Skip to content

Navigation Menu

Appearance settings

Explore
By company size
By use case
By industry
View all solutions
Topics
- AI
- DevOps
- Security
- Software Development
- View all
Explore
- GitHub Sponsors
  Fund open source developers
- The ReadME Project
  GitHub community articles
Repositories
- Enterprise platform
  AI-powered developer platform
Available add-ons
Pricing

Search code, repositories, users, issues, pull requests...

Search

Clear

Search syntax tips

Provide feedback

We read every piece of feedback, and take your input very seriously.

Include my email address so I can be contacted

Saved searches

Use saved searches to filter your results more quickly

Name

Query

To see all available qualifiers, see our documentation.

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

ggml-org / llama.cpp Public

Notifications You must be signed in to change notification settings
Fork 12.2k
Star 82.3k

Code
Issues 301
Pull requests 479
Discussions
Actions
Projects 10
Wiki
Security
Insights

Additional navigation options

Code
Issues
Pull requests
Discussions
Actions
Projects
Wiki
Security
Insights

Better 1.5 bit quantization #5971

Merged

ikawrakow merged 15 commits intomasterggml-org/llama.cpp:masterfrom

ik/iq1s_blocks16ggml-org/llama.cpp:ik/iq1s_blocks16Copy head branch name to clipboard

Mar 11, 2024

Conversation Commits15 (15)Checks Files changed

Merged

Better 1.5 bit quantization #5971
ikawrakow merged 15 commits intomasterfrom
ik/iq1s_blocks16

Commits

Commits on Mar 11, 2024

Trying blocvks of 16 for IQ1_S - seems slightly better
Kawrakow
committed
iq1s_blocks16: Adjust scale fudge factor to 1.125
Kawrakow
committed
iq1s_blocks16: going to blocks of 32
Kawrakow
committed
iq1s_blocks16: Use 2*<x^2> as sigma2 in weight adjustment
Kawrakow
committed
iq1s_blocks16: scalar and AVX2 dot products
Kawrakow
committed
iq1s_blocks16: CUDA dot product
Kawrakow
committed
iq1s_blocks16: Metal works, Neon does not
Kawrakow
committed
iq1s_blocks16: fixed Neon
Kawrakow
committed
iq1s_blocks16: very slightly faster TG on Metal
Kawrakow
committed
iq1s_blocks16: speedup Metal by packing codebook into uint32_t's
Kawrakow
committed
Formatting
Kawrakow
committed
iq1s_blocks16: uint32_t codebook is also better in CUDA
Kawrakow
committed
iq1s_blocks16: slightly faster Neon dot product
Kawrakow
committed
iq1s_blocks16: faster AVX2 dot product
Kawrakow
committed
iq1s_blocks16: adjust to ggml-common.h
Kawrakow
committed

Footer

© 2025 GitHub, Inc.

Footer navigation

Terms
Privacy
Security
Status
Docs
Contact

You can’t perform that action at this time.