[hotfix] quick fixes to make legacy tutorials runnable (#5559)

Co-authored-by: Edenzzzz <wtan45@wisc.edu>

colossalai/_analyzer/fx/tracer/tracer.py

@@ -237,7 +237,7 @@ class ColoTracer(Tracer):
         # override the tracer to support custom modules and checkpointing
         if self.trace_act_ckpt:
             orig_ckpt_func_apply = torch.utils.checkpoint.CheckpointFunction.apply
-            orig_ckpt_func_without_reentrant = torch.utils.checkpoint._checkpoint_without_reentrant
+            orig_ckpt_func_without_reentrant = torch.utils.checkpoint._checkpoint_without_reentrant_generator
 
             def checkpoint(run_function, preserve_rng_state=False, *args):
                 self.ckpt_regions.append(self.ckpt_idx)

colossalai/nn/layer/scaled_softmax.py

@@ -8,6 +8,14 @@ import torch.nn as nn
 
 from colossalai.kernel.kernel_loader import ScaledMaskedSoftmaxLoader, ScaledUpperTriangleMaskedSoftmaxLoader
 
+# NOTE: These kernels are compiled on specific GPU arch and not widely applicable.
+# try:
+#     from colossalai._C import scaled_masked_softmax as scaled_masked_softmax, scaled_upper_triangle_masked_softmax_cuda as scaled_upper_triang_masked_softmax
+# except ImportError:
+
+scaled_masked_softmax = None
+scaled_upper_triang_masked_softmax = None
+
 
 class AttnMaskType(enum.Enum):
     padding = 1

examples/tutorial/auto_parallel/auto_parallel_with_resnet.py

@@ -17,7 +17,7 @@ def synthesize_data():
 
 
 def main():
-    colossalai.launch_from_torch(config="./config.py")
+    colossalai.legacy.launch_from_torch(config="./config.py")
 
     logger = get_dist_logger()
 

examples/tutorial/hybrid_parallel/train.py

@@ -41,9 +41,9 @@ class DummyDataloader:
 
 def main():
     # launch from torch
-    parser = colossalai.get_default_parser()
+    parser = colossalai.legacy.get_default_parser()
     args = parser.parse_args()
-    colossalai.launch_from_torch(config=args.config)
+    colossalai.legacy.launch_from_torch(config=args.config)
 
     # get logger
     logger = get_dist_logger()

examples/tutorial/large_batch_optimizer/train.py

@@ -37,14 +37,14 @@ class DummyDataloader:
 
 def main():
     # initialize distributed setting
-    parser = colossalai.get_default_parser()
+    parser = colossalai.legacy.get_default_parser()
     parser.add_argument(
         "--optimizer", choices=["lars", "lamb"], help="Choose your large-batch optimizer", required=True
     )
     args = parser.parse_args()
 
     # launch from torch
-    colossalai.launch_from_torch(config=args.config)
+    colossalai.legacy.launch_from_torch(config=args.config)
 
     # get logger
     logger = get_dist_logger()
@@ -73,7 +73,7 @@ def main():
     )
 
     # initialize
-    engine, train_dataloader, test_dataloader, _ = colossalai.initialize(
+    engine, train_dataloader, test_dataloader, _ = colossalai.legacy.initialize(
         model=model,
         optimizer=optimizer,
         criterion=criterion,

examples/tutorial/opt/opt/run_clm.py

@@ -72,7 +72,7 @@ def get_time_stamp():
 
 
 def parse_args():
-    parser = colossalai.get_default_parser()
+    parser = colossalai.legacy.get_default_parser()
     parser.add_argument("-s", "--synthetic", action="store_true")
     parser.add_argument(
         "--dataset_name",
@@ -289,7 +289,7 @@ class DummyDataloader:
 def main():
     args = parse_args()
     disable_existing_loggers()
-    colossalai.launch_from_torch(config=dict())
+    colossalai.legacy.launch_from_torch(config=dict())
     logger = get_dist_logger()
     is_main_process = dist.get_rank() == 0
 

examples/tutorial/sequence_parallel/model/layers/bert_layer.py

@@ -1,9 +1,9 @@
 import torch
 import torch.nn as nn
 
-from colossalai.kernel.cuda_native import LayerNorm
 from colossalai.kernel.jit import bias_dropout_add_fused_inference, bias_dropout_add_fused_train
 from colossalai.legacy.nn.layer.parallel_sequence import TransformerSelfAttentionRing
+from colossalai.nn.layer.layernorm import MixedFusedLayerNorm as LayerNorm
 
 from .dropout import get_bias_dropout_add
 from .mlp import TransformerMLP

examples/tutorial/sequence_parallel/train.py

@@ -48,7 +48,7 @@ def pipeline_data_process_func(stage_output, micro_batch_data):
 def main():
     # initialize
     parse_args()
-    colossalai.launch_from_torch(config="./config.py", seed=1234, backend="nccl")
+    colossalai.legacy.launch_from_torch(config="./config.py", seed=1234, backend="nccl")
 
     logger = get_dist_logger()
 
@@ -136,7 +136,7 @@ def main():
     logger.info(f"LR Scheduler is built with {warmup_steps} warmup steps and {gpc.config.DECAY_ITERS} decay steps")
 
     # # init
-    engine, *dummy = colossalai.initialize(model, optimizer, criterion, verbose=True)
+    engine, *dummy = colossalai.legacy.initialize(model, optimizer, criterion, verbose=True)
 
     # build timer
     timer = MultiTimer()