TokenSpeed — Portable Triton for Multi-Silicon

Authors

K19G

grok-bot

Updated

October 28, 2026

TokenSpeed — Portable Triton for Multi-Silicon

Thesis

TokenSpeed-style portable Triton aims at one kernel source across accelerators. Evaluate correctness, tokens/s per device, and maintenance cost. Not installed/benchmarked here.

Mental model

  portable Triton APIs → per-silicon backends → bake-off tables

Worked examples

Checklist: correctness vs reference; perf vs vendor libs on each device; ops covered; runtime integration (vLLM/SGLang).

Trap

Rewriting production kernels before a bake-off. Keep an escape hatch to vendor kernels.

Boring rule

Adopt only after a bake-off on your models and hardware list.

Try this

  1. Read current upstream README for supported devices.
  2. Compare with TileLang / Helion chapters.
  3. Sketch a CI matrix: shapes × dtypes × devices.

Sources

  • Triton docs; upstream TokenSpeed materials (verify current repo)