Feature request: GLM-5.2 support / roadmap?

Hi TileRT team,

Thanks for releasing TileRT. We are evaluating low-latency GLM-5.2 serving on 8xB200 and wanted to ask whether GLM-5.2 support is on the roadmap, and if there is any tentative ETA or blocker list.

Context:

• I saw the helpful response in #41 that GLM-5.1 NVFP4 support is not straightforward because unlocking the B200 NVFP4 tensor core path requires W4A4 / model-side support.
• GLM-5.2 appears to differ from GLM-5.1 architecturally, including shared DSA-related weights, and public deployments have mentioned FP8 weights plus NVFP4 quantization for Blackwell.
• We are mainly trying to understand whether TileRT could support GLM-5.2 decode on Blackwell, even if the first supported path is FP8 or requires a specific calibrated NVFP4/W4A4 checkpoint.

A few questions:

1. Is GLM-5.2 support planned for TileRT?
2. Would support require an official Z.ai W4A4/NVFP4 checkpoint, or could an FP8/NVFP4 conversion path be supported?
3. If it is planned, is there any rough ETA?
4. If community contributions are welcome, which pieces would be most useful: weight conversion, shared DSA handling, MoE routing, MTP/speculation support, benchmark harnesses, or something else?

We can help test on 8xB200 and share logs/benchmarks if useful.

Thanks again for the work on TileRT.

Start by reviewing issue #41 and the current GLM-5.1, Blackwell, and quantization support described in the repository. Determine whether GLM-5.2 support would require shared DSA handling, MoE routing, MTP/speculation, weight conversion, or benchmark work; done would require an agreed implementation scope and validation on 8xB200.