& cache) {
const int64_t prefix_length = cache.mode == QwenImage21PrefixCache::Mode::REUSE ? 0 : layout.prefix_length;
auto h = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm1"])->forward(ctx, x);
h = modulate(ctx->ggml_ctx, h, modulation[0], prefix_length);
h = modulate(ctx, h, modulation[0], prefix_length);
h = std::dynamic_pointer_cast<QwenImage21Attention>(blocks["attn"])->forward(ctx, h, pe, layout.segments, masks, cache);
x = ggml_add(ctx->ggml_ctx, x, modulate(ctx->ggml_ctx, h, modulation[1], prefix_length, true));
x = ggml_add(ctx->ggml_ctx, x, modulate(ctx, h, modulation[1], prefix_length, true));
h = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm2"])->forward(ctx, x);
h = modulate(ctx->ggml_ctx, h, modulation[2], prefix_length);
ggml_tensor* gate;
auto fused = blocks.find("img_mlp.gate_up");
h = modulate(ctx, h, modulation[2], prefix_length);
auto fused = blocks.find("img_mlp.gate_up");
if (fused != blocks.end()) {
auto gate_up = std::dynamic_pointer_cast<Linear>(fused->second)->forward(ctx, h);
auto parts = ggml_ext_chunk(ctx->ggml_ctx, gate_up, 2, 0);
gate = parts[0];
h = parts[1];
h = ggml_swiglu(ctx->ggml_ctx, gate_up);
} else {
gate = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.gate_layer"])->forward(ctx, h);
h = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.proj"])->forward(ctx, h);
auto gate = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.gate_layer"])->forward(ctx, h);
h = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.proj"])->forward(ctx, h);
h = ggml_swiglu_split(ctx->ggml_ctx, gate, h);
}
h = ggml_mul(ctx->ggml_ctx, h, ggml_silu(ctx->ggml_ctx, gate));
h = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.out"])->forward(ctx, h);
return ggml_add(ctx->ggml_ctx, x, modulate(ctx->ggml_ctx, h, modulation[3], prefix_length, true));
return ggml_add(ctx->ggml_ctx, x, modulate(ctx, h, modulation[3], prefix_length, true));
}
};