This is the first stable release of mini-tensor, a PyTorch-inspired C++ tensor library designed for systems-level learning and performance experimentation.

🔧 Key Features

• 2D and 3D tensors with slicing, reshaping, and broadcasting
• Arithmetic and matrix ops (manual, Eigen, and CUDA-based)
• Neural network layers: Linear, ReLU, Softmax, Sequential
• IR tracing system with named tensor IDs and introspection
• CUDA support for mat_mul_cuda() and bmm_cuda()
• Fused CUDA kernel: bmm_add_cuda() for matmul + bias
• Benchmarks, reproducible builds, and device-safe memory

📊 Performance

• Up to 600×–700× speedup on batched matmul (GPU vs CPU)
• Benchmarks shown in README and demo.md

🧠 Why This Exists

Built as a solo project to explore inference system internals, GPU kernel integration, and forward-pass execution from first principles.

🔗 Full code, tests, and benchmarks: README