[AMD]: fix AITER flags for vllm v0.14.0 docker image

[Rohan138]

#496 updated the MI300/MI325 docker to the recent upstream v0.14.0 release, without updating the corresponding environment variables for AITER. The runs in #496 are using the default vLLM Triton attention kernel instead of the recommended AITER unified attention backend for gpt-oss.

Lower curves on both MI300 and MI325 are on 01/21, as opposed to previous perf.

[functionstackx]

thanks @Rohan138

can u add this to https://github.com/InferenceMAX/InferenceMAX/blob/main/perf-changelog.yaml and once ur PR is ready enable the sweep-enabled label to start the sweep?

additionally can u document the recommended flags in https://github.com/vllm-project/recipes/tree/main

Klaud reasonable reads the vllm-project/recipes repo & the release notes to determine what flags to change. If it isn't well documented then, just like the "average" human ROCm user, Klaud unfortunately won't change it.

[Rohan138]

Thanks @functionstackx, can you add the sweep-enabled label/give me perms? We'll document the updates in vllm-recipes as well.

[functionstackx]

hmm @cquil11

why did the sweeper bug out?

  File "/home/runner/work/InferenceMAX/InferenceMAX/utils/process_changelog.py", line 141, in <module>
    main()
  File "/home/runner/work/InferenceMAX/InferenceMAX/utils/process_changelog.py", line 107, in main
    result = subprocess.run(
             ^^^^^^^^^^^^^^^
  File "/usr/lib/python3.12/subprocess.py", line 571, in run
    raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['python3', 'utils/matrix_logic/generate_sweep_configs.py', 'test-config', '--config-keys', 'gptoss-fp4-mi325x-vllm', 'gptoss-fp4-mi300x-vllm', '--config-files', '.github/configs/amd-master.yaml', '.github/configs/nvidia-master.yaml', '--run-evals']' returned non-zero exit status 1.
Error: Process completed with exit code 1.

[functionstackx]

@Rohan138 since this is ur ROCm fork, i cant rebase myself. can u rebase again to include this commit? InferenceMAX/InferenceMAX@38546bb

[functionstackx]

@cquil11 the github action still doesnt work on forks? should we just manually import this to run the github action?

[cquil11]

/sweep test-config --config-keys gptoss-fp4-mi300x-vllm gptoss-fp4-mi325x-vllm --config-files .github/configs/amd-master.yaml --runner-config .github/configs/runners.yaml

[github-actions]

@cquil11 Kicking off a sweep.

Run: https://github.com/InferenceMAX/InferenceMAX/actions/runs/21367875017
Command: test-config --config-keys gptoss-fp4-mi300x-vllm gptoss-fp4-mi325x-vllm --config-files .github/configs/amd-master.yaml --runner-config .github/configs/runners.yaml
Pinned ref: fb14b96
Approval: not required (trusted collaborator).

[cquil11]

@functionstackx @Rohan138 the above command accomplishes the same sweep test

[functionstackx]

validated pass lgtm https://github.com/InferenceMAX/InferenceMAX/actions/runs/21367875017

@Rohan138 please update vllm-projects/recipes when u get the chance! thanks!