[Core][Bugfix] Fix Offline MM Beam Search (#16390)

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
This commit is contained in:
Alex Brooks
2025-04-15 10:33:02 +08:00
committed by GitHub
co-authored by Cyrus Leung
parent d2020acac7
commit 6b40996ae8
4 changed files with 140 additions and 30 deletions
+20 -12
View File
@@ -29,12 +29,11 @@ from vllm.distributed import (cleanup_dist_env_and_memory,
init_distributed_environment,
initialize_model_parallel)
from vllm.inputs import (ExplicitEncoderDecoderPrompt, TextPrompt,
TokensPrompt, to_enc_dec_tuple_list,
zip_enc_dec_prompts)
to_enc_dec_tuple_list, zip_enc_dec_prompts)
from vllm.logger import init_logger
from vllm.outputs import RequestOutput
from vllm.sampling_params import BeamSearchParams
from vllm.utils import cuda_device_count_stateless, is_list_of
from vllm.utils import cuda_device_count_stateless
logger = init_logger(__name__)
@@ -469,12 +468,19 @@ class HfRunner:
prompts: list[str],
beam_width: int,
max_tokens: int,
images: Optional[PromptImageInput] = None,
videos: Optional[PromptVideoInput] = None,
audios: Optional[PromptAudioInput] = None,
) -> list[tuple[list[list[int]], list[str]]]:
outputs = self.generate(prompts,
do_sample=False,
max_new_tokens=max_tokens,
num_beams=beam_width,
num_return_sequences=beam_width)
num_return_sequences=beam_width,
images=images,
videos=videos,
audios=audios)
for i in range(len(outputs)):
output_ids, output_str = outputs[i]
for j in range(len(output_ids)):
@@ -936,18 +942,20 @@ class VllmRunner:
def generate_beam_search(
self,
prompts: Union[list[str], list[list[int]]],
prompts: list[str],
beam_width: int,
max_tokens: int,
images: Optional[PromptImageInput] = None,
videos: Optional[PromptVideoInput] = None,
audios: Optional[PromptAudioInput] = None,
) -> list[tuple[list[list[int]], list[str]]]:
if is_list_of(prompts, str, check="all"):
prompts = [TextPrompt(prompt=prompt) for prompt in prompts]
else:
prompts = [
TokensPrompt(prompt_token_ids=tokens) for tokens in prompts
]
inputs = self.get_inputs(prompts,
images=images,
videos=videos,
audios=audios)
outputs = self.model.beam_search(
prompts,
inputs,
BeamSearchParams(beam_width=beam_width, max_tokens=max_tokens))
returned_outputs = []
for output in outputs: