fix CI

Author: foldl

models/qwen.cpp

@@ -1259,7 +1259,7 @@ namespace chatllm::qwen::vit
         :
         GRAPH_SIZE(GRAPH_SIZE), _ctx(&backend_context),
         n_threads(runtime_config.n_threads),
-        vis_config(0),
+        vis_config(),
         max_patches(max_patches)
     {
         _ctx.cache_dtype = runtime_config.cache_type;
@@ -1425,8 +1425,6 @@ namespace chatllm::qwen::v2_5_vl
         ids.push_back(vision_end_token_id);
     }
 
-    static BlockParams::PadEmbedding *pad_arg = nullptr;
-
     struct ImageGridSize
     {
         int w, h;
@@ -1440,8 +1438,8 @@ namespace chatllm::qwen::v2_5_vl
     {
     public:
         TensorPosHelper3D(int max_length, int image_id_start, ConditionalGeneration *gen)
-            : original_length(max_length), image_id_start(image_id_start),
-              BaseTensorPosHelper(max_length * 4),
+            : BaseTensorPosHelper(max_length * 4),
+              original_length(max_length), image_id_start(image_id_start),
               gen(gen)
         {
         }
@@ -1598,7 +1596,7 @@ namespace chatllm::qwen::v2_5_vl
                 continue;
             }
 
-            CHATLLM_CHECK(mm_index < images_grid.size());
+            CHATLLM_CHECK(mm_index < (int)images_grid.size());
 
             auto &dim = images_grid[mm_index++];
             for (int f = 0; f < dim.frame_num; f++, t += token_n_inc)

src/layers.cpp

@@ -2500,10 +2500,10 @@ namespace chatllm
             num_local_experts(num_local_experts), num_experts_per_tok(num_experts_per_tok),
             gate(ctx, hidden_size, num_local_experts, gate_use_bias),
             mover(new CPUMover(ctx, ctx->user_options.moe_on_cpu)),
-            experts(ctx, hidden_size, intermediate_size, num_local_experts, num_experts_per_tok, act, experts_use_bias),
             gate_score_correction_bias(gate_score_use_bias ? ggml::new_tensor_1d(ctx, GGML_TYPE_F32, num_local_experts) : nullptr),
             group_indices(grouped_max ? ggml::new_tensor_2d(ctx, GGML_TYPE_I32, 1, num_experts_per_tok) : nullptr),
             router_scale(router_scale ? ggml::new_tensor_1d(ctx, GGML_TYPE_F32, num_local_experts) : nullptr),
+            experts(ctx, hidden_size, intermediate_size, num_local_experts, num_experts_per_tok, act, experts_use_bias),
             norm_topk_prob(true),
             score_func(ScoreFunc::Softmax),
             routed_scaling_factor(-1.0f),

src/layers.h

@@ -1272,10 +1272,10 @@ namespace chatllm
               attn_scaling(true),
               causal(true),
               last_attn_scores(nullptr),
-              pos_helper(helper ? helper : &def_pos_helper),
               sinks(BlockParams::CoreAttentionUseSinks::get() > 0 ?
                     ggml::new_tensor_1d(ctx, ggml::type::GGML_TYPE_F32, BlockParams::CoreAttentionUseSinks::get())
-                    : nullptr)
+                    : nullptr),
+              pos_helper(helper ? helper : &def_pos_helper)
         {
             allocate_pos_tensor(ctx);
         }