Motivation

Found this issue in a MUSA container where MATE is not installed. Although the validation in ServerArgs._adjust_attention_backend() checks at the string level and will pick fa for Ring Attention (if attention backend is not set), when actually resolving the attention backend via get_attn_backend(), Torch SDPA could still be selected as it is the fallback solution. This leads to silent incorrect behavior or confusing errors downstream.

> sglang generate --model-path /home/dist/diffusion/yeahdongcn/Qwen/Qwen-Image \
    --sp-degree 2 --ulysses-degree 1 --ring-degree 2 --num-gpus 2 \
    --warmup --prompt "Doraemon is eating dorayaki"

Modifications

Validate attention backend for Ring Attention in USPAttention.

Accuracy Tests

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.