Sequence-parallel amax reduction fix (#74)

* Fix no reduce_amax option for SP case
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* add warning about overriding reduce_amax
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

transformer_engine/pytorch/fp8.py

@@ -493,6 +493,7 @@ def reduce_tensor_across_group_op_max(
 
 def global_amax_reduction(
     fp8_meta: Dict[str, Any],
+    reduce_amax: bool = False,
     reduce_amax_across_tp_group: bool = False,
     tp_group: Optional[dist_group_type] = None,
     forward: bool = True,
@@ -508,6 +509,7 @@ def global_amax_reduction(
     chunk_sizes = [x.numel() for x in _global_fp8_buffer[amax_buffer_key]]
     contiguous_amax = torch.cat(_global_fp8_buffer[amax_buffer_key])
 
+    if reduce_amax:
         reduce_tensor_across_group_op_max(contiguous_amax, fp8_meta["fp8_group"])
     if reduce_amax_across_tp_group:
         reduce_tensor_across_group_op_max(contiguous_amax, tp_group)

transformer_engine/pytorch/module.py

@@ -109,7 +109,7 @@ def _prepare_backward(fp8: bool,
     """Checks and prep for BWD."""
     if fp8:
         # Update amax and scale; Skip all setup for global amax reduction
-        if not fp8_meta["recipe"].reduce_amax:
+        if not (fp8_meta["recipe"].reduce_amax or reduce_amax_across_tp_group):
             amax_and_scale_update(fp8_meta, False)
         else:
             # From previous iteration
@@ -125,12 +125,14 @@ def _prepare_backward(fp8: bool,
     with torch.cuda.nvtx.range(name + " backward"):
         yield
 
-    if not fp8 or not fp8_meta["recipe"].reduce_amax:
-        return
-
+    if fp8 and (fp8_meta["recipe"].reduce_amax or reduce_amax_across_tp_group):
         if fp8_meta["first_module"]:
             global_amax_reduction(
-            fp8_meta, reduce_amax_across_tp_group, tp_group, forward=False
+                fp8_meta,
+                fp8_meta["recipe"].reduce_amax,
+                reduce_amax_across_tp_group,
+                tp_group,
+                forward=False,
             )
             delete_key_from_amax_buffer(forward=False)
 
@@ -454,10 +456,16 @@ class TransformerEngineBaseModule(torch.nn.Module, ABC):
             self.set_fp8_weights()
 
             update_weight_scale_inv = is_first_microbatch is None or is_first_microbatch
+            reduce_amax = self.fp8_meta["recipe"].reduce_amax or self.sequence_parallel
+            if self.fp8 and self.sequence_parallel and not self.fp8_meta["recipe"].reduce_amax:
+                warnings.warn(
+                    "Amax reduction across tensor parallel group is necessary "
+                    "when using sequence parallelism with FP8."
+                )
 
             # Previous iteration was grad_enabled
             if self.fp8_meta.get("update_amax_and_scale_fwd", False):
-                if self.fp8_meta["recipe"].reduce_amax:
+                if reduce_amax:
                     copy_amax_from_global_buffer(self.fp8_meta, forward=True)
                     amax_and_scale_update(
                         self.fp8_meta, True, update_weight_scale_inv=update_weight_scale_inv
@@ -470,7 +478,7 @@ class TransformerEngineBaseModule(torch.nn.Module, ABC):
 
             if self.fp8 and self.training:
                 # Setup for amax reduction
-                if self.fp8_meta["recipe"].reduce_amax:
+                if reduce_amax:
                     self.fp8_meta["first_module"] = is_first_fp8_module()
                     if self.fp8_meta["first_module"]:
                         self.fp8_meta["autocast_id_fwd"] = new_fp8_context_id()
@@ -501,11 +509,12 @@ class TransformerEngineBaseModule(torch.nn.Module, ABC):
             restore_fp8_meta_tensors(self.fp8_meta)
             return
 
-        if self.fp8 and self.training and self.fp8_meta["recipe"].reduce_amax:
+        if self.fp8 and self.training and reduce_amax:
             set_fp8_context_id(self.fp8_meta["autocast_id_fwd"])
             reduce_func = partial(
                 global_amax_reduction,
                 self.fp8_meta,
+                self.fp8_meta["recipe"].reduce_amax,
                 self.sequence_parallel,
                 self.tp_group,
                 forward=True,