BlockManager copy constructor fix

Remove unwanted copying of BlockManager in UT.
Also Delete BlockManager copy constructor & operator= so this mistake is
never repeated.

Signed-off-by: Tomer Shmilovich <[email protected]>

Author: Tomer Shmilovich

cpp/include/tensorrt_llm/batch_manager/kvCacheManager.h

@@ -854,6 +854,9 @@ class BlockManager
         std::shared_ptr<KVCacheEventManager> eventManager = nullptr, bool enableHashKey = false,
         bool enablePartialReuse = true, bool copyOnPartialReuse = true);
 
+    BlockManager(BlockManager const&) = delete;
+    BlockManager& operator=(BlockManager const&) = delete;
+
     //! \brief Calculate the proportional share each window size receives of the total memory pool
     //! \details Example:       (uniqueWindowSizeToLayers={1024: [1], 4096: [0, 4, 5], 8192: [2, 3]})
     //!          Would Return:  {1024: 0.0345, 4096: 0.4138, 8192: 0.5517} [sums to 1.0].

cpp/tests/unit_tests/batch_manager/kvCacheManagerTest.cpp

@@ -2115,7 +2115,7 @@ TEST_F(KVCacheManagerTest, KVCacheManagerLeafBlockWithDependentTest)
     EXPECT_THAT(cacheBlockIds0, ::testing::ElementsAreArray({0, 1, 2}));
 
     // Lower priority of middle block to prevent offloading
-    auto const blockManager = kvCacheManager.getBlockManager();
+    auto const& blockManager = kvCacheManager.getBlockManager();
     auto middleBlock = blockManager.getBlockById(cacheBlockIds0[1], maxAttentionWindow);
     middleBlock->setPriority(0);
 
@@ -2215,7 +2215,7 @@ TEST_P(KVCacheManagerTest, DISABLED_KVCacheManagerAllocationTest)
             std::nullopt, nvinfer1::DataType::kHALF, sinkTokenLength, stream, std::nullopt, enableBlockReuse,
             onboardBlocks);
 
-    auto const blockManager = kvCacheManager.getBlockManager();
+    auto const& blockManager = kvCacheManager.getBlockManager();
     auto const& bufferManager = blockManager.getBufferManager(theOnlyWindowSize(kvCacheManager));
     auto const memoryPoolUsedBefore = bufferManager.memoryPoolUsed();
     kvCacheManager.allocatePools(useUvm);