Function TEST_F

tests/test_tensor_stress.cpp:116–158

// ============= Memory Management Tests =============

TEST_F(TensorStressTest, MaxMemoryAllocation) {
    // Try to allocate tensors until we approach memory limit
    std::vector<Tensor> tensors;
    const size_t chunk_size = 10 * 1024 * 1024; // 10M floats = 40MB per tensor

    size_t total_allocated = 0;

    for (int i = 0; i < 50; ++i) { // Max 50 tensors (2GB) to avoid hanging
        // Check available memory before allocating
        size_t free_mem, total;
        cudaMemGetInfo(&free_mem, &total);
        if (free_mem < chunk_size * sizeof(float) * 3) {
            // Stop before we run too low
            break;
        }

        auto t = Tensor::empty({chunk_size}, Device::CUDA);
        if (!t.is_valid()) {
            break;
        }
        tensors.emplace_back(std::move(t));
        total_allocated += chunk_size * sizeof(float);
    }

    EXPECT_GT(tensors.size(), 0) << "Should be able to allocate at least one tensor";

    // Clear tensors to free memory
    tensors.clear();
    cudaDeviceSynchronize();

    // Clear PyTorch cache to actually release memory
    c10::cuda::CUDACachingAllocator::emptyCache();

    // Clear tensor library's memory pool cache
    Tensor::trim_memory_pool();
    cudaDeviceSynchronize();

    // Memory should be freed
    size_t free_after, total_after;
    cudaMemGetInfo(&free_after, &total_after);
    EXPECT_GT(free_after, initial_free_mem_ - 100 * 1024 * 1024)
        << "Memory not properly freed after clearing tensors";
}

TEST_F(TensorStressTest, RapidAllocationDeallocation) {
    // Rapidly allocate and deallocate tensors