MXFP4 MoE quantization: per-expert scales and block scaling for mixture-of-experts

MoE quantization is challenging because different experts have different activation distributions. The existing moe-mixed-precision record documents that shared experts (accessed by all tokens) may need higher precision than token-conditioned experts. For MXFP4 MoE, the question is whether the 32-element block size is appropriate for MoE layers. The SGLang docs show that MXFP4 is supported for MoE on Ascend NPUs (950PR/DT Series) with both online and offline quantization. The docs also show that modelopt_fp4 supports 'load-time BF16/FP16/FP8 MoE conversion with per-tensor FP32 activation scales'. This suggests that per-tensor scales (not per-expert) are used for MoE, which may be suboptimal. The block scaling in MXFP4 (32 elements) may not align well with MoE's expert structure. RELEVANCE TO THE REPO'S THEME: MoE quantization is another case where the format may not match the model structure. The 32-element block size is designed for dense layers, but MoE layers have different characteristics. This is a potential source of silent accuracy degradation that has not been well-documented.