https://pengchengneo.github.io/Recent content on Pengcheng's BlogHugo -- gohugo.iozh-CN© 2026 PengchengSat, 04 Apr 2026 00:00:00 +0000https://pengchengneo.github.io/posts/gemm-kernel-optimization-10-questions/Sat, 04 Apr 2026 00:00:00 +0000https://pengchengneo.github.io/posts/gemm-kernel-optimization-10-questions/基于 Pallas GMM FP8 blockwise 量化内核开发的实战问答，涵盖量化粒度、子通道循环、分阶段 tiling、编译链路、精度对齐方法论等核心话题。https://pengchengneo.github.io/projects/sglang-jax/Fri, 20 Mar 2026 00:00:00 +0000https://pengchengneo.github.io/projects/sglang-jax/SGLang 的 JAX 后端实现，支持在 TPU 上运行高效推理。https://pengchengneo.github.io/posts/sglang-jax-tpu-inference/Wed, 29 Oct 2025 00:00:00 +0000https://pengchengneo.github.io/posts/sglang-jax-tpu-inference/SGLang-JAX 是基于 JAX 和 XLA 构建的开源推理引擎，支持 continuous batching 和 speculative decoding，在 TPU 上实现高效原生推理。