Skip to content

deepseekv3 开启 tp + ep 的混合模式的时候,精度会下降。 #1223

Description

@hiworldwzj

LOADWORKER=14 python -m lightllm.server.api_server --model_dir /mtc/sufubao/DeepSeek-V3.2 --tp 8 --graph_max_batch_size 32 --tool_call_parser deepseekv32 --mem_fraction 0.8 --reasoning_parser deepseek-v3 --dp 4 --enable_ep_moe

HF_ALLOW_CODE_EVAL=1 HF_DATASETS_OFFLINE=0 lm_eval --model local-completions --model_args '{"model":"deepseek-ai/DeepSeek-V3.2", "base_url":"http://localhost:8000/v1/completions", "max_length": 16384}' --tasks gsm8k --batch_size 500 --confirm_run_unsafe_code

上面是评测的命令,也做了tp 8 然后 dp 2, dp 1的消融分析,也是一样的现象,可能是ep 模式下,qkv 部分的支持切分可能存在精度问题。只有tp 8, dp 8 的时候,精度是正常的。

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions