Introduce TornadoWeights for Qwen2

Author: orionpapadakis

src/main/java/org/beehive/gpullama3/inference/weights/standard/Qwen2StandardWeights.java

@@ -6,6 +6,7 @@
 import org.beehive.gpullama3.inference.weights.Weights;
 
 public class Qwen2StandardWeights extends StandardWeights {
+    // Qwen2-specific weights
     public final FloatTensor[] q_bias, k_bias, v_bias;
 
     public Qwen2StandardWeights(

src/main/java/org/beehive/gpullama3/inference/weights/tornado/Qwen2TornadoWeights.java

@@ -0,0 +1,42 @@
+package org.beehive.gpullama3.inference.weights.tornado;
+
+import org.beehive.gpullama3.core.model.GGMLType;
+import uk.ac.manchester.tornado.api.types.arrays.FloatArray;
+import uk.ac.manchester.tornado.api.types.arrays.HalfFloatArray;
+
+public class Qwen2TornadoWeights extends TornadoWeights {
+
+    // Qwen2-specific tornado weights
+    FloatArray[] q_biasLayered;
+    FloatArray[] k_biasLayered;
+    FloatArray[] v_biasLayered;
+
+    public Qwen2TornadoWeights(FloatArray tokenEmbeddingTable, FloatArray[] rms_att_weightLayered, HalfFloatArray[] wqLayered, HalfFloatArray[] wkLayered, HalfFloatArray[] wvLayered,
+            FloatArray[] wqBiasLayered,
+            FloatArray[] wkBiasLayered,
+            FloatArray[] wvBiasLayered,
+            HalfFloatArray[] woLayered, FloatArray[] rms_ffn_weightLayered, HalfFloatArray[] w1Layered,
+            HalfFloatArray[] w2Layered, HalfFloatArray[] w3Layered, FloatArray rms_final_weight_as_floatArray, FloatArray freq_cis_realFlat, FloatArray freq_cis_imagFlat, HalfFloatArray wclsByteArray,
+            GGMLType weightType) {
+        // call to TornadoWeights constructor
+        super(tokenEmbeddingTable,
+                rms_att_weightLayered,
+                wqLayered,
+                wkLayered,
+                wvLayered,
+                woLayered,
+                rms_ffn_weightLayered,
+                w1Layered,
+                w2Layered,
+                w3Layered,
+                rms_final_weight_as_floatArray,
+                freq_cis_realFlat,
+                freq_cis_imagFlat,
+                wclsByteArray,
+                weightType);
+        // init qwen2-specific fields
+        this.q_biasLayered = wqBiasLayered;
+        this.k_biasLayered = wkBiasLayered;
+        this.v_biasLayered = wvBiasLayered;
+    }
+}

src/main/java/org/beehive/gpullama3/model/loader/Qwen2ModelLoader.java

@@ -20,6 +20,7 @@
 import org.beehive.gpullama3.tokenizer.impl.Qwen3Tokenizer;
 import org.beehive.gpullama3.tokenizer.impl.Tokenizer;
 import org.beehive.gpullama3.tokenizer.vocabulary.Vocabulary;
+import uk.ac.manchester.tornado.api.types.arrays.FloatArray;
 
 import java.io.IOException;
 import java.nio.channels.FileChannel;
@@ -112,7 +113,6 @@ public Weights loadWeights(Map<String, GGMLTensorEntry> tensorEntries, Configura
             return createStandardWeights(tensorEntries, config, ropeFreqs, tokenEmbeddings, outputWeight);
         }
     }
-    // @formatter:on
 
     @Override
     public Weights createStandardWeights(Map<String, GGMLTensorEntry> tensorEntries, Configuration config, Pair<float[], float[]> ropeFreqs, GGMLTensorEntry tokenEmbeddings,
@@ -141,6 +141,30 @@ public Weights createStandardWeights(Map<String, GGMLTensorEntry> tensorEntries,
     }
 
     @Override
+    public Weights createTornadoVMWeights(Map<String, GGMLTensorEntry> tensorEntries, Configuration config, Pair<float[], float[]> ropeFreqs, GGMLTensorEntry tokenEmbeddings,
+            GGMLTensorEntry outputWeight) {
+        return new Qwen2TornadoWeights(
+                loadTensorAsFloatArray(tokenEmbeddings),
+                loadArrayAsFloatArrayFromBuffer(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".attn_norm.weight")),
+                loadArrayAsHalfFloatArray(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".attn_q.weight")),
+                loadArrayAsHalfFloatArray(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".attn_k.weight")),
+                loadArrayAsHalfFloatArray(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".attn_v.weight")),
+                // Qwen2-specific: qkv bias
+                loadArrayAsFloatArray(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".attn_q.bias")),
+                loadArrayAsFloatArray(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".attn_k.bias")),
+                loadArrayAsFloatArray(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".attn_v.bias")),
+
+                loadArrayAsHalfFloatArray(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".attn_output.weight")),
+                loadArrayAsFloatArrayFromBuffer(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".ffn_norm.weight")),
+                loadArrayAsHalfFloatArray(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".ffn_gate.weight")),            // w1
+                loadArrayAsHalfFloatArray(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".ffn_down.weight")),            // w2
+                loadArrayAsHalfFloatArray(config.numberOfLayers(), i -> tensorEntries.get("blk." + i + ".ffn_up.weight")),              // w3
+                floatBufferToFloatArray(tensorEntries.get("output_norm.weight")),
+                FloatArray.fromArray(ropeFreqs.first()),
+                FloatArray.fromArray(ropeFreqs.second()),
+                loadTensorAsHalfFloatArray(outputWeight),
+                outputWeight.ggmlType()
+        );
     }
     // @formatter:on