GLM5.1 nvfp4 support

[drdaliang]

Any plan on releasing a version that supports GLM-5.1 nvfp4 to unlock the compute power of B200?

[xysmlx]

Hi, thanks for the interest!

One thing worth flagging: actually unlocking B200's NVFP4 tensor core path requires W4A4 (4-bit weights and activations). Weight-only NVFP4 would still fall back to a higher-precision matmul and wouldn't use the NVFP4 compute path. W4A4 in turn depends on model-side support (e.g., calibration, QAT) from the model provider to preserve quality.

Currently the only available NVFP4 checkpoint for GLM-5.1 is from third-party, and there's no official W4A4 / NVFP4 release from Z.ai. So, NVFP4 inference for it isn't on our roadmap at the moment. If Z.ai officially releases it in the future, we'll revisit this.