biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Fix qwen2.5-vl overflow issue (#13968)

Browse Source 
Signed-off-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
Isotr0py
2025-02-28 01:30:39 +08:00
committed by GitHub
parent 1dd422b64a
commit 7864875879
4 changed files with 22 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
vllm/model_executor/models/minicpmo.py
View File

@@ -47,7 +47,7 @@ from .minicpmv import (MiniCPMV2_6, MiniCPMVDummyInputsBuilder,
MiniCPMVMultiModalDataParser,
MiniCPMVMultiModalProcessor, MiniCPMVProcessingInfo,
_minicpmv_field_config)
from .utils import AutoWeightsLoader, maybe_prefix
from .utils import AutoWeightsLoader, cast_overflow_tensors, maybe_prefix
CPU_DEVICE = torch.device("cpu")
@@ -469,13 +469,8 @@ class MiniCPMWhisperEncoderLayer(nn.Module):
training=self.training)
hidden_states = residual + hidden_states
if hidden_states.dtype == torch.float16 and (
torch.isinf(hidden_states).any()
or torch.isnan(hidden_states).any()):
clamp_value = torch.finfo(hidden_states.dtype).max - 1000
hidden_states = torch.clamp(hidden_states,
min=-clamp_value,
max=clamp_value)
if hidden_states.dtype == torch.float16:
hidden_states = cast_overflow_tensors(hidden_states)
outputs = (hidden_states, )