Skip to content

[Bug] Mixtral & Grok2 Piecewise CUDA Graph Accuracy Drop #13469

@hebiao064

Description

@hebiao064

Checklist

  • I searched related issues but found no solution.
  • The bug persists in the latest version.
  • Issues without environment info and a minimal reproducible demo are hard to resolve and may receive no feedback.
  • If this is not a bug report but a general question, please start a discussion at https://github.com/sgl-project/sglang/discussions. Otherwise, it will be closed.
  • Please use English. Otherwise, it will be closed.

Describe the bug

Repro:

https://huggingface.co/xai-org/grok-2

When we enabled radix cache

GSM8K score will drop to 0.8 instead of 0.9

The second send_one will output garbage result

<pad><pad><pad><pad><pad><pad><pad><pad>
<pad><pad><pad><pad><pad><pad><pad><pad>
<pad><pad><pad><pad><pad><pad><pad><pad>
<pad><pad><pad><pad><pad><pad><pad><pad>

When we disable radix cache

The second send_one will still output garbage result

<pad><pad><pad><pad><pad><pad><pad><pad>
<pad><pad><pad><pad><pad><pad><pad><pad>
<pad><pad><pad><pad><pad><pad><pad><pad>
<pad><pad><pad><pad><pad><pad><pad><pad>

Reproduction

Download model from: https://huggingface.co/xai-org/grok-2

python3 -m sglang.launch_server --model /local/grok-2 --tokenizer-path /local/grok-2/tokenizer.tok.json --tp 8 --quantization fp8 --attention-backend triton --enable-piecewise-cuda-graph --disable-radix-cache
# First Run
python /sgl-workspace/sglang/python/sglang/test/send_one.py

# Second Run
python /sgl-workspace/sglang/python/sglang/test/send_one.py

Then send_one or run gsm8k

Environment

Latest main

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