Merge branch 'develop_v2.5_swap' into 'develop_v2.5'

add swap env

See merge request dcutoolkit/deeplearing/TransformerEngine!40

transformer_engine/pytorch/module/layernorm_linear.py

@@ -427,7 +427,7 @@ class _LayerNormLinear(torch.autograd.Function):
             )
             nvtx_range_pop(f"{nvtx_label}.fsdp_scatter")
 
-            if cpu_offloading:
+            if cpu_offloading or int(os.getenv("NVTE_SWAP_OVERLAP_GRAD", "0")):
                 ctx.grad_added_to_main_grad = hasattr(weight, "grad_added_to_main_grad")
 
                 if ctx.grad_added_to_main_grad:
@@ -556,7 +556,7 @@ class _LayerNormLinear(torch.autograd.Function):
 
             # For CPU offloading, we offloaded weight and weight.main_grad to different tensors,
             # we need to connect them into one.
-            if ctx.cpu_offloading:
+            if ctx.cpu_offloading or int(os.getenv("NVTE_SWAP_OVERLAP_GRAD", "0")):
                 if ctx.grad_added_to_main_grad:
                     origin_weight = ctx.weight_object
                 if ctx.requires_wgrad and ctx.fuse_wgrad_accumulation:

transformer_engine/pytorch/module/linear.py

@@ -368,7 +368,7 @@ class _Linear(torch.autograd.Function):
             )
             nvtx_range_pop(f"{nvtx_label}.fsdp_scatter")
 
-            if cpu_offloading:
+            if cpu_offloading or int(os.getenv("NVTE_SWAP_OVERLAP_GRAD", "0")):
                 ctx.grad_added_to_main_grad = hasattr(weight, "grad_added_to_main_grad")
 
                 if ctx.grad_added_to_main_grad:
@@ -459,7 +459,7 @@ class _Linear(torch.autograd.Function):
                 else None
             )
 
-            if ctx.cpu_offloading:
+            if ctx.cpu_offloading or int(os.getenv("NVTE_SWAP_OVERLAP_GRAD", "0")):
                 if ctx.grad_added_to_main_grad:
                     weight = ctx.weight_object
                 if ctx.requires_wgrad and ctx.fuse_wgrad_accumulation: