Exact BF16 GEMM on FP16 tensor cores (V100 / sm_70)

CUDA port + extensions of "Exact BF16 Product Packing into FP16 Matrix Multiplication". Volta has FP16 tensor cores but no BF16 support; this suite provides BF16 GEMM as a menu of speed/accuracy modes, all built on the same CUTLASS-sm70 mainloop with an empirically probed accumulator mapping.

• src/ implementation: common.cuh (helpers), exact_mma.cuh (CUTLASS config, rebase mainloop, probe), rowscan.cuh, one *_impl.cuh per mode.
• tests/ correctness: make test runs bit-exact/accuracy validation of all three modes (zeros, subnormals, Inf/NaN, odd shapes, permutation path).
• experiments/ benchmarks (bench_*.cu) and historical prototypes: bf16_pack_gemm.cu (paper's 9:8 K<=32 repro), deep_k_pack.cu, full_k_pack.cu (12:11 limbs), sig_gemm.cu (WMMA 1:1), cublas_roofline.cu, cutlass_hgemm_ref.cu (attribution baselines).
• analysis/ real-ML BF16 studies (SmolLM2 weights/activations).
• python/ the voltabf16 PyTorch bindings (see below), docs/ their Sphinx site.

Build on the GPU host: make -j4 && make test (needs CUDA 12.6 for sm_70 and a CUTLASS checkout; see CUTLASS/NVCC vars).

PyTorch bindings: voltabf16

pip install voltabf16 exposes the scaled_impl mode to PyTorch, so bfloat16 training on a V100 uses the FP16 tensor cores instead of torch's non-tensor-core fallback:

import voltabf16

with voltabf16.autobf16():
    loss = model(x.bfloat16()).sum()
    loss.backward()

Nothing in the model changes -- a TorchDispatchMode intercepts bfloat16 aten::mm/aten::addmm below autograd, so forward and both backward GEMMs are routed. MNIST MLP (784 -> 4096x3 -> 10), batch 2048, 5 epochs:

4.3x faster than torch's bfloat16, and FP32-class accurate (~1e-7 relative error against float64, versus ~1.7e-3 native) because the products are exact. It does not reach FP16 parity and cannot: the mainloop is ~78-83% of cuBLAS FP16, and the exponent scan and pack passes are inherent. docs/performance.md has the full accounting.

The binding is built against this repository's own src/*.cuh -- there is one copy of those headers, shared with make test. setup.py copies them into wheels at build time so an installed wheel is self-contained; an editable install or a run from the checkout uses src/ directly.

pip install -e '.[test,docs,bench]'
pytest                                    # needs a V100; skips otherwise
make -C docs html
python python/examples/bench_mnist.py --epochs 5 --batch 2048 --lr 0.15
make bin/tile_sweep && ./bin/tile_sweep 2048     # CUTLASS tile-shape sweep

.github/workflows/docs.yml builds and publishes the docs; .github/workflows/pypi.yml publishes voltabf16 to PyPI via Trusted Publishing when a GitHub Release is cut.