xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-03-24 15:45:48 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] MiDashengLM model contact error under concurrent testing (#24738)

Browse Source 
Signed-off-by: chenbing8 <chenbing8@xiaomi.com>
Signed-off-by: bingchen-mi <chenbing8@xiaomi.com>
This commit is contained in:
bingchen-mi 2025-09-15 14:38:12 +08:00 committed by GitHub
parent f4a948f33f
commit a8c0f59973
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 17 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
vllm/model_executor/models/midashenglm.py
View File

@ -497,8 +497,11 @@ class MiDashengLMDummyInputsBuilder(
hf_processor = self.info.get_hf_processor()
audio_token = hf_processor.audio_token
audio_bos_token = hf_processor.audio_bos_token
audio_eos_token = hf_processor.audio_eos_token
return audio_token * num_audios
single_audio_text = f"{audio_bos_token}{audio_token}{audio_eos_token}"
return single_audio_text * num_audios
def get_dummy_mm_data(
self,
@ -577,14 +580,7 @@ class MiDashengLMMultiModalProcessor(
vocab = tokenizer.get_vocab()
audio_token = getattr(processor, "audio_token", "<|AUDIO|>")
audio_bos_token = getattr(processor, "audio_bos_token",
"<|audio_bos|>")
audio_eos_token = getattr(processor, "audio_eos_token",
"<|audio_eos|>")
audio_token_id = vocab[audio_token]
audio_bos_id = vocab[audio_bos_token]
audio_eos_id = vocab[audio_eos_token]
out_mm_data = out_mm_kwargs.get_data()
audio_length = out_mm_data.get("audio_length")
@ -604,7 +600,7 @@ class MiDashengLMMultiModalProcessor(
audio_tokens = [audio_token_id] * num_features
return PromptUpdateDetails.select_token_id(
[audio_bos_id] + audio_tokens + [audio_eos_id],
audio_tokens,
embed_token_id=audio_token_id,
)
@ -670,8 +666,18 @@ class MiDashengLMModel(nn.Module, SupportsMultiModal, SupportsPP):
f"Got type: {type(mm_input)}")
if isinstance(mm_input, torch.Tensor):
return mm_input.reshape(-1, *mm_input.shape[2:])
else:
return torch.concat(mm_input)
if name == "input_values":
max_length = max(tensor.shape[1] for tensor in mm_input)
padded_mm_input = [
torch.nn.functional.pad(tensor,
(0, max_length - tensor.shape[1]))
if tensor.shape[1] < max_length else tensor
for tensor in mm_input
]
return torch.concat(padded_mm_input)
return torch.concat(mm_input)
def _parse_and_validate_audio_input(
self, **kwargs: object) -> Optional[MiDashengLMAudioInputs]: