DFlash 2 claims 16 to 25% more verified output while keeping speculative drafts parallel

I found the implementation PRs: vLLM has PR #52816, and llama.cpp has PR #27342.

I'm seeing about 27 tokens/s decoding on a DGX Spark with vLLM, Qwen 3.8 27B NVFP4, and DFlash 2.

The technology looks impressive, but the "an agent writes in an afternoon what a chatbot writes in a month" line is unnecessary. The performance numbers speak for themselves.

This has made low-memory-bandwidth model usage much nicer for me.

I watched the video closely, and DFlash 2's tool call appears to fail on Python syntax while the autoregressive side succeeds. That is normally a one-shot task for models in this class, so I want a more constrained test to rule out an odd regression.

DFlash is supposed to be lossless, so if the draft path really regresses versus the target model, that would be an implementation bug.