@tduncan / Nvidia-Optimization-Stack-VLLM / commits / ace9fda495

[CI/Build][BugFix][The Rock][AMD] Add spawn method in vision examples to avoid reinitialization (#47932)

Signed-off-by: Randall Smith <Randall.Smith@amd.com> Co-authored-by: Andreas Karatzas <akaratza@amd.com>

rasmith committed Jul 19, 2026 at 13:41 UTC ace9fda495bc9a132584e054cc909d504164141f

2 files changed +8

examples/generate/multimodal/vision_language_multi_image_offline.py

@@ -17,6 +17,7 @@ from transformers import AutoProcessor, AutoTokenizer

17 from vllm import LLM, EngineArgs, SamplingParams

18 from vllm.lora.request import LoRARequest

19 from vllm.multimodal.utils import fetch_image

20 +from vllm.platforms import current_platform

21 from vllm.utils.argparse_utils import FlexibleArgumentParser

23 QUESTION = "What is the content of each image?"

@@ -1443,6 +1444,8 @@ def run_generate(

1444 engine_args.seed = seed

1445 if tensor_parallel_size is not None:

1446 engine_args.tensor_parallel_size = tensor_parallel_size

1447 + if current_platform.is_rocm():

1448 + os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

1449 llm = LLM.from_engine_args(engine_args)

1450

1451 sampling_params = SamplingParams(

@@ -1484,6 +1487,8 @@ def run_chat(

1487 engine_args.seed = seed

1488 if tensor_parallel_size is not None:

1489 engine_args.tensor_parallel_size = tensor_parallel_size

1490 + if current_platform.is_rocm():

1491 + os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

1492 llm = LLM.from_engine_args(engine_args)

1493

1494 sampling_params = (

examples/generate/multimodal/vision_language_offline.py

@@ -21,6 +21,7 @@ from vllm.assets.image import ImageAsset

21 from vllm.assets.video import VideoAsset

22 from vllm.lora.request import LoRARequest

23 from vllm.multimodal.image import convert_image_mode

24 +from vllm.platforms import current_platform

25 from vllm.utils.argparse_utils import FlexibleArgumentParser

@@ -2646,6 +2647,8 @@ def main(args):

2647 if args.tensor_parallel_size is not None:

2648 engine_args.tensor_parallel_size = args.tensor_parallel_size

2649 engine_args = maybe_add_vit_cuda_graph_compilation_config(args, engine_args)

2650 + if current_platform.is_rocm():

2651 + os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"

2652 llm = LLM.from_engine_args(engine_args)

2653

2654 # Don't want to check the flag multiple times, so just hijack `prompts`.