byteark/Deepseek-V3-Lowrank80p

Deepseek-V3-Lowrank80p

---

This repository provides the low-rank version of Deepseek-V3, the route expert weight are recovered using low-rank approximation (reduce 20% weights).

	Average Score of MMLU (%)	Average Score of GSM8K (%)
deepseek/DeepSeek-V3	87.7	94.1
Deepseek-V3-Lowrank80p	86.7	94.5

Reference Implementations

• gh-efforts/DeepSeek-V3
• gh-efforts/sglang:
  • sample command:

  DEEPSEEK_RANK=1280 DEEPSEEK_SCALE_RANK=10 python3 -m sglang.launch_server --model-path /data1/asvd_dskv3_packed_63_backup --host 0.0.0.0 --port 40000 --tp-size 8 --enable-ep-moe --trust-remote-code --mem-fraction-static 0.9 --disable-cuda-graph