[Test/CI] remove test cases to reduce CI duration (#5753)

* [test] smaller gpt2 test case

* [test] reduce test cases: tests/test_zero/test_gemini/test_zeroddp_state_dict.py

* [test] reduce test cases: tests/test_zero/test_gemini/test_grad_accum.py

* [test] reduce test cases tests/test_zero/test_gemini/test_optim.py

* Revert "[test] smaller gpt2 test case"

Some tests might depend on the size of model (num of chunks)

This reverts commit df705a5.

* [test] reduce test cases: tests/test_checkpoint_io/test_gemini_checkpoint_io.py

* [CI] smaller test model for two mwo the two modifid cases

* [CI] hardcode gpt model for tests/test_zero/test_gemini/test_search.py since we need a fixed answer there

tests/kit/model_zoo/transformers/gpt.py

@@ -18,23 +18,8 @@ def data_gen():
     # tokenized_input = tokenizer(input, return_tensors='pt')
     # input_ids = tokenized_input['input_ids']
     # attention_mask = tokenized_input['attention_mask']
-    # input_ids = torch.tensor([[15496, 11, 616, 3290, 318, 13779, 318, 13779]], dtype=torch.int64)
-    # attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1]], dtype=torch.int64)
-    input_ids = torch.tensor(
-        [
-            [15496, 11, 616, 3290, 318, 13779, 318, 13779, 15496, 11, 616, 3290, 318, 13779, 318, 13779],
-            [15496, 11, 616, 3290, 318, 13779, 318, 13779, 15496, 11, 616, 3290, 318, 13779, 318, 13779],
-        ],
-        dtype=torch.int64,
-    )
-    attention_mask = torch.tensor(
-        [
-            [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
-            [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
-        ],
-        dtype=torch.int64,
-    )
-
+    input_ids = torch.tensor([[22, 11, 616, 4, 5, 13, 318, 345]], dtype=torch.int64)
+    attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1]], dtype=torch.int64)
     return dict(input_ids=input_ids, attention_mask=attention_mask)
 
 
@@ -50,9 +35,9 @@ def data_gen_for_question_answering():
     # question answering data gen
     # `labels` is the type not the token id for token classification, 0 or 1
     data = data_gen()
-    start_positions = torch.tensor([[0], [0]], dtype=torch.int64)
+    start_positions = torch.tensor([0], dtype=torch.int64)
     data["start_positions"] = start_positions
-    end_positions = torch.tensor([[1], [1]], dtype=torch.int64)
+    end_positions = torch.tensor([1], dtype=torch.int64)
     data["end_positions"] = end_positions
     return data
 
@@ -61,39 +46,27 @@ def data_gen_for_token_classification():
     # token classification data gen
     # `labels` is the type not the token id for token classification, 0 or 1
     data = data_gen()
-    data["labels"] = torch.tensor(
-        [
-            [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1],
-            [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1],
-        ],
-        dtype=torch.int64,
-    )
+    data["labels"] = torch.tensor([[0, 0, 0, 0, 0, 0, 0, 1]], dtype=torch.int64)
     return data
 
 
 def data_gen_for_sequence_classification():
     # sequence classification data gen
     data = data_gen()
-    data["labels"] = torch.tensor([[1], [1]], dtype=torch.int64)
+    data["labels"] = torch.tensor([1], dtype=torch.int64)
     return data
 
 
 def date_gen_for_double_heads():
     num_choices = 2
     batch_size = 2
     input_ids = torch.tensor(
-        [
-            [15496, 11, 616, 3290, 318, 13779, 318, 13779, 15496, 11, 616, 3290, 318, 13779, 318, 13779],
-            [15496, 11, 616, 3290, 318, 13779, 318, 13779, 15496, 11, 616, 3290, 318, 13779, 318, 13779],
-        ],
-        dtype=torch.int64,
-    )
-    attention_mask = torch.tensor(
-        [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]],
+        [[46, 11, 616, 432, 318, 19, 318, 555], [777, 11, 235, 333, 318, 231, 468, 136]],
         dtype=torch.int64,
     )
-
+    attention_mask = torch.tensor([[1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1]], dtype=torch.int64)
     mc_labels = torch.zeros(input_ids.shape[0], dtype=torch.int64)
+
     mc_token_ids = torch.arange(0, num_choices, dtype=torch.int64)
     mc_token_ids = mc_token_ids.expand((batch_size, num_choices))
     multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, num_choices, -1).contiguous()
@@ -122,14 +95,14 @@ def date_gen_for_double_heads():
     n_layer=2,
     n_head=4,
     n_embd=128,
-    vocab_size=50258,
+    vocab_size=1024,
     attn_pdrop=0,
     embd_pdrop=0,
     resid_pdrop=0,
     summary_first_dropout=0,
     hidden_dropout=0,
     problem_type="single_label_classification",
-    pad_token_id=50256,
+    pad_token_id=1022,
     tie_word_embeddings=True,
 )
 

tests/test_checkpoint_io/test_gemini_checkpoint_io.py

@@ -21,14 +21,10 @@
 from tests.kit.model_zoo import model_zoo
 
 MODEL_PLACEMENT_CONFIGS = [
-    {"placement_policy": "static", "shard_param_frac": 0.0},  # zero2
-    {"placement_policy": "static", "shard_param_frac": 1.0},  # zero3
-    {"placement_policy": "static", "shard_param_frac": 0.5},  # zero3-half
+    {"placement_policy": "static", "shard_param_frac": 0.5},
 ]
 
 OPTIM_PLACEMENT_CONFIGS = [
-    {"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 0.0},  # zero2
-    {"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 1.0},  # zero2-offload
     {"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 0.5},  # zero2-offload-half
 ]
 

tests/test_zero/test_gemini/test_grad_accum.py

@@ -15,9 +15,7 @@
 from tests.kit.model_zoo import model_zoo, run_fwd
 
 PLACEMENT_CONFIGS = [
-    {"placement_policy": "static", "shard_param_frac": 0.0},  # zero2
-    {"placement_policy": "static", "shard_param_frac": 1.0},  # zero3
-    {"placement_policy": "static", "shard_param_frac": 0.5},  # zero3-half
+    {"placement_policy": "static", "shard_param_frac": 0.75},
     {"placement_policy": "auto"},
 ]
 
@@ -109,7 +107,7 @@ def exam_gemini_grad_acc(
     torch_model = DDP(torch_model, device_ids=[rank])
 
     set_seed(rank)
-    accum_iter = 4
+    accum_iter = 2
     train_dataloader = DummyDataloader(data_gen_fn)
     for i, data in enumerate(train_dataloader):
         delay_unscale = False if (i + 1) % accum_iter == 0 else True

tests/test_zero/test_gemini/test_optim.py

@@ -15,17 +15,7 @@
 from tests.kit.model_zoo import model_zoo, run_fwd_bwd
 
 PLACEMENT_CONFIGS = [
-    {"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 0.0},  # zero2
-    {"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 1.0},  # zero2-offload
-    {"placement_policy": "static", "shard_param_frac": 0.0, "offload_optim_frac": 0.5},  # zero2-offload-half
-    {"placement_policy": "static", "shard_param_frac": 1.0},  # zero3
-    {"placement_policy": "static", "shard_param_frac": 0.5},  # zero3-half
-    {
-        "placement_policy": "static",
-        "shard_param_frac": 1.0,
-        "offload_optim_frac": 1.0,
-        "offload_param_frac": 1.0,
-    },  # zero3-offload-all
+    {"placement_policy": "static", "shard_param_frac": 0.3, "offload_param_frac": 0.3, "offload_optim_frac": 0.3},
     {"placement_policy": "auto"},
 ]
 
@@ -73,7 +63,7 @@ def check_param(model: GeminiDDP, torch_model: torch.nn.Module, dtype: torch.dty
 @parameterize("model_name", TEST_MODELS)
 @parameterize("mixed_precision", [torch.half, torch.bfloat16])
 @parameterize("master_weights", [True, False])
-@parameterize("enable_async_reduce", [False, True])
+@parameterize("enable_async_reduce", [True])
 def exam_model_step(
     placement_config, model_name: str, mixed_precision: torch.dtype, master_weights: bool, enable_async_reduce=True
 ):
@@ -136,7 +126,7 @@ def exam_model_step(
             check_param(model, torch_model, mixed_precision)
 
 
-@parameterize("placement_config", [PLACEMENT_CONFIGS[3]])
+@parameterize("placement_config", [{"placement_policy": "static", "shard_param_frac": 1.0}])
 @parameterize("model_name", EXAMPLE_MODELS)
 @parameterize("mixed_precision", [torch.half])
 def exam_tiny_example(placement_config, model_name: str, mixed_precision: torch.dtype):
@@ -197,7 +187,7 @@ def run_dist(rank, world_size, port):
 
 
 @pytest.mark.dist
-@pytest.mark.parametrize("world_size", [1, 4])
+@pytest.mark.parametrize("world_size", [4])
 @rerun_if_address_is_in_use()
 def test_optim(world_size):
     spawn(run_dist, world_size)

tests/test_zero/test_gemini/test_search.py

@@ -1,18 +1,31 @@
 import pytest
 import torch
+import transformers
 
 import colossalai
 from colossalai.accelerator import get_accelerator
 from colossalai.testing import rerun_if_address_is_in_use, spawn
 from colossalai.zero.gemini.chunk import init_chunk_manager, search_chunk_configuration
-from tests.kit.model_zoo import model_zoo
 
+CONFIG = transformers.GPT2Config(
+    n_layer=2,
+    n_head=4,
+    n_embd=128,
+    vocab_size=50258,
+    attn_pdrop=0,
+    embd_pdrop=0,
+    resid_pdrop=0,
+    summary_first_dropout=0,
+    hidden_dropout=0,
+    problem_type="single_label_classification",
+    pad_token_id=50256,
+    tie_word_embeddings=True,
+)
+
+model_builder = lambda: transformers.GPT2LMHeadModel(CONFIG)
 
-def exam_search_chunk_size():
-    model_builder, data_gen_fn, output_transform_fn, *_ = next(
-        iter(model_zoo.get_sub_registry("transformers_gpt_lm").values())
-    )
 
+def exam_search_chunk_size():
     # make sure torch_model and model has the same parameter values
     model = model_builder()
     config_dict, *_ = search_chunk_configuration(
@@ -27,10 +40,6 @@ def exam_search_chunk_size():
 def exam_chunk_manager():
     world_size = torch.distributed.get_world_size()
 
-    model_builder, data_gen_fn, output_transform_fn, *_ = next(
-        iter(model_zoo.get_sub_registry("transformers_gpt_lm").values())
-    )
-
     sharded_ddp_model = model_builder()
     chunk_manager = init_chunk_manager(
         sharded_ddp_model,

tests/test_zero/test_gemini/test_zeroddp_state_dict.py

@@ -10,9 +10,7 @@
 from tests.kit.model_zoo import model_zoo
 
 PLACEMENT_CONFIGS = [
-    {"placement_policy": "static", "shard_param_frac": 0.0},  # zero2
-    {"placement_policy": "static", "shard_param_frac": 1.0},  # zero3
-    {"placement_policy": "static", "shard_param_frac": 0.5},  # zero3-half
+    {"placement_policy": "static", "shard_param_frac": 0.75},
     {"placement_policy": "auto"},
 ]
 
@@ -26,8 +24,8 @@ def ignore_the_first_parameter(model: torch.nn.Module):
 
 @parameterize("placement_config", PLACEMENT_CONFIGS)
 @parameterize("keep_gathered", [True, False])
-@parameterize("model_name", ["transformers_gpt_lm", "transformers_bert_for_sequence_classification"])
-@parameterize("master_weights", [False, True])
+@parameterize("model_name", ["transformers_gpt_lm"])
+@parameterize("master_weights", [True, False])
 def exam_state_dict(placement_config, keep_gathered, model_name: str, master_weights: bool):
     set_seed(431)
     model_builder, data_gen_fn, output_transform_fn, *_ = next(iter(model_zoo.get_sub_registry(model_name).values()))
@@ -81,7 +79,7 @@ def run_dist(rank, world_size, port):
 
 
 @pytest.mark.dist
-@pytest.mark.parametrize("world_size", [1, 4])
+@pytest.mark.parametrize("world_size", [4])
 @rerun_if_address_is_in_use()
 def test_zero_ddp(world_size):
     spawn(run_dist, world_size)