feat(qwen36): routed experts from a qpack container through bounded Metal slots - #1323

Open

Avicennasis wants to merge 14 commits into

JustVugg:devfrom

Avicennasis:feat/qwen36-qpack-engine

Commits on Sep 4, 2026

Add MLX affine quantization reference contract
Avicennasis
committed
c36a54f
Add packed affine Metal GEMV path
Avicennasis
committed
75f77d2
Add strict qpack expert reader
Avicennasis
committed
8f433db
Add qpack inspection utility
Avicennasis
committed
0f540ff
fix(qwen36): read tokenizer.json merges in string and pair spellings
Avicennasis
committed
6cdd552
fix(build): fall back to the Homebrew prefixes when brew is off the PATH
Avicennasis
committed
5b4a844
feat(affine): add a loader-side dequant oracle to the MLX affine contract
Avicennasis
committed
57db3b1
feat(qwen36): load MLX-affine dense weights from a qpack container
Avicennasis
committed
088e9a5
feat(qwen36): unshift MLX full-gamma norm weights to the zero-centered form
Avicennasis
committed
32c1525
tools(qwen36): snap-view generator to run qwen36 over a qpack container
Avicennasis
committed
05868e5
feat(metal): whole-slot buffers with offset-addressed affine dispatch
Avicennasis
committed
46e3c3b
feat(qwen36): routed experts from a qpack container through bounded slots
Avicennasis
committed
d4df8e9
test(qwen36): Apple parity gates for the qpack Metal path under eviction
Avicennasis
committed
d1d9d07
feat(qwen36): report qpack dispatch counters at the end of engine runs
Avicennasis
committed
121c663