[Doc] Relax Deep Dive (#17380)

* [Doc] Relax Deep Dive

Similar as TensorIR Deep Dive, we also have Relax Deep Dive.

docs/conf.py

@@ -424,6 +424,7 @@ def jupyter_notebook(script_blocks, gallery_conf, target_dir, real_func):
     # New tutorial structure under docs folder
     tvm_path.joinpath("docs", "get_started", "tutorials"),
     tvm_path.joinpath("docs", "how_to", "tutorials"),
+    tvm_path.joinpath("docs", "deep_dive", "relax", "tutorials"),
     tvm_path.joinpath("docs", "deep_dive", "tensor_ir", "tutorials"),
 ]
 
@@ -443,6 +444,7 @@ def jupyter_notebook(script_blocks, gallery_conf, target_dir, real_func):
     # New tutorial structure under docs folder
     "get_started/tutorials/",
     "how_to/tutorials/",
+    "deep_dive/relax/tutorials/",
     "deep_dive/tensor_ir/tutorials/",
 ]
 
@@ -598,10 +600,10 @@ def force_gc(gallery_conf, fname):
 ## Setup header and other configs
 import tlcpack_sphinx_addon
 
-footer_copyright = "© 2023 Apache Software Foundation | All rights reserved"
+footer_copyright = "© 2024 Apache Software Foundation | All rights reserved"
 footer_note = " ".join(
     """
-Copyright © 2023 The Apache Software Foundation. Apache TVM, Apache, the Apache feather,
+Copyright © 2024 The Apache Software Foundation. Apache TVM, Apache, the Apache feather,
 and the Apache TVM project logo are either trademarks or registered trademarks of
 the Apache Software Foundation.""".split(
         "\n"
@@ -614,7 +616,6 @@ def force_gc(gallery_conf, fname):
 header_links = [
     ("Community", "https://tvm.apache.org/community"),
     ("Download", "https://tvm.apache.org/download"),
-    ("VTA", "https://tvm.apache.org/vta"),
     ("Blog", "https://tvm.apache.org/blog"),
     ("Docs", "https://tvm.apache.org/docs"),
     ("Conference", "https://tvmconf.org"),

docs/deep_dive/relax/abstraction.rst

@@ -0,0 +1,73 @@
+..  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.
+
+.. _relax-abstraction:
+
+Graph Abstraction for ML Models
+-------------------------------
+Graph abstraction is a key technique used in machine learning (ML) compilers
+to represent and reason about the structure and data flow of ML models. By
+abstracting the model into a graph representation, the compiler can perform
+various optimizations to improve performance and efficiency. This tutorial will
+cover the basics of graph abstraction, its key elements of Relax IR, and how it enables optimization in ML compilers.
+
+What is Graph Abstraction?
+~~~~~~~~~~~~~~~~~~~~~~~~~~
+Graph abstraction is the process of representing an ML model as a directed graph,
+where the nodes represent computational operations (e.g., matrix multiplication,
+convolution) and the edges represent the flow of data between these operations.
+This abstraction allows the compiler to analyze the dependencies and
+relationships between different parts of the model.
+
+.. code:: python
+
+    from tvm.script import relax as R
+
+    @R.function
+    def main(
+        x: R.Tensor((1, 784), dtype="float32"),
+        weight: R.Tensor((784, 256), dtype="float32"),
+        bias: R.Tensor((256,), dtype="float32"),
+    ) -> R.Tensor((1, 256), dtype="float32"):
+        with R.dataflow():
+            lv0 = R.matmul(x, weight)
+            lv1 = R.add(lv0, bias)
+            gv = R.nn.relu(lv1)
+            R.output(gv)
+        return gv
+
+Key Features of Relax
+~~~~~~~~~~~~~~~~~~~~~
+Relax, the graph representation utilized in Apache TVM's Unity strategy,
+facilitates end-to-end optimization of ML models through several crucial
+features:
+
+- **First-class symbolic shape**: Relax employs symbolic shapes to represent
+  tensor dimensions, enabling global tracking of dynamic shape relationships
+  across tensor operators and function calls.
+
+- **Multi-level abstractions**: Relax supports cross-level abstractions, from
+  high-level neural network layers to low-level tensor operations, enabling
+  optimizations that span different hierarchies within the model.
+
+- **Composable transformations**: Relax offers a framework for composable
+  transformations that can be selectively applied to different model components.
+  This includes capabilities such as partial lowering and partial specialization,
+  providing flexible customization and optimization options.
+
+These features collectively empower Relax to offer a powerful and adaptable approach
+to ML model optimization within the Apache TVM ecosystem.

docs/deep_dive/relax/index.rst

@@ -0,0 +1,34 @@
+..  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.
+
+.. _relax:
+
+Relax
+=====
+Relax is a high-level abstraction for graph optimization and transformation in Apache TVM stack.
+Additionally, Apache TVM combine Relax and TensorIR together as a unity strategy for cross-level
+optimization. Hence, Relax is usually working closely with TensorIR for representing and optimizing
+the whole IRModule
+
+
+.. toctree::
+    :maxdepth: 2
+
+    abstraction
+    learning
+    tutorials/relax_creation
+    tutorials/relax_transformation