HabanaAI/vllm-fork reached EOL at v1.23.0; vllm-gaudi plugin is the only supported path
[major] framework
Fields
id
gotchas/xpu-gaudi-fork-eolClass
framework
Severity
major
Concepts
gaudihpuvllm
Symptom
The HabanaAI/vllm-fork repository reached end-of-life (EOL) at v1.23.0 and is deprecated in v1.24.0, remaining functional only for legacy use cases until then. The vllm-gaudi plugin is production-ready in v1.23.0 and becomes the default by v1.24.0.
Root cause
Intel migrated from a fork-based approach to a plugin-based architecture for vLLM on Gaudi. The plugin approach follows the vLLM hardware pluggable RFC and provides better maintainability.
Workaround
Migrate to the vllm-gaudi plugin (https://github.com/vllm-project/vllm-gaudi). The plugin is community-maintained by the Intel Gaudi team. Version 0.26.0 is built on vLLM 0.26.0 and compatible with Gaudi v1.24.1 and PyTorch 2.11.
notes
Anyone still using the HabanaAI/vllm-fork should migrate to the plugin. The fork is no longer receiving updates. The plugin supports all major models (Llama 3.1, Mistral, Mixtral, Qwen2.5, Qwen3, DeepSeek) and includes custom ops for MoE, FP8 quantization, and Mamba architectures.
Notes
Anyone still using the HabanaAI/vllm-fork should migrate to the plugin. The fork is no longer receiving updates. The plugin supports all major models (Llama 3.1, Mistral, Mixtral, Qwen2.5, Qwen3, DeepSeek) and includes custom ops for MoE, FP8 quantization, and Mamba architectures.