Fix OOM in DEQUANTIZE_TO_FP32 E2E benchmark test

Use MemorySegmentTensorDataFactory for off-heap FP32 weight storage and
increase MaxDirectMemorySize to 12g so multiple large-model tests can
run in the same JVM without exhausting direct buffer memory.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: michalharakal

skainet-apps/skainet-kllama/build.gradle.kts

@@ -113,7 +113,7 @@ kotlin {
 }
 
 tasks.withType<Test>().configureEach {
-    jvmArgs("--enable-preview", "--add-modules", "jdk.incubator.vector")
+    jvmArgs("--enable-preview", "--add-modules", "jdk.incubator.vector", "-XX:MaxDirectMemorySize=12g")
     maxHeapSize = "6g"
 }
 

skainet-apps/skainet-kllama/src/jvmTest/kotlin/sk/ainet/apps/kllama/E2EQuantizedInferenceTest.kt

@@ -205,8 +205,10 @@ class E2EQuantizedInferenceTest {
     fun `benchmark DEQUANTIZE_TO_FP32 baseline inference`() {
         if (skipIfNoModel()) return
         runBlocking {
-            // Use default heap-based factory for dequantized FP32 weights
-            val ctx = DirectCpuExecutionContext()
+            // Use MemorySegment factory so dequantized FP32 weights live off-heap
+            val arena = Arena.ofShared()
+            val memSegFactory = MemorySegmentTensorDataFactory()
+            val ctx = DirectCpuExecutionContext(tensorDataFactory = memSegFactory)
 
             val ingestion = LlamaIngestion<FP32>(
                 ctx = ctx,
@@ -259,6 +261,9 @@ class E2EQuantizedInferenceTest {
             println("==============================================")
 
             assertTrue(tokPerSec > 0, "Should produce positive throughput")
+
+            arena.close()
+            memSegFactory.close()
         }
     }
 }