[LMI] Bump vLLM wheel to patched v0.28.0 and torch to 2.13.0 for LMI 29.0.0

[shiehbeth]

Description

Point the LMI container at the patched vLLM v0.28.0 wheel (base tag v0.28.0 2cf0a691 + 6 cherry-picked PRs: vllm-project/vllm#44675 vllm-project/vllm#48210 vllm-project/vllm#48363 vllm-project/vllm#48402 vllm-project/vllm#50040 vllm-project/vllm#51742) for the DJL 0.37.0 / LMI 29.0.0 release.

Wheel: https://publish.djl.ai/vllm/vllm-0.28.0-cp312-cp312-linux_x86_64.whl
sha256 100e71ec575de176e717fb396f97a5e3cb207b41d5457117b8690ca2fa55a9b3, built py3.12 / CUDA 13.0 (cu130), TORCH_CUDA_ARCH_LIST=8.0;8.6;8.9;9.0;10.0;12.0.

torch 2.11.0 → 2.13.0 is required, not optional. vLLM 0.28.0 pins torch==2.13.0 (requirements/cuda.txt), and the wheel's compiled .so files are torch-ABI-locked — so the container runtime must move together with the wheel. torchvision moves 0.26.0 → 0.28.0 for the same reason. This is the one difference in shape from the LMI 28.0.0 bump, which was wheel-URL only.

Two notes for reviewers:

• torchaudio stays at 2.11.0 and is deliberately not pinned here. 2.11.0 is the newest released torchaudio — the torch/torchaudio version numbers have decoupled this cycle, so torchaudio==2.13.0 does not exist. This file pins neither torchaudio nor torchcodec, so both resolve transitively; all four (torch 2.13.0+cu130, torchvision 0.28.0+cu130, torchaudio 2.11.0+cu130, torchcodec 0.14+) are confirmed present on the cu130 index.
• The pyproject.toml [test] pin bump is not cosmetic. A stale vllm== pin there breaks test_vllm_embedding with module 'djl_python.lmi_vllm' has no attribute 'vllm_async_service'.

--extra-index-url .../whl/cu130 is unchanged, and lmi.Dockerfile needs no change (already nvidia/cuda:13.0.2-devel-ubuntu24.04 / cuda_version=cu130 / python_version=3.12).

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• New feature (non-breaking change which adds functionality)
• This change requires a documentation update

Checklist:

• Please add the link of Integration Tests Executor run with related tests.
• Have you manually built the docker image and verify the change?
• Have you run related tests? Check how to set up the test environment here; One example would be pytest tests.py -k "TestVllm1" -m "vllm"
• Have you added tests that prove your fix is effective or that this feature works?
• Has code been commented, particularly in hard-to-understand areas?
• Have you made corresponding changes to the documentation?

Feature/Issue validation/testing

Pre-merge verification completed against the built artifact (not predicted):

• Wheel published and reachable. https://publish.djl.ai/vllm/vllm-0.28.0-cp312-cp312-linux_x86_64.whl returns HTTP 200, content-length: 625092004. The object was read back out of S3 and re-hashed to 100e71ec…, matching the build byte-for-byte.
• Pin bumps taken from the wheel's own METADATA, not inferred: Requires-Dist: torch==2.13.0, torchaudio==2.11.0, torchvision==0.28.0, torchcodec>=0.14.
• Dependency pin cross-check clean. All 5 pins overlapping between lmi-container-requirements.txt and vLLM 0.28.0's requirements/{common,cuda}.txt agree — torch/torchvision now match exactly, transformers >= 5.5.3 and compressed-tensors == 0.17.0 were already byte-identical, huggingface-hub resolves. 0 conflicts.
• djl_python ↔ vLLM 0.28.0 API compatibility check clean (static, AST). 35/35 distinct vllm.* import statements resolve (module found and imported symbol present), and 4/4 constructor call sites accept every kwarg passed against the real __init__ signatures (OnlineRenderer, OpenAIServingChat, OpenAIServingCompletion, ServingEmbedding). No glue adaptation appears to be needed for this bump.
• GPU arch coverage confirmed from the binary with cuobjdump: _C_stable_libtorch.abi3.so and _moe_C_stable_libtorch.abi3.so both carry compiled SASS for sm_80, sm_86, sm_89, sm_90, sm_90a, sm_100 and sm_120.

Full container integration testing runs on the nightly build that this merge enables.

[shiehbeth]

Superseded by #3075 — same commit, reopened from an in-repo branch so CI runs without the fork-PR approval gate. Closing this one to keep the trail clean.