Working

Libraries/MLXVLM/Models/Qwen2VL.swift

@@ -686,37 +686,22 @@ public class Qwen2VLProcessor: UserInputProcessor {
         return (flattenedPatches, .init(gridT, gridH, gridW))
     }
 
-    private func prepareMessages(_ messages: [Message], imageTHW: [THW]?) -> [Message] {
+    private func prepareMessages(_ messages: [Message]) -> [Message] {
         var messages = messages
         print(messages)
         // Add system message if not present
         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"] 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
     }
 
-    public func prepare(prompt: UserInput.Prompt, imageTHW: [THW]?) throws -> String {
-        let messages = prepareMessages(prompt.asMessages(), imageTHW: imageTHW)
-        let tokens = try tokenizer.applyChatTemplate(messages: messages)
-        return tokenizer.decode(tokens: tokens)
-    }
+    //    public func prepare(prompt: UserInput.Prompt, imageTHW: [THW]?) throws -> String {
+    //        let messages = prepareMessages(prompt.asMessages())
+    //        let tokens = try tokenizer.applyChatTemplate(messages: messages)
+    //        return tokenizer.decode(tokens: tokens)
+    //    }
 
     public func prepare(input: UserInput) throws -> LMInput {
         // Text-only input
@@ -725,15 +710,34 @@ public class Qwen2VLProcessor: UserInputProcessor {
             let promptTokens = try tokenizer.applyChatTemplate(messages: messages)
             return LMInput(tokens: MLXArray(promptTokens))
         }
+
         // Input with images
         let images = try input.images.map {
             try preprocess(images: [$0.asCIImage()], processing: input.processing)
         }
         let pixels = concatenated(images.map { $0.0 })
         let image = LMInput.ProcessedImage(pixels: pixels, imageGridThw: images.map { $0.1 })
-        // Prepare messages with image markers
-        let messages = prepareMessages(input.prompt.asMessages(), imageTHW: image.imageGridThw)
-        let promptTokens = try tokenizer.applyChatTemplate(messages: messages)
+
+        // Get tokens from messages
+        let messages = prepareMessages(input.prompt.asMessages())
+        var promptTokens = try tokenizer.applyChatTemplate(messages: messages)
+
+        // Replace single image pad token with correct number for each image
+        let imagePadToken = try tokenizer.encode(text: "<|image_pad|>").first!
+        let mergeLength = config.mergeSize * config.mergeSize
+
+        // TODO: This assumes that there is only one image. A better solution is needed for the case when multiple images are included.
+        if let imageGridThw = image.imageGridThw {
+            for thw in imageGridThw {
+                if let padIndex = promptTokens.firstIndex(of: imagePadToken) {
+                    let paddingCount = thw.product / mergeLength
+                    promptTokens.replaceSubrange(
+                        padIndex ... (padIndex),
+                        with: Array(repeating: imagePadToken, count: paddingCount)
+                    )
+                }
+            }
+        }
 
         // TODO: For debugging. Remove later.
         let promptTokensDecoded = try tokenizer.decode(tokens: promptTokens)