DeepSeek V4.1 Flash DSpark drafter (GGUF)

The speculative-decoding drafter for DeepSeek V4.1 Flash in the Lucebox engine. It is the draft block DeepSeek ships inside the checkpoint (the mtp.* tensors: three stages, 128 routed experts top-3, a block of five tokens), converted to GGUF. The target model drafts nothing itself; this file proposes tokens and the target verifies them in one pass, and every verified token is bit-identical to plain decoding.

Routed experts are repacked bit-exactly to MXFP4, dense projections are Q8_0. The Q2K-Q4K build requantizes the same drafter (routed experts Q2_K, dense projections Q4_K) for the 192 GB setup, where the drafter shares the R9700 with 12 GB of hot experts: it leaves room for larger prompt chunks and drafts faster, with about the same acceptance. Converted with server/scripts/convert_ds4_dspark_draft_to_gguf.py from the source checkpoint.

Run

With the target from Lucebox/DeepSeek-V4.1-Flash-ROCMFP23-GGUF on a Lucebox with 128 GB (AMD Radeon AI PRO R9700 + Ryzen AI MAX+ 395):

luce_server DeepSeek-V4.1-Flash-ROCMFP2S.gguf \
  --profile ds41-lucebox \
  --draft DeepSeek-V4.1-Flash-DSpark-draft-MXFP4-Q8.gguf

With 192 GB (Ryzen AI MAX+ PRO 495) every expert stays in memory, with the smaller drafter and an expert placement and router bias fitted to the box:

luce_server DeepSeek-V4.1-Flash-ROCMFP2S.gguf \
  --profile ds41-gorgon \
  --draft DeepSeek-V4.1-Flash-DSpark-draft-Q2K-Q4K.gguf \
  --ds4-expert-placement placement.json \
  --ds4-router-bias router_bias.bin

Lucebox Manage installs the files and the fitted layout, and picks the launch for the box's memory from its Engine page.

Measured

--profile ds41-gorgon with the Q2K-Q4K drafter, 192 GB, fresh server:

--profile ds41-lucebox at its 128K context, first request after start (DS41 guide):

The engine verifies three drafted tokens per step by default (LUCE_DS4_SPEC_Q, 2 to 6), the fastest on the Lucebox.

License

MIT, as the source checkpoint.