---
title: '57088'
pageLayout: default
slug: thoughts/inference/vllm-prs/57088
permalink: https://aarnphm.xyz/thoughts/inference/vllm-prs/57088.md
generator:
  quartz: v4.6.0
  hostedProvider: Cloudflare
  baseUrl: aarnphm.xyz
full: https://aarnphm.xyz/llms-full.txt
---
# [#57088](https://github.com/aarnphm/aarnphm.github.io/issues/57088) \[Docs\] Add ecosystem kernels index skill

[vllm-project/vllm#57088](https://github.com/vllm-project/vllm/pull/57088) · benchislett · OPEN · created 2026-09-15T23:14:08Z · merged <no value>

## Purpose

This PR adds a new skill: `kernels-ecosystem`. It is a one-off snapshot of all kernels GPT-6 could find in vLLM, SGLang, TRTLLM, TokenSpeed, and FlashInfer. It is grouped by operator at the top-level for easy browsing. The skill includes a workflow to update the index, which we will need to do periodically. Ideally this index can be hosted outside of vLLM and serve a more frequently crawled index, but for now I think there is a lot of value in merging this skill.

Already, I have discovered a dozen new kernels that I hadn’t previously known about.

---

## comments

**faradawn** (2026-09-21T23:16:34Z):

Tagging a few folks from the agent-readiness survey. Please react 👍 or comment

@dllehr-amd - you described agent code not using the MoE oracle. Does this PR help?

@tlrmchlsmth — you wanted it easier to tell which backends run for a given`vllm serve` + hardware.

@MatthewBonanni — mentioned attention backends (FlashInfer `prims-ts`,
TRTLLM CuteDSL fused GQA).

Also @WoosukKwon, @AndreasKaratzas, @tjtanaa, @jinyouzhi, @penli9 @esmeetu @aoshen02

