Skip to content

Treat NVFP4 as a quantization suffix when pairing drafters - #657

Merged
Lazarus-931 merged 2 commits into
Blaizzy:mainfrom
Lazarus-931:fix/nvfp4-drafter-pairing
Oct 3, 2026
Merged

Lazarus-931 merged 2 commits into
Blaizzy:mainfrom
Lazarus-931:fix/nvfp4-drafter-pairing

Conversation

@Lazarus-931

@Lazarus-931 Lazarus-931 commented Oct 3, 2026 •

Copy link
Copy Markdown
Contributor

this adds nvfp4 to list of drafter dtype support such that for example Qwen3.8-27B-NVFP4 can work with Qwen3.8-27B-DSpark

DrafterModelCompatibility strips conversion and quantization suffixes
before matching a drafter to its target, but the list covers mxfp4,
mxfp8, fp8 and int4/int8 while missing nvfp4. The suffix survives into
the pairing key, so an NVFP4 target only matches a drafter whose repo
name carries the same suffix.

The composer's Speed picker is built from compatibleDrafters, so
speculative decoding reads as unavailable on an NVFP4 target even when a
usable drafter is installed. Across 4,483 repos from mlx-community,
unsloth, nvidia, lmstudio-community and RedHatAI, nvfp4 is the only
quantization suffix in common use that the list does not already strip
(163 repos; `<N>bit` is handled separately by isNonTargetComponent).
@Lazarus-931
Lazarus-931 merged commit e374784 into Blaizzy:main Oct 3, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant