Added preliminary support for Qwen3.8-flash-next - #63

Merged

marcobambini merged 32 commits into

sqliteai:mainfrom

Skibisky:feature/qwen38-flash-next

Sep 18, 2026

Commits on Sep 4, 2026

convert: read Qwen3.8-Flash-Next into the existing container format
Skibisky
committed
c24e926
tokenizer: state the digit run, and accept Qwen's pattern
Skibisky
committed
6da8a47
qwen: the five architecture kernels, each against its own reference
Skibisky
committed
0b3a83e
model: the Qwen3.8-Flash-Next CPU forward pass
Skibisky
committed
1454a39
tests: a Qwen fixture, a container-native oracle, and the suite entries
Skibisky
committed
dd9ad87
docs: what Qwen3.8-Flash-Next support is, and what it measured
Skibisky
committed
c43170d
tests: two suite rules that a Qwen container walked into
Skibisky
committed
c8040c6

Commits on Sep 13, 2026

qwen: refuse a PLE conv kernel the ring cannot hold
mfethe1
committed
434f17a
Merge pull request #1 from mfethe1/pr63-ple-conv-fix
Skibisky
authored
0d1366f
profile: Qwen's phases, and what they said about its cache

Skibisky
and
claude
committed
6564cff
model: a Qwen layer's experts in one dispatch, once the cache holds them

Skibisky
and
claude
committed
d5a59f0
model: Qwen's trunk through i8mm, and the harness that decided it

Skibisky
and
claude
committed
7da84df
gitignore: the kernel_kl binary, beside sweep

Skibisky
and
claude
committed
0b92529
model: keep the pool awake through HyperConnection

Skibisky
and
claude
committed
fcc78e8
model: GDN's recurrence, one value head per task

Skibisky
and
claude
committed
3dc8570
model: QSA's attention, one query head per task

Skibisky
and
claude
committed
4f9e316
model: QSA's RoPE rows once, its blocks at once, its top-k as a sort

Skibisky
and
claude
committed
760b5dd

Commits on Sep 14, 2026

model: Qwen's routed experts staged by residency, not all-or-nothing

Skibisky
and
claude
committed
680219b
model: Qwen's router lookahead, from the next layer's normalized streams

Skibisky
and
claude
committed
7175bea
model: Qwen's routed experts in equal row ranges, three stages a layer

Skibisky
and
claude
committed
e46a74a

Commits on Sep 15, 2026

model: one quantization and one dispatch per vector Qwen projects

Skibisky
and
claude
committed
48867c3
model: fewer places a Qwen step lets the pool fall asleep

Skibisky
and
claude
committed
ceea3fa

Commits on Sep 16, 2026

Commits on Sep 17, 2026

qwen: reject unsafe container dimensions
Skibisky
committed
01e593c

Commits on Sep 18, 2026

qwen: normalize info name and document manifest fields
Skibisky
committed
68c4a40
Merge remote-tracking branch 'origin/main' into pr63-fixes
marcobambini
committed
ba1c3c0
qwen: size the shared expert's buffers from its width, not a bound on it

marcobambini
and
claude
committed
3b1c4f6
qwen: point the LEARNED references at the sections they now are

marcobambini
and
claude
committed
acc98c2
qwen: QSA's scalar score loop rounds the way the four-wide one does

marcobambini
and
claude
committed
84369e3
qwen: two Windows failures CI found on its first run of this branch

marcobambini
and
claude
committed
85d29e4
qwen: one multiply-add for QSA, fused wherever the target can fuse

marcobambini
and
claude
committed
46721e9