Ornith-1.5-122B-DFlash2

Blog | GitHub

This repository contains the DFlash 2 draft model for ornith-ai/Ornith-1.5-122B. It is not a standalone language model: it runs inside a speculative decoding server and drafts tokens for the target model to verify.

DFlash 2 is a block-diffusion drafter for speculative decoding. It predicts a whole block of tokens in a single pass and keeps the top candidates at every position. A lightweight selector then traces one coherent path through them. Two-tap dynamic convolutions in the backbone keep the draft from decaying toward the end of the block. Decoding is lossless: greedy output matches the target model exactly, and sampling preserves its distribution.

Quick Start

Serve with SGLang 0.5.21:

pip install "sglang[all]==0.5.21"

sglang serve \
  --model-path ornith-ai/Ornith-1.5-122B \
  --tp 2 \
  --enable-flashinfer-allreduce-fusion \
  --attention-backend trtllm_mha \
  --mamba-radix-cache-strategy extra_buffer \
  --mem-fraction-static 0.8 \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path incoai/Ornith-1.5-122B-DFlash2 \
  --speculative-num-draft-tokens 16

See the blog post for more details.

GGUF targets

The drafter reads the target's embedding of its mask token (248077). Ornith-1.5-122B stores that row at a scale of about 1e-6, below what the fp16 block scales of GGUF Q8_0 and K-quants hold, so GGUF builds of the target store it as zeros and draft acceptance roughly halves. mask_embedding.safetensors is an exact BF16 copy of the target's row, which an engine can read in its place; the drafter's weights are unchanged.

Evaluation

• Runtime: SGLang 0.5.21 on two NVIDIA GB300 GPUs (TP2), with TensorRT-LLM for target attention and FlashAttention 4 for DFlash 2 draft attention
• Speculation block size: 16 (15 draft tokens per verification step)
• Sampling: Ornith's default parameters (temperature 0.6, top-p 0.95, top-k 20), with thinking enabled
• Maximum new tokens: 4096
• Samples: 128 per task
• Prompts: benchmark formatting from z-lab/dflash; MT-Bench uses the first turn only

We compare autoregressive decoding and DFlash 2, which proposes fifteen draft tokens per verification step.

Acceptance Length

Acceptance length is the per-request mean of completion tokens divided by verification steps at concurrency 1. Higher is better.

Throughput

Throughput is total output tokens divided by end-to-end wall time. Each cell shows output tok/s (speedup vs. autoregressive).

Concurrency 1

Concurrency 8

Concurrency 32

Citation

If you find DFlash 2 useful, please cite:

@misc{inco2026dflash2,
  title  = {{DFlash 2: Keep Drafting Parallel}},
  author = {{Inco AI}},
  year   = {2026},
  month  = {August},
  url    = {https://inco.ai/blog/dflash2/}
}

Please also cite the original DFlash paper:

@inproceedings{chen2026dflash,
  title     = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author    = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026}
}