Fix Hardcoded localhost URLs

[MiguelCarpio]

What does this PR do?
Fix Hardcoded localhost URLs (IPv6/IPv4 Problem).

• TASK [gpu-validation : Wait for vllm API to appear]
• TASK [gpu-validation : Wait for vllm metrics endpoint to appear]
• TASK [gpu-validation : Run the model performance check script]

Why do we need this PR?
The playbook had hardcoded URLs scattered across multiple files using http://localhost:8000, which caused the following problems:

• localhost vs 127.0.0.1 resolution issue: localhost was resolving to IPv6 ::1, but vLLM was only listening on IPv4
• Different tasks used different URL formats
• Users couldn't easily change the API URL without editing multiple files

I added a new centralized variable gpu_validation_vllm_api_url

[csibbitt]

I'd like the s/uri/command/ change to be reverted unless you can show that ansible.builtin.uri actually runs from the controller as claimed.

After investigating this issue manually in the live environment, I believe the only meaningful change is s/localhost/127.0.0.1/ due to localhost resolving to ::1 and vllm only listening on IPV4.

Thanks for finding that!

[csibbitt]

👍

[csibbitt]

AFAICT This is not true, so I do not approve of this change from ansible.builtin.uri to ansible.builtin.command

Using the new variable is fine.

[bogdando]

both will run on the same target host.
If you want another host, you should have it in the inventory, and use delegate_to

[MiguelCarpio]

@csibbitt Thanks for the review, ansible.builtin.uri works fine and runs on the target VM. The issue was always the localhost resolving to ::1 and vllm only listening on IPV4. I updated the PR.

@bogdando setting gpu_validation_vllm_api_url we can check any URI reachable from the target VM.

[csibbitt]

Same as above

[bogdando]

LGTM