[CI] LLM tests integration

ghstack-source-id: 5f47a70
Pull-Request: #3216

Author: vmoens

.github/unittest/linux_libs/scripts_llm/environment.yml renamed to .github/unittest/llm/scripts_llm/environment.yml

@@ -23,14 +23,4 @@ lib_dir="${env_dir}/lib"
 
 conda deactivate && conda activate ./env
 
-python -c "import transformers, datasets"
-
-pytest test/test_rlhf.py --instafail -v --durations 200 --capture no --error-for-skips
-
-python examples/rlhf/train_rlhf.py \
-  sys.device=cuda:0 sys.ref_device=cuda:0 \
-  model.name_or_path=gpt2 train.max_epochs=2 \
-  data.batch_size=2 train.ppo.ppo_batch_size=2 \
-  train.ppo.ppo_num_epochs=1 reward_model.name_or_path= \
-  train.ppo.episode_length=8 train.ppo.num_rollouts_per_epoch=4 \
-  data.block_size=110 io.logger=csv
+pytest test/llm -vvv --instafail --durations 600 --capture no --error-for-skips

.github/unittest/linux_libs/scripts_llm/install.sh renamed to .github/unittest/llm/scripts_llm/install.sh

@@ -21,14 +21,15 @@ permissions:
 
 jobs:
   unittests:
+    if: ${{ github.event_name == 'push' || contains(github.event.pull_request.labels.*.name, 'llm/') }}
     strategy:
       matrix:
         python_version: ["3.9"]
         cuda_arch_version: ["12.8"]
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     with:
       repository: pytorch/rl
-      runner: "linux.g5.4xlarge.nvidia.gpu"
+      runner: "linux.g6.4xlarge.experimental.nvidia.gpu"
       # gpu-arch-type: cuda
       # gpu-arch-version: "11.7"
       docker-image: "nvidia/cudagl:11.4.0-base"