LLM inference in C/C++
C++ ggml
ggml : require input tensors to be GGML_OP_NONE (#29647)
Georgi Gerganovb5cf8ce02aaf
diff --git a/ggml/src/ggml.c b/ggml/src/ggml.cindex cdc10a7597..7c931afb99 100644--- a/ggml/src/ggml.c+++ b/ggml/src/ggml.c@@ -7992,6 +7992,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph////////////////////////////////////////////////////////////////////////////////void ggml_set_input(struct ggml_tensor * tensor) {+ GGML_ASSERT(tensor->op == GGML_OP_NONE);tensor->flags |= GGML_TENSOR_FLAG_INPUT;}diff --git a/src/llama-context.cpp b/src/llama-context.cppindex 88851c6cba..439beabdfa 100644--- a/src/llama-context.cpp+++ b/src/llama-context.cpp@@ -599,10 +599,7 @@ static int llama_graph_n_input_tensors(ggml_cgraph * gf) {}for (const auto & [tensor, nodes] : users) {- if (tensor->op != GGML_OP_NONE) {- LLAMA_LOG_WARN("%s: input tensor '%32s' has op %s, expected GGML_OP_NONE\n",- __func__, tensor->name, ggml_op_name(tensor->op));- }+ GGML_ASSERT(tensor->op == GGML_OP_NONE);for (const ggml_tensor * node : nodes) {LLAMA_LOG_DEBUG("%s: input tensor '%32s' [%s, ne = { %5" PRId64 ", %5" PRId64 ", %5" PRId64 ", %5" PRId64 " }] is used by node '%s' (%s)\n",__func__, tensor->name, ggml_type_name(tensor->type),