Merge branch 'main' into replace_deepcopy

Author: Svetlana Karslioglu

advanced_source/rpc_ddp_tutorial.rst

@@ -1,6 +1,6 @@
 Combining Distributed DataParallel with Distributed RPC Framework
 =================================================================
-**Authors**: `Pritam Damania <https://github.com/pritamdamania87>`_ and `Yi Wang <https://github.com/SciPioneer>`_
+**Authors**: `Pritam Damania <https://github.com/pritamdamania87>`_ and `Yi Wang <https://github.com/wayi1>`_
 
 .. note::
    |edit| View and edit this tutorial in `github <https://github.com/pytorch/tutorials/blob/main/advanced_source/rpc_ddp_tutorial.rst>`__.

beginner_source/Intro_to_TorchScript_tutorial.py

@@ -33,6 +33,7 @@
 
 import torch  # This is all you need to use both PyTorch and TorchScript!
 print(torch.__version__)
+torch.manual_seed(191009)  # set the seed for reproducibility
 
 
 ######################################################################
@@ -308,7 +309,7 @@ def forward(self, x, h):
 
 # New inputs
 x, h = torch.rand(3, 4), torch.rand(3, 4)
-traced_cell(x, h)
+print(scripted_cell(x, h))
 
 
 ######################################################################

conf.py

@@ -34,6 +34,7 @@
 import pytorch_sphinx_theme
 import torch
 import glob
+import random
 import shutil
 from custom_directives import IncludeDirective, GalleryItemDirective, CustomGalleryItemDirective, CustomCalloutItemDirective, CustomCardItemDirective
 import distutils.file_util
@@ -85,6 +86,11 @@
 
 # -- Sphinx-gallery configuration --------------------------------------------
 
+def reset_seeds(gallery_conf, fname):
+    torch.manual_seed(42)
+    torch.set_default_device(None)
+    random.seed(10)
+
 sphinx_gallery_conf = {
     'examples_dirs': ['beginner_source', 'intermediate_source',
                       'advanced_source', 'recipes_source', 'prototype_source'],
@@ -94,7 +100,8 @@
     'backreferences_dir': None,
     'first_notebook_cell': ("# For tips on running notebooks in Google Colab, see\n"
                             "# https://pytorch.org/tutorials/beginner/colab\n"
-                            "%matplotlib inline")
+                            "%matplotlib inline"),
+    'reset_modules': (reset_seeds)
 }
 
 if os.getenv('GALLERY_PATTERN'):

intermediate_source/mario_rl_tutorial.py

@@ -53,6 +53,8 @@
 # Super Mario environment for OpenAI Gym
 import gym_super_mario_bros
 
+from tensordict import TensorDict
+from torchrl.data import TensorDictReplayBuffer, LazyMemmapStorage
 
 ######################################################################
 # RL Definitions
@@ -348,7 +350,7 @@ def act(self, state):
 class Mario(Mario):  # subclassing for continuity
     def __init__(self, state_dim, action_dim, save_dir):
         super().__init__(state_dim, action_dim, save_dir)
-        self.memory = deque(maxlen=100000)
+        self.memory = TensorDictReplayBuffer(storage=LazyMemmapStorage(100000))
         self.batch_size = 32
 
     def cache(self, state, next_state, action, reward, done):
@@ -373,14 +375,15 @@ def first_if_tuple(x):
         reward = torch.tensor([reward], device=self.device)
         done = torch.tensor([done], device=self.device)
 
-        self.memory.append((state, next_state, action, reward, done,))
+        # self.memory.append((state, next_state, action, reward, done,))
+        self.memory.add(TensorDict({"state": state, "next_state": next_state, "action": action, "reward": reward, "done": done}, batch_size=[]))
 
     def recall(self):
         """
         Retrieve a batch of experiences from memory
         """
-        batch = random.sample(self.memory, self.batch_size)
-        state, next_state, action, reward, done = map(torch.stack, zip(*batch))
+        batch = self.memory.sample(self.batch_size)
+        state, next_state, action, reward, done = (batch.get(key) for key in ("state", "next_state", "action", "reward", "done"))
         return state, next_state, action.squeeze(), reward.squeeze(), done.squeeze()
 
 

intermediate_source/seq2seq_translation_tutorial.py

@@ -45,7 +45,7 @@
    :alt:
 
 To improve upon this model we'll use an `attention
-mechanism <https://arxiv.org/abs/1409.0473>`__, which lets the decoder
+mechanism <https://arxiv.org/abs/1508.04025>`__, which lets the decoder
 learn to focus over a specific range of the input sequence.
 
 **Recommended Reading:**
@@ -66,8 +66,8 @@
    Statistical Machine Translation <https://arxiv.org/abs/1406.1078>`__
 -  `Sequence to Sequence Learning with Neural
    Networks <https://arxiv.org/abs/1409.3215>`__
--  `Neural Machine Translation by Jointly Learning to Align and
-   Translate <https://arxiv.org/abs/1409.0473>`__
+-  `Effective Approaches to Attention-based Neural Machine
+   Translation <https://arxiv.org/abs/1508.04025>`__
 -  `A Neural Conversational Model <https://arxiv.org/abs/1506.05869>`__
 
 You will also find the previous tutorials on

intermediate_source/torch_compile_tutorial.py

@@ -69,7 +69,7 @@
 
 def foo(x, y):
     a = torch.sin(x)
-    b = torch.cos(x)
+    b = torch.cos(y)
     return a + b
 opt_foo1 = torch.compile(foo)
 print(opt_foo1(torch.randn(10, 10), torch.randn(10, 10)))
@@ -80,7 +80,7 @@ def foo(x, y):
 @torch.compile
 def opt_foo2(x, y):
     a = torch.sin(x)
-    b = torch.cos(x)
+    b = torch.cos(y)
     return a + b
 print(opt_foo2(torch.randn(10, 10), torch.randn(10, 10)))
 
@@ -105,7 +105,7 @@ def forward(self, x):
 #
 # Let's now demonstrate that using ``torch.compile`` can speed
 # up real models. We will compare standard eager mode and 
-# ``torch.compile`` by evaluating and training ResNet-18 on random data.
+# ``torch.compile`` by evaluating and training a ``torchvision`` model on random data.
 #
 # Before we start, we need to define some utility functions.
 

prototype_source/fx_graph_mode_ptq_dynamic.py

@@ -239,9 +239,27 @@ def evaluate(model_, data_source):
     .set_object_type(nn.LSTM, default_dynamic_qconfig)
     .set_object_type(nn.Linear, default_dynamic_qconfig)
 )
-# Deepcopying the original model because quantization api changes the model inplace and we want
+# Load model to create the original model because quantization api changes the model inplace and we want
 # to keep the original model for future comparison
-model_to_quantize = copy.deepcopy(model)
+
+
+model_to_quantize = LSTMModel(
+    ntoken = ntokens,
+    ninp = 512,
+    nhid = 256,
+    nlayers = 5,
+)
+
+model_to_quantize.load_state_dict(
+    torch.load(
+        model_data_filepath + 'word_language_model_quantize.pth',
+        map_location=torch.device('cpu')
+        )
+    )
+
+model_to_quantize.eval()
+
+
 prepared_model = prepare_fx(model_to_quantize, qconfig_mapping, example_inputs)
 print("prepared model:", prepared_model)
 quantized_model = convert_fx(prepared_model)
@@ -289,4 +307,4 @@ def time_model_evaluation(model, test_data):
 # 3. Conclusion
 # -------------
 # This tutorial introduces the api for post training dynamic quantization in FX Graph Mode,
-# which dynamically quantizes the same modules as Eager Mode Quantization.
+# which dynamically quantizes the same modules as Eager Mode Quantization.

recipes_source/recipes/changing_default_device.py

@@ -43,9 +43,6 @@
 print(mod.weight.device)
 print(mod(torch.randn(128, 20)).device)
 
-# And then globally return it back to CPU
-torch.set_default_device('cpu')
-
 ################################################################
 # This function imposes a slight performance cost on every Python
 # call to the torch API (not just factory functions). If this