{ "python": "3.13.5", "pytorch": "2.10.0+cpu", "config": { "seq": 5, "batch": 2, "hidden": 16, "query_heads": 4, "kv_heads": 2, "head_dim": 4, "world_size": 2, "eps": 1e-06 }, "dense_shapes": { "x": [ 5, 2, 16 ], "q": [ 5, 2, 4, 4 ], "k_v": [ 5, 2, 2, 4 ] }, "rank_local": { "query_heads": 2, "kv_heads": 1 }, "rule": "keep complete query/KV groups on each rank; sum output-projection partials" }