Full title: Port the full GLM-5.3 DSA path to sm_80 inside the Morrowmake fork (Triton MQA logits, software E4M3 writers, PP-boundary all-reduce fix) so the 4 CMP + 2 Blackwell layout runs the Pollard INT4 build (2 to 4 day port)
Reported by / on it: @darkmaniac7
Wiki page: GLM 5.3 on 8x 8 GB (vllm-170hx-glm5)
Cards: both
Card D-097, seeded from the cmpunlocker Discord with credit (reference title:eaac867d2c155142).