DeepSpeed

94a5551f - test: add CPU unit tests for VLLMRollout and HybridEngineRollout

test: add CPU unit tests for VLLMRollout and HybridEngineRollout VLLMRollout (14 tests): - init validation (wrong engine, missing model path) - _extract_token_ids: all 4 code paths - _start_server: command construction, env vars, vllm_python - _wait_for_health: crash detection, timeout - utility helpers (_get_own_ip, _find_free_port) HybridEngineRollout (9 tests): - config defaults, constructor - _sample_top_p: shape, deterministic, top_p filtering, batch - sync_weights no-op - generate dispatches to CB vs graph+CB Signed-off-by: Guokai Ma <guokai.ma@intel.com>