[bug] fix early return (#5740)

* [bug] fix silly bug

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* [chore] add test for prefetch

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

tests/test_zero/test_gemini/test_fwd_bwd.py

@@ -40,12 +40,14 @@ def check_grad(model: GeminiDDP, torch_model: torch.nn.Module):
 @parameterize("model_name", ["transformers_gpt_lm"])
 @parameterize("use_grad_checkpoint", [False, True])
 @parameterize("master_weights", [False, True])
+@parameterize("max_prefetch", [0, 1, 4])
 def exam_gpt_fwd_bwd(
     placement_config,
     keep_gather,
     model_name: str,
     use_grad_checkpoint: bool = False,
     master_weights: bool = True,
+    max_prefetch: int = 0,
 ):
     init_device = get_accelerator().get_current_device()
     model_builder, data_gen_fn, output_transform_fn, loss_fn, *_ = next(
@@ -69,7 +71,13 @@ def exam_gpt_fwd_bwd(
     config_dict[world_size]["chunk_size"] = 5000
     config_dict[world_size]["keep_gathered"] = keep_gather
     model = GeminiDDP(
-        model, config_dict, init_device, pin_memory=True, **placement_config, master_weights=master_weights
+        model,
+        config_dict,
+        init_device,
+        pin_memory=True,
+        **placement_config,
+        master_weights=master_weights,
+        max_prefetch=max_prefetch,
     )
     optimizer = HybridAdam(model.parameters(), lr=1e-3)
     zero_optim = GeminiOptimizer(optimizer, model, initial_scale=1)

tests/test_zero/test_gemini/test_grad_accum.py

@@ -50,8 +50,14 @@ def check_grad(model: GeminiDDP, torch_model: torch.nn.Module):
 @parameterize("model_name", ["transformers_gpt_lm"])
 @parameterize("master_weights", [False, True])
 @parameterize("use_grad_checkpoint", [False, True])
+@parameterize("max_prefetch", [0, 1, 4])
 def exam_gemini_grad_acc(
-    placement_config, keep_gathered: bool, model_name: str, master_weights: bool, use_grad_checkpoint: bool
+    placement_config,
+    keep_gathered: bool,
+    model_name: str,
+    master_weights: bool,
+    use_grad_checkpoint: bool,
+    max_prefetch: int,
 ):
     init_device = get_accelerator().get_current_device()
     model_builder, data_gen_fn, output_transform_fn, loss_fn, *_ = next(
@@ -81,6 +87,7 @@ def exam_gemini_grad_acc(
         pin_memory=True,
         enable_gradient_accumulation=True,
         master_weights=master_weights,
+        max_prefetch=max_prefetch,
         **placement_config,
     )
     optimizer = HybridAdam(gemini_model.parameters(), lr=1e-3)

tests/test_zero/test_gemini/test_grad_clip.py

@@ -52,7 +52,8 @@ def check_param(model: GeminiDDP, torch_model: torch.nn.Module):
 @parameterize("placement_config", PLACEMENT_CONFIGS)
 @parameterize("model_name", ["transformers_gpt_lm"])
 @parameterize("master_weights", [True, False])
-def exam_grad_clipping(placement_config, model_name: str, master_weights: bool):
+@parameterize("max_prefetch", [0, 1, 4])
+def exam_grad_clipping(placement_config, model_name: str, master_weights: bool, max_prefetch: int):
     set_seed(1912)
     model_builder, data_gen_fn, output_transform_fn, loss_fn, *_ = next(
         iter(model_zoo.get_sub_registry(model_name).values())
@@ -84,6 +85,7 @@ def exam_grad_clipping(placement_config, model_name: str, master_weights: bool):
         chunk_init_device=init_device,
         pin_memory=True,
         master_weights=master_weights,
+        max_prefetch=max_prefetch,
         **placement_config,
     )
 

tests/test_zero/test_gemini/test_optim.py

@@ -71,7 +71,10 @@ 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])
-def exam_model_step(placement_config, model_name: str, mixed_precision: torch.dtype, master_weights: bool):
+@parameterize("max_prefetch", [0, 1, 4])
+def exam_model_step(
+    placement_config, model_name: str, mixed_precision: torch.dtype, master_weights: bool, max_prefetch: int
+):
     set_seed(42)
     model_builder, data_gen_fn, output_transform_fn, loss_fn, *_ = next(
         iter(model_zoo.get_sub_registry(model_name).values())
@@ -94,7 +97,12 @@ def exam_model_step(placement_config, model_name: str, mixed_precision: torch.dt
     config_dict[world_size]["chunk_size"] = 5000
     config_dict[world_size]["keep_gathered"] = False
     model = GeminiDDP(
-        model, config_dict, **placement_config, mixed_precision=mixed_precision, master_weights=master_weights
+        model,
+        config_dict,
+        **placement_config,
+        mixed_precision=mixed_precision,
+        master_weights=master_weights,
+        max_prefetch=max_prefetch,
     )
 
     optimizer = HybridAdam(model.parameters(), lr=1e-3)

tests/test_zero/test_gemini/test_zeroddp_state_dict.py

@@ -28,7 +28,8 @@ def ignore_the_first_parameter(model: torch.nn.Module):
 @parameterize("keep_gathered", [True, False])
 @parameterize("model_name", ["transformers_gpt_lm", "transformers_bert_for_sequence_classification"])
 @parameterize("master_weights", [False, True])
-def exam_state_dict(placement_config, keep_gathered, model_name: str, master_weights: bool):
+@parameterize("max_prefetch", [0, 1, 4])
+def exam_state_dict(placement_config, keep_gathered, model_name: str, master_weights: bool, max_prefetch: int):
     set_seed(431)
     model_builder, data_gen_fn, output_transform_fn, *_ = next(iter(model_zoo.get_sub_registry(model_name).values()))
 
@@ -44,7 +45,14 @@ def exam_state_dict(placement_config, keep_gathered, model_name: str, master_wei
     config_dict, *_ = search_chunk_configuration(model, search_range_m=1, search_interval=100)
     config_dict[world_size]["chunk_size"] = 5000
     config_dict[world_size]["keep_gathered"] = keep_gathered
-    model = GeminiDDP(model, config_dict, **placement_config, pin_memory=True, master_weights=master_weights)
+    model = GeminiDDP(
+        model,
+        config_dict,
+        **placement_config,
+        pin_memory=True,
+        master_weights=master_weights,
+        max_prefetch=max_prefetch,
+    )
     model.train()
 
     zero_dict = model.state_dict(only_rank_0=False)

tests/test_zero/test_gemini/test_zerooptim_state_dict.py

@@ -20,7 +20,8 @@ PLACEMENT_CONFIGS = [
 
 @parameterize("placement_config", PLACEMENT_CONFIGS)
 @parameterize("keep_gathered", [True, False])
-def exam_zero_optim_state_dict(placement_config, keep_gathered):
+@parameterize("max_prefetch", [0, 1, 4])
+def exam_zero_optim_state_dict(placement_config, keep_gathered, max_prefetch):
     set_seed(431)
     model_builder, data_gen_fn, output_transform_fn, *_ = next(
         iter(model_zoo.get_sub_registry("transformers_gpt_lm").values())
@@ -35,7 +36,7 @@ def exam_zero_optim_state_dict(placement_config, keep_gathered):
     config_dict[world_size]["chunk_size"] = 5000
     config_dict[world_size]["keep_gathered"] = keep_gathered
 
-    model = GeminiDDP(model, config_dict, **placement_config, pin_memory=True)
+    model = GeminiDDP(model, config_dict, **placement_config, pin_memory=True, max_prefetch=max_prefetch)
 
     optimizer = HybridAdam(model.parameters())
     optim = GeminiOptimizer(optimizer, model, initial_scale=32)  # initialize the link between chunk16 and chunk32