When running an SGLang inference with tp=2, SGLang spawns 2 child processes, each running on a separate GPU.
Using cuda-checkpoint on one of the child processes succeeds, but restoring it results in an error. “Could not restore on process ID 3406143: "invalid argument"”
When running an SGLang inference with tp=2, SGLang spawns 2 child processes, each running on a separate GPU.
Using cuda-checkpoint on one of the child processes succeeds, but restoring it results in an error. “Could not restore on process ID 3406143: "invalid argument"”