Issues

• Request: DFlash / DFlash2 draft checkpoint for Qwen/Qwen3.8-Flash-Next

  z-lab/dflash#176 · Filter by author stuchapin909 opened on Oct 9, 2026
  1

• Qwen3-1.7B-DFlash2 (4.91 accept length on math_reasoning; 220 --> 738 tokens/sec; up to 3.35× speedup)

  z-lab/dflash#175 · Filter by author rasyosef opened on Oct 4, 2026

• Training code for Dflash-2

  z-lab/dflash#174 · Filter by author parjanya20 opened on Sep 30, 2026

• A DFlash2 speculative decoding draft model for gemma-4-E2B-it (167 --> 626 tokens/s; up to 3.74x speedup on Math)

  z-lab/dflash#173 · Filter by author rasyosef opened on Sep 26, 2026

• dflash_generate is not lossless on Gated-DeltaNet hybrid targets (Qwen3-Next / Qwen3.5): Cache.crop leaves rejected draft tokens in the recurrent state

  z-lab/dflash#172 · Filter by author 0xWhiteMage opened on Sep 23, 2026

• Advantage over DSpark not reproduced

  z-lab/dflash#171 · Filter by author team109 opened on Sep 21, 2026

• Qwen3.8-27B DFlash2: accepted length ~3 on llama.cpp vs ~5.5 on vLLM with the same drafter — expected from target pairing, or a selector setting?

  z-lab/dflash#170 · Filter by author AiwendilInTheWoods opened on Sep 18, 2026

• About reproducing the results in the blog

  z-lab/dflash#166 · Filter by author sglucas opened on Aug 30, 2026

• DFlash2 + vLLM prefix caching: hits drop to 0, full re-prefill each request — wall-clock loses to MTP+cache on multi-turn workloads

  z-lab/dflash#165 · Filter by author cgoudie opened on Aug 29, 2026
  1

• Multilingual (Danish) acceptance collapse in DFlash/DFlash2 drafters — request: multilingual training mix or training pipeline release

  z-lab/dflash#163 · Filter by author ThomasQvist opened on Aug 27, 2026

• How hard is it to add DFlash / DFlash 2 to transformers, and is it even worth it?

  z-lab/dflash#162 · Filter by author ilperev opened on Aug 20, 2026

• Request: DFlash draft checkpoint for ornith-ai/Ornith-1.5-35B-A3B

  z-lab/dflash#161 · Filter by author nolanwang1996-cmd opened on Aug 20, 2026

• [Apple M1 Max] DFlash2 w4a16 acceptance collapses to 6.25%; w4a32 restores 75%

  z-lab/dflash#160 · Filter by author SakuraToErii opened on Aug 19, 2026

• [MLX / DFlash2] Qwen3.8-27B greedy output diverges from pure AR with the official block size 5

  z-lab/dflash#159 · Filter by author fgomsan opened on Aug 19, 2026
  1

• bug: build_target_layer_ids never runs, and its guess would be wrong for 14 of 20 checkpoints

  z-lab/dflash#156 · Filter by author aqilmarwan opened on Aug 17, 2026

• Question about block size choice in DFlash

  z-lab/dflash#154 · Filter by author OliverZhou131 opened on Aug 13, 2026

• Guidance on target hidden-layer selection

  z-lab/dflash#153 · Filter by author sleepyshep opened on Aug 10, 2026

• Qwen3.5-122B-A10B drafter: 2–12% acceptance under llama.cpp with a quantized target, while Qwen3.6-35B-A3B holds 50–59% in an identical setup

  z-lab/dflash#152 · Filter by author olliehm opened on Aug 2, 2026
  1

• MLX: default block_size 16 is 24-29% below the throughput optimum on Apple Silicon

  z-lab/dflash#151 · Filter by author PhilipJohnBasile opened on Aug 1, 2026

• Data point: drafts transfer across a model lineage (98% acceptance cross-model)

  z-lab/dflash#149 · Filter by author gsrunion opened on Jul 24, 2026

• Request access to z-lab/MiniMax-M2.7-DFlash model

  z-lab/dflash#148 · Filter by author wangxj12 opened on Jul 13, 2026

• Question about drafter for diffusion models themselves

  z-lab/dflash#147 · Filter by author hsamrat007 opened on Jul 9, 2026

• [Bug] DFlash speculative decoding causes cudaErrorIllegalAddress under concurrent load when CUDA graphs are enabled

  z-lab/dflash#146 · Filter by author Love-learning-Li opened on Jul 9, 2026

• Docker image is not publicly available - in contrast to README.md

  z-lab/dflash#144 · Filter by author kyr0 opened on Jun 28, 2026

• KIND MODEL REQUEST: gemma4:12b-it-qat

  z-lab/dflash#143 · Filter by author ttx89-dev opened on Jun 25, 2026
  1