Use quantization-specific activation init in Qwen2 and Deepseek models

Author: orionpapadakis

src/main/java/org/beehive/gpullama3/inference/state/Qwen2State.java

@@ -41,7 +41,11 @@ protected StateFields createStateFields(Configuration configuration) {
         fields.valueCache = Stream.generate(() -> ArrayFloatTensor.allocate(config.contextLength(), nEmbdGqa)).limit(config.numberOfLayers()).toArray(FloatTensor[]::new);
 
         // TornadoVM wrappers with Qwen2 dimensions
-        fields.embeddingX = new HalfFloatArray(config.dim());
+        switch (config.modelType()) {
+            case "FP16" -> fields.createActivationFP16(config.dim());
+            case "Q8_0" -> fields.createActivationQ8_0(config.dim());
+            default -> throw new UnsupportedOperationException("Quantization format " + config.modelType());
+        }
         fields.wrapX = new FloatArray(config.dim());
         fields.wrapXb = new FloatArray(config.dim());
         fields.wrapXb2 = new FloatArray(config.dim());