[ci] fixed booster test (#5251)

* [ci] fixed booster test

* [ci] fixed booster test

* [ci] fixed booster test

.github/workflows/build_on_pr.yml

@@ -90,7 +90,7 @@ jobs:
     runs-on: [self-hosted, gpu]
     container:
       image: hpcaitech/pytorch-cuda:2.1.0-12.1.0
-      options: --gpus all --rm -v /data/scratch/cifar-10:/data/scratch/cifar-10 -v /data/scratch/llama-tiny:/data/scratch/llama-tiny
+      options: --gpus all --rm -v /dev/shm -v /data/scratch/llama-tiny:/data/scratch/llama-tiny
     timeout-minutes: 60
     defaults:
       run:
@@ -165,7 +165,6 @@ jobs:
           --ignore tests/test_checkpoint_io \
           tests/
         env:
-          NCCL_SHM_DISABLE: 1
           LD_LIBRARY_PATH: /github/home/.tensornvme/lib:/usr/local/nvidia/lib:/usr/local/nvidia/lib64
           LLAMA_PATH: /data/scratch/llama-tiny
 
@@ -205,4 +204,3 @@ jobs:
         with:
           name: report
           path: report/
-

.github/workflows/build_on_schedule.yml

@@ -13,15 +13,16 @@ jobs:
     runs-on: [self-hosted, gpu]
     container:
       image: hpcaitech/pytorch-cuda:2.0.0-11.7.0
-      options: --gpus all --rm -v /data/scratch/cifar-10:/data/scratch/cifar-10 -v /data/scratch/llama-tiny:/data/scratch/llama-tiny
+      options: --gpus all --rm -v /dev/shm -v /data/scratch/llama-tiny:/data/scratch/llama-tiny
     timeout-minutes: 90
     steps:
       - name: Check GPU Availability # ensure all GPUs have enough memory
         id: check-avai
         run: |
           avai=true
-          for i in $(seq 0 3);
+          ngpu=$(nvidia-smi --query-gpu=name --format=csv,noheader | wc -l)
-          do
+          endIndex=$(($ngpu-1))
+          for i in $(seq 0 $endIndex);
             gpu_used=$(nvidia-smi -i $i --query-gpu=memory.used --format=csv,noheader,nounits)
             [ "$gpu_used" -gt "2000" ] && avai=false
           done
@@ -74,7 +75,7 @@ jobs:
         if: ${{ failure() }}
         run: |
           url=$SERVER_URL/$REPO/actions/runs/$RUN_ID
-          msg="Scheduled Build and Test failed on 8 GPUs, please visit $url for details"
+          msg="Scheduled Build and Test failed, please visit $url for details"
           echo $msg
           python .github/workflows/scripts/send_message_to_lark.py -m "$msg" -u $WEBHOOK_URL
         env:

tests/kit/model_zoo/transformers/chatglm2.py

@@ -2,7 +2,6 @@ import torch
 
 from colossalai.shardformer.modeling.chatglm2_6b.configuration_chatglm import ChatGLMConfig
 from colossalai.shardformer.modeling.chatglm2_6b.modeling_chatglm import ChatGLMForConditionalGeneration, ChatGLMModel
-
 from ..registry import ModelAttribute, model_zoo
 
 # ================================

tests/test_booster/test_plugin/test_3d_plugin.py

@@ -10,10 +10,11 @@ from colossalai.booster.plugin import HybridParallelPlugin
 from colossalai.fx import is_compatible_with_meta
 from colossalai.lazy.lazy_init import LazyInitContext
 from colossalai.nn.optimizer import HybridAdam
-from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
+from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
 from tests.kit.model_zoo import model_zoo
 
 
+@clear_cache_before_run()
 def run_fn(init_method, model_fn, data_gen_fn, output_transform_fn) -> Optional[str]:
     try:
         if init_method == "lazy":
@@ -69,7 +70,6 @@ def check_3d_plugin(init_method: str = "none", early_stop: bool = True):
         "transformers_llama_for_casual_lm"
     ).items():
         err = run_fn(init_method, model_fn, data_gen_fn, output_transform_fn)
-        torch.cuda.empty_cache()
 
         if err is None:
             passed_models.append(name)

tests/test_booster/test_plugin/test_gemini_plugin.py

@@ -12,10 +12,11 @@ from colossalai.fx import is_compatible_with_meta
 from colossalai.lazy.lazy_init import LazyInitContext
 from colossalai.nn.optimizer import HybridAdam
 from colossalai.tensor.colo_parameter import ColoParameter
-from colossalai.testing import parameterize, rerun_if_address_is_in_use, spawn
+from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
-from tests.kit.model_zoo import model_zoo, COMMON_MODELS, IS_FAST_TEST
+from tests.kit.model_zoo import COMMON_MODELS, IS_FAST_TEST, model_zoo
 
 
+@clear_cache_before_run()
 def run_fn(init_method, model_fn, data_gen_fn, output_transform_fn, zero_size, tp_size) -> Optional[str]:
     try:
         if init_method == "lazy":
@@ -116,7 +117,7 @@ def check_gemini_plugin(
             "transformers_falcon_for_sequence_classification",
             "transformers_falcon_for_token_classification",
             "transformers_falcon_for_question_answering",
-            "transformers_gptj_lm", # lead to OOM when running in ci
+            "transformers_gptj_lm",  # lead to OOM when running in ci
             "transformers_gptj_for_question_answering",
             "transformers_gptj_for_sequence_classification",
         ]:
@@ -145,7 +146,6 @@ def check_gemini_plugin(
             tp_size = 1
 
         err = run_fn(init_method, model_fn, data_gen_fn, output_transform_fn, zero_size, tp_size)
-        torch.cuda.empty_cache()
         if err is None:
             passed_models.append(name)
         else: