Debug and add to-do for next steps

Author: DePasqualeOrg

Libraries/MLXVLM/Models/Qwen2VL.swift

@@ -693,18 +693,22 @@ public class Qwen2VLProcessor: UserInputProcessor {
         if let role = messages[0]["role"] as? String, role != "system" {
             messages.insert(["role": "system", "content": "You are a helpful assistant."], at: 0)
         }
-        //        // Add image markers to last message if needed
-        //        if let imageTHW {
-        //            let lastIndex = messages.count - 1
-        //            var content = messages[lastIndex]["content"] ?? ""
-        //            let mergeLength = config.mergeSize * config.mergeSize
-        //            for thw in imageTHW {
-        //                content += "<|vision_start|>"
-        //                content += Array(repeating: "<|image_pad|>", count: thw.product / mergeLength).joined()
-        //                content += "<|vision_end|>"
-        //            }
-        //            messages[lastIndex]["content"] = content
-        //        }
+
+        //  // Add image markers to last message if needed
+        //  if let imageTHW {
+        //      let lastIndex = messages.count - 1
+        //      var content = messages[lastIndex]["content"] as? String ?? ""
+        //      let mergeLength = config.mergeSize * config.mergeSize
+        //      for thw in imageTHW {
+        //          content += "<|vision_start|>"
+        //          content += Array(repeating: "<|image_pad|>", count: thw.product / mergeLength).joined()
+        //          content += "<|vision_end|>"
+        //      }
+        //      messages[lastIndex]["content"] = content
+        //  }
+
+        // TODO: Instead of the above, replace the single `<|image_pad|>` with repeated padding, using the same logic as above to determine the number of repeats.
+
         return messages
     }
 
@@ -730,6 +734,11 @@ public class Qwen2VLProcessor: UserInputProcessor {
         // Prepare messages with image markers
         let messages = prepareMessages(input.prompt.asMessages(), imageTHW: image.imageGridThw)
         let promptTokens = try tokenizer.applyChatTemplate(messages: messages)
+
+        // TODO: For debugging. Remove later.
+        let promptTokensDecoded = try tokenizer.decode(tokens: promptTokens)
+        print(promptTokensDecoded)
+
         let promptArray = MLXArray(promptTokens).expandedDimensions(axis: 0)
         let mask = ones(like: promptArray).asType(.int8)
         return LMInput(text: .init(tokens: promptArray, mask: mask), image: image)