Fixed offloading for PyT version/ Added Attention activation offloading...

Fixed offloading for PyT version/ Added Attention activation offloading support/ Native FP8 support (#632)

* Fixed offloading for PyT version/ Added Attention activation offloading support/ Native FP8 support
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Removed activation offloading for fused attention
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Fixed the illegal memory access issue for activation offloading of attention
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Removed the version guard
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Pipeline failures fix
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Fixed lint erros
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

* Lint error fix
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

---------
Signed-off-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>
Co-authored-by: Selvaraj Anandaraj <selvaraja@login-eos01.eos.clusters.nvidia.com>

transformer_engine/pytorch/attention.py

@@ -1752,6 +1752,14 @@ class FlashAttention(torch.nn.Module):
                     deterministic=self.deterministic
                 )
         else:
+
+            from .cpu_offload import CPUOffloadEnabled
+            if CPUOffloadEnabled:
+                tensor_list = [query_layer, key_layer, value_layer, cu_seqlens_q, cu_seqlens_kv]
+                for tensor in tensor_list:
+                    if tensor is not None:
+                        tensor.activation_offloading = True
+
             with self.attention_dropout_ctx():
                 fa_optional_forward_kwargs = {}
                 if _flash_attn_2_3_plus:
@@ -1938,6 +1946,15 @@ class FusedAttnFunc(torch.autograd.Function):
             attn_scale, dropout_p, fast_zero_fill, qkv_layout, attn_bias_type, attn_mask_type,
             rng_gen)
 
+        from .cpu_offload import CPUOffloadEnabled
+        if CPUOffloadEnabled:
+            tensor_list = [q, k, v, out, cu_seqlens_q, cu_seqlens_kv]
+            qkv_layout = 'sbhd_sbhd_sbhd'
+            for tensor in tensor_list:
+                if tensor is not None:
+                    tensor.activation_offloading = True
+
+
         ctx.save_for_backward(q, k, v, out, cu_seqlens_q, cu_seqlens_kv)
         ctx.aux_ctx_tensors = aux_ctx_tensors
         ctx.max_seqlen_q = max_seqlen_q
@@ -2818,6 +2835,13 @@ class DotProductAttention(torch.nn.Module):
         assert (not context_parallel), \
             "Context parallelism is only implemented with Flash Attention and Fused Attention!"
 
+        from .cpu_offload import CPUOffloadEnabled
+        if CPUOffloadEnabled:
+            warnings.warn(
+                           "Attention activation Offloading is only implemented"
+                           "with Flash Attention and Fused Attention!"
+                         )
+
         if _NVTE_DEBUG:
             print("[DotProductAttention]: using unfused DPA")
         if use_unfused_attention:

transformer_engine/pytorch/cpu_offload.py

@@ -184,6 +184,7 @@ class SynchronizedGroupOffloadHandler(OffloadHandler):
         # the tensor back to gpu and deletes the cpu tensor.
         # These will increment whenever `group_commit()` is invoked
         self.current_group, self.tensor_count_current_group = (0, 0)
+        self.torch_tensor_count = 0
         self.tensor_tag_to_state = {}
 
     def on_group_commit_forward(self):
@@ -310,6 +311,11 @@ class AsyncDoubleBufferGroupOffloadHandler(SynchronizedGroupOffloadHandler):
 
 
     def tensor_push(self, tensor: torch.Tensor, **kwargs) -> Any:
+
+        torch_stray_tensor = isinstance(tensor,(torch._subclasses.fake_tensor.FakeTensor,
+                                        torch._subclasses.functional_tensor.FunctionalTensor))
+
+        if not torch_stray_tensor:
             # obtain a unique tensor tag
             tensor_tag = (self.current_group, self.tensor_count_current_group)
             self.tensor_count_current_group += 1
@@ -317,7 +323,8 @@ class AsyncDoubleBufferGroupOffloadHandler(SynchronizedGroupOffloadHandler):
 
             if (self.current_group < self.num_offload_group
                 and self.tensor_need_offloading_checker(tensor)):
-            # first copy the tensor to tensorbuf, so that the original tensor will not be deleted
+                # first copy the tensor to tensorbuf,
+                # so that the original tensor will not be deleted
                 tensor_buf = self.get_tensor_buf_for_offloaded_tensor(tensor, tensor_tag)
                 tensor_buf.copy_(tensor)
                 if hasattr(tensor,"weight_offloading"):
@@ -328,6 +335,11 @@ class AsyncDoubleBufferGroupOffloadHandler(SynchronizedGroupOffloadHandler):
                 self.tensor_tag_to_state[tensor_tag] = tensor_buf
             else:
                 self.tensor_tag_to_state[tensor_tag] = tensor
+        else:
+            tensor_tag = (-1,self.torch_tensor_count)
+            self.torch_tensor_count += 1
+            self.tensor_tag_to_state[tensor_tag] = tensor
+
         return tensor_tag
 
     def tensor_pop(self, tensor_tag, **kwargs):
@@ -350,6 +362,10 @@ class AsyncDoubleBufferGroupOffloadHandler(SynchronizedGroupOffloadHandler):
 
                     # if offload, return the reference to cpu copy
                     if self.tensor_need_offloading_checker(tensor_on_device):
+                        if hasattr(tensor_on_device,"weight_offloading"):
+                            delattr(tensor_on_device,"weight_offloading")
+                        if hasattr(tensor_on_device,"activation_offloading"):
+                            delattr(tensor_on_device,"activation_offloading")
                         state = SynchronizedGroupOffloadHandler.offload(tensor_on_device)
                         self.tensor_tag_to_state[tensor_tag] = state
 

transformer_engine/pytorch/module/layernorm_linear.py

@@ -242,7 +242,7 @@ class _LayerNormLinear(torch.autograd.Function):
             if cpu_offloading:
                 if fuse_wgrad_accumulation:
                     weight.main_grad.weight_offloading = True
-                if fp8:
+                if fp8 and weight_t_fp8 is not None:
                     weight_t_fp8.weight_offloading = True
                 ln_weight.weight_offloading = True
                 weight.weight_offloading = True

transformer_engine/pytorch/module/layernorm_mlp.py

@@ -424,8 +424,9 @@ class _LayerNormMLP(torch.autograd.Function):
                 if fuse_wgrad_accumulation:
                     fc1_weight.main_grad.weight_offloading = True
                     fc2_weight.main_grad.weight_offloading = True
-                if fp8:
+                if fp8 and fc1_weight_t_fp8 is not None:
                     fc1_weight_t_fp8.weight_offloading = True
+                if fp8 and fc2_weight_t_fp8 is not None:
                     fc2_weight_t_fp8.weight_offloading = True
                 ln_weight.weight_offloading = True
                 fc1_weight.weight_offloading = True

transformer_engine/pytorch/module/linear.py

@@ -275,7 +275,7 @@ class _Linear(torch.autograd.Function):
                 if cpu_offloading:
                     if fuse_wgrad_accumulation:
                         weight.main_grad.weight_offloading = True
-                    if fp8:
+                    if fp8 and weight_t_fp8 is not None:
                         weight_t_fp8.weight_offloading = True
                     weight.weight_offloading = True