correct handling of no-image case for qwen2-vl (#179)

davidkoski · web-flow · commit 4b67a794798d · 2025-01-27T15:36:05.000-08:00
- fixes #177
diff --git a/Libraries/MLXVLM/Models/Qwen2VL.swift b/Libraries/MLXVLM/Models/Qwen2VL.swift
@@ -783,7 +783,7 @@ public class Qwen2VL: Module, VLMModel, KVCacheDimensionProvider {
         -> MLXArray
     {
         guard let pixelValues, let gridThw else {
-            return languageModel(inputIds).logits
+            return languageModel.model.embedTokens(inputIds[.newAxis, .ellipsis])
         }
 
         // Get the input embeddings from the language model

Original file line number	Diff line number	Diff line change
`@@ -783,7 +783,7 @@ public class Qwen2VL: Module, VLMModel, KVCacheDimensionProvider {`
`783`	`783`	`-> MLXArray`
`784`	`784`	`{`
`785`	`785`	`guard let pixelValues, let gridThw else {`
`786`		`- return languageModel(inputIds).logits`
	`786`	`+ return languageModel.model.embedTokens(inputIds[.newAxis, .ellipsis])`
`787`	`787`	`}`
`788`	`788`
`789`	`789`	`// Get the input embeddings from the language model`