[fluid_ops]clean c_allreduce_max config (#72034)

co63oc · web-flow · commit 2180d71eae92 · 2025-04-03T14:30:41.000+08:00
diff --git a/paddle/fluid/framework/new_executor/pir_interpreter.cc b/paddle/fluid/framework/new_executor/pir_interpreter.cc
@@ -536,7 +536,6 @@ void PirInterpreter::UpdateNcclOpNum() {
       "pd_op.c_softmax_with_multi_label_cross_entropy",
       "pd_op.c_allgather",
       "pd_op.c_allreduce_avg",
-      "pd_op.c_allreduce_max",
       "pd_op.c_allreduce_min",
       "pd_op.c_allreduce_sum",
       "pd_op.c_allreduce_prod",
@@ -568,7 +567,6 @@ void PirInterpreter::UpdateNcclOpNum() {
       "pd_op.c_softmax_with_cross_entropy_grad",
       "pd_op.c_softmax_with_multi_label_cross_entropy_grad",
       "pd_op.c_allgather_grad",
-      "pd_op.c_allreduce_max_grad",
       "pd_op.c_allreduce_min_grad",
       "pd_op.c_allreduce_sum_grad",
       "pd_op.c_allreduce_prod_grad",
@@ -602,7 +600,6 @@ void PirInterpreter::UpdateNcclOpNum() {
       "pd_op.c_softmax_with_multi_label_cross_entropy_",
       "pd_op.c_allgather_",
       "pd_op.c_allreduce_avg_",
-      "pd_op.c_allreduce_max_",
       "pd_op.c_allreduce_min_",
       "pd_op.c_allreduce_sum_",
       "pd_op.c_allreduce_prod_",
@@ -634,7 +631,6 @@ void PirInterpreter::UpdateNcclOpNum() {
       "pd_op.c_softmax_with_cross_entropy_grad_",
       "pd_op.c_softmax_with_multi_label_cross_entropy_grad_",
       "pd_op.c_allgather_grad_",
-      "pd_op.c_allreduce_max_grad_",
       "pd_op.c_allreduce_min_grad_",
       "pd_op.c_allreduce_sum_grad_",
       "pd_op.c_allreduce_prod_grad_",
diff --git a/paddle/fluid/inference/tensorrt/convert/c_allreduce_op.cc b/paddle/fluid/inference/tensorrt/convert/c_allreduce_op.cc
@@ -20,7 +20,6 @@ namespace paddle::inference::tensorrt {
 using ReduceType = paddle::inference::tensorrt::plugin::ReduceType;
 std::map<std::string, ReduceType> op_to_reduce_type = {
     {"c_allreduce_sum", paddle::inference::tensorrt::plugin::kRedSum},
-    {"c_allreduce_max", paddle::inference::tensorrt::plugin::kRedMax},
     {"c_allreduce_min", paddle::inference::tensorrt::plugin::kRedMin},
     {"c_allreduce_prod", paddle::inference::tensorrt::plugin::kRedProd}};
 
@@ -89,6 +88,5 @@ class CAllReduceOpConverter : public OpConverter {
 }  // namespace paddle::inference::tensorrt
 
 REGISTER_TRT_OP_CONVERTER(c_allreduce_sum, CAllReduceOpConverter);
-REGISTER_TRT_OP_CONVERTER(c_allreduce_max, CAllReduceOpConverter);
 REGISTER_TRT_OP_CONVERTER(c_allreduce_min, CAllReduceOpConverter);
 REGISTER_TRT_OP_CONVERTER(c_allreduce_prod, CAllReduceOpConverter);
diff --git a/paddle/fluid/inference/tensorrt/op_teller.cc b/paddle/fluid/inference/tensorrt/op_teller.cc
@@ -3193,7 +3193,6 @@ struct SimpleOpTypeSetTeller : public Teller {
       "fused_bias_dropout_residual_layer_norm",
       "c_allreduce_sum",
       "c_allreduce_min",
-      "c_allreduce_max",
       "c_allreduce_prod",
       "roll",
       "cast",
@@ -3370,7 +3369,6 @@ struct SimpleOpTypeSetTeller : public Teller {
       "fused_bias_dropout_residual_layer_norm",
       "c_allreduce_sum",
       "c_allreduce_min",
-      "c_allreduce_max",
       "c_allreduce_prod",
       "roll",
       "cast",
diff --git a/python/paddle/distributed/auto_parallel/static/mapper.py b/python/paddle/distributed/auto_parallel/static/mapper.py
@@ -29,7 +29,6 @@ def is_collective_comm_op(op):
     comm_list = [
         "c_allreduce_sum",
         "c_allreduce_min",
-        "c_allreduce_max",
         "c_allreduce_prod",
         "all_gather",
         "all_reduce",
diff --git a/python/paddle/distributed/auto_parallel/static/operators/common.py b/python/paddle/distributed/auto_parallel/static/operators/common.py
@@ -687,7 +687,8 @@ def is_data_parallel_reduce_op(op):
 
 def is_amp_flag_sync_op(op):
     return (
-        op.type == "c_allreduce_max"
+        op.type == "all_reduce"
+        and op.desc.attr("op_type") == paddle.distributed.ReduceOp.MAX
         and op.desc.has_attr("op_namescope")
         and SyncMode.AmpFlagSync in op.desc.attr("op_namescope")
     )
diff --git a/python/paddle/distributed/auto_parallel/static/operators/dist_check_finite_and_unscale.py b/python/paddle/distributed/auto_parallel/static/operators/dist_check_finite_and_unscale.py
@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License
 
+import paddle
 from paddle.distributed.auto_parallel.static.process_group import (
     get_world_process_group,
 )
@@ -157,12 +158,12 @@ def backward(ctx, *args, **kwargs):
             },
         )
         allreduce_op = main_block.append_op(
-            type='c_allreduce_max',
-            inputs={'X': inf_var_int32},
-            outputs={'Out': inf_var_int32},
+            type='all_reduce',
+            inputs={'x': inf_var_int32},
+            outputs={'out': inf_var_int32},
             attrs={
                 'ring_id': group.id,
-                'use_calc_stream': True,
+                'op_type': paddle.distributed.ReduceOp.MAX,
                 OP_ROLE_KEY: OpRole.Optimize,
             },
         )
diff --git a/python/paddle/distributed/fleet/meta_optimizers/sharding/fp16_helper.py b/python/paddle/distributed/fleet/meta_optimizers/sharding/fp16_helper.py
@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import paddle
 from paddle.distributed.fleet.meta_optimizers.common import (
     OP_ROLE_KEY,
     OpRole,
@@ -182,12 +183,12 @@ def prune_fp16(block, shard, reduced_grads_to_param, ring_ids):
             # this allreduce communication should not overlap with calc
             block._insert_op_without_sync(
                 update_loss_scaling_op_idx,
-                type='c_allreduce_max',
-                inputs={'X': inf_var_int32},
-                outputs={'Out': inf_var_int32},
+                type='all_reduce',
+                inputs={'x': inf_var_int32},
+                outputs={'out': inf_var_int32},
                 attrs={
                     'ring_id': ring_id,
-                    'use_calc_stream': True,
+                    'op_type': paddle.distributed.ReduceOp.MAX,
                     OP_ROLE_KEY: OpRole.Optimize,
                 },
             )
@@ -249,12 +250,12 @@ def sync_amp_check_nan_inf(block, ring_ids):
                 continue
             block._insert_op_without_sync(
                 update_loss_scaling_op_idx,
-                type='c_allreduce_max',
-                inputs={'X': inf_var_int32},
-                outputs={'Out': inf_var_int32},
+                type='all_reduce',
+                inputs={'x': inf_var_int32},
+                outputs={'out': inf_var_int32},
                 attrs={
                     'ring_id': ring_id,
-                    'use_calc_stream': True,
+                    'op_type': paddle.distributed.ReduceOp.MAX,
                     OP_ROLE_KEY: OpRole.Optimize,
                 },
             )
diff --git a/python/paddle/distributed/fleet/meta_optimizers/sharding/utils.py b/python/paddle/distributed/fleet/meta_optimizers/sharding/utils.py
@@ -149,7 +149,10 @@ def check_allreduce_sum(block, shard, sharding_ring_id, dp_ring_id=-1):
                 elif "@GRAD":
                     idx_last_grad_allreduce = idx
 
-        if op.type == "c_allreduce_max":
+        if (
+            op.type == "all_reduce"
+            and op.desc.attr("op_type") == paddle.distributed.ReduceOp.MAX
+        ):
             idx_gradient_clip_allreduce = idx
 
     for op in block.ops:
diff --git a/python/paddle/distributed/passes/fuse_all_reduce.py b/python/paddle/distributed/passes/fuse_all_reduce.py
@@ -150,7 +150,6 @@ def filter_all_collective_op_indices(block):
     all_collective_ops = {
         "c_allreduce_sum",
         "c_allreduce_prod",
-        "c_allreduce_max",
         "c_allreduce_min",
         "c_allgather",
         "c_broadcast",

Original file line number	Diff line number	Diff line change
`@@ -687,7 +687,8 @@ def is_data_parallel_reduce_op(op):`
`687`	`687`
`688`	`688`	`def is_amp_flag_sync_op(op):`
`689`	`689`	`return (`
`690`		`- op.type == "c_allreduce_max"`
	`690`	`+ op.type == "all_reduce"`
	`691`	`+ and op.desc.attr("op_type") == paddle.distributed.ReduceOp.MAX`
`691`	`692`	`and op.desc.has_attr("op_namescope")`
`692`	`693`	`and SyncMode.AmpFlagSync in op.desc.attr("op_namescope")`
`693`	`694`	`)`