Heaps do lie: debugging a memory leak in vLLM.
Mistral AI
Mistral AI's engineering team debugged a memory leak in vLLM that caused 400 MB per minute of memory growth under specific conditions involving disaggregated serving and graph compilation. The leak occurred outside the heap in anonymous memory mappings managed by mremap system calls, which reallocated memory without properly releasing it, and was traced using BPFtrace after standard heap profiling tools proved inadequate. The investigation required moving beyond Python and heap-level analysis to kernel-level tracing to identify that NIXL's KV Cache transfer through UCX was triggering uncontrolled memory resizing.
