vLLM reuses its inference kernels in the training forward pass for matching results

vLLM says its training forward pass now calls its inference kernels directly.

The stated aim is that the same token gets the same result across validated batch sizes, phases and parallel layouts. The kernels aligned are FP8 GEMM, sparse-attention Top-K, FP8 KV-cache reads, MoE communication and CUDA Graph execution.

The backward pass does not share these kernels. vLLM says it uses separate checks for the backward pass.