Unverified50% confidenceFactExact time

Model parallelism consists of two approaches: tensor parallelism, which splits tensors within a single layer across multiple cards, and pipeline parallelism, which assigns different layers to different devices.

1

Sources

50%

Confidence

Long-term

Relevance

9/1/2026

First Seen

Sources

ML System Design: The Critical Leap from Model Theory to Production Practice

Related Entities

Model Parallelismconcept Tensor Parallelismconcept Pipeline Parallelismconcept

Related Claims

Unverified流水线并行按模型的不同层将计算分配到不同设备，通信量仅为相邻层的激活张量，但存在流水线气泡问题77% similar Unverified张量并行将单层权重切分到多卡，延迟低但通信开销大；流水线并行按层切分，吞吐高但首token延迟长66% similar UnverifiedChiplet技术将不同功能模块拆分为多个独立的小芯片，再通过先进封装技术高密度互连组合成系统级封装（SiP）61% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/840644

API

curl https://kongchang.com/api/v1/knowledge/claims/840644

MCP

get_claim(id=840644)