Update requirements-hpu.txt for open telemetry tracing support

[louie-tsai]

OPEA Project starts having OpenTelemetry tracing feature as below PR.
opea-project/GenAIExamples#1488
We have TGI/TEI enabled for the Open Telemetry tracing, and would like to also enable for vllm.
Current runtime issue is missing opentelemetry-api package, so have the PR to fix it.

[michalkuligowski]

I am not sure if we want to add it here, those packages are not needed for other workloads. Also as I look into vllm project repo, they dont add it either to requirements file and leave it to user, also please see:

vllm-fork/examples/online_serving/opentelemetry/Otel.md

Line 4 in 7c55e66

@kzawora-intel can you comment?

[louie-tsai]

@michalkuligowski
Both TGI Gaudi and TEI Gaudi have those packages in their requirements.
here are related lines for those packages installation in TGI.
https://github.com/huggingface/tgi-gaudi/blob/habana-main/server/requirements.txt#L40

We also saw those tracing enabled in OPEA once we gave related OTLP endpoint URL correctly.
opea-project/GenAIExamples#1316

However, vLLM doesn't have that packages, so this is a gap for vLLM Gaudi comparing to TGI Gaudi.
Good to have that enabled like TGI Gaudi.

thanks

[michalkuligowski]

I understand it's in tgi, but it seems that vllm decided to do it this way. Also did you check

vllm-fork/examples/online_serving/opentelemetry/Otel.md

Line 4 in 7c55e66

I linked in my previous comment? It shows requirements for opentelemetry, so I think that should suffice.

[louie-tsai]

@michalkuligowski
all other instructions in the otel.md could be covered by launch scripts like docker compose yaml fie, but no package installation need to handle inside Dockerfile, so the otel.md won't help when users deploy the vllm directly into their cluster. install those packages won't impact performance. open telemetry won't be enabled without setting those arguments for vllm server.

[eero-t]

IMHO doing non-Gaudi change like this in Gaudi fork seems wrong; it should belong either into upstream vLLM project, or OPEA vLLM build.

(OPEA vLLM build script can easily concatenate the needed modules to vllm-fork requirements file.)

[louie-tsai]

IMHO doing non-Gaudi change like this in Gaudi fork seems wrong; it should belong either into upstream vLLM project, or OPEA vLLM build.

(OPEA vLLM build script can easily concatenate the needed modules to vllm-fork requirements file.)

OPEA vLLM use Dockerfile.hpu directly. no other wrapper. also this fork maintain its own requirement.txt, so upstream won't help...

[eero-t]

OPEA vLLM use Dockerfile.hpu directly. no other wrapper. also this fork maintain its own requirement.txt, so upstream won't help...

@louie-tsai Adding something like cat ../requirements-otel.txt >> requirements-hpu.txt just before docker build -f Dockerfile.hpu ... line here should suffice:
https://github.com/opea-project/GenAIComps/blob/main/comps/third_parties/vllm/src/build_docker_vllm.sh

[adobrzyn]

Hi @louie-tsai is this PR still needed? Can we close it?

[michalkuligowski]

Please re-open if still needed, closing as stale