chore(vllm_batch): bump batch image to vLLM 0.17.1 for NVFP4 TP>1 fix

[lilyz-ai]

Summary

• Bumps vllm_batch and vllm_batch_v2 image tags from 0.17.0-rc1-batch → 0.17.1-batch
• vLLM 0.17.1 adds explicit Nemotron-3-Super NVFP4 support and fixes accuracy with tensor parallelism > 1 (required since batch jobs use num_shards=4)
• Image built and pushed: 692474966980.dkr.ecr.us-west-2.amazonaws.com/llm-engine/batch-infer-vllm:0.17.1-batch

Build command

cd llm-engine
./model-engine/model_engine_server/inference/vllm/build_and_upload_image.sh \
    batch vllm_batch \
    --vllm-version=0.17.1 \
    --vllm-base-image=vllm/vllm-openai:v0.17.1 \
    --account=692474966980

Test plan

• After merge + Helm redeploy, configmap shows 0.17.1-batch
• Run test_nemotron_super_120b_batch.py end-to-end

🤖 Generated with Claude Code

Greptile Summary

This PR bumps the vllm_batch and vllm_batch_v2 Helm values from 0.17.0-rc1-batch to 0.17.1-batch to pick up vLLM's official NVFP4 tensor-parallelism fix, and pairs the image bump with a small backward-compatibility shim in vllm_server.py that re-adds --disable-log-requests as a silent no-op whenever the flag is absent from vLLM's own arg parser (it was removed in v0.17).

• charts/model-engine/values.yaml: both vllm_batch and vllm_batch_v2 tags updated from 0.17.0-rc1-batch → 0.17.1-batch.
• model-engine/model_engine_server/inference/vllm/vllm_server.py: defensive if not any(...) guard checks parser._actions before adding the deprecated flag, preventing a double-registration conflict if vLLM ever re-introduces the argument in a future version.
• The shim is well-commented and follows existing patterns in the file; no functional logic is changed.

Confidence Score: 5/5

• This PR is safe to merge — it is a targeted version bump with a minimal, well-guarded backward-compat shim and no logic regressions.
• Both changes are minimal and low-risk: the values.yaml change is a single-line tag bump, and the vllm_server.py change is a defensive no-op shim that only activates when the flag is absent from the parser. The guard prevents double-registration, the help text is accurate, and the comment clearly explains the rationale. No existing behaviour is altered for any caller that does not pass the deprecated flag.
• No files require special attention.

Important Files Changed

Filename	Overview
charts/model-engine/values.yaml	Bumps vllm_batch and vllm_batch_v2 image tags from 0.17.0-rc1-batch to 0.17.1-batch. Straightforward version pin change with no logic impact.
model-engine/model_engine_server/inference/vllm/vllm_server.py	Adds a backward-compatibility shim that re-registers --disable-log-requests as a no-op if vLLM v0.17+ dropped it, preventing arg-parse failures in existing callers. The guard against double-registration and use of getattr are correct defensive patterns.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[parse_args called] --> B[make_arg_parser builds vLLM parser]
    B --> C{Is '--disable-log-requests'\nalready in parser._actions?}
    C -- Yes\n(vLLM re-added it) --> E[Skip — no duplicate registration]
    C -- No\n(vLLM v0.17+ removed it) --> D[Add '--disable-log-requests'\nas no-op action=store_true]
    D --> F[parser.parse_args]
    E --> F
    F --> G[Return args to caller]

Loading

_{Reviews (1): Last reviewed commit: "chore(vllm_batch): bump batch image to v..." | Re-trigger Greptile}