refactor: consolidate graph helpers in sd.cpp

docs/build.md

@@ -33,8 +33,9 @@ cmake --build build-upstream --config Release
 ```
 
 Both patched and upstream builds use the selected library and matching private
-headers for backend integration. Backend options such as `-DSD_CUDA=ON` apply
-to the selected tree as usual.
+headers for application graph helpers and backend integration. The generic
+operator registration API remains in `ggml-ext.h`. Backend options such as
+`-DSD_CUDA=ON` apply to the selected tree as usual.
 
 `SD_USE_UPSTREAM_GGML` defaults to `OFF`, which enables the patched GGML
 extensions. Set it to `ON` when using upstream GGML; it selects the compatibility
@@ -52,8 +53,8 @@ quantization types remain available, subject to backend operator support.
 
 `SD_USE_SYSTEM_GGML=ON` instead links an installed GGML CMake package, located
 with `ggml_DIR` or `CMAKE_PREFIX_PATH`. An installed patched package must include
-the extension and graph APIs. Set `SD_GGML_SOURCE_DIR` to its matching source
-tree for private backend headers, including CUDA/HIP/MUSA vendor headers.
+the extension API. Set `SD_GGML_SOURCE_DIR` to its matching source tree for
+private graph and backend headers, including CUDA/HIP/MUSA vendor headers.
 The installed library must use the same ABI settings as sd.cpp,
 including `GGML_MAX_NAME`. For an upstream package, also set
 `SD_USE_UPSTREAM_GGML=ON` and point `SD_GGML_SOURCE_DIR` to its matching source

docs/ggml_extensions.md

@@ -448,22 +448,18 @@ kernel compiled for SM80 or newer. CPU, HIP, MUSA, and Vulkan do not advertise
 support. ggml no longer contains a Sol-Attn opcode, constructor, allocation rule,
 or backend kernel.
 
-## Public graph access
+## Internal graph access
 
-`ggml-graph.h` exposes leaf enumeration, append-only leaf construction, recorded
-use counts, mapped inference-graph copying, use-count copying for segments, and
-backend execution of a node range. Mapped copies retain node order and consumer
-counts from the original graph, including consumers outside a graph view. The
-caller owns tensor copies and remaps their storage and edges; invalid mappings
-leave the destination graph unchanged. Range execution uses already allocated
-tensors and retains the original graph's consumer counts.
+`src/core/ggml_compat.cpp` uses `ggml-impl.h` for leaf enumeration, graph copying,
+use-count preservation and range execution. Patched and upstream builds share
+this implementation; ggml does not expose a separate public graph API for these
+application helpers. Mapped copies retain node order and consumer counts,
+including consumers outside a graph view. Invalid mappings leave the destination
+unchanged. Workspace measurement, graph cuts and evaluation callbacks use these
+helpers.
 
-Workspace measurement, graph cuts, and evaluation callbacks use these APIs.
-Backend extensions require matching private headers. Compatibility with an
-upstream ggml lacking this header is isolated in `src/core/ggml_compat.cpp`; that
-optional build still requires matching upstream private headers. The compatibility
-layer also retains the workaround for old Metal backends wrapping failed host
-mappings. The patched Metal backend returns null directly when mapping fails.
+The compatibility layer also retains the workaround for old Metal backends
+wrapping failed host mappings. The patched Metal backend returns null directly.
 
 Vulkan command recording services are declared in the private
 `ggml/src/ggml-ext-vulkan.h`. Pipeline ownership, submission and synchronization

src/core/compute_workspace.cpp

@@ -69,7 +69,7 @@ namespace sd {
 
     bool ComputeWorkspace::allocate(ggml_cgraph* graph, const AssignNodes& assign_nodes) {
         GGML_ASSERT(!active_);
-        const size_t capacity = static_cast<size_t>(ggml_graph_n_nodes(graph) + ggml_graph_n_leafs(graph)) + 8;
+        const size_t capacity = static_cast<size_t>(ggml_graph_n_nodes(graph) + sd_ggml_graph_n_leafs(graph)) + 8;
         if (scheduler_ == nullptr || capacity > scheduler_capacity_) {
             release();
             scheduler_          = make_scheduler(capacity);
@@ -104,8 +104,8 @@ namespace sd {
         for (int i = 0; i < ggml_graph_n_nodes(graph); ++i) {
             visit(ggml_graph_node(graph, i));
         }
-        for (int i = 0; i < ggml_graph_n_leafs(graph); ++i) {
-            visit(ggml_graph_leaf(graph, i));
+        for (int i = 0; i < sd_ggml_graph_n_leafs(graph); ++i) {
+            visit(sd_ggml_graph_leaf(graph, i));
         }
         for (size_t i = 0; i < tensors.size(); ++i) {
             visit(tensors[i]->view_src);
@@ -121,7 +121,7 @@ namespace sd {
                 key.insert(key.end(), bytes, bytes + sizeof(value));
             };
             append(ggml_graph_n_nodes(graph));
-            append(ggml_graph_n_leafs(graph));
+            append(sd_ggml_graph_n_leafs(graph));
             for (auto tensor : tensors) {
                 append(tensor);
                 append(tensor->type);
@@ -150,7 +150,7 @@ namespace sd {
                     const auto alignment_offset = reinterpret_cast<uintptr_t>(tensor->data) % ggml_backend_get_alignment(external);
                     append(alignment_offset);
                 }
-                const int32_t uses = ggml_graph_use_count(graph, tensor);
+                const int32_t uses = sd_ggml_graph_use_count(graph, tensor);
                 append(uses);
             }
             // Match the full allocation metadata, including topology and external placement.
@@ -197,7 +197,7 @@ namespace sd {
             const auto found   = mapped.find(tensor);
             return found == mapped.end() ? nullptr : found->second;
         };
-        GGML_ASSERT(ggml_graph_cpy_mapped(graph, copy_graph, map_tensor, &copies));
+        GGML_ASSERT(sd_ggml_graph_cpy_mapped(graph, copy_graph, map_tensor, &copies));
         Measurement result;
         auto scheduler = make_scheduler(graph_size);
         if (scheduler != nullptr) {

src/core/ggml_compat.cpp

@@ -1,8 +1,6 @@
 #include "core/ggml_compat.h"
 #include "core/ggml_extend_backend.h"
 
-#ifdef SD_GGML_GRAPH_COMPAT
-// Upstream ggml still requires its private graph layout for these operations.
 #include "ggml-impl.h"
 
 #include <unordered_map>
@@ -10,11 +8,11 @@
 #include <utility>
 #include <vector>
 
-int ggml_graph_n_leafs(const ggml_cgraph* graph) {
+int sd_ggml_graph_n_leafs(const ggml_cgraph* graph) {
     return graph->n_leafs;
 }
 
-ggml_tensor* ggml_graph_leaf(const ggml_cgraph* graph, int index) {
+ggml_tensor* sd_ggml_graph_leaf(const ggml_cgraph* graph, int index) {
     if (index < 0) {
         index += graph->n_leafs;
     }
@@ -22,12 +20,12 @@ ggml_tensor* ggml_graph_leaf(const ggml_cgraph* graph, int index) {
     return graph->leafs[index];
 }
 
-void ggml_graph_add_leaf(ggml_cgraph* graph, ggml_tensor* tensor) {
+void sd_ggml_graph_add_leaf(ggml_cgraph* graph, ggml_tensor* tensor) {
     GGML_ASSERT(tensor && graph->n_leafs < graph->size);
     graph->leafs[graph->n_leafs++] = tensor;
 }
 
-int32_t ggml_graph_use_count(const ggml_cgraph* graph, const ggml_tensor* tensor) {
+int32_t sd_ggml_graph_use_count(const ggml_cgraph* graph, const ggml_tensor* tensor) {
     if (!graph || !tensor || !graph->visited_hash_set.size) {
         return 0;
     }
@@ -37,7 +35,7 @@ int32_t ggml_graph_use_count(const ggml_cgraph* graph, const ggml_tensor* tensor
                : 0;
 }
 
-bool ggml_graph_cpy_mapped(const ggml_cgraph* src, ggml_cgraph* dst, ggml_graph_tensor_map_fn map, void* user_data) {
+bool sd_ggml_graph_cpy_mapped(const ggml_cgraph* src, ggml_cgraph* dst, sd_ggml_graph_tensor_map_fn map, void* user_data) {
     if (!src || !dst || src == dst || !map || dst->grads || dst->grad_accs ||
         src->n_nodes > dst->size || src->n_leafs > dst->size) {
         return false;
@@ -104,7 +102,7 @@ bool ggml_graph_cpy_mapped(const ggml_cgraph* src, ggml_cgraph* dst, ggml_graph_
         ggml_graph_add_node(dst, tensor);
     }
     for (auto* tensor : leafs) {
-        ggml_graph_add_leaf(dst, tensor);
+        sd_ggml_graph_add_leaf(dst, tensor);
     }
     for (const auto& entry : counts) {
         const size_t index     = ggml_hash_insert(&dst->visited_hash_set, entry.first);
@@ -113,7 +111,7 @@ bool ggml_graph_cpy_mapped(const ggml_cgraph* src, ggml_cgraph* dst, ggml_graph_
     return true;
 }
 
-bool ggml_graph_cpy_use_counts(const ggml_cgraph* src, ggml_cgraph* dst) {
+bool sd_ggml_graph_cpy_use_counts(const ggml_cgraph* src, ggml_cgraph* dst) {
     if (!src || !dst || !src->visited_hash_set.size || !dst->visited_hash_set.size) {
         return false;
     }
@@ -152,7 +150,7 @@ bool ggml_graph_cpy_use_counts(const ggml_cgraph* src, ggml_cgraph* dst) {
     return true;
 }
 
-ggml_status ggml_backend_graph_compute_async_range(ggml_backend_t backend, ggml_cgraph* graph, int first, int last) {
+ggml_status sd_ggml_backend_graph_compute_async_range(ggml_backend_t backend, ggml_cgraph* graph, int first, int last) {
     if (!backend || !graph || first < 0 || last < first || last > graph->n_nodes) {
         return GGML_STATUS_FAILED;
     }
@@ -162,8 +160,6 @@ ggml_status ggml_backend_graph_compute_async_range(ggml_backend_t backend, ggml_
     auto view = ggml_graph_view(graph, first, last);
     return ggml_backend_graph_compute_async(backend, &view);
 }
-#endif
-
 #ifdef SD_USE_UPSTREAM_GGML
 #include "ggml-backend-impl.h"
 #endif

src/core/ggml_compat.h

@@ -3,19 +3,14 @@
 
 #include "ggml-backend.h"
 
-#if __has_include("ggml-graph.h")
-#include "ggml-graph.h"
-#elif defined(SD_USE_UPSTREAM_GGML)
-#define SD_GGML_GRAPH_COMPAT
-
-int ggml_graph_n_leafs(const struct ggml_cgraph* graph);
-struct ggml_tensor* ggml_graph_leaf(const struct ggml_cgraph* graph, int index);
+int sd_ggml_graph_n_leafs(const struct ggml_cgraph* graph);
+struct ggml_tensor* sd_ggml_graph_leaf(const struct ggml_cgraph* graph, int index);
 // Like ggml_graph_add_node, appends metadata without traversing tensor dependencies.
-void ggml_graph_add_leaf(struct ggml_cgraph* graph, struct ggml_tensor* tensor);
+void sd_ggml_graph_add_leaf(struct ggml_cgraph* graph, struct ggml_tensor* tensor);
 // Returns zero when the tensor has no recorded use count. Includes uses outside graph views.
-int32_t ggml_graph_use_count(const struct ggml_cgraph* graph, const struct ggml_tensor* tensor);
+int32_t sd_ggml_graph_use_count(const struct ggml_cgraph* graph, const struct ggml_tensor* tensor);
 
-typedef struct ggml_tensor* (*ggml_graph_tensor_map_fn)(const struct ggml_tensor* tensor, void* user_data);
+typedef struct ggml_tensor* (*sd_ggml_graph_tensor_map_fn)(const struct ggml_tensor* tensor, void* user_data);
 
 // Copies forward node/leaf order and recorded use counts, including external consumers.
 // The callback must be injective and may return NULL for dependencies outside the
@@ -23,21 +18,17 @@ typedef struct ggml_tensor* (*ggml_graph_tensor_map_fn)(const struct ggml_tensor
 // Tensor storage and edges are mapped by the caller.
 // The destination must have no gradients and enough capacity. Gradients are not copied.
 // Invalid mappings leave the destination unchanged. src and dst must be distinct.
-bool ggml_graph_cpy_mapped(const struct ggml_cgraph* src, struct ggml_cgraph* dst, ggml_graph_tensor_map_fn map, void* user_data);
+bool sd_ggml_graph_cpy_mapped(const struct ggml_cgraph* src, struct ggml_cgraph* dst, sd_ggml_graph_tensor_map_fn map, void* user_data);
 
 // Copies recorded counts for tensors in dst's node/leaf lists. Other metadata is unchanged.
 // This preserves external consumers when constructing an inference graph segment.
-bool ggml_graph_cpy_use_counts(const struct ggml_cgraph* src, struct ggml_cgraph* dst);
+bool sd_ggml_graph_cpy_use_counts(const struct ggml_cgraph* src, struct ggml_cgraph* dst);
 
 // Executes nodes [first, last) with the original graph's use counts. No graph ownership
 // is transferred. Tensor storage must already be allocated; invalid bounds fail.
-enum ggml_status ggml_backend_graph_compute_async_range(ggml_backend_t backend,
+enum ggml_status sd_ggml_backend_graph_compute_async_range(ggml_backend_t backend,
                                                            struct ggml_cgraph* graph,
                                                            int first,
                                                            int last);
 
-#else
-#error "Patched ggml requires ggml-graph.h; select SD_USE_UPSTREAM_GGML for upstream compatibility"
-#endif
-
 #endif

src/core/ggml_extend_backend.cpp

@@ -647,7 +647,7 @@ ggml_status sd_backend_graph_compute_with_eval_callback(ggml_backend_t backend,
             need = callback_eval(t, true, callback_eval_user_data);
         }
 
-        status = ggml_backend_graph_compute_async_range(backend, gf, j0, j1 + 1);
+        status = sd_ggml_backend_graph_compute_async_range(backend, gf, j0, j1 + 1);
         if (status != GGML_STATUS_SUCCESS) {
             break;
         }

src/core/ggml_graph_cut.cpp

@@ -35,8 +35,8 @@ namespace sd::ggml_graph_cut {
     static int graph_leaf_index(ggml_cgraph* gf, const ggml_tensor* tensor) {
         GGML_ASSERT(gf != nullptr);
         GGML_ASSERT(tensor != nullptr);
-        for (int i = 0; i < ggml_graph_n_leafs(gf); ++i) {
-            if (ggml_graph_leaf(gf, i) == tensor) {
+        for (int i = 0; i < sd_ggml_graph_n_leafs(gf); ++i) {
+            if (sd_ggml_graph_leaf(gf, i) == tensor) {
                 return i;
             }
         }
@@ -389,7 +389,7 @@ namespace sd::ggml_graph_cut {
                         return fail(segment_label + " references an unavailable cut node " +
                                     std::to_string(input.node_index));
                     }
-                } else if (input.leaf_index < 0 || input.leaf_index >= ggml_graph_n_leafs(gf)) {
+                } else if (input.leaf_index < 0 || input.leaf_index >= sd_ggml_graph_n_leafs(gf)) {
                     return fail(segment_label + " references an invalid leaf " +
                                 std::to_string(input.leaf_index));
                 }
@@ -444,15 +444,15 @@ namespace sd::ggml_graph_cut {
 
     int leaf_count(ggml_cgraph* gf) {
         GGML_ASSERT(gf != nullptr);
-        return ggml_graph_n_leafs(gf);
+        return sd_ggml_graph_n_leafs(gf);
     }
 
     ggml_tensor* leaf_tensor(ggml_cgraph* gf, int leaf_index) {
         GGML_ASSERT(gf != nullptr);
-        if (leaf_index < 0 || leaf_index >= ggml_graph_n_leafs(gf)) {
+        if (leaf_index < 0 || leaf_index >= sd_ggml_graph_n_leafs(gf)) {
             return nullptr;
         }
-        return ggml_graph_leaf(gf, leaf_index);
+        return sd_ggml_graph_leaf(gf, leaf_index);
     }
 
     ggml_backend_buffer_t tensor_buffer(const ggml_tensor* tensor) {
@@ -548,11 +548,11 @@ namespace sd::ggml_graph_cut {
     };
 
     std::vector<uint64_t> graph_layout(ggml_cgraph* graph, bool include_bindings) {
-        const size_t graph_size = static_cast<size_t>(ggml_graph_n_leafs(graph)) + ggml_graph_n_nodes(graph);
+        const size_t graph_size = static_cast<size_t>(sd_ggml_graph_n_leafs(graph)) + ggml_graph_n_nodes(graph);
         GraphLayoutTensors layout_tensors(graph_size);
         const auto& tensors = layout_tensors.tensors;
-        for (int i = 0; i < ggml_graph_n_leafs(graph); ++i) {
-            layout_tensors.add(ggml_graph_leaf(graph, i));
+        for (int i = 0; i < sd_ggml_graph_n_leafs(graph); ++i) {
+            layout_tensors.add(sd_ggml_graph_leaf(graph, i));
         }
         for (int i = 0; i < ggml_graph_n_nodes(graph); ++i) {
             layout_tensors.add(ggml_graph_node(graph, i));
@@ -568,9 +568,9 @@ namespace sd::ggml_graph_cut {
                                      GGML_MAX_OP_PARAMS / sizeof(int32_t) + (include_bindings ? 2 : 0);
         signature.reserve(2 + graph_size + tensors.size() * tensor_fields);
         signature.push_back(ggml_graph_n_nodes(graph));
-        signature.push_back(ggml_graph_n_leafs(graph));
-        for (int i = 0; i < ggml_graph_n_leafs(graph); ++i) {
-            signature.push_back(layout_tensors.index(ggml_graph_leaf(graph, i)));
+        signature.push_back(sd_ggml_graph_n_leafs(graph));
+        for (int i = 0; i < sd_ggml_graph_n_leafs(graph); ++i) {
+            signature.push_back(layout_tensors.index(sd_ggml_graph_leaf(graph, i)));
         }
         for (int i = 0; i < ggml_graph_n_nodes(graph); ++i) {
             signature.push_back(layout_tensors.index(ggml_graph_node(graph, i)));
@@ -606,12 +606,12 @@ namespace sd::ggml_graph_cut {
                                    const Plan& plan,
                                    const std::vector<uint64_t>& layout) {
         GGML_ASSERT(gf != nullptr);
-        if (plan.leaf_names.size() != static_cast<size_t>(ggml_graph_n_leafs(gf)) ||
+        if (plan.leaf_names.size() != static_cast<size_t>(sd_ggml_graph_n_leafs(gf)) ||
             plan.layout != layout) {
             return false;
         }
-        for (int i = 0; i < ggml_graph_n_leafs(gf); ++i) {
-            if (plan.leaf_names[i] != ggml_graph_leaf(gf, i)->name) {
+        for (int i = 0; i < sd_ggml_graph_n_leafs(gf); ++i) {
+            if (plan.leaf_names[i] != sd_ggml_graph_leaf(gf, i)->name) {
                 return false;
             }
         }
@@ -654,7 +654,7 @@ namespace sd::ggml_graph_cut {
             }
             return ggml_graph_node(gf, input_ref.node_index);
         }
-        if (input_ref.leaf_index < 0 || input_ref.leaf_index >= ggml_graph_n_leafs(gf)) {
+        if (input_ref.leaf_index < 0 || input_ref.leaf_index >= sd_ggml_graph_n_leafs(gf)) {
             return nullptr;
         }
         return leaf_tensor(gf, input_ref.leaf_index);
@@ -735,7 +735,7 @@ namespace sd::ggml_graph_cut {
         GGML_ASSERT(segment_graph != nullptr);
 
         for (ggml_tensor* leaf : leaves) {
-            ggml_graph_add_leaf(segment_graph, leaf);
+            sd_ggml_graph_add_leaf(segment_graph, leaf);
         }
 
         for (int output_node_index : segment.output_node_indices) {
@@ -752,7 +752,7 @@ namespace sd::ggml_graph_cut {
         for (int node_idx : segment.internal_node_indices) {
             ggml_graph_add_node(segment_graph, ggml_graph_node(gf, node_idx));
         }
-        GGML_ASSERT(ggml_graph_cpy_use_counts(gf, segment_graph));
+        GGML_ASSERT(sd_ggml_graph_cpy_use_counts(gf, segment_graph));
         *graph_ctx_out = graph_ctx;
         return segment_graph;
     }
@@ -888,8 +888,8 @@ namespace sd::ggml_graph_cut {
             return plan;
         }
         plan.layout = graph_layout(gf, false);
-        for (int i = 0; i < ggml_graph_n_leafs(gf); ++i) {
-            plan.leaf_names.emplace_back(ggml_graph_leaf(gf, i)->name);
+        for (int i = 0; i < sd_ggml_graph_n_leafs(gf); ++i) {
+            plan.leaf_names.emplace_back(sd_ggml_graph_leaf(gf, i)->name);
         }
         plan.compute_buffer_size =
             measure_graph_compute_buffer(backend, gf, params_tensor_set);