Skip to content

quantization: complete MXFP4 and NVFP4#25878

Open
misfit-alpaca wants to merge 1 commit into
ggml-org:masterfrom
misfit-alpaca:nvfp4-mxfp4-wire
Open

quantization: complete MXFP4 and NVFP4#25878
misfit-alpaca wants to merge 1 commit into
ggml-org:masterfrom
misfit-alpaca:nvfp4-mxfp4-wire

Conversation

@misfit-alpaca

@misfit-alpaca misfit-alpaca commented Jul 19, 2026

Copy link
Copy Markdown

This PR/commit implements dense MXFP4 support (all tensors), and MoE NVFP4 (similar to MXFP4). It also implements the quantization routine for these types:

  • MXFP4
  • MXFP4_MOE
  • NVFP4
  • NVFP4_MOE

Overview

See above.

More specifically, it also adds NVFP4_MOE and MXFP4 quantization types which are new. The engine could already process these but with warnings, and this also is fixed.

Additional information

I found NVFP4 on my hardware to be decently fast. I'm using an i3 14100 with OpenBLAS on musl libc, with 8 GB DDR5.

Unlike other implementations of finished MXFP4 and NVFP4 support, this gives the best of all the worlds. People who want a dense MXFP4 get it, people who want MoE NVFP4 get it, and now people can have NVFP4 too. This PR kills multiple birds with one stone.

Requirements

This PR/commit implements dense MXFP4 support (all tensors),
and MoE NVFP4 (similar to MXFP4). It implements the quantization
routine for all of these types.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant