[AMD] CI - Detect the aiter version and rebuild if needed

[yctseng0211]

Motivation

https://github.com/sgl-project/sglang/actions/runs/20021784695/job/57446087825
Our AMD CI workflow always pulls the latest pre-built daily Docker image and then installs dependencies based on the PR code. However, this installation step does not rebuild or update AITER, which is already baked into the pre-built image.

As a result, when a PR upgrades the AITER version, the CI does not actually test against the new AITER, but instead tests against whatever version was pre-built into the daily image. This leads to a blind spot where AITER-related regressions cannot be caught at PR time.

Example:
In the PR #14497 , we upgraded AITER to v0.1.7.post5.
In our later manual tests we found that this version introduces an accuracy regression in:
accuracy-test-2-gpu-amd (linux-mi325-gpu-2)

However, the CI for this PR incorrectly passed(see in https://github.com/sgl-project/sglang/actions/runs/20021784695/job/57446087825), because it was still using the older AITER version bundled in the daily image, not the upgraded version specified in the PR.

Conclusion:
We need CI to verify whether the AITER version in the PR matches the one inside the pulled image. If they differ, CI must rebuild/install the correct AITER version inside the container before running tests.

cc: @bingxche @michael-amd @saienduri

Modifications

Accuracy Tests

Simulation - aiter version upgraded to v0.1.8

https://github.com/sgl-project/sglang/actions/runs/20368597241/job/58529505129

Simulation - If aiter version unchanged

https://github.com/sgl-project/sglang/actions/runs/20388611486/job/58594282444?pr=15460

Benchmarking 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.
• Work with maintainers to merge your PR. See the PR Merge Process

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[michael-amd]

LGTM