"""Canonical shape sweep for FP8 GEMM. Mix of: - square aligned (the easy case) - off-alignment K (common real-world failure mode for tile-quantized kernels) - skinny (decode-like, memory-bound) - rectangular (prefill with grouped attention) """ SHAPES = [ {"M": 4096, "N": 4096, "K": 4096}, # square aligned {"M": 4096, "N": 4096, "K": 4127}, # K not multiple of 128 -> forces predicated tails {"M": 32, "N": 8192, "K": 8192}, # skinny M (decode) {"M": 4096, "N": 14336, "K": 4096}, # Llama3 up-proj shape ]