Skip to content
Navigation Menu
Sign in
Products
Platform
GitHub
Repository hosting, code review, and collaboration
Copilot
AI coding agent for tasks, projects, and automated work
Build & Ship
Actions
CI/CD and automation
Code Quality
Quality analysis, fixes, and enforcement
Codespaces
Cloud development environments
Secure
Code Security
Vulnerability detection and remediation
Secret Protection
Secret scanning and leak prevention
All Features
Developers
Learn
Documentation
Training
Certifications
Build
Build with Copilot
Customize Copilot
Integrations
Explore
Trending repos
GitHub Sponsors
Developer articles
Changelog
Changelog
Your source for the latest features, improvements, and fixes across GitHub products, all in one place.
Businesses
By Stage
Enterprises
Team
Startups
Nonprofits
By Industry
Healthcare
Financial services
Manufacturing
Government
See all industries
Evaluate GitHub
Why GitHub
Customer Stories
Business insights
Trust Center
Add-ons
GitHub Advanced Security
Enterprise-grade code security
Copilot for business
Enterprise-grade AI features
Premium Support
Enterprise-grade 24/7 support
Resources
Learn & Connect
Documentation
Resource center
Blog
Events & webinars
Community Forum
Customer Stories
Integrations
Support
Customer support
Training
GitHub Status
Company
About
Partners
Changelog
GitHub Universe
GitHub Universe
The event for developers, by developers — talks, workshops, and what’s next for GitHub.
Pricing
Search
/
Sign in
Sign up
Appearance settings
You signed in with another tab or window.
Reload
to refresh your session.
You signed out in another tab or window.
Reload
to refresh your session.
You switched accounts on another tab or window.
Reload
to refresh your session.
Dismiss alert
{{ message }}
deepspeedai
DeepSpeed
Repository navigation
Code
Issues
1.2k
(1.2k)
Pull requests
375
(375)
Discussions
Actions
Projects
Security and quality
Insights
More
items
SuperOffload - PCIe Optimization
- #9021
#9021
Draft
xylian86
wants to merge 13 commits into
deepspeedai:master
deepspeedai/DeepSpeed:master
from
xylian86:xinyu/superoffload-pcie
xylian86/DeepSpeed:xinyu/superoffload-pcie
Copy head branch name to clipboard
Conversation
Commits
13
(13)
Checks
Files changed
Draft
SuperOffload - PCIe Optimization
#9021
xylian86
wants to merge 13 commits into
deepspeedai:master
deepspeedai/DeepSpeed:master
from
xylian86:xinyu/superoffload-pcie
xylian86/DeepSpeed:xinyu/superoffload-pcie
Copy head branch name to clipboard
Commits
Commits on Nov 1, 2026
SuperOffload: checkpoint the CPU Adam state, fix unused params and rollbacks, speculate with clipping
Show description for 3846f50
xylian86
committed
3846f50
View commit details
Copy full SHA for 3846f50
Browse repository at this point
SuperOffload: follow hyperparameter changes and loss scale increases, commit decoupled saves
Show description for 983cba3
xylian86
committed
983cba3
View commit details
Copy full SHA for 983cba3
Browse repository at this point
ZeRO-3 partial offload: follow hyperparameter schedules on the device sub-groups and checkpoint their Adam state
Show description for afded17
xylian86
committed
afded17
View commit details
Copy full SHA for afded17
Browse repository at this point
ZeRO-3 partial offload: load the device sub-groups' moments in the param dtype
Show description for d53ffb6
xylian86
committed
d53ffb6
View commit details
Copy full SHA for d53ffb6
Browse repository at this point
SuperOffload: do not step during backward when a beta is 0
Show description for 9867efa
xylian86
committed
9867efa
View commit details
Copy full SHA for 9867efa
Browse repository at this point
Merge branch 'master' into xinyu/superoffload-fixes
xylian86
authored
80a21ae
View commit details
Copy full SHA for 80a21ae
Browse repository at this point
Commits on Nov 5, 2026
Merge deepspeedai/DeepSpeed#8972 (xinyu/superoffload-fixes at 80a21ae)
xylian86
committed
0eae034
View commit details
Copy full SHA for 0eae034
Browse repository at this point
Merge deepspeedai/DeepSpeed#8973 (xinyu/zero3-partial-offload-optimizer at d53ffb6)
xylian86
committed
febec28
View commit details
Copy full SHA for febec28
Browse repository at this point
CPU Adam: release the GIL during steps, step fp32 params on bf16 gradients
Show description for 68d1086
xylian86
committed
68d1086
View commit details
Copy full SHA for 68d1086
Browse repository at this point
SuperOffload: run CPU Adam in a thread of the training process, copy gradients to the CPU once
Show description for f90d9ea
xylian86
committed
f90d9ea
View commit details
Copy full SHA for f90d9ea
Browse repository at this point
Commits on Nov 6, 2026
SuperOffload: copy large gradients to the CPU in pieces on two streams
Show description for 2ce5bcc
xylian86
committed
2ce5bcc
View commit details
Copy full SHA for 2ce5bcc
Browse repository at this point
Commits on Nov 8, 2026
CPU Adam: scale gradients as the step reads them, change a step to another gradient scale in one pass
Show description for ad833ac
xylian86
committed
ad833ac
View commit details
Copy full SHA for ad833ac
Browse repository at this point
SuperOffload: change the CPU steps of a clipped step in one pass instead of undoing and redoing them
Show description for b2b46ae
xylian86
committed
b2b46ae
View commit details
Copy full SHA for b2b46ae
Browse repository at this point
You can’t perform that action at this time.
