DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-11 03:24:30 +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/diffusion/qwen_image_2_1.hpp
+19 -21
View File
@@ -146,7 +146,7 @@ namespace Qwen {
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
auto weight = params["weight"];
if (ctx->weight_adapter) {
weight = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, weight, prefix + "weight");
weight = ctx->weight_adapter->patch_weight(ctx, weight, prefix + "weight");
}
weight = ggml_scale_bias(ctx->ggml_ctx, weight, 1.f, 1.f);
return ggml_mul(ctx->ggml_ctx, ggml_rms_norm(ctx->ggml_ctx, x, eps), weight);
@@ -164,7 +164,7 @@ namespace Qwen {
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
x = std::dynamic_pointer_cast<QwenImage21ZeroCenterRMSNorm>(blocks["text_norm"])->forward(ctx, x);
x = std::dynamic_pointer_cast<Linear>(blocks["in_layer"])->forward(ctx, x);
x = ggml_ext_gelu(ctx->ggml_ctx, x);
x = ggml_ext_gelu(ctx, x);
return std::dynamic_pointer_cast<Linear>(blocks["out_layer"])->forward(ctx, x);
}
};
@@ -196,7 +196,7 @@ namespace Qwen {
// prefix before its full-sequence K/V can accumulate across layers.
ctx->expand_graph(q);
auto persist = [&](ggml_tensor* tensor, int axis, const char* name) {
auto part = ggml_ext_slice(ctx->ggml_ctx, tensor, axis, 0, cache.prefix_length);
auto part = ggml_ext_slice(ctx, tensor, axis, 0, cache.prefix_length);
// Pack the contiguous data into wider rows so quantization blocks
// can exceed head_dim without padding or changing element order.
part = ggml_reshape_2d(ctx->ggml_ctx, part, x->ne[0], cache.prefix_length);
@@ -236,17 +236,15 @@ namespace Qwen {
} else {
for (size_t i = 0; i < segments.size(); ++i) {
const auto& segment = segments[i];
auto sq = ggml_ext_slice(ctx->ggml_ctx, q, 1, segment.start, segment.end);
auto sk = ggml_ext_slice(ctx->ggml_ctx, k, 1, 0, segment.end);
auto sv = ggml_ext_slice(ctx->ggml_ctx, v, 2, 0, segment.end);
auto sq = ggml_ext_slice(ctx, q, 1, segment.start, segment.end);
auto sk = ggml_ext_slice(ctx, k, 1, 0, segment.end);
auto sv = ggml_ext_slice(ctx, v, 2, 0, segment.end);
auto out = attend(sq, sk, sv, masks[i]);
result = result == nullptr ? out : ggml_concat(ctx->ggml_ctx, result, out, 1);
}
}
auto to_out = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
if (sd_backend_is(ctx->backend, "Vulkan") || sd_backend_is(ctx->backend, "ROCm")) {
to_out->set_force_prec_f32(true);
}
to_out->set_precision_flags(GGML_PRECISION_F32_VULKAN | GGML_PRECISION_F32_ROCM);
return to_out->forward(ctx, result);
}
};
@@ -268,7 +266,7 @@ namespace Qwen {
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 rows = ggml_ext_chunk(runner_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);
@@ -277,8 +275,8 @@ namespace Qwen {
if (prefix_length == 0) {
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]);
auto target = apply(ggml_ext_slice(runner_ctx, x, 1, prefix_length, x->ne[1]), rows[0]);
auto prefix = apply(ggml_ext_slice(runner_ctx, x, 1, 0, prefix_length), rows[1]);
return ggml_concat(ctx, prefix, target, 1);
}
@@ -323,27 +321,27 @@ namespace Qwen {
}
ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* timestep, ggml_tensor* context, const std::vector<ggml_tensor*>& refs, ggml_tensor* pe, const QwenImage21Layout& layout, const std::vector<ggml_tensor*>& masks, const QwenImage21PrefixCache& cache) {
auto time = ggml_concat(ctx->ggml_ctx, timestep, ggml_ext_zeros_like(ctx->ggml_ctx, timestep), 0);
auto time = ggml_concat(ctx->ggml_ctx, timestep, ggml_ext_zeros_like(ctx, timestep), 0);
// Runtime flow timesteps already use the [0, 1000] scale.
time = ggml_ext_timestep_embedding(ctx->ggml_ctx, time, 256, 10000, 1.f);
time = ggml_ext_timestep_embedding(ctx, time, 256, 10000, 1.f);
time = std::dynamic_pointer_cast<TimestepEmbedding>(blocks["time_text_embed.timestep_embedder"])->forward(ctx, time);
time = ggml_silu(ctx->ggml_ctx, time);
auto modulation = std::dynamic_pointer_cast<Linear>(blocks["modulation.1"])->forward(ctx, time);
auto mod = ggml_ext_chunk(ctx->ggml_ctx, modulation, 4, 0);
auto mod = ggml_ext_chunk(ctx, modulation, 4, 0);
auto img_in = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
ggml_tensor* joint = nullptr;
if (cache.mode == QwenImage21PrefixCache::Mode::REUSE) {
joint = img_in->forward(ctx, DiT::patchify(ctx->ggml_ctx, x, 1, 1));
joint = img_in->forward(ctx, DiT::patchify(ctx, x, 1, 1));
} else {
auto text = std::dynamic_pointer_cast<QwenImage21TextProjection>(blocks["txt_in"])->forward(ctx, context);
for (const auto& segment : layout.segments) {
ggml_tensor* h;
if (segment.image_index < 0) {
h = ggml_ext_slice(ctx->ggml_ctx, text, 1, segment.context_start,
h = ggml_ext_slice(ctx, text, 1, segment.context_start,
segment.context_start + segment.end - segment.start);
} else {
auto image = segment.image_index == static_cast<int>(refs.size()) ? x : refs[segment.image_index];
h = img_in->forward(ctx, DiT::patchify(ctx->ggml_ctx, image, 1, 1));
h = img_in->forward(ctx, DiT::patchify(ctx, image, 1, 1));
}
joint = joint == nullptr ? h : ggml_concat(ctx->ggml_ctx, joint, h, 1);
}
@@ -359,13 +357,13 @@ namespace Qwen {
sd::ggml_graph_cut::mark_graph_cut(joint, layer_cache.cut_group, "joint");
}
if (cache.mode != QwenImage21PrefixCache::Mode::REUSE) {
joint = ggml_ext_slice(ctx->ggml_ctx, joint, 1, layout.prefix_length, joint->ne[1]);
joint = ggml_ext_slice(ctx, joint, 1, layout.prefix_length, joint->ne[1]);
}
auto scale = std::dynamic_pointer_cast<Linear>(blocks["norm_out.linear"])->forward(ctx, ggml_ext_chunk(ctx->ggml_ctx, time, 2, 1)[0]);
auto scale = std::dynamic_pointer_cast<Linear>(blocks["norm_out.linear"])->forward(ctx, ggml_ext_chunk(ctx, time, 2, 1)[0]);
joint = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_out.norm"])->forward(ctx, joint);
joint = ggml_mul(ctx->ggml_ctx, joint, ggml_scale_bias(ctx->ggml_ctx, scale, 1.f, 1.f));
joint = std::dynamic_pointer_cast<Linear>(blocks["proj_out"])->forward(ctx, joint);
return DiT::unpatchify_and_crop(ctx->ggml_ctx, joint, x->ne[1], x->ne[0], 1, 1);
return DiT::unpatchify_and_crop(ctx, joint, x->ne[1], x->ne[0], 1, 1);
}
};