mct-nightly 2.2.0.20240930.532__py3-none-any.whl → 2.2.0.20241002.500__py3-none-any.whl

{mct_nightly-2.2.0.20240930.532.dist-info → mct_nightly-2.2.0.20241002.500.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.2.0.20240930.532
+Version: 2.2.0.20241002.500
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct_nightly-2.2.0.20240930.532.dist-info → mct_nightly-2.2.0.20241002.500.dist-info}/RECORD

@@ -1,4 +1,4 @@
-model_compression_toolkit/__init__.py,sha256=t1wNg0lS5JpWEPKyQf-PoKxWqEgc_58HfooVwjVCFsQ,1573
+model_compression_toolkit/__init__.py,sha256=0OYwjkiM5Okt4kzkKaRTqc3Iq-TCsVN1uOUzYDgffog,1573
 model_compression_toolkit/constants.py,sha256=i4wYheBkIdQmsQA-axIpcT3YiSO1USNc-jaNiNE8w6E,3920
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=3DByV41XHRR3kLTJNbpaMmikL8icd9e1N-nkQAY9oDk,4567
@@ -15,7 +15,7 @@ model_compression_toolkit/core/common/framework_implementation.py,sha256=kSg2f7w
 model_compression_toolkit/core/common/framework_info.py,sha256=1ZMMGS9ip-kSflqkartyNRt9aQ5ub1WepuTRcTy-YSQ,6337
 model_compression_toolkit/core/common/memory_computation.py,sha256=ixoSpV5ZYZGyzhre3kQcvR2sNA8KBsPZ3lgbkDnw9Cs,1205
 model_compression_toolkit/core/common/model_builder_mode.py,sha256=jll9-59OPaE3ug7Y9-lLyV99_FoNHxkGZMgcm0Vkpss,1324
-model_compression_toolkit/core/common/model_collector.py,sha256=ofcepKtxc3j2Ouz6BpAKXTzPgjABnpRP47ndmJCXAkk,8352
+model_compression_toolkit/core/common/model_collector.py,sha256=T0J3hLmqJI8eQEXlBfqbnPNJ4XpPUp0zfRSjL0CQYu8,8381
 model_compression_toolkit/core/common/model_validation.py,sha256=LaG8wd6aZl0OJgieE3SeiVDEPxtk8IHq9-3wSnmWhY4,1214
 model_compression_toolkit/core/common/node_prior_info.py,sha256=WXX_PrGVG9M9I_REG5ZzFBohwmV4yf356sZnrja_FLo,2832
 model_compression_toolkit/core/common/similarity_analyzer.py,sha256=FikcIqgQQpfiXr9VJvgl-wk8OyH7-LvC8ku7TkhJfJM,9200
@@ -219,7 +219,7 @@ model_compression_toolkit/core/pytorch/__init__.py,sha256=Rf1RcYmelmdZmBV5qOKvKW
 model_compression_toolkit/core/pytorch/constants.py,sha256=YwD_joIF0vK8UG2vW1NVvg36pCNWA0vHOXjAgy_XWn0,2794
 model_compression_toolkit/core/pytorch/default_framework_info.py,sha256=-Vls1P_8Ckm_18nnOsmQkZ71SmzHwtQLbQ383Z4Rb-U,4365
 model_compression_toolkit/core/pytorch/pytorch_device_config.py,sha256=S25cuw10AW3SEN_fRAGRcG_I3wdvvQx1ehSJzPnn-UI,4404
-model_compression_toolkit/core/pytorch/pytorch_implementation.py,sha256=xmcJyU-rkIDX1a_X9LILzf2Ko2z_4I4xnlHkezKH-2w,27669
+model_compression_toolkit/core/pytorch/pytorch_implementation.py,sha256=2RGf4ii9zxJwGLA3mp-qzDp4khFaYNUNN95bNuNNZ0c,27868
 model_compression_toolkit/core/pytorch/pytorch_node_prior_info.py,sha256=2LDQ7qupglHQ7o1Am7LWdfYVacfQnl-aW2N6l9det1w,3264
 model_compression_toolkit/core/pytorch/resource_utilization_data_facade.py,sha256=xpKj99OZKT9NT0vKIl_cOe8d89d2gef1gKoNT6PFElE,4989
 model_compression_toolkit/core/pytorch/utils.py,sha256=GE7T8q93I5C4As0iOias_dk9HpOvXM1N6---dJlyD60,3863
@@ -249,6 +249,7 @@ model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/remove_
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/reshape_with_static_shapes.py,sha256=hAZXzrEinHa-dJHLj39Hy_9Q-13QyO95rtYVSLrhvT8,4915
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/residual_collapsing.py,sha256=DcJEIkGvBdIMOelNIwaJUZ5UsAHiGnDJPR20I464vWo,2929
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/scale_equalization.py,sha256=XFtU9yuBmoZlX0f0mS6otMPWMk-RcWs94XdvvTNhW8Y,3303
+model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/scaled_dot_product_attention.py,sha256=ziL7jwTnjzTf7BHPRPYgWBSCUrSXSyjZnvQqsJhD1nM,12466
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/shift_negative_activation.py,sha256=3WCLvPyx7tVkM0rwYhYq-gntCzW9R_DcImR1ucKlPac,10772
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/softmax_shift.py,sha256=05lV4pIL3hJkZl4JQPV4wk_EFD0eYLG5b8cdzvZk4P8,1588
 model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/transform_function_call_method.py,sha256=EC9Dvp-_UlpDWnipnf8ds65wh_Y-T8pXAFIwRScWpiY,2044
@@ -550,8 +551,8 @@ tests_pytest/pytorch/gptq/test_annealing_cfg.py,sha256=hGC7L6mp3N1ygcJ3OctgS_Fz2
 tests_pytest/pytorch/gptq/test_gradual_act_quantization.py,sha256=tI01aFIUaiCILL5Qn--p1E_rLBUelxLdSY3k52lwcx0,4594
 tests_pytest/pytorch/trainable_infrastructure/__init__.py,sha256=RAe8mgIr1V8dRIQtLf_dSG5zTUCKuQzxyybYx1dzEAs,697
 tests_pytest/pytorch/trainable_infrastructure/test_linear_annealing.py,sha256=eNOpSp0GoLxtEdiRypBp8jaujXfdNxBwKh5Rd-P7WLs,1786
-mct_nightly-2.2.0.20240930.532.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-mct_nightly-2.2.0.20240930.532.dist-info/METADATA,sha256=p4oG8xi2574mBzHWDgHCKzuBAz49q9DLDya346NWcYc,20830
-mct_nightly-2.2.0.20240930.532.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
-mct_nightly-2.2.0.20240930.532.dist-info/top_level.txt,sha256=csdfSXhtRnpWYRzjZ-dRLIhOmM2TEdVXUxG05A5fgb8,39
-mct_nightly-2.2.0.20240930.532.dist-info/RECORD,,
+mct_nightly-2.2.0.20241002.500.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+mct_nightly-2.2.0.20241002.500.dist-info/METADATA,sha256=XUo1iMNL1fh6tGsBz-kglXfHvxhfWdOBebgokDVQJ4A,20830
+mct_nightly-2.2.0.20241002.500.dist-info/WHEEL,sha256=eOLhNAGa2EW3wWl_TU484h7q1UNgy0JXjjoqKoxAAQc,92
+mct_nightly-2.2.0.20241002.500.dist-info/top_level.txt,sha256=csdfSXhtRnpWYRzjZ-dRLIhOmM2TEdVXUxG05A5fgb8,39
+mct_nightly-2.2.0.20241002.500.dist-info/RECORD,,

model_compression_toolkit/__init__.py

@@ -27,4 +27,4 @@ from model_compression_toolkit import data_generation
 from model_compression_toolkit import pruning
 from model_compression_toolkit.trainable_infrastructure.keras.load_model import keras_load_quantized_model
 
-__version__ = "2.2.0.20240930.000532"
+__version__ = "2.2.0.20241002.000500"

model_compression_toolkit/core/common/model_collector.py

@@ -158,7 +158,7 @@ class ModelCollector:
         for td, sc in zip(tensor_data, self.stats_containers_list):
             if isinstance(sc, (list, tuple)):
                 if not isinstance(td, (list, tuple)):
-                    Logger.critical('\'tensor_data\' must be a list or a tuple if \'stats_containers_list\' contains lists or tuples.') # pragma: no cover
+                    Logger.critical(f"\'tensor_data\' is of type {type(td)} but must be of the same type as \'stats_containers_list\', which is of type {type(sc)}") # pragma: no cover
                 if len(sc) != len(td):
                     Logger.critical('\'tensor_data\' and \'stats_containers_list\' must have matching lengths') # pragma: no cover
                 for tdi, sci in zip(td, sc):

model_compression_toolkit/core/pytorch/graph_substitutions/substitutions/scaled_dot_product_attention.py

@@ -0,0 +1,231 @@
+# Copyright 2024 Sony Semiconductor Israel, Inc. All rights reserved.
+#
+# Licensed 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 torch.nn as nn
+import torch
+import math
+from copy import copy
+import numpy as np
+from model_compression_toolkit.core.common.graph.functional_node import FunctionalNode
+from model_compression_toolkit.core.common import BaseSubstitution
+from model_compression_toolkit.core.common.graph.graph_matchers import NodeOperationMatcher
+from model_compression_toolkit.core.common.graph.base_graph import Graph, BaseNode, OutTensor
+from model_compression_toolkit.core.pytorch.constants import DIM
+from model_compression_toolkit.core.pytorch.pytorch_device_config import get_working_device
+
+
+class ScaledDotProductDecomposition(BaseSubstitution):
+    """
+    Decompose torch.nn.scale_dot_product into its base operators:
+        Transpose (over k)
+        MatMul(over q and transposed k)
+        Mul (for scaling)
+        Add (for masking. optional operation, used in cases that attn_mask ig given)
+        Dropout
+        Softmax
+        Matmul.
+    """
+
+    def __init__(self):
+        """
+        Matches scaled_dot_product_attention node.
+        """
+        super().__init__(matcher_instance=NodeOperationMatcher(nn.functional.scaled_dot_product_attention))
+
+    def _get_input_by_name(self, attention_node: FunctionalNode, input_name: str,
+                           input_index: int, default_value: any) -> any:
+        """
+        Search for attention_node input value in op_call_kwargs (using input_name) and op_call_args (using input_index).
+        In case the input is not given, returns its default_value.
+
+        """
+        if input_name in attention_node.op_call_kwargs:
+            return attention_node.op_call_kwargs[input_name]
+        elif len(attention_node.op_call_args) > input_index:  # input order: [attn_mask, dropout_p, is_causal]
+            return attention_node.op_call_args[input_index]
+        return default_value
+
+    def _get_attention_input_nodes(self, graph: Graph, attention_node: FunctionalNode) -> dict:
+        q, k, v = 0, 1, 2
+        prev_nodes = graph.get_prev_nodes(attention_node, sink_index_sorted=True)
+        q_node, k_node, v_node = prev_nodes[q], prev_nodes[k], prev_nodes[v]
+        return {"q": q_node, "k": k_node, "v": v_node}
+
+    def _get_transpose_k_node(self, attention_node_name: str, key_node: BaseNode) -> BaseNode:
+        input_shape, output_shape = copy(key_node.output_shape[0]), copy(key_node.output_shape[0])
+        output_shape[-2], output_shape[-1] = input_shape[-1], input_shape[-2]
+        transpose_node = FunctionalNode(name=f"{attention_node_name}_{key_node.name}_transpose",
+                                        framework_attr={},
+                                        input_shape=input_shape,
+                                        output_shape=output_shape,
+                                        weights={},
+                                        layer_class=torch.transpose,
+                                        op_call_args=[-1, -2],  # axes to transpose
+                                        op_call_kwargs={},
+                                        functional_op=torch.transpose)
+        return transpose_node
+
+    def _get_scale_node(self, attention_node: FunctionalNode, q_node: BaseNode, matmul_node: BaseNode) -> FunctionalNode:
+        """
+        :return: multiplication node that represents multiplication by the scale factor
+        """
+        scale_name = f'{attention_node.name}_scale'
+        q_embd_axis = -1
+        input_scale = self._get_input_by_name(attention_node, "scale", 3, None)
+        scale_factor = input_scale if input_scale else (1 / math.sqrt(q_node.output_shape[0][q_embd_axis]))
+        scale_node = FunctionalNode(name=scale_name,
+                                    framework_attr={},
+                                    input_shape=(matmul_node.output_shape),
+                                    output_shape=matmul_node.output_shape,
+                                    weights={},
+                                    layer_class=torch.mul,
+                                    op_call_args=[scale_factor],
+                                    op_call_kwargs={},
+                                    functional_op=torch.mul)
+        return scale_node
+
+    def _get_matmul_node(self, attention_node_name: str, q_node: BaseNode, transposed_k_node: BaseNode) -> BaseNode:
+        matmul1_output_shape = copy(q_node.output_shape[0])
+        matmul1_output_shape[-2] = q_node.output_shape[0][-2]
+        matmul1_output_shape[-1] = transposed_k_node.output_shape[-1]
+        matmul_name = f'{attention_node_name}_matmul1'
+        return FunctionalNode(name=matmul_name,
+                              framework_attr={},
+                              input_shape=(tuple(q_node.output_shape[0]), tuple(transposed_k_node.output_shape)),
+                              output_shape=tuple(matmul1_output_shape),
+                              weights={},
+                              layer_class=torch.matmul,
+                              op_call_args=[],
+                              op_call_kwargs={},
+                              functional_op=torch.matmul)
+
+    def _get_mask_node(self, attention_node: FunctionalNode, scale_node: FunctionalNode) -> FunctionalNode:
+        """
+        :return: Add operator node with the mask tensor as input. In case there is no mask tensor, returns None.
+        """
+        attention_mask_tensor = self._get_attention_mask_tensor(attention_node)
+        if attention_mask_tensor is None:
+            return None
+        mask_node_name = f'{attention_node.name}_mask'
+        return FunctionalNode(name=mask_node_name,
+                              framework_attr={},
+                              input_shape=(scale_node.output_shape),
+                              output_shape=scale_node.output_shape,
+                              weights={},
+                              layer_class=torch.add,
+                              op_call_args=[],
+                              op_call_kwargs={'other': attention_mask_tensor},
+                              functional_op=torch.add)
+
+    def _get_softmax_node(self, attention_node_name: str, in_out_shape: tuple) -> BaseNode:
+        softmax_name = f'{attention_node_name}_softmax'
+        return BaseNode(name=softmax_name,
+                        framework_attr={DIM: -1},
+                        input_shape=in_out_shape,
+                        output_shape=in_out_shape,
+                        weights={},
+                        layer_class=nn.Softmax)
+
+    def _get_matmul2_node(self, attention_node_name: str, softmax_node: BaseNode, v_node: BaseNode) -> FunctionalNode:
+        matmul2_output_shape = list(copy(softmax_node.output_shape))
+        matmul2_output_shape[-2] = softmax_node.output_shape[-2]
+        matmul2_output_shape[-1] = v_node.output_shape[0][-1]
+        matmul2_name = f'{attention_node_name}_matmul2'
+        return FunctionalNode(name=matmul2_name,
+                              framework_attr={},
+                              input_shape=(tuple(softmax_node.output_shape), tuple(v_node.output_shape[0])),
+                              output_shape=tuple(matmul2_output_shape),
+                              weights={},
+                              layer_class=torch.matmul,
+                              op_call_args=[],
+                              op_call_kwargs={},
+                              functional_op=torch.matmul)
+
+    def _get_attention_mask_tensor(self, attention_node: FunctionalNode) -> torch.Tensor:
+        """
+        :return: mask tensor given as part of attention node input.
+        Since MCT doesn't support infinite values, we don't support is_causal (torch.nn.scale_dot_product_attention
+        argument) and boolean mask tensor, as they both require -inf values.
+        """
+        device = get_working_device()
+        is_causal = self._get_input_by_name(attention_node, "is_causal", 2, False)
+        if is_causal:
+            raise NotImplementedError("scaled_dot_product_attention is_causal feature is not implemented.")
+        input_weights = list(attention_node.weights.values())
+        attn_mask = input_weights[0] if len(input_weights) > 0 else None
+        if attn_mask is not None and (attn_mask.dtype == "bool"):
+            raise NotImplementedError(
+                "scaled_dot_product_attention attn_mask is of type boolean, which is not supported.")
+        if attn_mask is not None and (not np.isfinite(attn_mask).all()):
+            raise NotImplementedError(
+                "scaled_dot_product_attention attn_mask contains infinite value, which is not supported.")
+        return torch.from_numpy(attn_mask).to(device) if attn_mask is not None else None
+
+    def _get_dropout_node(self, attention_node: FunctionalNode, in_out_shape: tuple) -> BaseNode:
+        dropout_p = attention_node.op_call_kwargs.get('dropout_p', 0)
+        dropout_name = f'{attention_node.name}_dropout'
+        return BaseNode(name=dropout_name,
+                        framework_attr={"p": dropout_p},
+                        input_shape=in_out_shape,
+                        output_shape=in_out_shape,
+                        weights={},
+                        layer_class=nn.Dropout)
+
+    def substitute(self, graph: Graph, attention_node: FunctionalNode) -> Graph:
+        """
+        Removes a scaled_dot_product_attention node from the graph, and replaces it with a compatible graph that
+        consists of:
+            Transpose (over k)
+            MatMul(over q and transposed k)
+            Mul (for scaling)
+            Add (for masking. optional operation, used in cases that attn_mask ig given)
+            Dropout
+            Softmax
+            Matmul.
+        :param graph: A Graph to apply substitution on
+        :param attention_node: the node to replace
+        :return: A graph after the substitution
+        """
+        print("In scale_dot_product_attention substitution@@@@@@@@")
+        input_nodes = self._get_attention_input_nodes(graph, attention_node)
+        q_node, k_node, v_node = input_nodes["q"], input_nodes["k"], input_nodes["v"]
+        transpose_k_node = self._get_transpose_k_node(attention_node.name, k_node)
+        matmul_node = self._get_matmul_node(attention_node.name, q_node, transpose_k_node)
+        scale_node = self._get_scale_node(attention_node, q_node, matmul_node)
+        mask_node = self._get_mask_node(attention_node, scale_node)
+        softmax_node = self._get_softmax_node(attention_node.name, matmul_node.output_shape)
+        dropout_node = self._get_dropout_node(attention_node, softmax_node.output_shape)
+        matmul2_node = self._get_matmul2_node(attention_node.name, softmax_node, v_node)
+
+        graph.add_node_with_in_edges(transpose_k_node, [k_node])
+        graph.add_node_with_in_edges(matmul_node, [q_node, transpose_k_node])
+        graph.add_node_with_in_edges(scale_node, [matmul_node])
+        if mask_node:
+            graph.add_node_with_in_edges(mask_node, [scale_node])
+        graph.add_node_with_in_edges(softmax_node, [mask_node if mask_node else scale_node])
+        graph.add_node_with_in_edges(dropout_node, [softmax_node])
+        graph.add_node_with_in_edges(matmul2_node, [dropout_node if dropout_node else softmax_node, v_node])
+
+        graph_outputs = graph.get_outputs()
+        for i, g_out in enumerate(graph_outputs):
+            if g_out.node == attention_node:
+                graph_outputs[i] = OutTensor(node=matmul2_node, node_out_index=g_out.node_out_index)
+
+        graph.reconnect_out_edges(current_node=attention_node, new_node=matmul2_node)
+        graph.remove_edge(q_node, attention_node)
+        graph.remove_edge(k_node, attention_node)
+        graph.remove_edge(v_node, attention_node)
+        graph.remove_node(attention_node, new_graph_outputs=graph_outputs)
+        return graph

model_compression_toolkit/core/pytorch/pytorch_implementation.py

@@ -53,6 +53,8 @@ from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.li
     pytorch_linear_collapsing
 from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.multi_head_attention_decomposition \
     import MultiHeadAttentionDecomposition
+from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.scaled_dot_product_attention import \
+    ScaledDotProductDecomposition
 from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.transform_function_call_method import \
     TransformFunctionCallMethod
 from model_compression_toolkit.core.pytorch.graph_substitutions.substitutions.const_holder_conv import \
@@ -237,6 +239,7 @@ class PytorchImplementation(FrameworkImplementation):
         """
         return [ReshapeWithStaticShapes(),
                 MultiHeadAttentionDecomposition(),
+                ScaledDotProductDecomposition(),
                 TransformFunctionCallMethod(),
                 FunctionalConvSubstitution(fw_info),
                 FunctionalBatchNorm(),