introduce JZ's ggml-hexagon.

[Jeff Zhou (zhouwg)]

Could technical experts from Qualcomm take a look at this project(ggml-hexagon/ggml-hexagon#18)? On Snapdragon 8 Elite (aka 8 Gen 4) smartphones, JZ's ggml-hexagon achieves higher PP and TG performance than Qualcomm’s official ggml-hexagon implementation.

build v0.99.3.9-dev, device: Snapdragon 8 Elite(aka 8Gen4), HTP arch v79, Vendor OnePlus 13, command: /data/local/tmp/llama-completion -ngl 99 -t 6 -n 256 --ctx-size 8192 --ubatch-size 64 --poll 1000 --no-warmup --no-mmap -fa on -st -no-cnv -m /sdcard/gemma-4-E2B-it-Q4_0.gguf -p "Hello, good morning, you are a powerful domain expert and know many things, now pls help to introduce the movie Once Upon a Time in America briefly, pls pay attention short then 1000 words\n").

PP&TG in JZ's ggml-hexagon:

PP&TG in Qualcomm's ggml-hexagon:

[Mengsheng Wu (mengshengwu)]

Thanks for putting this in front of us, and thanks to JZ for the work — the numbers are impressive.

That said, ggml-hexagon lives in the upstream ggml / llama.cpp tree, not in GenieX. The right place to have this discussion — and to compare approaches with Qualcomm's ggml-hexagon maintainer — is on the llama.cpp repo itself:

• https://github.com/ggml-org/llama.cpp

I'd suggest opening an issue or discussion there and tagging Max Krasnyansky (@max-krasnyansky), who owns the current ggml-hexagon backend. That's where any adoption / merge / benchmarking discussion needs to happen; we'll happily follow along downstream once the upstream picture is clearer.

[Jeff Zhou (zhouwg)]

Thanks for your reply, and sorry for interrupting the GenieX team.
My contributions were blocked within the upstream llama.cpp community on 07/14/2025 due to complex circumstances.
I understand Max serves as Senior Tech Director at Qualcomm. I have already reached out to him here:
https://github.com/zhouwg/ggml-hexagon/discussions/18#discussioncomment-17708015
However, I have not yet received a response.

[Zack Li (zhiyuan8)]

Jeff Zhou (@zhouwg) Thanks Jeff for reaching out, we closely work with Max on llama.cpp Hexagon support and have routine meetings to go over all technical planning and implementations. Will escalate and get back to you.

[Jeff Zhou (zhouwg)]

Thanks for your reply. I look forward to hearing good news from you.