add __index_timestep_zero__ support

qwen_image.hpp

@@ -561,11 +561,13 @@ namespace Qwen {
             : GGMLRunner(backend, offload_params_to_cpu) {
             qwen_image_params.num_layers  = 0;
             qwen_image_params.zero_cond_t = zero_cond_t;
-            LOG_DEBUG("zero_cond_t: %d", zero_cond_t);
             for (auto pair : tensor_storage_map) {
                 std::string tensor_name = pair.first;
                 if (tensor_name.find(prefix) == std::string::npos)
                     continue;
+                if (tensor_name.find("__index_timestep_zero__") != std::string::npos) {
+                    qwen_image_params.zero_cond_t = true;
+                }
                 size_t pos = tensor_name.find("transformer_blocks.");
                 if (pos != std::string::npos) {
                     tensor_name = tensor_name.substr(pos);  // remove prefix
@@ -580,6 +582,9 @@ namespace Qwen {
                 }
             }
             LOG_INFO("qwen_image_params.num_layers: %ld", qwen_image_params.num_layers);
+            if (qwen_image_params.zero_cond_t) {
+                LOG_INFO("use zero_cond_t");
+            }
             qwen_image = QwenImageModel(qwen_image_params);
             qwen_image.init(params_ctx, tensor_storage_map, prefix);
         }

stable-diffusion.cpp

@@ -750,6 +750,7 @@ class StableDiffusionGGML {
         }
         ignore_tensors.insert("model.diffusion_model.__x0__");
         ignore_tensors.insert("model.diffusion_model.__32x32__");
+        ignore_tensors.insert("model.diffusion_model.__index_timestep_zero__");
 
         if (vae_decode_only) {
             ignore_tensors.insert("first_stage_model.encoder");