Add DeepSeek-V4-Flash without MTP (yet) - #1797

Open

nh13 wants to merge 8 commits into

ml-explore:mainfrom

nh13:dsv4/01-core-model

Commits on Aug 29, 2026

feat(cache): resolve prompt-cache classes by module-qualified name
nh13
committed
2faf873
feat(convert): route model-declared sensitive modules to the high bit width
nh13
committed
53303cc
feat(models): add DeepSeek-V4-Flash
nh13
committed
438494c
test(models): add DeepSeek-V4-Flash tests
nh13
committed
b2ef0bb
chore(deps): require transformers>=5.8.0 for deepseek_v4
nh13
committed
70819eb
fix(deepseek_v4): snapshot cache offset so batched decode matches sequential
nh13
committed
89fc267
fix(deepseek_v4): do not trim CompressedKVCache once compressed state exists
nh13
committed
df41d38

Commits on Aug 30, 2026

fix(deepseek_v4): handle a per-batch offset in the compressed-pool mask
nh13
committed
ffa5959