GLM-5.3 SFT claimed to reach 1.8× the throughput of an earlier recipe

Prime Intellect says Qwen3-235B-A22B SFT ran 1.15× faster than torchtitan on the same 4×8 B300 nodes at 16k sequence length, with the same memory use. It credits most of the gain to mixture-of-experts work. For GLM-5.3 full SFT, it reports 1.8× its previous recipe’s throughput. Its 12-node, 16k recipe uses FlashMLA and cuDNN for sparse attention’s forward and backward passes, respectively, without context parallelism or offloading.