vLLM 0.33 makes CUDA 13 the default and removes two legacy execution paths

vLLM 0.33.0 is a broad serving-stack release with a practical migration consequence: its default CUDA wheels and images now require CUDA 13.0, and the project no longer publishes the CUDA 12.9 variants. The release also upgrades its core stack to PyTorch 2.14.0 and Triton 3.8.0 (vLLM).

That packaging change lands alongside a wider accelerator refresh. The default ROCm artifacts move to ROCm 10.0, while ROCm 7.2 remains available under explicitly versioned tags. vLLM also adds an AWS EFA image that bundles EFA libraries and the Mooncake transfer engine for multi-node serving, and publishes shorter vllm/vllm aliases for its release images.

What changed in serving

The release adds an OpenAI-compatible /v1/decisions endpoint for supported generation models and enables it by default. The embeddings API gains token embeddings with optional late chunking, while the Rust frontend closes more parity gaps, including returning a controlled result rather than an HTTP 500 for long conversations and rejecting unsupported flags during startup.

Operationally, workers can now start concurrently when tensor parallelism is greater than one. A watchdog can also dump stack traces from stuck engine and worker processes, giving operators a more direct failure artifact than a silent hang.

The project reports substantial accelerator-specific work as well. Speculative decoding gains the XPress refiner, an opt-in FLy verification policy and quantized draft output heads. NVIDIA paths add latency-focused kernels and prefill improvements; AMD paths add MonoKernel decode support, a compact KV-cache format and broader GPU enablement. Those benchmark numbers are hardware- and workload-specific, so they are useful test targets rather than portable capacity promises.

Who needs to act

Teams pinning CUDA 12.9 images must select a supported CUDA 13.0 artifact before upgrading. ROCm deployments should similarly decide whether to move to the new 10.0 default or stay temporarily on the published 7.2 variant.

The release removes both Medusa and the MLP speculator, along with the legacy Ray compiled-graph executor. It deprecates --mamba-cache-mode, LM Format Enforcer and --ir-op-priority; NIXL connectors now require NIXL 1.5 or newer. Requests received while generation is paused now return HTTP 503, which may also affect health checks, retry rules and client-side error handling.

Before rollout, platform teams should inventory those flags and executors, test image and driver compatibility, and verify clients against the new pause behavior. The release contains 927 commits from 389 contributors, so a staged canary is the safer upgrade path than treating 0.33.0 as a routine image swap.