AMD accelerates GLM-5.2 FP8 on Instinct MI355X

Researchers benchmarked and accelerated GLM-5.2 FP8 inference on eight AMD Instinct MI355X GPUs using end-to-end optimizations including prefill context parallelism, AITER sparse-attention kernels, and FP8 KV-cache handling. EvalScope evaluations demonstrated peak prefill throughput increases up to 629K tok/s at a 64K input length and a 7.16x improvement at a 1M-token input length. GSM8K validation confirmed that these optimizations preserved model accuracy without regression. Open article →