更换图床
This commit is contained in:
@@ -6,7 +6,7 @@ date = "2026-06-07"
|
||||
description = ""
|
||||
categories = ["AI"]
|
||||
tags = ["SGLang", "GLM", "HiCache"]
|
||||
image = "https://assets.mitsea.cn/blog/posts/2026/06/GLM-5.1%20Coding%20%E5%9C%BA%E6%99%AF%E9%83%A8%E7%BD%B2%E4%BC%98%E5%8C%96%EF%BC%9A%20HiCache%20%E9%85%8D%E7%BD%AE%E6%8C%87%E5%8D%97%E4%B8%8E%E5%9C%BA%E6%99%AF%E5%8E%8B%E6%B5%8B/jimmy-liu-k2zuVhKjA4o-unsplash.avif"
|
||||
image = "https://assets.flinty.moe/blog/posts/2026/06/GLM-5.1%20Coding%20%E5%9C%BA%E6%99%AF%E9%83%A8%E7%BD%B2%E4%BC%98%E5%8C%96%EF%BC%9A%20HiCache%20%E9%85%8D%E7%BD%AE%E6%8C%87%E5%8D%97%E4%B8%8E%E5%9C%BA%E6%99%AF%E5%8E%8B%E6%B5%8B/jimmy-liu-k2zuVhKjA4o-unsplash.avif"
|
||||
+++
|
||||
|
||||
## 背景
|
||||
@@ -330,11 +330,11 @@ HiCache 路径确实在工作:约 290 万 token 落在 host cache,约 2320
|
||||
|
||||
**HiCache** — 输出平稳,几乎无 ≥1s 的 chunk 间隔:
|
||||
|
||||

|
||||

|
||||
|
||||
**MTP** — 平均吞吐更高,但 decode/revisit 请求 **100% 出现 ≥1s 的 chunk gap**,P95 达 12.84s,即「写着写着突然停十几秒」:
|
||||
|
||||

|
||||

|
||||
|
||||
### Mixed 多并发容量
|
||||
|
||||
@@ -345,9 +345,9 @@ HiCache 路径确实在工作:约 290 万 token 落在 host cache,约 2320
|
||||
|
||||
HiCache 在 mixed 场景下可稳定支撑 **4 路并发** coding agent;MTP 在并发 2 就已因 chunk gap 超标被判不可用。
|
||||
|
||||

|
||||

|
||||
|
||||

|
||||

|
||||
|
||||
分档细节(mixed 模式):
|
||||
|
||||
|
||||
Reference in New Issue
Block a user