Artifact v2: DFlash support with NVFP4 quantized drafter weights
#4
by cometkim - opened
I've updated this artifact to add DFlash2 feature.
Upstream's registered schema stores the module's matrices as W8G32_F16S. This image stores the 34 drafter matrices weight-only NVFP4 instead (norms and conv base kernels stay BF16), shrinking the module payload from 2,226,805,248 to 1,082,882,820 bytes (−1.07 GiB); Text weights are byte-identical between the two encodings.
To run this artifact with DFlash2, the feat/dflash2 patch from cometkim's fork and ID registration are required. (Details are also available in the model card)
The minimal patch set, including the model converter, is maintained in the feat/qwen3.8-nvfp4full branch. (Includes some Windows-specific patches.)
cometkim changed discussion title from Artifact v2: DFlash support with NVFP4 re-formatted drafter weights to Artifact v2: DFlash support with NVFP4 quantized drafter weights