ggml : add ggml_build_forward_select (#18550)

* ggml : add ggml_build_forward_select

* cuda : adapt CUDA graph compat to new feature

* vulkan : update logic to handle command buffer closing

* ggml : check compute for fusion

* ggml : add comment
This commit is contained in:
Georgi Gerganov
2026-01-19 20:03:19 +02:00
committed by GitHub
parent 3d55846a5c
commit 365a3e8c31
17 changed files with 148 additions and 28 deletions
+4 -1
View File
@@ -12191,6 +12191,9 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
if (ggml_is_empty(node) || ggml_op_is_empty(node->op) || !node->buffer) {
return false;
}
if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
return false;
}
VK_LOG_DEBUG("ggml_vk_build_graph(" << node << ", " << ggml_op_name(node->op) << ")");
ctx->semaphore_idx = 0;
@@ -13645,7 +13648,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
int last_node = cgraph->n_nodes - 1;
// If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
while (last_node > 0 && ggml_vk_is_empty(cgraph->nodes[last_node])) {
while (last_node > 0 && (ggml_vk_is_empty(cgraph->nodes[last_node]) || ((cgraph->nodes[last_node]->flags & GGML_TENSOR_FLAG_COMPUTE) == 0))) {
last_node -= 1;
}