[ci] fix shardformer tests. (#5255)

* fix ci

fix

* revert: revert p2p

* feat: add enable_metadata_cache option

* revert: enable t5 tests

---------
Co-authored-by: Wenhao Chen <cwher@outlook.com>

colossalai/booster/plugin/hybrid_parallel_plugin.py

@@ -919,6 +919,7 @@ class HybridParallelPlugin(PipelinePluginBase):
         custom_policy (Policy, optional): Custom policy for Shardformer. Defaults to None.
         pp_style (str, optional): The style for pipeline parallelism. Defaults to '1f1b'.
         num_model_chunks (int, optional): The number of model chunks for interleaved pipeline parallelism. Defaults to 1.
+        enable_metadata_cache (bool, optional): Whether to enable metadata cache for pipeline parallelism. Defaults to True.
     """
 
     def __init__(
@@ -956,6 +957,7 @@ class HybridParallelPlugin(PipelinePluginBase):
         custom_policy: Policy = None,
         pp_style: str = "1f1b",
         num_model_chunks: int = 1,
+        enable_metadata_cache: bool = True,
     ) -> None:
         super().__init__()
         assert (
@@ -1002,10 +1004,14 @@ class HybridParallelPlugin(PipelinePluginBase):
                     num_model_chunks=num_model_chunks,
                     num_microbatch=num_microbatches,
                     microbatch_size=microbatch_size,
+                    enable_metadata_cache=enable_metadata_cache,
                 )
             elif pp_style == "1f1b":
                 self.schedule = OneForwardOneBackwardSchedule(
-                    self.stage_manager, num_microbatches=num_microbatches, microbatch_size=microbatch_size
+                    stage_manager=self.stage_manager,
+                    num_microbatches=num_microbatches,
+                    microbatch_size=microbatch_size,
+                    enable_metadata_cache=enable_metadata_cache,
                 )
             else:
                 raise NotImplementedError()

tests/test_shardformer/test_model/test_shard_gpt2.py

@@ -165,7 +165,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
 )
 @clear_cache_before_run()
 def run_gpt2_test(test_config):
-    sub_model_zoo = model_zoo.get_sub_registry("transformers_gpt")
+    sub_model_zoo = model_zoo.get_sub_registry("transformers_gpt", exclude="transformers_gptj")
 
     for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
         check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config)
@@ -200,7 +200,7 @@ def run_gpt2_test(test_config):
 )
 @clear_cache_before_run()
 def run_gpt2_3d_test(test_config):
-    sub_model_zoo = model_zoo.get_sub_registry("transformers_gpt")
+    sub_model_zoo = model_zoo.get_sub_registry("transformers_gpt", exclude="transformers_gptj")
 
     for name, (model_fn, data_gen_fn, output_transform_fn, loss_fn, _) in sub_model_zoo.items():
         check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn, test_config)

tests/test_shardformer/test_model/test_shard_t5.py

@@ -86,6 +86,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
             "tp_size": 2,
             "pp_size": 2,
             "num_microbatches": 2,
+            "enable_metadata_cache": False,
             "enable_all_optimization": True,
             "use_lazy_init": True,
             "precision": "fp16",
@@ -95,6 +96,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
             "tp_size": 1,
             "pp_size": 2,
             "num_microbatches": 4,
+            "enable_metadata_cache": False,
             "use_lazy_init": False,
             "precision": "fp16",
             "initial_scale": 1,
@@ -110,6 +112,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
             "tp_size": 1,
             "pp_size": 4,
             "num_microbatches": 4,
+            "enable_metadata_cache": False,
             "enable_all_optimization": False,
             "use_lazy_init": False,
             "precision": "fp32",
@@ -128,6 +131,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
             "tp_size": 1,
             "pp_size": 2,
             "num_microbatches": 2,
+            "enable_metadata_cache": False,
             "enable_all_optimization": True,
             "use_lazy_init": True,
             "zero_stage": 1,
@@ -159,6 +163,7 @@ def run_t5_test(test_config):
             "tp_size": 2,
             "pp_size": 2,
             "num_microbatches": 4,
+            "enable_metadata_cache": False,
             "enable_all_optimization": False,
             "use_lazy_init": False,
             "precision": "fp32",
@@ -168,6 +173,7 @@ def run_t5_test(test_config):
             "tp_size": 2,
             "pp_size": 2,
             "num_microbatches": 4,
+            "enable_metadata_cache": False,
             "enable_all_optimization": False,
             "use_lazy_init": False,
             "precision": "fp16",

tests/test_shardformer/test_model/test_shard_whisper.py

@@ -114,6 +114,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
             "tp_size": 2,
             "pp_size": 2,
             "num_microbatches": 2,
+            "enable_metadata_cache": False,
             "enable_all_optimization": True,
             "use_lazy_init": True,
             "precision": "fp32",
@@ -123,6 +124,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
             "tp_size": 1,
             "pp_size": 2,
             "num_microbatches": 4,
+            "enable_metadata_cache": False,
             "use_lazy_init": False,
             "precision": "fp32",
             "initial_scale": 1,
@@ -138,6 +140,7 @@ def check_forward_backward(model_fn, data_gen_fn, output_transform_fn, loss_fn,
             "tp_size": 1,
             "pp_size": 4,
             "num_microbatches": 4,
+            "enable_metadata_cache": False,
             "use_lazy_init": False,
             "precision": "fp32",
         },
@@ -163,6 +166,7 @@ def run_whisper_test(test_config):
             "tp_size": 2,
             "pp_size": 2,
             "num_microbatches": 4,
+            "enable_metadata_cache": False,
             "enable_all_optimization": False,
             "use_lazy_init": False,
             "precision": "fp32",
@@ -172,6 +176,7 @@ def run_whisper_test(test_config):
             "tp_size": 2,
             "pp_size": 2,
             "num_microbatches": 2,
+            "enable_metadata_cache": False,
             "enable_all_optimization": False,
             "use_lazy_init": False,
             "precision": "fp32",