b12x 0.1.0

Unapologetically SM120-only CuTe DSL kernels for NVFP4 GEMM and MoE.

b12x

b12x is an unapologetically SM120-only CuTe DSL kernel library for Blackwell-class NVFP4 GEMM and routed Mixture-of-Experts inference.

The project is intentionally narrow. It is not a generic CUDA kernel collection for LLM inference. It is a focused package for shipping a small number of high-performance kernels and the runtime glue needed to launch them cleanly from PyTorch and sglang/vLLM.

What It Includes

• Clean, standalone CuTe DSL dense NVFP4 GEMM kernels
• A fused tensor-parallel MoE path for NVFP4 expert weights
• CuTe DSL FP4 packing, scaling, and quantization helpers
• Torch reference implementations for correctness checking
• Lightweight integration surfaces for PyTorch and sglang

What It Does Not Try To Be

• Multi-architecture
• Backward-compatible with pre-SM120 GPUs
• A model-serving framework
• A wrapper around inherited FlashInfer runtime code

Requirements

• NVIDIA Blackwell SM120 GPU
• CUDA 13 toolchain
• CUDA 13 PyTorch build, torch>=2.10.0
• nvidia-cutlass-dsl[cu13]==4.4.1

Package Layout

• b12x.gemm
  • Dense NVFP4 GEMM kernels
• b12x.integration
  • Public runtime entrypoints such as b12x_moe_fp4
• b12x.moe.fused
  • The fused MoE kernel, scheduler, and reference paths
• b12x.quant
  • Expert-weight quantization helpers
• b12x.sglang
  • sglang integration shims

The published wheel only contains the b12x package tree. Benchmarks, experiments, and tests remain in the source repo.