【Operator Mechanism】Pr68945 _C_ops.c_concat in dynamic graph bug fix (#…

…70870)

paddle/fluid/framework/new_executor/pir_interpreter.cc

@@ -593,6 +593,7 @@ void PirInterpreter::UpdateNcclOpNum() {
       "pd_op.barrier_grad",
       "pd_op.alltoall_grad",
       "pd_op.global_gather_grad",
+      "pd_op.c_concat_grad",
       "pd_op.distributed_fused_lamb_grad",
       "pd_op.margin_cross_entropy_grad",
       "pd_op.sync_batch_norm_grad",

paddle/phi/api/generator/api_gen.py

@@ -455,7 +455,13 @@ def source_include(header_file_path):
 #include "paddle/phi/api/profiler/event_tracing.h"
 #include "paddle/phi/api/profiler/supplement_tracing.h"
 
+#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
+#include "paddle/phi/core/distributed/comm_context_manager.h"
+#include "paddle/phi/core/distributed/nccl_comm_context.h"
+#endif
+
 #ifdef PADDLE_WITH_DISTRIBUTE
+#include "paddle/phi/core/distributed/store/store_utils.h"
 #include "paddle/phi/infermeta/spmd_rules/rules.h"
 #include "paddle/phi/core/distributed/auto_parallel/reshard/reshard_utils.h"
 #endif

paddle/phi/api/generator/dist_api_gen.py

@@ -82,6 +82,46 @@
   }}
 """
 
+NCCL_COMMCONTEXT_INIT = """
+#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
+  const auto & comm_context_manager_ = phi::distributed::CommContextManager::GetInstance();
+  if (nranks > 1 && !comm_context_manager_.Has(std::to_string(ring_id))) {{
+    auto store = phi::distributed::CreateOrGetGlobalTCPStore();
+    phi::distributed::CommContextManager::CreateNCCLCommContext(
+            store, std::to_string(ring_id), rank, nranks);
+  }}
+#endif
+"""
+
+SET_NCCL_COMMCONTEXT = """
+#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
+  const auto & comm_context_manager = phi::distributed::CommContextManager::GetInstance();
+  phi::distributed::NCCLCommContext* comm_context = nullptr;
+  if (comm_context_manager.Has(std::to_string(ring_id))) {{
+    comm_context = static_cast<phi::distributed::NCCLCommContext *>(
+          comm_context_manager.Get(std::to_string(ring_id)));
+    PADDLE_ENFORCE_NE(
+        comm_context,
+        nullptr,
+        common::errors::Unavailable(
+            "NCCLCommContext is nullptr, collective op should "
+            "has ring_id(%d) attr.",
+            std::to_string(ring_id)));
+    if (!comm_context->GetDevContext() || !comm_context->GetDevContext()->GetCommContext())
+    {{
+        auto kernel_res = phi::KernelFactory::Instance().SelectKernelOrThrowError(
+            "{}", {{kernel_backend, kernel_layout, kernel_data_type}}, true);
+        if (FLAGS_low_precision_op_list) {{
+        phi::KernelFactory::Instance().AddToLowPrecisionKernelList("{}", kernel_data_type);
+        }}
+        Backend act_kernel_backend = kernel_res.has_fallback_cpu ? Backend::CPU : kernel_backend;
+        auto* dev_context = GetDeviceContextByBackend(act_kernel_backend);
+        dev_context->SetCommContext(comm_context);
+    }}
+  }}
+#endif
+"""
+
 # 1. InferSPMD
 SINGLE_DIST_META_IN_TEMPLATE = """
     auto meta_dist_input_{name} = MakeDistMetaTensor(*{name}.impl());"""
@@ -861,6 +901,24 @@ def process_data_type_args(args_item):
             input_args=input_args, mesh=mesh, kernel_code=kernel_select_code
         )
 
+        # Current initialization only consider the case where the parameters of op contain ring_id, nranks and rank.
+        # Other cases will be addressed in the future.
+        if 'ring_id' in self.attrs['names']:
+            if (
+                'rank' in self.attrs['names']
+                and 'nranks' in self.attrs['names']
+            ):
+                if_condition_code = (
+                    if_condition_code
+                    + '\n'
+                    + self.generate_nccl_commcontext_init_code()
+                )
+            if_condition_code = (
+                if_condition_code
+                + '\n'
+                + self.generate_set_nccl_commcontext_code()
+            )
+
         return kernel_key_item_init + if_condition_code
 
     def generate_specialized_infer_spmd_code(self) -> str:
@@ -1322,6 +1380,12 @@ def generate_kernel_selection_code(self) -> str:
             self.api, self.kernel['func'][0], self.kernel['func'][0]
         )
 
+    def generate_nccl_commcontext_init_code(self) -> str:
+        return NCCL_COMMCONTEXT_INIT.format(self.kernel['func'][0])
+
+    def generate_set_nccl_commcontext_code(self) -> str:
+        return SET_NCCL_COMMCONTEXT.format(self.kernel['func'][0], self.api)
+
     def generate_reshard_input_code(self) -> str:
         input_reshard_code = ""
         if self.generate_infer_spmd is True:

paddle/phi/api/generator/dist_bw_api_gen.py

@@ -517,7 +517,13 @@ def source_include(header_file_path, fw_header_file_path):
 #include "paddle/phi/api/profiler/event_tracing.h"
 #include "paddle/phi/api/profiler/supplement_tracing.h"
 
+#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
+#include "paddle/phi/core/distributed/comm_context_manager.h"
+#include "paddle/phi/core/distributed/nccl_comm_context.h"
+#endif
+
 #ifdef PADDLE_WITH_DISTRIBUTE
+#include "paddle/phi/core/distributed/store/store_utils.h"
 #include "paddle/phi/infermeta/spmd_rules/rules.h"
 #include "paddle/phi/core/distributed/auto_parallel/reshard/reshard_utils.h"
 #endif

paddle/phi/ops/yaml/backward.yaml

@@ -321,6 +321,12 @@
     data_type : out_grad
   no_need_buffer : input
 
+- backward_op : c_concat_grad
+  forward : c_concat (Tensor x, int rank, int nranks, int ring_id, bool use_calc_stream, bool use_model_parallel) -> Tensor(out)
+  args : (Tensor out_grad, int rank = 0, int nranks = 1, int ring_id = 0, bool use_model_parallel = true)
+  output : Tensor(x_grad)
+  invoke: c_split(out_grad, rank, nranks, ring_id, use_model_parallel)
+
 - backward_op : cast_grad
   forward : cast (Tensor x, DataType dtype) -> Tensor(out)
   args : (Tensor x, Tensor out_grad)

paddle/phi/ops/yaml/op_compat.yaml

@@ -521,6 +521,7 @@
   drop_empty_grad : [input_grad]
 
 - op : c_concat
+  backward: c_concat_grad
   inputs :
     x : X
   outputs :

paddle/phi/ops/yaml/ops.yaml

@@ -762,7 +762,7 @@
     param : [x, nranks]
   kernel :
     func : c_concat
-  traits : paddle::dialect::ForwardOnlyTrait
+  backward: c_concat_grad
 
 - op : c_identity
   args : (Tensor x, int ring_id, bool use_calc_stream, bool use_model_parallel)

python/paddle/distributed/fleet/layers/mpu/mp_ops.py

@@ -25,6 +25,7 @@
     LayerHelper,
     _create_tensor,
     in_dynamic_mode,
+    in_dynamic_or_pir_mode,
     in_pir_mode,
 )
 from paddle.nn import Layer
@@ -139,21 +140,7 @@ def _c_concat(tensor, group=None):
     rank = group.rank
     nranks = group.nranks
 
-    if in_dynamic_mode():
-        return _legacy_C_ops.c_concat(
-            tensor,
-            'ring_id',
-            ring_id,
-            'use_calc_stream',
-            True,
-            'rank',
-            rank,
-            'nranks',
-            nranks,
-            'use_model_parallel',
-            True,
-        )
-    elif in_pir_mode():
+    if in_dynamic_or_pir_mode():
         return _C_ops.c_concat(tensor, rank, nranks, ring_id, True, True)
     else:
         op_type = 'c_concat'