1. pip uninstall setuptools

2. pip install setuptools==59.5.0

--- a/VisCPM/models/cpmbee.py
+++ b/VisCPM/models/cpmbee.py
@@ -1,4 +1,5 @@
import torch
+import torch_mlu
from typing import Optional, Tuple, List
from typing_extensions import TypedDict

@@ -260,7 +261,7 @@ class CPMBeeTorch(torch.nn.Module):
attention_mask = attention_mask & sample_mask_2d
# length mask
mask_1d = present_num_segments != 0
- attention_mask = mask_1d.view(batch, 1, len_buffer) & attention_mask
+ attention_mask = mask_1d.view(batch, 1, len_buffer) & attention_mask.bool()
if hidden_states is None:
hidden_states = self.input_ ding(input, input_sub)

diff --git a/VisCPM/generation/vllm_bee.py b/VisCPM/generation/vllm_bee.py
old mode 100644
new mode 100755
index d2820da..e628956
--- a/VisCPM/generation/vllm_bee.py
+++ b/VisCPM/generation/vllm_bee.py
@@ -1,6 +1,7 @@
from typing import Any, Dict, List, Tuple, Optional
import numpy as np
import torch
+import torch_mlu
import torch.nn.functional as F
import os

@@ -235,7 +236,8 @@ class VLLMCPMBeeGeneration:

model_inputs, other_info = self._process_list(data_list, max_inp_length)

- with torch.inference_mode():
+ with torch.no_grad():
+ #with torch.inference_mode():
if vision_hidden_states is None:
pixel_values = torch.stack(pixel_values).to(self.device)
model_inputs['pixel_values'] = pixel_values

xxx/python3.7/site-packages/torch/utils/data/__init__.py

from torch.utils.data.dataloader import DataLoader, _DatasetKind, get_worker_info, default_collate