All issues

Search results

About reproducing the results in the blog

Status: Open.

#166 In z-lab/dflash;

· 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

Status: Open.

#165 In z-lab/dflash;

· cgoudie opened on Aug 29, 2026

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

Status: Open.

#163 In z-lab/dflash;

· ThomasQvist opened on Aug 27, 2026

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

Status: Open.

#162 In z-lab/dflash;

· ilperev opened on Aug 20, 2026

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

Status: Open.

#161 In z-lab/dflash;

· nolanwang1996-cmd opened on Aug 20, 2026

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

Status: Open.

#160 In z-lab/dflash;

· SakuraToErii opened on Aug 19, 2026

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

Status: Open.

#159 In z-lab/dflash;

· fgomsan opened on Aug 19, 2026

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

Status: Open.

#156 In z-lab/dflash;

· aqilmarwan opened on Aug 17, 2026

Question about block size choice in DFlash

Status: Open.

#154 In z-lab/dflash;

· OliverZhou131 opened on Aug 13, 2026

Guidance on target hidden-layer selection

Status: Open.

#153 In z-lab/dflash;

· 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

Status: Open.

#152 In z-lab/dflash;

· olliehm opened on Aug 2, 2026

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

Status: Open.

#151 In z-lab/dflash;

· PhilipJohnBasile opened on Aug 1, 2026