Async amax reduction (#118)

* async amax reduction

add env knob to enable async amax reduction
Signed-off-by: slym <slym@login-preos01.a51.clusters.nvidia.com>

* Style fixes
Signed-off-by: Tim Moon <tmoon@nvidia.com>

* remove is_last_model
Signed-off-by: slym <slym@login-preos01.a51.clusters.nvidia.com>

* fix naming
Signed-off-by: slym <slym@login-preos01.a51.clusters.nvidia.com>

* revert var name
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* revert var name
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: slym <slym@login-preos01.a51.clusters.nvidia.com>
Signed-off-by: Tim Moon <tmoon@nvidia.com>
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: Tim Moon <tmoon@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>
Co-authored-by: slym <slym@login-preos01.a51.clusters.nvidia.com>

transformer_engine/pytorch/fp8.py

@@ -27,6 +27,7 @@ _fp8_tensors_recompute_buffer = []
 _amax_forward_global_reduce_func = None
 _buffer_delete_key_fwd = None
 _buffer_delete_key_bwd = None
+_amax_reduce_handle_fwd = None
 _is_fp8_available = None
 _reason_for_no_fp8 = ""
 
@@ -73,6 +74,12 @@ def get_autocast_key(forward: bool = True) -> str:
     return "autocast_id_bwd"
 
 
+def get_amax_reduce_handle_fwd() -> Union[bool, None]:
+    """Return AMAX reduction wait handle of forward prop."""
+    global _amax_reduce_handle_fwd
+    return _amax_reduce_handle_fwd
+
+
 def get_global_fp8_buffer() -> Dict[str, List[torch.Tensor]]:
     """Returns global fp8 buffer."""
     return _global_fp8_buffer
@@ -264,6 +271,7 @@ def fp8_autocast(
     global _FP8_ENABLED, _FP8_CALIBRATION,  _FP8_RECIPE, _FP8_DISTRIBUTED_GROUP, _FP8_AUTOCAST_DEPTH
     global _IS_FIRST_FP8_MODULE, _FP8_AUTOCAST_COUNTER
     global _global_fp8_buffer, _buffer_delete_key_fwd
+    global _amax_reduce_handle_fwd
     fp8_state = (_FP8_ENABLED, _FP8_CALIBRATION, _FP8_RECIPE, _FP8_DISTRIBUTED_GROUP)
     try:
         _FP8_ENABLED = enabled
@@ -287,7 +295,7 @@ def fp8_autocast(
 
         if _FP8_AUTOCAST_DEPTH == 0:
             if callable(_amax_forward_global_reduce_func):
-                _amax_forward_global_reduce_func()
+                _amax_reduce_handle_fwd = _amax_forward_global_reduce_func()
             delete_key_from_amax_buffer(forward=True)
 
 
@@ -521,16 +529,18 @@ def get_fp8_te_dtype(
 
 
 def reduce_tensor_across_group_op_max(
-    tensor: torch.Tensor, group: dist_group_type
+    tensor: torch.Tensor, group: dist_group_type, async_op: bool
 ) -> None:
     """Reduce tensor across given group."""
     if torch.distributed.is_initialized():
-        torch.distributed.all_reduce(
+        wait_handle = torch.distributed.all_reduce(
             tensor,
             op=torch.distributed.ReduceOp.MAX,
             group=group,
-            async_op=False,
+            async_op=async_op,
         )
+        return wait_handle
+    return None
 
 
 def global_amax_reduction(
@@ -543,14 +553,19 @@ def global_amax_reduction(
 
     # Key already deleted.
     if amax_buffer_key not in _global_fp8_buffer:
-        return
+        return None
 
     chunk_sizes = [x.numel() for x in _global_fp8_buffer[amax_buffer_key]]
     contiguous_amax = torch.cat(_global_fp8_buffer[amax_buffer_key])
 
-    reduce_tensor_across_group_op_max(contiguous_amax, fp8_meta["fp8_group"])
+    wait_handle = reduce_tensor_across_group_op_max(
+        contiguous_amax,
+        fp8_meta["fp8_group"],
+        fp8_meta["async_amax_reduction"],
+    )
 
     _global_fp8_buffer[amax_buffer_key] = list(contiguous_amax.split(chunk_sizes))
+    return wait_handle
 
 
 def delete_key_from_amax_buffer(forward: bool = True) -> None:

transformer_engine/pytorch/module.py

@@ -41,6 +41,7 @@ from .fp8 import (
     copy_forward_fp8_meta_tensors_for_recompute,
     get_old_fp8_meta_tensors_for_recompute,
     restore_fp8_meta_tensors,
+    get_amax_reduce_handle_fwd,
 )
 from .jit import (
     bias_gelu_fused,
@@ -84,6 +85,7 @@ _2X_ACC_FPROP = False
 _2X_ACC_DGRAD = True
 _2X_ACC_WGRAD = True
 _cublas_workspace = None
+_amax_reduce_handle_bwd = None
 
 
 def get_cublas_workspace_size_bytes() -> None:
@@ -106,6 +108,11 @@ def get_workspace() -> torch.Tensor:
 def _prepare_backward(fp8: bool, fp8_meta: Dict[str, Any],  name: str = "") -> None:
     """Checks and prep for BWD."""
     if fp8:
+        global _amax_reduce_handle_bwd
+        if _amax_reduce_handle_bwd is not None:
+            _amax_reduce_handle_bwd.wait()
+            _amax_reduce_handle_bwd = None
+
         # Update amax and scale; Skip all setup for global amax reduction
         if not fp8_meta["recipe"].reduce_amax:
             amax_and_scale_update(fp8_meta, False)
@@ -125,7 +132,7 @@ def _prepare_backward(fp8: bool, fp8_meta: Dict[str, Any],  name: str = "") -> N
 
     if fp8 and fp8_meta["recipe"].reduce_amax:
         if fp8_meta["first_module"]:
-            global_amax_reduction(fp8_meta, forward=False)
+            _amax_reduce_handle_bwd = global_amax_reduction(fp8_meta, forward=False)
             delete_key_from_amax_buffer(forward=False)
 
 
@@ -184,6 +191,9 @@ class TransformerEngineBaseModule(torch.nn.Module, ABC):
         self.sequence_parallel = False
         self.fp8_weight_shapes = []
         self.fp8_meta["autocast_id_fwd_stack"] = []
+        self.fp8_meta["async_amax_reduction"] = bool(
+            int(os.getenv("NVTE_ASYNC_AMAX_REDUCTION", "1"))
+        )
 
     def set_meta_tensor(self, fwd: bool) -> None:
         """Init scales and amaxes for fwd | bwd."""
@@ -497,6 +507,10 @@ class TransformerEngineBaseModule(torch.nn.Module, ABC):
                 if self.fp8_meta["recipe"].reduce_amax:
                     self.fp8_meta["first_module"] = is_first_fp8_module()
                     if self.fp8_meta["first_module"]:
+                        # Wait for the prior AMAX reduction to finish
+                        amax_reduce_handle_fwd = get_amax_reduce_handle_fwd()
+                        if amax_reduce_handle_fwd is not None:
+                            amax_reduce_handle_fwd.wait()
                         self.fp8_meta["autocast_id_fwd"] = new_fp8_context_id()
                         set_fp8_context_id(self.fp8_meta["autocast_id_fwd"])
                     else: