Initialization fixes: allowing simple case like pytest pass, also making apex optional

Signed-off-by: Boris Fomitchev <bfomitchev@nvidia.com>

megatron/initialize.py

@@ -39,6 +39,12 @@ def initialize_megatron(extra_args_provider=None, args_defaults={},
         # Make sure cuda is available.
         assert torch.cuda.is_available(), 'Megatron requires CUDA.'
 
+    # This is temporary WAR to make simple case like pytest calling with same args twice
+    # Need to implement clean factory init.
+    if mpu.model_parallel_is_initialized():
+        return
+    
+    
     # Parse args, build tokenizer, and set adlr-autoresume,
     # tensorboard-writer, and timers.
     set_global_variables(extra_args_provider=extra_args_provider,

megatron/model/transformer.py

@@ -18,7 +18,13 @@
 import math
 
 import torch
-from apex.normalization.fused_layer_norm import FusedLayerNorm as LayerNorm
+try:
+    from apex.normalization.fused_layer_norm import FusedLayerNorm as LayerNorm
+    # Try to use FusedLayerNorm from Apex - this will trigger an error.
+    _ = FusedLayerNorm(8, eps=1e-5)
+
+except Exception as e:
+    from torch.nn import LayerNorm
 
 from megatron import get_args
 from megatron import mpu