[Feature] RLHF dataloading (pytorch#1309)

Co-authored-by: Alessandro Pietro Bardelli <apbard@users.noreply.github.com> Co-authored-by: Tom Begley <tomcbegley@gmail.com>
ai4co · Jun 27, 2023 · 2c47fa3 · 2c47fa3
1 parent 56518a7
commit 2c47fa3
Show file tree

Hide file tree

Showing 23 changed files with 1,875 additions and 5 deletions.
diff --git a/.circleci/unittest/linux/scripts/run_test.sh b/.circleci/unittest/linux/scripts/run_test.sh
@@ -21,6 +21,6 @@ export CKPT_BACKEND=torch
 
 python .circleci/unittest/helpers/coverage_run_parallel.py -m pytest test/smoke_test.py -v --durations 200
 python .circleci/unittest/helpers/coverage_run_parallel.py -m pytest test/smoke_test_deps.py -v --durations 200 -k 'test_gym or test_dm_control_pixels or test_dm_control or test_tb'
-python .circleci/unittest/helpers/coverage_run_parallel.py -m pytest --instafail -v --durations 200 --ignore test/test_distributed.py
+python .circleci/unittest/helpers/coverage_run_parallel.py -m pytest --instafail -v --durations 200 --ignore test/test_distributed.py test/test_rlhf.py
 coverage combine
 coverage xml -i
diff --git a/.circleci/unittest/linux_libs/scripts_rlhf/environment.yml b/.circleci/unittest/linux_libs/scripts_rlhf/environment.yml
@@ -0,0 +1,21 @@
+channels:
+  - pytorch
+  - defaults
+dependencies:
+  - pip
+  - pip:
+    - hypothesis
+    - future
+    - cloudpickle
+    - pytest
+    - pytest-cov
+    - pytest-mock
+    - pytest-instafail
+    - pytest-rerunfailures
+    - pytest-error-for-skips
+    - expecttest
+    - pyyaml
+    - scipy
+    - hydra-core
+    - transformers
+    - datasets
diff --git a/.circleci/unittest/linux_libs/scripts_rlhf/install.sh b/.circleci/unittest/linux_libs/scripts_rlhf/install.sh
@@ -0,0 +1,51 @@
+#!/usr/bin/env bash
+
+unset PYTORCH_VERSION
+# For unittest, nightly PyTorch is used as the following section,
+# so no need to set PYTORCH_VERSION.
+# In fact, keeping PYTORCH_VERSION forces us to hardcode PyTorch version in config.
+apt-get update && apt-get install -y git wget gcc g++
+#apt-get update && apt-get install -y git wget freeglut3 freeglut3-dev
+
+set -e
+
+eval "$(./conda/bin/conda shell.bash hook)"
+conda activate ./env
+
+if [ "${CU_VERSION:-}" == cpu ] ; then
+    version="cpu"
+else
+    if [[ ${#CU_VERSION} -eq 4 ]]; then
+        CUDA_VERSION="${CU_VERSION:2:1}.${CU_VERSION:3:1}"
+    elif [[ ${#CU_VERSION} -eq 5 ]]; then
+        CUDA_VERSION="${CU_VERSION:2:2}.${CU_VERSION:4:1}"
+    fi
+    echo "Using CUDA $CUDA_VERSION as determined by CU_VERSION ($CU_VERSION)"
+    version="$(python -c "print('.'.join(\"${CUDA_VERSION}\".split('.')[:2]))")"
+fi
+
+
+# submodules
+git submodule sync && git submodule update --init --recursive
+
+printf "Installing PyTorch with %s\n" "${CU_VERSION}"
+if [ "${CU_VERSION:-}" == cpu ] ; then
+    # conda install -y pytorch torchvision cpuonly -c pytorch-nightly
+    # use pip to install pytorch as conda can frequently pick older release
+#    conda install -y pytorch cpuonly -c pytorch-nightly
+    pip3 install --pre torch --extra-index-url https://download.pytorch.org/whl/nightly/cpu --force-reinstall
+else
+    pip3 install --pre torch --extra-index-url https://download.pytorch.org/whl/nightly/cu116 --force-reinstall
+fi
+
+# install tensordict
+pip install git+https://github.com/pytorch-labs/tensordict.git
+
+# smoke test
+python -c "import tensordict"
+
+printf "* Installing torchrl\n"
+pip3 install -e .
+
+# smoke test
+python -c "import torchrl"
diff --git a/.circleci/unittest/linux_libs/scripts_rlhf/post_process.sh b/.circleci/unittest/linux_libs/scripts_rlhf/post_process.sh
@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+
+set -e
+
+eval "$(./conda/bin/conda shell.bash hook)"
+conda activate ./env