feat: prepare native text requests and Q8 GPU projection - #167

Merged

forkwright merged 17 commits into

mainfrom

feat/native-workload-boundaries

Sep 8, 2026

Commits on Sep 8, 2026

feat(text): prepare native generation requests
forkwright
committed
47ba668
docs: record prepared native text ownership
forkwright
committed
ad9e7b1
fix(text): release selected logits before the next decoder step
forkwright
committed
51a4333
fix(text): remove obsolete tokenizer test setup
forkwright
committed
7ab1ed5
docs: define Q8 projection ownership and qualification limits
forkwright
committed
9817e6f
docs: clarify kernel-local launcher ownership
forkwright
committed
c96c50f
Add checked Q8_0 GEMV kernel baseline
forkwright
committed
5a9a8cc
docs(kernels): state raw projection caller obligations precisely
forkwright
committed
464bfb6
refactor(kernels): compile raw Q8 ABI only in the HIP build
forkwright
committed
2b71ec7
fix(kernels): retain contextual Q8 format error propagation
forkwright
committed
7be9fb9
fix(kernels): return validated layout guard result
forkwright
committed
e395507
test(text): falsify prepared request boundaries
forkwright
committed
b4417ce
fix(kernels): bind refusal tests to the quant error owner
forkwright
committed
b8b158f
test(text): expect checked embedding row refusal
forkwright
committed
2ee68f2
test(kernels): derive the allocation boundary without panic helpers
forkwright
committed
e0d8205
fix(kernels): close Q8 representation and numerical-domain contracts
forkwright
committed
f6369ca
refactor(text): elide unused prepared implementation lifetime names
forkwright
committed
e260576