feat/cuda: I4 - cuBLAS linear core (linear_f32 + linear_bf16) - #24

Open

magiodev wants to merge 11 commits into

antirez:mainfrom

magiodev:cuda/04-linear

Commits on Aug 12, 2026

feat/cuda: I1 scaffold - CUDA backend skeleton (probe+create/free real, 100 op stubs)
magiodev
committed
69aa9b4
feat/cuda: build CLI objects as .cuda.o with -D_GNU_SOURCE (Linux CLOCK_MONOTONIC/strdup)
magiodev
committed
c0d92e1
feat/cuda: fix nvcc strictness - seterr takes const h3_gpu*, stub returns cast to exact type
magiodev
committed
cd0e18b
feat/cuda: probe uses cudaDeviceGetAttribute(UnifiedMemory) (CUDA 13 removed prop.unifiedMemory)
magiodev
committed
a626261
feat/cuda: unified_memory=1 (GB10 unified; CUDA13 removed unifiedMemory attr)
magiodev
committed
2e42886
feat/cuda: extern "C" guards in h3_gpu.h/h3_cuda.h + include h3_cuda.h in h3_cuda.cu
magiodev
committed
65c8ac3
feat/cuda: I2 tensor layer - real device alloc/copies/file streaming, command-buffer no-ops, stats
magiodev
committed
9c1d598
feat/cuda: I3 elementwise/norm/activation/embedding CUDA kernels
magiodev
committed
e174ac4
feat/cuda: BF16 helpers -> __device__ with CUDA intrinsics (kernels call them)
magiodev
committed
f9477ba
feat/cuda: I4 cuBLAS linear core (linear_f32 + linear_bf16 real)
magiodev
committed
0987aab
feat/cuda: I4 fix - BF16 helpers before bias kernels, CUBLAS_GEMM_DEFAULT algo
magiodev
committed
3e23b2b