inference bug The expanded size of the tensor (1575) must match the existing size (788) at non-singleton dimension 3. Target sizes: [1, 28, 788, 1575]. Tensor sizes: [1, 1, 788, 788] #144

ljeff97 · 2025-02-26T11:55:37Z

model = Qwen2VLForConditionalGeneration.from_pretrained(
pwd, torch_dtype="auto", device_map="auto"
)
如果使用原qwenvl2 7b模型，可以正常推理
如果使用RL训练后保存下来的qwenvl2 7b模型，会报错

transformers版本已经使用requirements.txt里的版本

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

inference bug The expanded size of the tensor (1575) must match the existing size (788) at non-singleton dimension 3. Target sizes: [1, 28, 788, 1575]. Tensor sizes: [1, 1, 788, 788] #144

inference bug The expanded size of the tensor (1575) must match the existing size (788) at non-singleton dimension 3. Target sizes: [1, 28, 788, 1575]. Tensor sizes: [1, 1, 788, 788] #144

ljeff97 commented Feb 26, 2025

inference bug The expanded size of the tensor (1575) must match the existing size (788) at non-singleton dimension 3. Target sizes: [1, 28, 788, 1575]. Tensor sizes: [1, 1, 788, 788] #144

inference bug The expanded size of the tensor (1575) must match the existing size (788) at non-singleton dimension 3. Target sizes: [1, 28, 788, 1575]. Tensor sizes: [1, 1, 788, 788] #144

Comments

ljeff97 commented Feb 26, 2025