eplb draft - #38

Open

YzhouKwen wants to merge 12 commits into

DeepLink-org:zcx/eplb_layer_idxfrom

YzhouKwen:liudy/eplb_draft

Commits on May 8, 2025

test.py
YzhouKwen
committed
7651481
adjust: change model layer count
YzhouKwen
committed
37d234c
add _load_weight_experts_with_eplb in deepseek_v2.py
YzhouKwen
committed
afc7c5e
add enable_eplb in nn/moe.py
YzhouKwen
committed
da496cf
add layer_idx in nn/moe.py
YzhouKwen
committed
1cebe01
add rank_expert_list print in nn/moe.py
YzhouKwen
committed
d5dd6df
add expert_list to FusedMoENormal class and FusedMoELowLatency class in backends/cuda/moe.py
YzhouKwen
committed
a3e5531
add layer_idx to FusedMoENormal class and ep_expert_list()
YzhouKwen
committed
112e0aa
generate eplb.json
YzhouKwen
committed
983bf97
add expert_list() to FusedDeepEpMoEBlockedF8Impl class and run well for num_hidden_layer=4
YzhouKwen
committed
571e2e0

Commits on May 9, 2025

add print and set num_hidden_layers=8
YzhouKwen
committed
1ae2409

Commits on May 12, 2025

Commented out the code for modifying model layers; kept the original layer structure
YzhouKwen
committed
4491c89