happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vulkan : update logic to handle command buffer closing

This commit is contained in:
Georgi Gerganov 2026-01-16 17:44:44 +02:00
parent ced0693190
commit 9696c3e748
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
1 changed files with 4 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
ggml/src/ggml-vulkan/ggml-vulkan.cpp
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;
}
@ -13719,10 +13722,6 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
submit_node_idx = i;
}
if ((cgraph->nodes[i]->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
continue;
}
if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
auto bytes = ggml_nbytes(cgraph->nodes[i]->src[0]);
mul_mat_bytes += bytes;