• Sources: primary
  • Summary: Meta's design adds a plugin interface to lldb-server that sets internal breakpoints on GPU runtime events, initializes the vendor debug API, and opens a second LLDB target for the device. New GDB remote packets handle plugin discovery and connection setup, and a mock accelerator plugin tests the flow without GPU hardware. The post shows a HIP kernel breakpoint resolving when the GPU module loads, and it supports GPU core dumps.
  • Why it matters: Engineers debugging AI workloads across host and kernel code may get one LLDB or IDE session instead of switching between ROCgdb and CUDA-GDB.
  • Background: GPU vendors have shipped their own debuggers, such as AMD's ROCgdb and NVIDIA's CUDA-GDB, because GPUs use different instruction sets, lane-based execution and separate address spaces. lldb-server is the LLDB component that controls the debugged process and speaks the GDB remote protocol to the client.