[amrit110]

Supersedes #289 (the Dependabot PR for the same bump, which is far behind main and has merge conflicts). This branch is cut from current main and relocked, so the Docker workflow runs here and builds both images with the import canary under the pr-<number> tag.

What changed

Only uv.lock, relocked with uv lock --upgrade-package vllm==0.24.0:

• vllm 0.22.0 to 0.24.0
• torch stays at 2.11.0 (same CUDA 13 wheels as main), so the CUDA 12.8 base image situation is unchanged from main
• transitive bumps required by vllm 0.24: flashinfer-python/cubin 0.6.12, compressed-tensors 0.17.0, fastsafetensors 0.5.0, humming-kernels 0.1.6, mistral-common 1.12.0, nvidia-cudnn-frontend 1.28.0, nvidia-cuda-nvcc/nvvm 13.2.86, new nvidia-cuda-tileiras
• gguf is no longer pulled in (vllm moved GGUF support to a separate plugin)
• the sglang extra still resolves in the same lock (sglang 0.5.12.post1, torch 2.11.0); it shares mistral-common 1.12.0 and nvidia-cudnn-frontend 1.28.0 with the vllm side

No code changes were needed. I checked vec_inf against the vllm 0.23 and 0.24 changes:

• No EngineArgs fields were removed between 0.22.0 and 0.24.0, so every vllm serve flag vec_inf emits or accepts (--served-model-name, --host, --port, --tensor-parallel-size, --pipeline-parallel-size, --max-model-len, --trust-remote-code, --distributed-executor-backend, the -tp/-pp/-dp* short aliases) still exists.
• Prometheus metric names are identical between 0.22.0 and 0.24.0.
• None of the models in vec_inf/config/models.yaml use an architecture that 0.24 removed.
• Removed env vars (VLLM_RPC_TIMEOUT, VLLM_USE_FBGEMM, VLLM_FLASHINFER_MOE_BACKEND, VLLM_MORIIO_*) are not used anywhere in this repo.

Behaviour changes users may notice

• Removed model architectures: ERNIE, Xverse, Dots1, Bamba, Mono-InternVL (InternLM2VE), the InternLMForCausalLM alias, MiniMax-Text-01/M1/VL-01, JAIS, and first-generation Qwen / QwenVL (QWenLMHeadModel). Custom configs that point at these will fail to load.
• GGUF models now need vllm-gguf-plugin, which is not in the image.
• vllm no longer sets CUDA_VISIBLE_DEVICES internally; there is a new --device-ids flag. Slurm still sets CUDA_VISIBLE_DEVICES for the job, so the normal path should be unaffected.
• Model Runner V2 is now the default for Llama and Mistral dense models (0.23) and for quantized models (0.24). Async EPLB and the Triton MoE backend on Hopper are also on by default.
• Transformers v4 support is deprecated (the lock already uses transformers 5.9.0).

Testing

• Unit tests pass locally on Python 3.10, 3.11 and 3.12 using uv 0.5.21 with the same commands as CI (uv sync --dev --prerelease=allow, then pytest, then uv sync --no-dev and the import test).
• pre-commit run --all-files passes (ruff, mypy, pytest).
• Not verified: actually serving a model on a GPU. The Docker builds here only prove the locked set installs and imports together.