Adapt to latest vllm changes

[lianhao]

Description

• Remove --eager-enforce on hpu to improve performance
• Refactor to the upstream docker entrypoint changes

Issues

Fixes #631.

Type of change

List the type of change like below. Please delete options that are not relevant.

• Bug fix (non-breaking change which fixes an issue)

Dependencies

List the newly introduced 3rd party dependency if exists.

Tests

Describe the tests that you ran to verify your changes.

[eero-t]

Approved, matches changes for "GenAIExamples" in opea-project/GenAIExamples#1210 (and corresponding PR for "GenAIComps" repo).

@poussa ?

[eero-t]

Investigating the CI failure for "agent, gaudi, ci-gaudi-values, common" test, I see 2 bugs:

• agent values.yaml specifies (huge) meta-llama/Meta-Llama-3.1-70B-Instruct model for CPU version of TGI
  • https://github.com/opea-project/GenAIInfra/blob/main/helm-charts/common/agent/values.yaml#L10
  • Something that heavy should be used only for sharded accelerator configs, not for CPU!
• CI test does not provide HF token to access that for the sharded Gaudi TGI pods:
  • [pod/agent20241211092531-tgi-6f6b65dc97-cjhjh/model-downloader] Access to model meta-llama/Llama-3.1-70B-Instruct is restricted. You must have access to it and be authenticated to access it. Please log in.

(Besides the size, I think another model would be nicer as default due to license used on Meta's models.)

[eero-t]

My vLLM PR includes same agent and (relevant) vLLM component changes as yours, but strangely that same CI agent test succeeded for it: https://github.com/opea-project/GenAIInfra/actions/runs/12262626198/job/34212355870?pr=610 ?

EDIT: today's push on my PR got the same issue.

[eero-t]

--tensor-parallel-size option can be dropped, as 1 value is the default:
https://docs.vllm.ai/en/latest/usage/engine_args.html

[lianhao]

This specific failure is caused by 2 different bugs: #639, #641

[lianhao]

we need to wait for PR #642 to land-in first

[eero-t]

This is included to #610, so reviewing & merging that instead is another option.

[lianhao]

Let's use #610 instead.