fix(T2V-Turbo): fix the bug of lora save&load in T2V-Turbo (#830)

* bugfix num_train_epochs

* replace ops.repeat() with repeat_interleave() or tile()

* fix lora save&load

* bugfix and aligns tile

* bugfix lora

examples/t2v_turbo/intern_vid2/models/backbones/bert/xbert.py

@@ -985,7 +985,7 @@ def get_extended_attention_mask(self, attention_mask: Tensor, input_shape: Tuple
             if is_decoder:
                 batch_size, seq_length = input_shape
                 seq_ids = ops.arange(seq_length)
-                causal_mask = seq_ids[None, None, :].repeat(batch_size, seq_length, 1) <= seq_ids[None, :, None]
+                causal_mask = seq_ids[None, None, :].tile((batch_size, seq_length, 1)) <= seq_ids[None, :, None]
                 # in case past_key_values are used we need to add a prefix ones mask to the causal mask
                 # causal and attention masks must have same type with pytorch version < 1.3
                 causal_mask = causal_mask.to(attention_mask.dtype)

examples/t2v_turbo/intern_vid2/models/backbones/internvideo2/internvideo2.py

@@ -908,7 +908,7 @@ def pretrain_internvideo2_6b_patch14_224(config):
         dim=-1,
     ).to(ms.bool)
 
-    output = model(ops.rand((4, 3, num_frames, img_size, img_size)), mask.repeat(4, 1))
+    output = model(ops.rand((4, 3, num_frames, img_size, img_size)), mask.repeat_interleave(4, 1))
     print(output[0].shape)
     print(output[1].shape)
     print(output[2].shape)

examples/t2v_turbo/lvdm/modules/encoders/ip_resampler.py

@@ -132,7 +132,7 @@ def __init__(
             )
 
     def construct(self, x):
-        latents = self.latents.repeat(x.shape[0], 1, 1)
+        latents = self.latents.tile((x.shape[0], 1, 1))
 
         x = self.proj_in(x)
 

examples/t2v_turbo/pipeline/t2v_turbo_ms_pipeline.py

@@ -67,7 +67,7 @@ def _encode_prompt(
 
         bs_embed, seq_len, _ = prompt_embeds.shape
         # duplicate text embeddings for each generation per prompt, using mps friendly method
-        prompt_embeds = prompt_embeds.repeat(num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.tile((1, num_videos_per_prompt, 1))
         prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, seq_len, -1)
 
         # Don't need to get uncond prompt embedding because of LCM Guided Distillation
@@ -177,7 +177,7 @@ def __call__(
         bs = batch_size * num_videos_per_prompt
 
         # 6. Get Guidance Scale Embedding
-        w = ms.Tensor(guidance_scale).repeat(bs)
+        w = ms.Tensor(guidance_scale).tile((bs,))
         w_embedding = self.get_w_embedding(w, embedding_dim=256)
 
         # 7. LCM MultiStep Sampling Loop:

examples/t2v_turbo/pipeline/t2v_turbo_vc2_pipeline.py

@@ -56,7 +56,7 @@ def _encode_prompt(
 
         bs_embed, seq_len, _ = prompt_embeds.shape
         # duplicate text embeddings for each generation per prompt, using mps friendly method
-        prompt_embeds = prompt_embeds.repeat(num_videos_per_prompt, 1)
+        prompt_embeds = prompt_embeds.tile((1, num_videos_per_prompt, 1))
         prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, seq_len, -1)
 
         # Don't need to get uncond prompt embedding because of LCM Guided Distillation
@@ -171,7 +171,7 @@ def __call__(
         bs = batch_size * num_videos_per_prompt
 
         # 6. Get Guidance Scale Embedding
-        w = ms.Tensor(guidance_scale).repeat(bs)
+        w = ms.Tensor(guidance_scale).tile((bs,))
         w_embedding = self.get_w_embedding(w, embedding_dim=256)
 
         # 7. LCM MultiStep Sampling Loop:

examples/t2v_turbo/utils/lora.py

@@ -4,6 +4,8 @@
 from itertools import groupby
 from typing import Dict, List, Optional, Set, Tuple, Type, Union
 
+import numpy as np
+
 import mindspore as ms
 import mindspore.common.initializer as init
 from mindspore import nn, ops
@@ -19,6 +21,8 @@
 def load_lora_from_pkl(file_path):
     with open(file_path, "rb") as file:
         loras = pickle.load(file)
+
+    loras = [ms.Tensor(lora_weight) if isinstance(lora_weight, np.ndarray) else lora_weight for lora_weight in loras]
     return loras
 
 
@@ -571,8 +575,8 @@ def save_lora_weight(
 ):
     weights = []
     for _up, _down in extract_lora_ups_down(model, target_replace_module=target_replace_module):
-        weights.append(_up.weight.value().to(ms.float32))
-        weights.append(_down.weight.value().to(ms.float32))
+        weights.append(_up.weight.value().asnumpy())
+        weights.append(_down.weight.value().asnumpy())
 
     import pickle
 

examples/t2v_turbo/viclip/viclip_vision.py

@@ -219,11 +219,11 @@ def inflate_weight(weight_2d, time_dim, center=True):
     logger.info(f"Init center: {center}")
     if center:
         weight_3d = ops.zeros(*weight_2d.shape)
-        weight_3d = weight_3d.unsqueeze(2).repeat(1, 1, time_dim, 1, 1)
+        weight_3d = weight_3d.unsqueeze(2).tile((1, 1, time_dim, 1, 1))
         middle_idx = time_dim // 2
         weight_3d[:, :, middle_idx, :, :] = weight_2d
     else:
-        weight_3d = weight_2d.unsqueeze(2).repeat(1, 1, time_dim, 1, 1)
+        weight_3d = weight_2d.unsqueeze(2).tile((1, 1, time_dim, 1, 1))
         weight_3d = weight_3d / time_dim
     return weight_3d