Les heaps mentent bien : déboguer une fuite mémoire dans vLLM.

A memory leak was suspected in vLLM, appearing only under specific conditions with disaggregated prefill/decode serving and graph compilation enabled. Initial debugging attempts with standard Python memory profilers like Memray failed to identify the source of the leak. The investigation then moved to Heaptrack, a more advanced tool, which showed that the heap memory was stable but the Resident Set Size (RSS) was increasing. This discrepancy suggested the leak was occurring outside the heap, likely in anonymous memory mappings managed directly by system calls, prompting a deeper dive into lower-level system analysis.