GLM agent built its own inference infra in two weeks, tripling end-to-end throughput

jietang · x · 2026-09-17

Zhipu engineer jietang shares how GLM-5.3-Flash went from first run on domestic accelerators to serving all production traffic in two weeks, with 3.2× end-to-end throughput — with most optimization done by an Infra Agent powered by GLM-5.3 itself.

The constraints were tough: limited memory and interconnect bandwidth, 1M-token context, multimodal requests, and an immature stack with missing kernels. Key trade-offs: compute for memory (ReplaySSM), communication for memory (intra-node tensor parallelism), precision for capacity (mixed INT8/FP8/BF16 caching), and disaggregation for scheduling freedom (Encode–Prefill–Decode).

The most important lesson wasn't any single optimization but dense feedback: local correctness tests, execution traces, microbenchmarks, and end-to-end measurements enabling targeted hypothesis testing instead of relying on aggregate metrics alone.

Original post →

More from coding & agent

coding & agent channel →