Summary

Adds our fork for the NVIDIA DGX Spark (GB10 chip, Blackwell sm_121a, 128GB unified memory).

The upstream code targets H100 and uses FA3 kernels + kernels package, neither of which work on the GB10's sm_121a architecture. This fork gets autoresearch running out-of-the-box on DGX Spark with the following changes:

• Replace FA3 with flex_attention — FA3 kernels are Hopper-only (sm_90). flex_attention uses Triton-generated kernels that work on any architecture including Blackwell sm_121a, with sliding window support via create_block_mask
• Switch PyTorch wheels from cu128 to cu130 — cu128 only supports up to sm_120. cu130 ships CUDA 13.0 with native sm_121a codegen and cuDNN 9.13, which gave us a 2.86x throughput improvement (49K → 140K tok/sec)
• Python 3.10 → 3.13 — better compatibility with the cu130 torch wheels and other updated dependencies
• Fix MFU calculation — report against GB10's 125 TFLOPS BF16 peak instead of H100's 989.5 TFLOPS (otherwise MFU reads ~3% when it's actually ~27%)
• Drop kernels dependency — not needed once FA3 is replaced with flex_attention

We also benchmarked FA4 (blake-snc sm120 fork) and cuDNN SDPA as alternative attention backends — both showed parity or worse vs flex_attention on this hardware. The GB10 is memory-bandwidth limited at 273 GB/s, so most kernel-level optimizations don't move the needle.