Cherry pick Fara text only fix for 0.11.4 release (#1913)

### Description

This PR cherry-picks the Fara text-only fix PR into rel-0.11.4.

### Motivation and Context

This cherry-pick needs to be included for the 0.11.4 patch release.

Co-authored-by: Akshay Sonawane <111780983+apsonawane@users.noreply.github.com>

Author: kunal-vaishnavi

src/models/qwen2_5_vl_image_processor.cpp

@@ -27,10 +27,6 @@ Qwen2_5VLImageProcessor::Qwen2_5VLImageProcessor(Config& config, const SessionIn
 }
 
 std::unique_ptr<NamedTensors> Qwen2_5VLImageProcessor::Process(const Tokenizer& tokenizer, const Payload& payload) const {
-  if (!payload.images) {
-    throw std::runtime_error("No images provided to Qwen2.5VLImageProcessor");
-  }
-
   std::string prompt = std::string(payload.prompt);
   Ort::Allocator& allocator{Ort::Allocator::GetWithDefaultOptions()};
   auto named_tensors = std::make_unique<NamedTensors>();
@@ -41,6 +37,11 @@ std::unique_ptr<NamedTensors> Qwen2_5VLImageProcessor::Process(const Tokenizer&
   std::copy(input_ids.begin(), input_ids.end(), input_ids_value->GetTensorMutableData<int32_t>());
   named_tensors->emplace(Config::Defaults::InputIdsName, std::make_shared<Tensor>(std::move(input_ids_value)));
 
+  if (!payload.images) {
+    // No images provided - return text-only tensors
+    return named_tensors;
+  }
+
   // Run image preprocessing using onnxruntime-extensions
   // This will execute the full pipeline from processor_config.json:
   // DecodeImage -> ConvertRGB -> Resize (smart_resize) -> Rescale -> Normalize -> PatchImage

src/models/qwen_vl_model.cpp

@@ -70,7 +70,7 @@ void Qwen2_5_VL_PipelineState::SetExtraInputs(const std::vector<ExtraInput>& ext
     }
   }
   if (!pixel_values_val) {
-    throw std::runtime_error("Vision pipeline: pixel_values input not found in extra_inputs");
+    return;
   }
 
   auto pixel_type_info = pixel_values_val->GetTensorTypeAndShapeInfo();