[Model][SM70] Adapt Qwen3.8 Flash Next NVFP4 - #345

Merged

yangzhuxinyzx merged 31 commits into

mainfrom

codex/v100-qwen38-flash-next-nvfp4-20260826-140311

Aug 27, 2026

Commits on Aug 26, 2026

[Model] Start Qwen3.8 Flash Next SM70 adaptation
yangzhuxinyzx
committed
a3cc771
[Model] Add Qwen3.8 Flash Next SM70 V2 route
yangzhuxinyzx
committed
a90419b
[Doc] Record verified ModelScope snapshot
yangzhuxinyzx
committed
6b23eea
[Bugfix] Attach Qwen4Exp expert mapping
yangzhuxinyzx
committed
3fc247d
[Core] Report checkpoint name on weight-load failure
yangzhuxinyzx
committed
09f68f8
[Bugfix] Preserve QKV shard metadata in auto loading
yangzhuxinyzx
committed
86b1912
[Model] Pass 1Cat GDN output buffer in Qwen4Exp
yangzhuxinyzx
committed
78225e6
[Model] Honor custom V2 model state providers
yangzhuxinyzx
committed
d174f3f
[Bugfix] Zero heterogeneous hybrid KV pages
yangzhuxinyzx
committed
7fb6068
[Bugfix] Fall back for unsupported custom AR dtypes
yangzhuxinyzx
committed
f73cedb
[Core] Invoke custom KV cache bind hooks
yangzhuxinyzx
committed
d64150f
[Bugfix] Match QSA to local Flash KV layout
yangzhuxinyzx
committed
b37cd74
[Bugfix] Keep PLE scales in model dtype
yangzhuxinyzx
committed
7349fe2
[Kernel] Decode offloaded PLE FP8 rows on SM70
yangzhuxinyzx
committed
c228412
[Bugfix] Preserve dynamic PLE request shapes
yangzhuxinyzx
committed
de10e2e

Commits on Aug 27, 2026