TensorCraft-HPC is a header-only C++/CUDA kernel library and technical whitepaper / architecture showcase for learning, validating, and packaging modern AI operators. The repository now presents itself as a three-part public surface: a whitepaper for the thesis, an academy for the learning path, and an evidence layer for benchmarks, references, and research framing.

cmake --preset cpu-smoke
cmake --build --preset cpu-smoke --parallel 2
cmake --install build/cpu-smoke --prefix /tmp/tensorcraft-install
python3 -m build --wheel

cmake --preset dev
cmake --build --preset dev --parallel $(nproc)
ctest --preset dev --output-on-failure

#include "tensorcraft/kernels/gemm.hpp"
#include "tensorcraft/memory/tensor.hpp"

// Create GPU tensors (RAII-managed)
tensorcraft::FloatTensor A({4096, 4096});
tensorcraft::FloatTensor B({4096, 4096});
tensorcraft::FloatTensor C({4096, 4096});

// Optimized GEMM (92% cuBLAS performance)
tensorcraft::kernels::gemm(A.data(), B.data(), C.data(), 4096, 4096, 4096);

import tensorcraft_ops as tc
import numpy as np

# GPU-accelerated GEMM
A = np.random.randn(4096, 4096).astype(np.float16)
B = np.random.randn(4096, 4096).astype(np.float16)
C = tc.gemm(A, B)

# FlashAttention
Q, K, V = [np.random.randn(32, 128, 64).astype(np.float16) for _ in range(3)]
output = tc.flash_attention(Q, K, V)

Benchmarks on A100 80GB, CUDA 12.4, FP16 Tensor Core

• Showcase home: https://aicl-lab.github.io/modern-ai-kernels/
• Whitepaper: https://aicl-lab.github.io/modern-ai-kernels/en/whitepaper/
• Academy: https://aicl-lab.github.io/modern-ai-kernels/en/academy/
• Evidence: https://aicl-lab.github.io/modern-ai-kernels/en/evidence/
• Kernel Atlas / API: https://aicl-lab.github.io/modern-ai-kernels/en/api/gemm
• 中文文档: https://aicl-lab.github.io/modern-ai-kernels/zh/

This repository uses OpenSpec as the active development workflow:

1. Review accepted specs in openspec/specs/
2. Create or update changes under openspec/changes/
3. Implement against that change
4. Run validation before merge
5. Use /review before merging structural changes

modern-ai-kernels/
├── include/tensorcraft/   # Header-only C++/CUDA library
├── src/python_ops/        # Python bindings
├── tests/                 # Validation
├── benchmarks/            # Benchmark binaries
├── docs/                  # GitHub Pages + documentation
├── openspec/              # Active spec workflow
└── .github/               # Workflows, templates

• Build system: CMake presets
• Formatting: .clang-format, .clang-tidy, pre-commit
• LSP: clangd with compile_commands.json
• GitHub automation: CI, Pages, release workflow

Contributions are welcome! Please:

1. Read the OpenSpec workflow
2. Follow the code style (run pre-commit hooks)
3. Add tests for new functionality
4. Update documentation

Released under the MIT License.