The report used eight NVIDIA H200 GPUs, CUDA 13.0, PyTorch 2.14.0+cu130, humming-kernels 0.1.18, and a vLLM build from main. Startup fails during profiling with an AssertionError because the model's MoE intermediate size of 2304 becomes a per-rank K dimension of 576 at TP=4 or 288 at TP=8; neither is divisible by the 128-element input-scale group. TP=1 and TP=2 are listed as working, as is expert parallelism.
The issue says vLLM already pads expert K to a multiple of 128, but Humming must support the remaining partial scale group. A Humming GEMM fix is merged and a quantization fix remains draft, with a Humming release and vLLM dependency bump still planned. Main-branch Hopper serving defaults to BF16 activations, so that default path is described as unaffected. The report does not establish a release date for the fix or confirm the behavior across other models, GPU architectures, or deployments.