Ripple-65B-Preview-DFlash2 GGUF

GGUF conversions of ricursiveAI/Ripple-65B-Preview-DFlash2, the DFlash2 drafter for Ripple-65B-Preview. A drafter proposes several tokens per step, and the target model checks them, so decoding runs faster with the same output.

File Type Size
Ripple-65B-Preview-DFlash2-Q8_0.gguf Q8_0 4.30 GB
Ripple-65B-Preview-DFlash2-Q4_K_M.gguf Q4_K_M 2.39 GB
Ripple-65B-Preview-DFlash2-Q4_0.gguf Q4_0 2.28 GB

Use it with gmlx

gmlx run Ripple-65B-Preview-Q4_K_M.gguf \
  --draft-gguf Ripple-65B-Preview-DFlash2-Q8_0.gguf

Any GGUF of Ripple-65B-Preview works as the target. The drafter reads the target's token embeddings and output head, so it adds only its own layers to memory.

llama.cpp

llama.cpp loads these files as dflash drafters. It does not yet match the reference recipe in two ways:

  • It applies RoPE to draft layer 4, which the recipe runs without RoPE.
  • It lets the full-attention layer see the whole context. The recipe limits it to the last 16384 tokens.

The files carry both settings in dflash.attention.rope_pattern and dflash.attention.context_window, which gmlx reads.

How these files were made

uv run scripts/convert_dflash2.py \
  --checkpoint Ripple-65B-Preview-DFlash2 \
  --target-gguf Ripple-65B-Preview-Q4_K_M.gguf \
  --outtype q8_0 --out Ripple-65B-Preview-DFlash2-Q8_0.gguf

The script is in the gmlx repository. Source revision: 35609b3be3dd08a822b3ba1257c1ea5dd6f87632.

Downloads last month
1,264
GGUF
Model size
4B params
Architecture
dflash
Hardware compatibility
Log In to add your hardware

4-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for ashbash/Ripple-65B-Preview-DFlash2-GGUF

Quantized
(5)
this model