Is your feature request related to a problem? Please describe.
Currently, Omlx does not support [DFlash: Block Diffusion for Flash Speculative Decoding](https://arxiv.org/html/2602.06036v1) natively. This limits speculative decoding efficiency, resulting in lower acceptance lengths and throughput compared to what DFlash can achieve. For workloads requiring high performance inference, this becomes a bottleneck.
Describe the solution you'd like
Add native support for DFlash within the Omlx project. This would allow acceptance lengths to increase to ~6–8 tokens per token verified by target model and throughput improvements of up to 3.5×–6× compared to standard autoregressive decoding.
Describe alternatives you've considered
- Using standard autoregressive decoding (baseline, but limited to acceptance length of 1).
- Implementing speculative decoding with other draft models, though these approaches typically yield lower acceptance lengths and less throughput improvement compared to DFlash.
Additional context
- DFlash paper: https://arxiv.org/html/2602.06036v1
- Benchmarks show DFlash consistently achieves acceptance lengths between 6–8 tokens per block, leading to significant speedups in inference throughput.
Is your feature request related to a problem? Please describe.
Currently, Omlx does not support [DFlash: Block Diffusion for Flash Speculative Decoding](https://arxiv.org/html/2602.06036v1) natively. This limits speculative decoding efficiency, resulting in lower acceptance lengths and throughput compared to what DFlash can achieve. For workloads requiring high performance inference, this becomes a bottleneck.
Describe the solution you'd like
Add native support for DFlash within the Omlx project. This would allow acceptance lengths to increase to ~6–8 tokens per token verified by target model and throughput improvements of up to 3.5×–6× compared to standard autoregressive decoding.
Describe alternatives you've considered
Additional context