The chat template here matches `zai-org/GLM-5.3` at `aca966e4`, one commit behind its main. With it, render time grows with the square of the parallel tool calls in one turn, so a large fan-out passes vLLM's 30 s render timeout. Upstream fixed this in https://huggingface.co/zai-org/GLM-5.3/discussions/31.
This replaces the file with upstream's current one (`ba32f8c1`). Nothing else changes.
| Parallel tool calls | main | this PR |
|---:|---:|---:|
| 400 | 1.70 s | 0.02 s |
| 800 | 6.7 s | 0.04 s |
| 1,600 | 27.5 s | 0.09 s |
transformers 5.22.1 with jinja2 3.1.6, one M4 core. Output is byte-identical to main on 18,000 randomized conversations.