This is the first post on blog.higcp.com. The blog is built with Jekyll on GitHub Pages, with a custom skin mimicking Google Cloud Console design: white background, Google Sans typography, Google Blue accents, no gradients, no decorative emoji.

What I’ll write about

  • TPU v7 (Ironwood) — training and inference experience: model loading, checkpoint conversion, sharding strategies, performance optimization.
  • GPU inference — vLLM and SGLang deployment notes: model registration quirks, MoE prefetch deadlocks, KV cache tuning, FP8/FP4 trade-offs.
  • Multi-agent systems — running multiple LLM-powered bots on the same infrastructure, IPC patterns, debugging cold-path bugs.
  • Cloud infra — GCP Cloud DNS, GKE topology, Cloud Storage gotchas, cross-project IAM headaches.

Why Jekyll

Three reasons:

  1. Markdown all the way down. Source files are just .md text under _posts/. No CMS, no DB, no auth. git push is the publish button.
  2. GitHub Pages handles hosting + HTTPS. Let’s Encrypt certificate provisioned automatically for the blog.higcp.com custom domain. Zero server maintenance.
  3. The default theme minima is solid. With a small SCSS override file (_sass/gcp-overrides.scss), it ports cleanly to Material Design without forking a heavy theme.

Code style example

import jax
import jax.numpy as jnp

@jax.jit
def matmul(x: jax.Array, y: jax.Array) -> jax.Array:
    return x @ y

# Trillium-class TPU (v5p) — 4096 chips, BF16
x = jnp.ones((8192, 8192), dtype=jnp.bfloat16)
y = jnp.ones((8192, 8192), dtype=jnp.bfloat16)
out = matmul(x, y)
print(out.shape)  # (8192, 8192)

Tables for hardware specs

Chip HBM BF16 TFLOPS FP8 TFLOPS Pod scale
TPU v5p 95 GB 459 — 8,960 chips
TPU v7 (Ironwood) 192 GB ~2,307 4,614 9,216 chips
NVIDIA B200 192 GB 1,125 4,500 per node

Specs sourced from Google Cloud official announcement (Ironwood, 2025) and NVIDIA Blackwell datasheet.

That’s it for now. More posts will follow as I write them.