GPU backend: GB10 (DGX Spark) — mxfp4 decode + cuBLAS, equivalence-gated

[rhCat]

Milestone: GPU compute backend on the GB10 (DGX Spark)

Hardware decoded from the box (nvidia-smi): NVIDIA GB10 — Grace Blackwell
superchip, 128 GB unified LPDDR5X, 20-core Grace, Blackwell tensor cores
(FP4/E2M1 native), CUDA 13.0. No discrete VRAM — the GPU reads the same
memory the CPU caches.

What that means for the architecture:

Lever	Effect
Unified memory	no PCIe bottleneck, zero-copy: the cache slot buffers are already GPU-visible; the mxfp4 pool + trunk can sit fully resident (memory gate becomes the residency planner)
Blackwell FP4	the pool format we already emit (E2M1 values + E8M0 block scales) is the GPU's native math type
Bandwidth wall	~273 GB/s unified; ~residency -> GB/token ~0; estimate 5-10 ms/token vs 7 s/token today (verify on the box, don't trust the estimate)

Plan:

1. Toolchain probe + equivalence gate on the box (small, run first):
   • nvcc --version, cuBLAS presence
   • a ~40-line CUDA kernel that decodes the mxfp4 pool (2 values/byte, E8M0
     scales, block32) to bf16/fp32, compared element-wise against the CPU
     reference decode on the fixture — the scalar kernels (step 2) are the
     oracle; GPU must match within fp32 accumulation tolerance
   • make gpu-probe target + fixture run on the box, paste back
2. Engine --backend gpu v1: mxfp4 decode kernel (zero-copy over the
   existing cache slots) + cuBLAS GEMM for the expert chain and router;
   --dump-state equivalence vs --backend cpu on the fixture (the
   determinism gate extended to cross-backend)
3. Residency tuning: with ~125 GB available, the gate plans pin-all-trunk
   • ~90% pool residency; measure GB/token -> ~0 and the new sec/token
4. Real-pool run on the box: --backend gpu --preset server on the
   existing ~/ds4f-mxfp4 pool

Verification policy: everything runs on the box (per the project rule —
verification on the compute machine); the Mac develops + gates the CPU
reference only.

Later (not this milestone): native FP4 GEMM (skip the decode), DSpark draft
running concurrently on the Grace CPU as the real prefetcher.