vLLM reuses GPU memory across users for throughput. In its GGUF dequantize kernels, the output tensor is allocated full-size with torch::empty (uninitialized), but the element count is passed through a 32-bit int. Load a model whose tensor dimensions multiply past INT_MAX and that count wraps, so the kernel fills only part of the buffer. The rest keeps stale GPU memory, which in multi-tenant serving can be another user's data. No crash, no error. One typedef is the root cause, one fix corrects a...