Can dflash support the Vulkan backend?

[zc317414]

Can dflash support the Vulkan backend? @cryptopoly

[cryptopoly]

Hey thanks for your message.

It's not supported today and unlikely soon unfortunately.

Two upstream forks:

bstnxbt/dflash-mlx (we ship this) — kernels written against MLX = Apple Silicon Metal only.
z-lab/dflash (CUDA reference) — kernels written against CUDA = NVIDIA only.

Vulkan would need either:

Upstream Vulkan port — neither team has signalled this. dflash kernels (recurrent rollback cache + draft KV cache + branchless masks) are tightly coupled to backend-specific primitives (Metal MMA / CUDA tensor cores). Porting = months of work.
Our own port — would require rewriting the kernels in GLSL/SPIR-V + a Vulkan compute pipeline. Massive effort, no upstream maintenance.
llama.cpp Vulkan path — llama.cpp's Vulkan backend handles inference, but dflash isn't a llama.cpp feature. dflash is a separate Python+Metal/CUDA stack.

Practical answer: AMD ROCm / Intel Arc / older NVIDIA users get llama.cpp Vulkan inference instead of dflash. dflash speedup (2-4× on Apple Silicon Metal MMA / Hopper tensor cores) doesn't transfer.

The closest path is upstream llama.cpp's tree-spec decode + Vulkan backend together — works, but no dflash-specific gains.