feat: upgrade pytorch to 2.3.1, brevitas to 0.10 (#788)

Makefile

@@ -57,6 +57,7 @@ setup_env:
 	poetry run python -m pip install -U --pre --extra-index-url https://pypi.zama.ai/cpu "$(CONCRETE_PYTHON_VERSION)"
 	"$(MAKE)" fix_omp_issues_for_intel_mac
 	poetry run python -c "import skorch" || true # Details above
+	poetry run python -c "from brevitas.core.scaling import AccumulatorAwareParameterPreScaling" || true # Details above
 
 .PHONY: sync_env # Synchronise the environment
 sync_env: 

deps_licenses/licenses_linux_user.txt

@@ -1,32 +1,42 @@
 Name, Version, License
+Jinja2, 3.1.4, BSD License
+MarkupSafe, 2.1.5, BSD License
 PyYAML, 6.0.1, MIT License
-brevitas, 0.8.0, UNKNOWN
-certifi, 2024.6.2, Mozilla Public License 2.0 (MPL 2.0)
+brevitas, 0.10.2, UNKNOWN
+certifi, 2024.7.4, Mozilla Public License 2.0 (MPL 2.0)
 charset-normalizer, 3.3.2, MIT License
 coloredlogs, 15.0.1, MIT License
 concrete-python, 2.7.0, BSD-3-Clause
 dependencies, 2.0.1, BSD License
 dill, 0.3.8, BSD License
-filelock, 3.15.3, The Unlicense (Unlicense)
+filelock, 3.15.4, The Unlicense (Unlicense)
 flatbuffers, 24.3.25, Apache Software License
-fsspec, 2024.6.0, BSD License
+fsspec, 2024.6.1, BSD License
 huggingface-hub, 0.23.4, Apache Software License
 humanfriendly, 10.0, MIT License
-hummingbird-ml, 0.4.8, MIT License
+hummingbird-ml, 0.4.11, MIT License
 idna, 3.7, BSD License
 importlib_resources, 6.4.0, Apache Software License
 joblib, 1.4.2, BSD License
 jsonpickle, 3.2.2, BSD License
 mpmath, 1.3.0, BSD License
 networkx, 3.1, BSD License
 numpy, 1.23.5, BSD License
-nvidia-cublas-cu11, 11.10.3.66, Other/Proprietary License
-nvidia-cuda-nvrtc-cu11, 11.7.99, Other/Proprietary License
-nvidia-cuda-runtime-cu11, 11.7.99, Other/Proprietary License
-nvidia-cudnn-cu11, 8.5.0.96, Other/Proprietary License
+nvidia-cublas-cu12, 12.1.3.1, Other/Proprietary License
+nvidia-cuda-cupti-cu12, 12.1.105, Other/Proprietary License
+nvidia-cuda-nvrtc-cu12, 12.1.105, Other/Proprietary License
+nvidia-cuda-runtime-cu12, 12.1.105, Other/Proprietary License
+nvidia-cudnn-cu12, 8.9.2.26, Other/Proprietary License
+nvidia-cufft-cu12, 11.0.2.54, Other/Proprietary License
+nvidia-curand-cu12, 10.3.2.106, Other/Proprietary License
+nvidia-cusolver-cu12, 11.4.5.107, Other/Proprietary License
+nvidia-cusparse-cu12, 12.1.0.106, Other/Proprietary License
+nvidia-nccl-cu12, 2.20.5, Other/Proprietary License
+nvidia-nvjitlink-cu12, 12.5.82, Other/Proprietary License
+nvidia-nvtx-cu12, 12.1.105, Other/Proprietary License
 onnx, 1.16.1, Apache License v2.0
 onnxconverter-common, 1.13.0, MIT License
-onnxmltools, 1.11.0, Apache Software License
+onnxmltools, 1.12.0, Apache Software License
 onnxoptimizer, 0.3.13, Apache License v2.0
 onnxruntime, 1.18.0, MIT License
 packaging, 24.1, Apache Software License; BSD License
@@ -39,16 +49,19 @@ requests, 2.32.3, Apache Software License
 scikit-learn, 1.1.3, BSD License
 scipy, 1.10.1, BSD License
 six, 1.16.0, MIT License
-skl2onnx, 1.12, Apache Software License
+skl2onnx, 1.17.0, Apache Software License
 skops, 0.5.0, MIT
 skorch, 0.11.0, new BSD 3-Clause
-sympy, 1.12.1, BSD License
+sympy, 1.13.0, BSD License
 tabulate, 0.8.10, MIT License
 threadpoolctl, 3.5.0, BSD License
-torch, 1.13.1, BSD License
+torch, 2.3.1, BSD License
 tqdm, 4.66.4, MIT License; Mozilla Public License 2.0 (MPL 2.0)
-typing_extensions, 4.5.0, Python Software Foundation License
+triton, 2.3.1, MIT License
+typing_extensions, 4.12.2, Python Software Foundation License
 tzdata, 2024.1, Apache Software License
+unfoldNd, 0.2.2, MIT License
 urllib3, 2.2.2, MIT License
 xgboost, 1.6.2, Apache Software License
 z3-solver, 4.13.0.0, MIT License
+zipp, 3.19.2, MIT License

deps_licenses/licenses_linux_user.txt.md5

@@ -1 +1 @@
-adb925c3b7be3e651975febcf49b6543
+6d367701c3ef5eff8763f4e994e03681

docs/deep-learning/onnx_support.md

@@ -15,54 +15,22 @@ This example uses Post-Training Quantization, i.e., the quantization is not perf
 ```python
 import numpy
 import onnx
-import tensorflow
-import tf2onnx
 
 from concrete.ml.torch.compile import compile_onnx_model
 from concrete.fhe.compilation import Configuration
 
 
-class FC(tensorflow.keras.Model):
-    """A fully-connected model."""
-
-    def __init__(self):
-        super().__init__()
-        hidden_layer_size = 10
-        output_size = 5
-
-        self.dense1 = tensorflow.keras.layers.Dense(
-            hidden_layer_size,
-            activation=tensorflow.nn.relu,
-        )
-        self.dense2 = tensorflow.keras.layers.Dense(output_size, activation=tensorflow.nn.relu6)
-        self.flatten = tensorflow.keras.layers.Flatten()
-
-    def call(self, inputs):
-        """Forward function."""
-        x = self.flatten(inputs)
-        x = self.dense1(x)
-        x = self.dense2(x)
-        return self.flatten(x)
-
 
 n_bits = 6
-input_output_feature = 2
+input_output_feature = 5
 input_shape = (input_output_feature,)
 num_inputs = 1
 n_examples = 5000
 
-# Define the Keras model
-keras_model = FC()
-keras_model.build((None,) + input_shape)
-keras_model.compute_output_shape(input_shape=(None, input_output_feature))
-
 # Create random input
 input_set = numpy.random.uniform(-100, 100, size=(n_examples, *input_shape))
 
-# Convert to ONNX
-tf2onnx.convert.from_keras(keras_model, opset=14, output_path="tmp.model.onnx")
-
-onnx_model = onnx.load("tmp.model.onnx")
+onnx_model = onnx.load(f"tests/data/tf_onnx/fc_{input_output_feature}.onnx")
 onnx.checker.check_model(onnx_model)
 
 # Compile
@@ -83,7 +51,7 @@ print("Equality:           ", numpy.sum(y_clear == y_fhe), "over", numpy.size(y_
 ```
 
 {% hint style="warning" %}
-While Keras was used in this example, it is not officially supported. Additional work is needed to test all of Keras's types of layers and models.
+While a Keras ONNX model was used in this example, Keras/Tensorflow support in Concrete ML is only partial and experimental.
 {% endhint %}
 
 ## Quantization Aware Training
@@ -106,7 +74,7 @@ quantized_numpy_module = compile_onnx_model(
 
 ## Supported operators
 
-The following operators are supported for evaluation and conversion to an equivalent FHE circuit. Other operators were not implemented, either due to FHE constraints or because they are rarely used in PyTorch activations or scikit-learn models.
+The following ONNX operators are supported for evaluation and conversion to an equivalent FHE circuit. Other operators were not implemented, either due to FHE constraints or because they are rarely used in PyTorch activations or scikit-learn models.
 
 <!--- gen_supported_ops.py: inject supported operations for evaluation [BEGIN] -->
 

docs/deep-learning/torch_support.md

@@ -221,7 +221,6 @@ Concrete ML also supports some of their QAT equivalents from Brevitas.
 - [`torch.nn.CELU`](https://pytorch.org/docs/stable/generated/torch.nn.CELU.html)
 - [`torch.nn.ELU`](https://pytorch.org/docs/stable/generated/torch.nn.ELU.html)
 - [`torch.nn.GELU`](https://pytorch.org/docs/stable/generated/torch.nn.GELU.html)
-- [`torch.nn.Hardshrink`](https://pytorch.org/docs/stable/generated/torch.nn.Hardshrink.html)
 - [`torch.nn.HardSigmoid`](https://pytorch.org/docs/stable/generated/torch.nn.Hardsigmoid.html)
 - [`torch.nn.Hardswish`](https://pytorch.org/docs/stable/generated/torch.nn.Hardswish)
 - [`torch.nn.HardTanh`](https://pytorch.org/docs/stable/generated/torch.nn.Hardtanh.html)