[Patch] Add loss for ORT inference

.github/workflows/test_onnxruntime_train.yml

@@ -54,14 +54,14 @@ jobs:
       - name: Build Image
         working-directory: examples/onnxruntime/training/docker
         run: |
-            docker build -f Dockerfile-cu10  -t ort/cu10 .
+            docker build -f Dockerfile-ort1.9.0-cu102 -t ort9/cu10 .
       - name: Run Test
         run: |
             docker run --rm -p 80:8888 \
               --gpus all \
               -v $PWD:/workspace \
               --workdir=/workspace \
-              ort/cu10:latest /bin/bash /run.sh    
+              ort9/cu10:latest /bin/bash /run.sh    
 
   stop-runner:
     name: Stop self-hosted EC2 runner

examples/onnxruntime/training/docker/Dockerfile-ort1.11.1-cu113

@@ -0,0 +1,39 @@
+# Use nvidia/cuda image
+FROM nvidia/cuda:11.3.0-cudnn8-devel-ubuntu20.04
+CMD nvidia-smi
+
+# Ignore interactive questions during `docker build`
+ENV DEBIAN_FRONTEND noninteractive
+
+# Bash shell
+RUN chsh -s /bin/bash
+SHELL ["/bin/bash", "-c"]
+
+RUN apt-get update && apt-get install -y --no-install-recommends apt-utils
+
+# Install and update tools to minimize security vulnerabilities
+RUN apt-get update
+RUN apt-get install -y software-properties-common wget apt-utils patchelf git libprotobuf-dev protobuf-compiler cmake \
+    bzip2 ca-certificates libglib2.0-0 libxext6 libsm6 libxrender1 mercurial subversion libopenmpi-dev && \
+    apt-get clean
+RUN unattended-upgrade
+RUN apt-get autoremove -y
+
+# Install Pythyon (3.8 as default)
+RUN update-alternatives --install /usr/bin/python python /usr/bin/python3 1
+RUN apt-get install -y python3-pip
+RUN update-alternatives --install /usr/bin/pip pip /usr/bin/pip3 1
+RUN pip install --upgrade pip
+
+# Install onnxruntime-training dependencies
+# This version doesn't support running gpt2 with fp16
+RUN pip install onnx==1.11.0 ninja
+RUN pip install torch==1.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
+RUN pip install onnxruntime-training==1.11.1+cu113 -f https://download.onnxruntime.ai/onnxruntime_stable_cu113.html
+RUN pip install torch-ort
+# RUN python -m torch_ort.configure
+# RUN pip install coloredlogs seqeval scipy transformers datasets sacrebleu
+
+WORKDIR .
+
+CMD ["/bin/bash"]

examples/onnxruntime/training/docker/Dockerfile-ort1.11.0-cu113 → examples/onnxruntime/training/docker/Dockerfile-ort1.12.0-cu113

@@ -3,7 +3,7 @@ FROM nvidia/cuda:11.3.0-cudnn8-devel-ubuntu20.04
 CMD nvidia-smi
 
 # Ignore interactive questions during `docker build`
-ENV DEBIAN_FRONTEND noninteractive 
+ENV DEBIAN_FRONTEND noninteractive
 
 # Bash shell
 RUN chsh -s /bin/bash
@@ -26,15 +26,15 @@ RUN update-alternatives --install /usr/bin/pip pip /usr/bin/pip3 1
 RUN pip install --upgrade pip
 
 # Install onnxruntime-training dependencies
-# Temporary workaround, wait for a more stable version on the next release of onnxruntime-training 
+# Temporary workaround, wait for a more stable version on the next release of onnxruntime-training
 # (unable to run gpt2 with fp16 for now)
-RUN pip install onnx==1.10.2 ninja
+RUN pip install onnx==1.11.0 ninja
 RUN pip install torch==1.11.0+cu113 -f https://download.pytorch.org/whl/torch_stable.html
-RUN pip install onnxruntime-training==1.11.0+cu113 -f https://download.onnxruntime.ai/onnxruntime_stable_cu113.html
+RUN pip install onnxruntime-training==1.12.0+cu113 -f https://download.onnxruntime.ai/onnxruntime_stable_cu113.html
 RUN pip install torch-ort
 # RUN python -m torch_ort.configure
 # RUN pip install coloredlogs seqeval scipy transformers datasets sacrebleu
 
 WORKDIR .
 
-CMD ["/bin/bash"]
+CMD ["/bin/bash"]

examples/onnxruntime/training/docker/run.sh

@@ -4,14 +4,24 @@ CMD=${1:-/bin/bash}
 GPU_DEVICES=${2:-"all"}
 
 # Build Image
-# docker build -f Dockerfile-cu10  -t ort/cu10 .
-# docker build -f Dockerfile-cu11  -t ort/cu11 .
+# docker build -f Dockerfile-ort1.9.0-cu102  -t ort9/cu10 .
+# docker build -f Dockerfile-ort1.9.0-cu111  -t ort9/cu11 .
 
 # Run Image
-# docker run -it --rm -p 80:8888 --gpus $GPU_DEVICES ort/cu10:latest $CMD
-# docker run -it --rm -p 80:8888 --gpus $GPU_DEVICES ort/cu11:latest $CMD
+# docker run -it --rm -p 80:8888 --gpus $GPU_DEVICES ort9/cu10:latest $CMD
+# docker run -it --rm -p 80:8888 --gpus $GPU_DEVICES ort9/cu11:latest $CMD
 
 # Install dependencies
-pip install coloredlogs transformers>=4.15.0 datasets>=1.8.0
-pip install scipy sklearn
+pip install transformers>=4.15.0 datasets>=1.8.0
+pip install coloredlogs absl-py rouge_score seqeval scipy sacrebleu transformers datasets nltk sklearn
+pip install deepspeed mpi4py
+
+# Install apex
+git clone https://github.com/NVIDIA/apex \
+    && cd apex \
+    && pip install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./
+
+# Install fairscale
+pip install fairscale
+
 python -m unittest tests/onnxruntime/nightly_test_onnxruntime.py

optimum/onnx/configuration.py

@@ -17,8 +17,9 @@
 from collections import OrderedDict
 from typing import TYPE_CHECKING, Any, Mapping, Optional, Union
 
+from transformers.file_utils import TensorType, is_tf_available, is_torch_available
 from transformers.onnx.utils import compute_effective_axis_dimension
-from transformers.utils import TensorType, is_tf_available, is_torch_available, logging
+from transformers.utils import logging
 
 
 if TYPE_CHECKING:

optimum/onnxruntime/quantization.py

@@ -294,7 +294,7 @@ def export(
                 The dictionary mapping the nodes name to their quantization ranges, used and required only when applying
                 static quantization.
             use_external_data_format (`bool`, defaults to `False`):
-                Whether uto se external data format to store model which size is >= 2Gb.
+                Whether to use external data format to store model which size is >= 2Gb.
             preprocessor (`QuantizationPreprocessor`, *optional*):
                 The preprocessor to use to collect the nodes to include or exclude from quantization.