11-quack-gemm-sm120.md

Patch 11: quack SM120 GEMM

Priority: Medium — Custom BF16/FP16 dense GEMM for SM120.

Fix

quack/gemm_sm120.py — A GemmSm120(GemmSm90) subclass implementing SM120-compatible GEMM using:

• warp.MmaF16BF16Op (32 threads) instead of WGMMA (128 threads)
• ldmatrix for SMEM→register copies
• 99 KB SMEM budget
• Optional pingpong scheduling

Based on CUTLASS example blackwell_geforce/dense_gemm.py.

Limitation: BF16/FP16 only. FP8 not supported via warp-level MMA on SM120.