Skip to content

Why "reshard" becomes heavily slower for quantized model? #40

Description

@LZC-BELIEVER

Greetings!
I have successfully run fp8 quant on Qwen-7B model and int8 quant on Qwen-0.5B model in docker "hiyouga/verl:ngc-th2.6.0-cu126-vllm0.8.4-flashinfer0.2.2-cxx11abi0", but I found a problem:
the reshard time of both is heavily higher than base bf16 model.

Image Image

I found your example log of FP8 Rollout for Qwen2.5-0.5B-Instruct on DAPO has the same appearance:

Image

Is that normal? Does quant operation give rise to that a heavy time increase during resharding? Or could you explain that in some other way?

Looking forward to your reply!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions