DeepSpeed

d9abdac2 - feat(kernels): grouped GEMM for decode batches 2..32 — one launch, no weight copies

feat(kernels): grouped GEMM for decode batches 2..32 — one launch, no weight copies External bench feedback: at bs>1 the weight-bound projections run as separate GEMMs (SGLang/vLLM use merged gate|up and qkv weights), and the gap to SGLang grows with batch (81% at bs=1 -> 66% at bs=32). Add a weight-streaming grouped GEMM: one launch covers up to four original weight matrices (each element read exactly once, activation tile in shared memory, fp32 accumulation), so the launch structure of a merged GEMM is achieved without copying or reordering weights — preserving the read-original-Parameters contract for live-weight co-location. Entry points: triple_gemm (q/k/v), quad_gemm (GDN in_proj, concat output), dual_gemm_silu_mul (MLP with silu(g)*u epilogue). Dispatched for decode batches 2..32 alongside the existing b=1 GEMV path; larger batches stay on cuBLAS composite. Signed-off-by: Guokai Ma <guokai.ma@intel.com>