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 6, 14:30 UTC
[Bugfix] Fix loading Music Flamingo (#35535)
Signed-off-by: Nick Cao <ncao@redhat.com>
Nick Cao committed
Mar 17, 2026 at 01:24 UTC
20b14095a4e64e0cba71a40b264d0bc96ffb9c07
2 files changed
+10
-7
vllm/model_executor/models/audioflamingo3.py
-6
@@ -128,12 +128,6 @@ class AudioFlamingo3Encoder(Qwen2AudioEncoder):
128
super().__init__(config)
129
self.avg_pooler = nn.AvgPool1d(kernel_size=2, stride=2)
130
# self.layer_norm is already initialized in super().__init__
131
- # Keep a dummy freqs parameter for MusicFlamingo checkpoints.
132
- self.pos_emb = nn.Module()
133
- freqs = torch.empty(getattr(config, "num_mel_bins", 128))
134
- self.pos_emb.register_parameter(
135
- "freqs", nn.Parameter(freqs, requires_grad=False)
136
- )
131
132
def forward(
133
self,
vllm/model_executor/models/musicflamingo.py
+10
-1
@@ -21,6 +21,7 @@ from vllm.multimodal.processing import BaseProcessingInfo
21
from .audioflamingo3 import (
22
AudioFlamingo3DummyInputsBuilder,
23
AudioFlamingo3ForConditionalGeneration,
24
+ AudioFlamingo3MultiModalDataParser,
25
AudioFlamingo3MultiModalProcessor,
26
)
27
@@ -53,8 +54,16 @@ class MusicFlamingoProcessingInfo(BaseProcessingInfo):
54
hf_processor = self.get_hf_processor(**kwargs)
55
return hf_processor.feature_extractor
56
57
+ def get_data_parser(self):
58
+ feature_extractor = self.get_feature_extractor()
59
+
60
+ return AudioFlamingo3MultiModalDataParser(
61
+ target_sr=feature_extractor.sampling_rate,
62
+ expected_hidden_size=self._get_expected_hidden_size(),
63
+ )
64
+
65
def get_supported_mm_limits(self) -> Mapping[str, int | None]:
57
- return {"audio": None}
66
+ return {"audio": 1}
67
68
69
class MusicFlamingoDummyInputsBuilder(AudioFlamingo3DummyInputsBuilder):