feat(phi4): run Phi-4-mini-instruct Q8_0 GGUF through ryzenai-corelib - #748

Open

chizamd wants to merge 17 commits into

chizamd:feature/phi4-rai

Commits on Sep 22, 2026

feat(phi4): run Phi-4-mini-instruct Q8_0 GGUF through ryzenai-corelib

committed
1313ef8

Commits on Sep 23, 2026

feat(rai): build the ryzenai-corelib backend on Linux

alfxu-amd
and
claude
committed
47cdd53

Commits on Sep 24, 2026

build(rai): make home_install.sh produce a working corelib install

alfxu-amd
and
claude
committed
14e4aa3
fix(rai): pin XILINX_XRT instead of letting XRT guess it

alfxu-amd
and
claude
committed
851dfe7
fix(rai): give the engines a real device, and say why when there is none

alfxu-amd
and
claude
committed
1efbdd2
feat(rai): decide what an install can run from two independent keys

alfxu-amd
and
claude
committed
40aa9fe
refactor(models): put backend and platform in every model path

alfxu-amd
and
claude
committed
bebaf79

Commits on Sep 25, 2026

test(phi4): expect the backend error 40aa9fe4 now raises

chizamd
and
cursoragent
committed
11ac533
fix(automodel): keep prefill logits owned until they are sampled

chizamd
and
cursoragent
committed
41a057a
fix(rai): drop the bring-up DEBUG prints 2bf92e8f meant to remove

chizamd
and
cursoragent
committed
19f1f19
fix(rai): bind the rotary tables as host views

alfxu-amd
and
claude
committed
cb8a19d
fix(phi4): say why inference failed instead of swallowing the reason

alfxu-amd
and
claude
committed
0fed8a9
fix(rai): do not gate a corelib build on the stock NPU stack checks

alfxu-amd
and
claude
committed
2bf92e8
docs(rai): add a Linux setup guide, and redact a stack identifier

alfxu-amd
and
claude
committed
8d8bb27
test(rai): follow the rotary tables onto host views

chizamd
and
cursoragent
committed
ecebd2e
test(phi4): make the every-tensor-role GGUF rejection test opt-in

chizamd
and
cursoragent
committed
f217211
feat: ensure main branch compatibility
alfxu-amd
committed
7ad2cdc