Gli heap mentono: debug di una perdita di memoria in vLLM.

A memory leak was investigated within the vLLM framework that occurred only under specific conditions involving a disaggregated prefill/decode setup. Initial debugging attempts using standard Python memory profilers like Memray and Guppy failed to identify the source of the leak. The team then turned to Heaptrack, a more advanced profiler, which revealed that while heap memory was stable, the Resident Set Size (RSS) was continuously increasing. This discrepancy indicated the leak was happening outside the heap, likely within anonymous memory mappings managed by the `mmap` system call. The investigation suggested the root cause was related to the KV Cache transfer mechanism involving NIXL and UCX.