Does BitNet-embedding-0.6B / BitNet-embedding-270M support GPU inference?

[yazdanifar]

The README's "Model Releases" section gives BitNet-b1.58-2B-4T an explicit
"GPU Support: Official GPU inference kernel available for accelerated
deployment" bullet, but the BitNet-embedding-0.6B and BitNet-embedding-270M
sections have no equivalent bullet — only CPU/prefill-speedup claims.

The gpu/README.md also states the W2A8 CUDA kernel "is tailored for use
with the BitNet-b1.58-2B-4T model," and the HF model cards for the two
embedding models only document bitnet.cpp/llama-embedding (CPU) usage.

Trying to run the embedding models' I2_S GGUF through a CUDA-enabled
llama-embedding build (even with -ngl 0) currently crashes with:

GGML_ASSERT(nb10 % sizeof(src1_t) == 0) failed

in ggml-cuda/binbcast.cu, because the CUDA backend still gets registered
and the scheduler routes some ops (e.g. certain flash-attention configs)
to CUDA0 regardless of -ngl.

Could you confirm:

1. Is GPU inference for the embedding models officially unsupported today?
2. Is there a plan to extend the gpu/ W2A8 kernel (or fix the llama.cpp
   CUDA binbcast bug) to support these embedding architectures
   (Qwen3-0.6B / Gemma3 backbones)?

Thanks!