{ "step": 50000, "qk_mode": "dmap", "variant": "folded", "operator": "mahalanobis: -1/2||mu_i-mu_j||^2, mu=R\u00b7W_M, W_M=(W_Q+W_K)/sqrt2 (== tied W_Q=W_K); potential = column term 1/2||mu_j||^2 (metric self-affinity)", "folded": true, "attn_proj": "wmv [2d,d] (W_N dropped)", "base": "SiT-XL-2-256x256.pt", "conditional": true, "provenance": "resume", "warm_start_from": null, "surgery_rel_shift": 0.2882988154888153, "qk_rmsnorm": false, "learn_logit_scale": false, "precision": "tf32", "lr": 0.0001, "objective": "SiT transport Linear/velocity (t in [0,1])" }