A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 18:30 UTC
[Bugfix] Explicitly disable Phi-4-multimodal in V1 (#14889)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Cyrus Leung committed
Mar 17, 2025 at 00:05 UTC
e53b1350f289d65011d9251fd826646c169018df
1 file changed
+3
-2
vllm/model_executor/models/phi4mm.py
+3
-2
@@ -33,7 +33,7 @@ from vllm.sequence import IntermediateTensors, SequenceData
33
from vllm.transformers_utils.tokenizer import cached_tokenizer_from_config
34
35
from .idefics2_vision_model import Idefics2VisionTransformer
36
-from .interfaces import SupportsLoRA, SupportsMultiModal
36
+from .interfaces import SupportsLoRA, SupportsMultiModal, SupportsV0Only
37
from .phi4mm_audio import AudioEmbedding
38
from .utils import AutoWeightsLoader, WeightsMapper, maybe_prefix
39
@@ -1433,7 +1433,8 @@ def cat_with_pad(tensors, dim, padding_value=0):
1433
"image", get_max_phi4mm_image_tokens)
1434
@INPUT_REGISTRY.register_dummy_data(dummy_data_for_phi4mm)
1435
@INPUT_REGISTRY.register_input_processor(input_processor_for_phi4mm)
1436
-class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal):
1436
+class Phi4MMForCausalLM(nn.Module, SupportsLoRA, SupportsMultiModal,
1437
+ SupportsV0Only):
1438
"""
1439
Implements the Phi-4-multimodal-instruct model in vLLM.
1440
"""