DeepSpeed

2da647d2 - perf(kernels): vectorized 16-byte staging in the grouped GEMM kernels

perf(kernels): vectorized 16-byte staging in the grouped GEMM kernels The smem staging loaded one bf16 per thread iteration (2-byte transactions); stage 8 bf16 per uint4 load instead. Wrapper guards enforce K % 8 == 0 and 16B-aligned pointers; kn and the smem strides are multiples of 8 by construction. Measured at bs=32: 775.9 -> 1233.1 tok/s (+59%, now faster than vLLM's 1127.2 on the same GPU); 2-byte loads were wasting 7/8 of each memory transaction. Signed-off-by: Guokai Ma <guokai.ma@intel.com>