enhence mixtoken for qwen

[cocoshe]

优化了qwen在开启mixtoken时，构造mixtoken datatse时预先读入了所有样本图片，导致的显存易溢出的问题

测试：

# 修改之后，从dataloader中得到的单个step的输入
# batch_size=1, mixtoken=false
bs1_false_images = np.load('bs1_false_images.npy')
bs1_false_input_ids = np.load('bs1_false_input_ids.npy')

# batch_size=1, mixtoken=true
bs1_true_images = np.load('bs1_true_images.npy')
bs1_true_input_ids = np.load('bs1_true_input_ids.npy')

# batch_size=4, mixtoken=false
bs4_false_images = np.load('bs4_false_images.npy')
bs4_false_input_ids = np.load('bs4_false_input_ids.npy')

# batch_size=4, mixtoken=true
bs4_true_images = np.load('bs4_true_images.npy')
bs4_true_input_ids = np.load('bs4_true_input_ids.npy')

# 修改之前，从dataloader中得到的单个step的输入
# batch_size=1, mixtoken=false
origin_bs1_false_images = np.load('origin_bs1_false_images.npy')
origin_bs1_false_input_ids = np.load('origin_bs1_false_input_ids.npy')

# batch_size=1, mixtoken=true
origin_bs1_true_images = np.load('origin_bs1_true_images.npy')
origin_bs1_true_input_ids = np.load('origin_bs1_true_input_ids.npy')

# batch_size=4, mixtoken=false
origin_bs4_false_images = np.load('origin_bs4_false_images.npy')
origin_bs4_false_input_ids = np.load('origin_bs4_false_input_ids.npy')

# batch_size=4, mixtoken=true
origin_bs4_true_images = np.load('origin_bs4_true_images.npy')
origin_bs4_true_input_ids = np.load('origin_bs4_true_input_ids.npy')


# 验证等效
print(bs1_false_images.shape, origin_bs1_false_images.shape)
print(np.sum(bs1_false_images - origin_bs1_false_images))

print(bs1_false_input_ids.shape, origin_bs1_false_input_ids.shape)
print(np.sum(bs1_false_input_ids - origin_bs1_false_input_ids))

print(bs1_true_images.shape, origin_bs1_true_images.shape)
print(np.sum(bs1_true_images - origin_bs1_true_images))

print(bs1_true_input_ids.shape, origin_bs1_true_input_ids.shape)
print(np.sum(bs1_true_input_ids - origin_bs1_true_input_ids))

print(bs4_false_images.shape, origin_bs4_false_images.shape)
print(np.sum(bs4_false_images - origin_bs4_false_images))

print(bs4_false_input_ids.shape, origin_bs4_false_input_ids.shape)
print(np.sum(bs4_false_input_ids - origin_bs4_false_input_ids))

print(bs4_true_images.shape, origin_bs4_true_images.shape)
print(np.sum(bs4_true_images - origin_bs4_true_images))

print(bs4_true_input_ids.shape, origin_bs4_true_input_ids.shape)
print(np.sum(bs4_true_input_ids - origin_bs4_true_input_ids))

输出：

(1, 3, 448, 448) (1, 3, 448, 448)
0.0
(1, 2048) (1, 2048)
0
(6, 3, 448, 448) (6, 3, 448, 448)
0.0
(1, 2048) (1, 2048)
0
(4, 3, 448, 448) (4, 3, 448, 448)
0.0
(4, 2048) (4, 2048)
0
(24, 3, 448, 448) (24, 3, 448, 448)
0.0
(4, 2048) (4, 2048)
0

[paddle-bot]

Thanks for your contribution!

[cocoshe]

@LokeZhou 辛苦review~