vLLM 0.9: PyTorch 2.7, Blackwell kernels, large-scale inference

What happened

vLLM v0.9.0 was released, with 649 commits from 215 contributors (82 new).

The technical details

From the release notes:

• upgrade to PyTorch 2.7, and the default wheel moves from CUDA 12.4 to 12.8
• an initial set of optimised NVIDIA Blackwell kernels for attention and MLP, plus FlashInfer support
• initial support for large-scale inference:
  • EP (expert parallelism): MoE permute/unpermute kernels and an all-to-all interface
  • DP (data parallelism): engine process management decoupled from communication
  • PD (prefill/decode disaggregation): NIXL integration and multiple KV connectors
• documentation moved from Sphinx to MkDocs

Why it matters

As open models became huge MoEs (Llama 4 in the previous post), serving them meant spreading experts across GPUs and separating the prompt-processing (prefill) phase from token generation (decode). This release lays that groundwork.