Fix memory overhead of linear layer when all gather from sequence parallel (NVIDIA#2125)

* fix memory overhead of all gather from sequence parallel

Signed-off-by: Yuzhong Wang <[email protected]>

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

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

* Update transformer_engine/pytorch/tensor/_internal/float8_blockwise_tensor_base.py

Signed-off-by: Tim Moon <[email protected]>

* quick fix the errors when for UB buffers

Signed-off-by: Yuzhong Wang <[email protected]>

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

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

* Update transformer_engine/pytorch/module/linear.py

Signed-off-by: Tim Moon <[email protected]>

* Avoid deallocating FP8 scale-invs since they are reused

Signed-off-by: Tim Moon <[email protected]>

---------

Signed-off-by: Yuzhong Wang <[email protected]>
Signed-off-by: Tim Moon <[email protected]>
Signed-off-by: Tim Moon <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Moon <[email protected]>
Co-authored-by: Tim Moon <[email protected]>

Author: 3 people

transformer_engine/pytorch/module/layernorm_linear.py

@@ -353,8 +353,11 @@ def forward(
 
         # Deallocate GEMM input tensor if no longer needed
         if not weight.requires_grad and not return_layernorm_output:
-            ln_out = ln_out_total = None
             clear_tensor_data(ln_out, ln_out_total)
+            ln_out = ln_out_total = None
+        elif with_input_all_gather and not return_layernorm_output_gathered:
+            clear_tensor_data(ln_out_total)
+            ln_out_total = None
 
         # ------------------------------------------------------
         # Prepare output tensor
@@ -891,9 +894,19 @@ def wgrad_gemm(
                         grad_bias = grad_bias_
                     del grad_bias_
 
-                    # Deallocate input tensor if permitted
-                    if not ctx.return_layernorm_output:
+                    # Deallocate input tensors if permitted
+                    if not ctx.return_layernorm_output and not ctx.return_layernorm_output_gathered:
+                        # Input tensors have not been exposed externally
+                        clear_tensor_data(ln_out)
+                    elif ctx.ln_out_needs_gather and ctx.return_layernorm_output_gathered:
+                        # Non-gathered input has not been exposed externally
+                        clear_tensor_data(ln_out)
+                    if ctx.ln_out_needs_gather:
+                        # Gathered input is internal
                         clear_tensor_data(ln_out_total)
+                    if ctx.parallel_mode == "row" and ctx.sequence_parallel:
+                        # Gathered grad output tensor is internal
+                        clear_tensor_data(grad_output)
 
                 # Update grad input if overlapping reduce-scatter with wgrad GEMM
                 if ctx.ub_bulk_wgrad:
@@ -1169,7 +1182,9 @@ def __init__(
         self.return_bias = return_bias
         self.apply_bias = self.use_bias and not return_bias
         self.return_layernorm_output = return_layernorm_output
-        self.return_layernorm_output_gathered = return_layernorm_output_gathered
+        self.return_layernorm_output_gathered = (
+            return_layernorm_output_gathered if return_layernorm_output else False
+        )
         self.zero_centered_gamma = zero_centered_gamma
         self.symmetric_ar_type = symmetric_ar_type
 

transformer_engine/pytorch/module/linear.py

@@ -317,6 +317,13 @@ def forward(
         # Finished forward GEMM...
         # ------------------------------------------------------
 
+        # Deallocate GEMM input tensor if no longer needed
+        # TODO(yuzhongw, tmoon): Figure out why inputmat_total is not automatically
+        # deallocated by GC. Manually deallocating is a temporary hack.
+        if with_input_all_gather_nccl:
+            clear_tensor_data(inputmat_total)
+            inputmat_total = None
+
         # ------------------------------------------------------
         # Prepare output tensor
         # Note: Perform tensor-parallel communication
@@ -878,9 +885,16 @@ def wgrad_gemm(
                         grad_bias = grad_bias_
                     del grad_bias_
 
-                    # Deallocate input tensor if permitted
+                    # Deallocate tensors if permitted
                     if ctx.owns_input:
+                        # Input tensor is internal
+                        clear_tensor_data(inputmat_total)
+                    elif ctx.backward_input_needs_gather:
+                        # Gathered input tensor is internal
                         clear_tensor_data(inputmat_total)
+                    if ctx.parallel_mode == "row" and ctx.sequence_parallel:
+                        # Gathered grad output tensor is internal
+                        clear_tensor_data(grad_output)
 
                 # Update grad input if overlapping reduce-scatter with wgrad GEMM
                 if ctx.ub_bulk_wgrad:

transformer_engine/pytorch/tensor/_internal/float8_blockwise_tensor_base.py

@@ -349,9 +349,14 @@ def _create_columnwise(self):
     def _transpose_columnwise_data(self):
         """Plainly transpose the columnwise data and scale inv."""
         if self._columnwise_data is not None:
+            # TODO(yuzhongw, tmoon): Figure out why _old_data is not automatically
+            # deallocated by GC. Manually deallocating is a temporary hack.
+            _old_data = self._columnwise_data
             self._columnwise_data = tex.fp8_transpose(
                 self._columnwise_data, self._fp8_dtype, out=None
             )
+            _old_data.data = _empty_tensor()
+            del _old_data
 
     def __repr__(self):
         if self._rowwise_data is not None:

transformer_engine/pytorch/tensor/_internal/float8_tensor_base.py

@@ -95,8 +95,13 @@ def __new__(
         return instance
 
     def clear(self):
-        """Deallocate this tensor's memory. Typically not needed and must be used carefully."""
-        for t in (self._data, self._transpose, self._scale_inv):
+        """Deallocate this tensor's memory. Typically not needed and must be used carefully.
+
+        Scale-inv tensor is not deallocated because it's often shared
+        between multiple FP8 tensors.
+
+        """
+        for t in (self._data, self._transpose):
             if t is not None:
                 t.data = _empty_tensor()
         self._transpose_invalid = True