IslandYAN

About multi-gpus training

2

That's definitely an impressive work! I'm trying to reproduce some results on inpainting task and had some concern about the data_parallel mode. Referring to the codes, batch_size is 4 for...

推理视频时无法使用num_return_sequences参数

2

您好，我在尝试使用Qwen3-VL-4B进行视频推理，选择了官方的demo代码，但是发现如果更改num_return_sequences为2或者8等非1的其他值时，会有报错。在num_return_sequences=1时可以正常运行。在使用图像推理时也可以正常推理。 ``` model = AutoModelForImageTextToText.from_pretrained( "Qwen3-VL-4B-Instruct", dtype=torch.bfloat16, attn_implementation="flash_attention_2", device_map="auto" ).cuda() .... generated_ids = model.generate(**inputs, max_new_tokens=128, num_return_sequences=2) ``` 报错信息为： --- 不知道是否有解决办法呢？谢谢。