Support dpsk-4.1 sft and rl - #201

Merged

tastelikefeet merged 23 commits into

modelscope:mainfrom

tastelikefeet:feat/deepseek-v41

Sep 20, 2026

Commits on Sep 15, 2026

feat(deepseek-v41): add DeepSeek-V4.1-Flash bridge
z0o0ey
committed
fe8e498
feat(deepseek-v41): add DSpark and Engram support
z0o0ey
committed
50987dd
Merge remote-tracking branch 'origin/main' into feat/deepseek-v41
z0o0ey
committed
267e465

Commits on Sep 16, 2026

wip
z0o0ey
committed
3a31a08
Merge remote-tracking branch 'origin/main' into feat/deepseek-v41
z0o0ey
committed
4416b5b
feat(deepseek-v41): support pipeline parallelism via HybridModel (B1 text backbone)
z0o0ey
committed
4076d63
feat(deepseek-v41): map DSpark draft stack (mtp.*) under HybridModel (B3)
z0o0ey
committed
e01303b
feat(deepseek-v41): host HybridModel backbone in multimodal wrapper (B4)
z0o0ey
committed
02ac651
feat(deepseek-v41): default to HybridModel path for all layouts (B5)
z0o0ey
committed
33c5b80
feat(deepseek-v41): support virtual pipeline parallelism (VPP)
z0o0ey
committed
012ce0f

Commits on Sep 17, 2026

Commits on Sep 18, 2026

fix(deepseek-v41): gate per-head query RMS norm to V4 only
z0o0ey
committed
a34fcc4
fix(deepseek-v41): build the engram in the bridge's TransformerLayer fork
z0o0ey
committed
313c87c
fix(deepseek-v41): key DSpark off block_size and freeze the draft stack
z0o0ey
committed
870e32a
test(deepseek-v41): cover packed and CP backward
z0o0ey
committed
d1a8d74
fix(deepseek-v41): harden missing weight export
z0o0ey
committed
506c622

Commits on Sep 19, 2026

Commits on Sep 20, 2026

lint
z0o0ey
committed
0d1a9fc
fix
z0o0ey
committed
1cc4b6a
fix(deepseek-v41): harden Megatron 0.16-0.19 compatibility
z0o0ey
committed
74eea16
lint
z0o0ey
committed
da4c176