Add the LinearLora Class

examples/singa_peft/src/singa_peft/tuners/linear_lora/layer.py

@@ -0,0 +1,89 @@
+#
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+#
+
+import math
+from singa import tensor
+from singa import autograd
+from singa import layer
+
+
+class LinearLoRALayer(layer.Layer):
+    """
+    LinearLoRALayer: LoRA implemented in a linear layer
+    """
+    def __init__(
+            self,
+            base_layer: layer.Linear,
+            r: int = 8,
+            lora_alpha: int = 1,
+            lora_dropout: float = 0.,
+    ):
+        r"""
+        Args:
+            base_layer: a linear layer, The input and output channels of the linear lora layer are equal to this base layer.
+            r: the rank in LoRA, which determines the size of the low-rank matrix. An integer greater than 0 is required, default 8.
+            lora_alpha: learning rate scaling factor, default 1
+            lora_dropout: dropout ratio, default 0.
+        """
+        super().__init__()
+        if r <= 0:
+            raise ValueError(f"`r` should be a positive integer value but the value passed is {r}")
+        self.r = r
+        self.base_layer = base_layer
+        self.in_features = base_layer.in_features
+        self.out_features = base_layer.out_features
+        self.lora_alpha = lora_alpha
+        self.lora_dropout = lora_dropout
+        self.merged = False
+
+
+    def initialize(self, x):
+        # freeze weights of base layer
+        if self.base_layer._initialized is False:
+            self.base_layer.initialize(x)
+        self.freeze_pretrained_weight(True)
+        # actual trainable parameters
+        lora_A_shape = (self.r, self.in_features)
+        lora_B_shape = (self.out_features, self.r)
+        self.lora_A = tensor.Tensor(
+            shape=lora_A_shape,
+            dtype=x.dtype,
+            requires_grad=True,
+            stores_grad=True
+        )
+        self.lora_B = tensor.Tensor(
+            shape=lora_B_shape,
+            dtype=x.dtype,
+            requires_grad=True,
+            stores_grad=True
+        )
+        std = math.sqrt(2.0 / (self.in_features + self.out_features))
+        # initialize A the same way as the default for nn.Linear and B to zero
+        self.lora_A.gaussian(0.0, std)
+        self.lora_B.set_value(0.0)
+        self.scaling = tensor.Tensor(shape=(1,), requires_grad=False, stores_grad=False)
+        self.scaling.set_value(1.0 * self.lora_alpha / self.r)
+
+    def freeze_pretrained_weight(self, freeze: bool = True):
+        # freeze weights of base layer
+        self.base_layer.W.requires_grad = not freeze
+        self.base_layer.W.stores_grad = not freeze
+        if self.base_layer.b is not None:
+            self.base_layer.b.requires_grad = not freeze
+            self.base_layer.b.stores_grad = not freeze