Fix test-backend-ops

ggml/src/ggml-openvino/ggml-decoder.cpp

@@ -76,6 +76,9 @@ GgmlOvDecoder::GgmlOvDecoder(struct ggml_cgraph* cgraph) {
     m_cgraph = cgraph;
     for (int node_n = 0; node_n < cgraph->n_nodes; node_n++) {
         auto* cur_node = cgraph->nodes[node_n];
+        if (cur_node->op == GGML_OP_NONE) {
+            continue;
+        }
         m_nodes.push_back(cur_node);
         set_input_output(cur_node, true);
     }

ggml/src/ggml-openvino/ggml-openvino.cpp

@@ -258,12 +258,25 @@ static bool is_op_unsupported_case(const ggml_tensor* op) {
         }
     }
 
+    if (op->op == GGML_OP_PERMUTE) {
+        if (op->type == GGML_TYPE_BF16) {
+            // err msg: [GPU] Could not find a suitable kernel for transpose
+            GGML_LOG_WARN("OpenVINO backend does not support PERMUTE with BF16 type\n");
+            return true;
+        }
+    }
+
     if (op->op == GGML_OP_MUL_MAT) {
         if ((op->src[0]->view_src && op->src[0]->op != GGML_OP_PERMUTE) ||
             (op->src[1]->view_src && op->src[1]->op != GGML_OP_PERMUTE)) {
             GGML_LOG_WARN("OpenVINO backend does not support MUL_MAT with view_src tensors that are not PERMUTE\n");
             return true;
         }
+        if (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F16) {
+            // Has accuracy issue, try enabling this and see `test-backend-ops -o "MUL_MAT"`
+            GGML_LOG_WARN("OpenVINO backend does not support MUL_MAT with two F16 tensors\n");
+            return true;
+        }
     }
 
     if (op->op == GGML_OP_ROPE) {

ggml/src/ggml-openvino/openvino/op/soft_max.cpp

@@ -53,7 +53,7 @@ OutputVector translate_soft_max(const NodeContext& context) {
 
     auto mask_node = context.get_input(1);
 
-    auto token_len = context.get_input("token_len");
+    auto token_len = context.has_input("token_len") ? context.get_input("token_len") : get_dimensions(input_node, {1});
     auto zero = ov::op::v0::Constant::create(ov::element::i64, {1}, {0});
     auto one = ov::op::v0::Constant::create(ov::element::i64, {1}, {1});
     std::shared_ptr<ov::Node> mask_node_sliced =

ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.hpp

@@ -24,6 +24,6 @@ class TRANSFORMATIONS_API MarkCompressedFloatConstants;
 
 class ov::pass::MarkCompressedFloatConstants : public MatcherPass {
 public:
-    OPENVINO_MATCHER_PASS_RTTI("MarkCompressedFloatConstants");
+    OPENVINO_MATCHER_PASS_RTTI("MarkCompressedFloatConstants")
     MarkCompressedFloatConstants();
 };

ggml/src/ggml-openvino/utils.cpp

@@ -81,7 +81,7 @@ enum ggml_status openvino_frontend_compute(ggml_backend_t backend, struct ggml_c
         config = get_npu_config();
     }
 
-    if (cgraph->n_nodes == 1) {
+    if (is_naive(cgraph)) {
         return naive_compute(cgraph, core, device, config);
     }
 
@@ -250,11 +250,16 @@ ov::AnyMap get_npu_config() {
     return config;
 }
 
+bool is_naive(struct ggml_cgraph* cgraph) {
+    constexpr int naive_graph_size_threshold = 20;
+    return cgraph->n_nodes < naive_graph_size_threshold;
+}
+
 enum ggml_status naive_compute(struct ggml_cgraph* cgraph,
                                ov::Core& core,
                                const std::string& device,
                                const ov::AnyMap& config) {
-    if (cgraph->nodes[0]->op == GGML_OP_NONE) {
+    if (cgraph->n_nodes == 1 && cgraph->nodes[0]->op == GGML_OP_NONE) {
         return GGML_STATUS_SUCCESS;
     }
 
@@ -264,8 +269,6 @@ enum ggml_status naive_compute(struct ggml_cgraph* cgraph,
     auto model = ov::frontend::ggml::FrontEnd::convert(input_model, naive);
     auto infer_request = core.compile_model(model, device, config).create_infer_request();
 
-    ov::serialize(model, "IR.xml");
-
     auto ov_params = model->get_parameters();
     for (size_t i = 0; i < ov_params.size(); i++) {
         auto param_name = ov_params[i]->get_friendly_name();

ggml/src/ggml-openvino/utils.h

@@ -44,5 +44,7 @@ ov::AnyMap get_npu_config();
 
 ov::Tensor get_ov_input_tensor(std::shared_ptr<GgmlOvDecoder> ggml_decoder, const std::string& param_name);
 
+bool is_naive(struct ggml_cgraph* cgraph);
+
 enum ggml_status naive_compute(struct ggml_cgraph* cgraph, ov::Core& core, const std::string& device,
                                const ov::AnyMap& config);