【开源实习】技术公开课内容测试任务10

Season2.step_into_llm/13.musicgen/README.md

@@ -22,4 +22,96 @@ git clone https://github.com/mindspore-lab/mindnlp
 
 cd mindnlp/llm/inference/musicgen
 pip install gradio
-python app_zh.py
+python app_zh.py
+
+
+**如果使用mindspore2.6.0和mindnlp0.4.0运行run_musicgen.ipynb时则需要对mindnlp源码作以下修改**
+- 1. vim ~/anaconda3/envs/ms/lib/python3.9/site-packages/mindnlp/core/ops/creation.py
+注释这行代码：#from mindspore._c_expression import Tensor as CTensor # pylint: disable=no-name-in-module, import-error
+修改为：from mindspore import Tensor as CTensor
+- 2. vim ~/anaconda3/envs/ms/lib/python3.9/site-packages/mindnlp/transformers/models/jamba/modeling_jamba.py
+#from mindspore._c_expression import Tensor as RawTensor # pylint: disable=no-name-in-module
+from  mindspore  import Tensor as RawTenso
+- 3. vim ~/anaconda3/envs/ms/lib/python3.9/site-packages/mindnlp/core/ops/creation.py
+找到empty函数注释并修改为下面这个函数：
+def empty(*size, dtype=None):
+    import mindspore as ms
+    from mindspore import dtype as mstype
+    import numpy as np
+
+    # 处理size参数
+    if isinstance(size[0], (tuple, list)):
+        size = size[0]
+
+    # 处理dtype参数
+    if dtype is None:
+        dtype = mstype.float32
+    elif isinstance(dtype, (float, int)):
+        dtype = mstype.float32
+    elif not isinstance(dtype, mstype.Type):
+        try:
+            dtype = mstype.pytype_to_dtype(dtype)
+        except:
+            dtype = mstype.float32
+
+    # 使用numpy作为中介创建空数组
+    np_array = np.empty(size, dtype=np.float32)
+    return ms.Tensor(np_array, dtype=dtype)
+如果使用的是CPU还需要进行4.5.操作
+- 4. vim ~/anaconda3/envs/ms/lib/python3.9/site-packages/mindnlp/core/ops/other.py
+找到repeat_interleave函数注释并修改为如下函数：
+def repeat_interleave(input, repeats, dim=None):
+    if dim is None:
+        input = input.flatten()
+        dim = 0
+    if isinstance(repeats, int):
+        if input.dtype == ms.bool_:
+            return mindspore.Tensor(np.tile(input.asnumpy(), (repeats,) + (1,) * (len(input.shape) - 1)), dtype=mindspore.bool_)
+        else:
+            return mindspore.numpy.tile(input, (repeats,) + (1,) * (len(input.shape) - 1))
+    else:
+        raise NotImplementedError("Only integer `repeats` is supported.")
+- 5. vim ~/anaconda3/envs/ms/lib/python3.9/site-packages/mindnlp/transformers/modeling_attn_mask_utils.py
+找到 _make_causal_mask函数注释并修改为下面的函数：
+def _make_causal_mask(
+    input_ids_shape,
+    dtype,  # 注意：调用方必须保证传入正确的dtype
+    past_key_values_length: int = 0,
+    sliding_window: Optional[int] = None,
+):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz, tgt_len = input_ids_shape
+
+    # 1. 创建初始mask（直接使用0/1而不是极值）
+    mask = ops.zeros((tgt_len, tgt_len), dtype=dtype)
+    mask_cond = ops.arange(tgt_len)
+
+    # 2. 使用三角矩阵生成因果mask
+    causal_mask = ops.triu(
+        ops.ones((tgt_len, tgt_len), dtype=mstype.float32), 
+        diagonal=1
+    ).to(dtype)  # 先创建float矩阵再转换
+
+    # 3. 合并past key values长度
+    if past_key_values_length > 0:
+        mask = ops.cat([
+            ops.zeros((tgt_len, past_key_values_length), dtype=dtype),
+            causal_mask
+        ], dim=-1)
+    else:
+        mask = causal_mask
+
+    # 4. 滑动窗口处理（保持类型一致）
+    if sliding_window is not None:
+        diagonal = past_key_values_length - sliding_window + 1
+        context_mask = ops.triu(
+            ops.ones((tgt_len, tgt_len + past_key_values_length), dtype=mstype.float32),
+            diagonal=diagonal
+        ).to(dtype)
+        mask = mask + context_mask  # 使用加法代替masked_fill
+
+    return mask[None, None, :, :].broadcast_to(
+        (bsz, 1, tgt_len, tgt_len + past_key_values_length)
+    )