diff --git a/examples/generate/multimodal/vision_language_multi_image_offline.py b/examples/generate/multimodal/vision_language_multi_image_offline.py index c354142774..1b9fddc99e 100644 --- a/examples/generate/multimodal/vision_language_multi_image_offline.py +++ b/examples/generate/multimodal/vision_language_multi_image_offline.py @@ -17,6 +17,7 @@ from transformers import AutoProcessor, AutoTokenizer from vllm import LLM, EngineArgs, SamplingParams from vllm.lora.request import LoRARequest from vllm.multimodal.utils import fetch_image +from vllm.platforms import current_platform from vllm.utils.argparse_utils import FlexibleArgumentParser QUESTION = "What is the content of each image?" @@ -1443,6 +1444,8 @@ def run_generate( engine_args.seed = seed if tensor_parallel_size is not None: engine_args.tensor_parallel_size = tensor_parallel_size + if current_platform.is_rocm(): + os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" llm = LLM.from_engine_args(engine_args) sampling_params = SamplingParams( @@ -1484,6 +1487,8 @@ def run_chat( engine_args.seed = seed if tensor_parallel_size is not None: engine_args.tensor_parallel_size = tensor_parallel_size + if current_platform.is_rocm(): + os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" llm = LLM.from_engine_args(engine_args) sampling_params = ( diff --git a/examples/generate/multimodal/vision_language_offline.py b/examples/generate/multimodal/vision_language_offline.py index bddf6388ae..661401046b 100644 --- a/examples/generate/multimodal/vision_language_offline.py +++ b/examples/generate/multimodal/vision_language_offline.py @@ -21,6 +21,7 @@ from vllm.assets.image import ImageAsset from vllm.assets.video import VideoAsset from vllm.lora.request import LoRARequest from vllm.multimodal.image import convert_image_mode +from vllm.platforms import current_platform from vllm.utils.argparse_utils import FlexibleArgumentParser @@ -2646,6 +2647,8 @@ def main(args): if args.tensor_parallel_size is not None: engine_args.tensor_parallel_size = args.tensor_parallel_size engine_args = maybe_add_vit_cuda_graph_compilation_config(args, engine_args) + if current_platform.is_rocm(): + os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" llm = LLM.from_engine_args(engine_args) # Don't want to check the flag multiple times, so just hijack `prompts`.