DeepSeek_v3 support#1735

Merged

regisss merged 21 commits into

huggingface:mainfrom

srajabos:DeepSeek_v3

Feb 21, 2025

Commits on Feb 18, 2025

Resolve rebase conflicts on DeepSeek_v3

srajabos
authored and
skavulya
committed
Update __init__.py

srajabos
authored and
skavulya
committed
Update __init__.py

srajabos
authored and
skavulya
committed
Update modeling_deepseek_v3.py

srajabos
authored and
skavulya
committed
Support optimized KV cache, static MOE and expert parallelism
skavulya
committed
Commented out attention_mask assertion for the mmlu tests

pallavijaini0525
authored and
skavulya
committed
Support optimized fusedDSPA, RoPE and RMS

srajabos
authored and
skavulya
committed
Commented out attention_mask assertionn

pallavijaini0525
authored and
skavulya
committed
Change references to deepseekv2 to deepseekv3
skavulya
committed
Override load_state_dict to support deepseek-R1
skavulya
committed
Added dynamic MoE changes

srajabos
authored and
skavulya
committed
Fix multicard expert parallelism for deepseekv3
skavulya
committed
Delete duplicate tests accidentally copied
skavulya
committed
Refactor deepseek_v3 and add clarifying comments
skavulya
committed
Fix edge case in expert slices in DeepSeek-V3
skavulya
committed
Add deepseekv3 to list of models supporting reuse_cache
skavulya
committed
Style fix in modeling_utils
skavulya
committed

Commits on Feb 19, 2025

Updated the README.md for the deepseek-r1-bf16
pallavijaini0525
committed

Commits on Feb 20, 2025