feat(opr-mm): add backend argument for remote send/recv

GitOrigin-RevId: 841a0e45ab

imperative/python/megengine/distributed/functional.py

@@ -265,6 +265,7 @@ def remote_send(inp: Tensor, dest_rank: int) -> Tensor:
     op.key = key
     op.addr, op.port = get_mm_server_addr()
     op.rank_to = dest_rank
+    op.backend = get_backend()
     (dummy,) = apply(_RemoteSend(op), inp)
 
     for g in grad_keys.values():
@@ -313,6 +314,7 @@ def remote_recv(
     op.dtype = dtype
     op.addr, op.port = get_mm_server_addr()
     op.rank_from = src_rank
+    op.backend = get_backend()
 
     (ret,) = apply(_RemoteRecv(op), inp)
     if _isscalar:

imperative/src/impl/ops/io_remote.cpp

@@ -35,7 +35,7 @@ cg::OperatorNodeBase* apply_on_var_node_remote_send(
     OperatorNodeConfig config{send.make_name()};
     cg::OperatorNodeBase* opr =
             graph->insert_opr(std::make_unique<mgb::opr::RemoteSend>(
-                    send.key, inputs[0], group_client, true, config));
+                    send.key, inputs[0], group_client, true, send.backend, config));
     return opr;
 }
 
@@ -49,7 +49,7 @@ cg::OperatorNodeBase* apply_on_var_node_remote_recv(
     auto&& graph = inputs[0]->owner_graph();
     return graph->insert_opr(std::make_unique<mgb::opr::RemoteRecv>(
             recv.key, inputs[0], *graph, group_client, config,
-            recv.shape, recv.dtype));
+            recv.shape, recv.dtype, recv.backend));
 }
 
 OP_TRAIT_REG(RemoteSend, RemoteSend, mgb::opr::RemoteSend)

imperative/src/test/io_remote.cpp

@@ -34,7 +34,7 @@ TEST(TestImperative, IORemote) {
 
     auto run_send = [&](std::shared_ptr<HostTensorND> hnd) {
         auto def = imperative::RemoteSend::make(
-            "io_remote_test", server_addr, port, 1);
+            "io_remote_test", server_addr, port, 1, "nccl");
         auto inp = Tensor::make(*hnd);
         auto oup = OpDef::apply_on_physical_tensor(*def, {inp});
     };
@@ -43,7 +43,7 @@ TEST(TestImperative, IORemote) {
         auto def = imperative::RemoteRecv::make(
             "io_remote_test", server_addr, port, 0,
             CompNode::load("gpu1"), TensorShape{vector_size},
-            dtype::Float32());
+            dtype::Float32(), "nccl");
         auto inp = Tensor::make(*hnd);
         auto oup = OpDef::apply_on_physical_tensor(*def, {inp});
         HostTensorND host_v;

src/core/include/megbrain/ir/ops.td

@@ -169,7 +169,8 @@ def RemoteSend : MgbHashableOp<"RemoteSend"> {
     MgbStringAttr:$key,
     MgbStringAttr:$addr,
     MgbUI32Attr:$port,
-    MgbUI32Attr:$rank_to
+    MgbUI32Attr:$rank_to,
+    MgbStringAttr:$backend
   );
 }
 
@@ -181,7 +182,8 @@ def RemoteRecv : MgbHashableOp<"RemoteRecv"> {
     MgbUI32Attr:$rank_from,
     MgbCompNodeAttr:$cn,
     MgbTensorShapeAttr:$shape,
-    MgbDTypeAttr:$dtype
+    MgbDTypeAttr:$dtype,
+    MgbStringAttr:$backend
   );
 }
 

src/opr-mm/impl/io_remote.cpp

@@ -24,8 +24,9 @@ MGB_DYN_TYPE_OBJ_FINAL_IMPL(RemoteSend);
 
 RemoteSend::RemoteSend(const std::string& key, VarNode* var,
                        std::shared_ptr<GroupClient> group_client,
-                       bool is_grad, const OperatorNodeConfig& config) :
+                       bool is_grad, std::string backend, const OperatorNodeConfig& config) :
         Super(var->owner_graph(), config, "remote_send", {var}),
+        m_backend(backend),
         m_is_grad(is_grad) {
     m_key = key;
     m_group_client = group_client;
@@ -41,9 +42,9 @@ RemoteSend::RemoteSend(const std::string& key, VarNode* var,
 
 SymbolVar RemoteSend::make(const std::string& key, SymbolVar var,
                            std::shared_ptr<GroupClient> group_client,
-                           bool is_grad, const OperatorNodeConfig& config) {
+                           bool is_grad, std::string backend, const OperatorNodeConfig& config) {
     return var.insert_single_output_opr<RemoteSend>(key, var.node(), group_client,
-                                                    is_grad, config);
+                                                    is_grad, backend, config);
 }
 
 void RemoteSend::scn_do_execute() {
@@ -64,7 +65,7 @@ void RemoteSend::scn_do_execute() {
         }
 
         m_megray_comm = MegRayCommBuilder::get_megray_comm(
-                reg_info.hash, m_key, 2, 0, MegRay::MEGRAY_NCCL, m_group_client);
+                reg_info.hash, m_key, 2, 0, get_megray_backend(m_backend), m_group_client);
 
         m_megray_ctx = get_megray_context(output(0)->comp_node());
 
@@ -122,7 +123,7 @@ MGB_IMPL_OPR_GRAD(RemoteSend) {
                             *opr.owner_graph(), opr.group_client(),
                             OperatorNodeConfig{opr.comp_node()}.name(
                                     opr.name() + ":grad_recv"),
-                            opr.input(0)->shape(), opr.input(0)->dtype())
+                            opr.input(0)->shape(), opr.input(0)->dtype(), opr.backend())
             .node();
 }
 #endif
@@ -134,9 +135,9 @@ MGB_DYN_TYPE_OBJ_FINAL_IMPL(RemoteRecv);
 RemoteRecv::RemoteRecv(const std::string& key, cg::ComputingGraph& graph,
                        std::shared_ptr<GroupClient> group_client,
                        const OperatorNodeConfig& config,
-                       const TensorShape& shape, DType dtype) :
+                       const TensorShape& shape, DType dtype, std::string backend) :
         Super(&graph, config, "remote_recv", {}),
-        m_shape(shape), m_dtype(dtype) {
+        m_shape(shape), m_dtype(dtype), m_backend(backend) {
     m_key = key;
     m_group_client = group_client;
 
@@ -150,9 +151,9 @@ RemoteRecv::RemoteRecv(const std::string& key, cg::ComputingGraph& graph,
 RemoteRecv::RemoteRecv(const std::string& key, VarNode* var, cg::ComputingGraph& graph,
                        std::shared_ptr<GroupClient> group_client,
                        const OperatorNodeConfig& config,
-                       const TensorShape& shape, DType dtype) :
+                       const TensorShape& shape, DType dtype, std::string backend) :
         Super(&graph, config, "remote_recv", {}),
-        m_shape(shape), m_dtype(dtype) {
+        m_shape(shape), m_dtype(dtype), m_backend(backend) {
     m_key = key;
     m_group_client = group_client;
 
@@ -167,18 +168,18 @@ RemoteRecv::RemoteRecv(const std::string& key, VarNode* var, cg::ComputingGraph&
 SymbolVar RemoteRecv::make(const std::string& key, cg::ComputingGraph& graph,
                            std::shared_ptr<GroupClient> group_client,
                            const OperatorNodeConfig& config,
-                           const TensorShape& shape, DType dtype) {
+                           const TensorShape& shape, DType dtype, std::string backend) {
     auto opr = graph.insert_opr(std::make_unique<RemoteRecv>(
-            key, graph, group_client, config, shape, dtype));
+            key, graph, group_client, config, shape, dtype, backend));
     return opr->output(0);
 }
 
 SymbolVar RemoteRecv::make(const std::string& key, SymbolVar var, cg::ComputingGraph& graph,
                            std::shared_ptr<GroupClient> group_client,
                            const OperatorNodeConfig& config,
-                           const TensorShape& shape, DType dtype) {
+                           const TensorShape& shape, DType dtype, std::string backend) {
     auto opr = graph.insert_opr(std::make_unique<RemoteRecv>(
-            key, var.node(), graph, group_client, config, shape, dtype));
+            key, var.node(), graph, group_client, config, shape, dtype, backend));
     return opr->output(0);
 }
 
@@ -201,7 +202,7 @@ void RemoteRecv::scn_do_execute() {
         }
 
         m_megray_comm = MegRayCommBuilder::get_megray_comm(
-                reg_info.hash, m_key, 2, 1, MegRay::MEGRAY_NCCL, m_group_client);
+                reg_info.hash, m_key, 2, 1, get_megray_backend(m_backend), m_group_client);
 
         m_megray_ctx = get_megray_context(output(0)->comp_node());
 
@@ -251,7 +252,7 @@ cg::OperatorNodeBase* opr_shallow_copy_remote_send(
     mgb_assert(inputs.size() == 1);
     auto&& opr = opr_.cast_final_safe<RemoteSend>();
     return RemoteSend::make(opr.key(), inputs[0], opr.group_client(),
-                            opr.is_grad(), config)
+                            opr.is_grad(), opr.backend(), config)
             .node()
             ->owner_opr();
 }
@@ -265,14 +266,14 @@ cg::OperatorNodeBase* opr_shallow_copy_remote_recv(
     if (inputs.size() == 1) {
         return RemoteRecv::make(opr.key(), inputs[0], *opr.owner_graph(),
                                 opr.group_client(), config, opr.shape(),
-                                opr.dtype())
+                                opr.dtype(), opr.backend())
                 .node()
                 ->owner_opr();
     } else {
         mgb_assert(inputs.size() == 0, "recv should have 1 or 0 input");
         return RemoteRecv::make(opr.key(), *opr.owner_graph(),
                                 opr.group_client(), config, opr.shape(),
-                                opr.dtype())
+                                opr.dtype(), opr.backend())
                 .node()
                 ->owner_opr();
     }

src/opr-mm/impl/io_remote.oprdecl

@@ -9,6 +9,8 @@ decl_raw_opr(
         Doc('key', 'key to bind send-recv pair', 'str'),
         Doc('var', 'variable to be sent', ':class:`.SymbolVar`'),
         Doc('is_grad', 'whether the send', 'bool'),
+        Doc('backend', 'Backend for collective communication, nccl or ucx',
+            'str', '\'nccl\''),
     ]
 )
 
@@ -24,7 +26,9 @@ decl_raw_opr(
             ':class:`.CompGraph`'),
         Doc('shape', 'output var shape'),
         Doc('dtype', 'data type of the output var; must match dtype at sender',
-            ':class:`numpy.dtype` compatible')
+            ':class:`numpy.dtype` compatible'),
+        Doc('backend', 'Backend for collective communication, nccl or ucx',
+            'str', '\'nccl\''),
     ]
 )
 

src/opr-mm/include/megbrain/opr/io_remote.h

@@ -48,17 +48,19 @@ MGB_DEFINE_OPR_CLASS(RemoteSend, RemoteIOBase) // {
     public:
         RemoteSend(const std::string& key, VarNode* var,
                    std::shared_ptr<GroupClient> group_client,
-                   bool is_grad, const OperatorNodeConfig& config);
+                   bool is_grad, std::string backend, const OperatorNodeConfig& config);
 
         static SymbolVar make(
                 const std::string& key, SymbolVar var,
                 std::shared_ptr<GroupClient> group_client,
-                bool is_grad, const OperatorNodeConfig& config = {});
+                bool is_grad, std::string backend, const OperatorNodeConfig& config = {});
 
+        const std::string& backend() const { return m_backend; }
         bool is_grad() const { return m_is_grad; }
 
     private:
         HostTensorND m_output_val;
+        std::string m_backend;
         bool m_is_grad;
 
         void scn_do_execute() override;
@@ -75,31 +77,33 @@ MGB_DEFINE_OPR_CLASS(RemoteRecv, RemoteIOBase) // {
         RemoteRecv(const std::string& key, cg::ComputingGraph& graph,
                    std::shared_ptr<GroupClient> group_client,
                    const OperatorNodeConfig& config, const TensorShape& shape,
-                   DType dtype);
+                   DType dtype, std::string backend);
 
         RemoteRecv(const std::string& key, VarNode* var, cg::ComputingGraph& graph,
                    std::shared_ptr<GroupClient> group_client,
                    const OperatorNodeConfig& config, const TensorShape& shape,
-                   DType dtype);
+                   DType dtype, std::string backend);
 
         static SymbolVar make(
                 const std::string& key, cg::ComputingGraph& graph,
                 std::shared_ptr<GroupClient> group_client,
                 const OperatorNodeConfig& config, const TensorShape& shape,
-                DType dtype);
+                DType dtype, std::string backend);
 
         static SymbolVar make(
                 const std::string& key, SymbolVar var, cg::ComputingGraph& graph,
                 std::shared_ptr<GroupClient> group_client,
                 const OperatorNodeConfig& config, const TensorShape& shape,
-                DType dtype);
+                DType dtype, std::string backend);
 
         const TensorShape& shape() const { return m_shape; }
         const DType& dtype() const { return m_dtype; }
+        const std::string& backend() const { return m_backend; }
 
     private:
         const TensorShape m_shape;
         const DType m_dtype;
+        const std::string m_backend;
         const CompNode m_comp_node;
         DeviceTensorND m_dev_buffer;
 

src/opr-mm/test/io_remote.cpp

@@ -33,10 +33,10 @@ TEST(TestOprIORemote, Identity) {
     auto graph = ComputingGraph::make();
 
     auto x = opr::Host2DeviceCopy::make(*graph, host_x, cn0);
-    auto xr = opr::RemoteSend::make("x", x, client, false);
+    auto xr = opr::RemoteSend::make("x", x, client, false, "nccl");
     auto y = opr::RemoteRecv::make("x", *graph.get(),
                                    client, {cn1}, host_x->shape(),
-                                   host_x->dtype());
+                                   host_x->dtype(), "nccl");
 
     auto func = graph->compile({{xr, {}}, make_callback_copy(y, host_y)});
 
@@ -57,7 +57,7 @@ TEST(TestOprIORemote, IdentityMultiThread) {
         auto graph = ComputingGraph::make();
         sys::set_thread_name("sender");
         auto x = opr::Host2DeviceCopy::make(*graph, host_x),
-             xr = opr::RemoteSend::make("x", x, client, false);
+             xr = opr::RemoteSend::make("x", x, client, false, "nccl");
         auto func = graph->compile({{xr, {}}});
         func->execute();
     };
@@ -67,7 +67,7 @@ TEST(TestOprIORemote, IdentityMultiThread) {
         auto graph = ComputingGraph::make();
         auto x = opr::RemoteRecv::make("x", *graph.get(),
                                        client, {cns[0]}, host_x->shape(),
-                                       host_x->dtype());
+                                       host_x->dtype(), "nccl");
         auto func = graph->compile({make_callback_copy(x, host_x_get)});
         func->execute();
     };
@@ -91,7 +91,7 @@ TEST(TestOprIORemote, IdentityWithGopt) {
         sys::set_thread_name("sender");
         auto graph = ComputingGraph::make();
         auto x = opr::Host2DeviceCopy::make(*graph, host_x) * 2 + 1,
-             xr = opr::RemoteSend::make("x", x, client, false);
+             xr = opr::RemoteSend::make("x", x, client, false, "nccl");
         auto func = graph->compile({{xr, {}}});
         func->execute();
     };
@@ -101,7 +101,7 @@ TEST(TestOprIORemote, IdentityWithGopt) {
         auto graph = ComputingGraph::make();
         auto x = opr::RemoteRecv::make("x", *graph.get(),
                                        client, {cns[0]}, host_x->shape(),
-                                       host_x->dtype());
+                                       host_x->dtype(), "nccl");
         auto func =
                 graph->compile({make_callback_copy((x - 1) / 2, host_x_get)});
         func->execute();
@@ -126,12 +126,12 @@ TEST(TestOprIORemote, APlusB) {
         auto graph = ComputingGraph::make();
         auto z = opr::RemoteRecv::make("z", *graph.get(),
                                        client, {cns[0]}, host_x->shape(),
-                                       host_x->dtype());
+                                       host_x->dtype(), "nccl");
         auto x = opr::Host2DeviceCopy::make(*graph, host_x).rename("x"),
              y = opr::Host2DeviceCopy::make(*graph, host_y).rename("y"),
-             xr = opr::RemoteSend::make("x", x, client, false)
+             xr = opr::RemoteSend::make("x", x, client, false, "nccl")
                           .rename("xr"),
-             yr = opr::RemoteSend::make("y", y, client, false)
+             yr = opr::RemoteSend::make("y", y, client, false, "nccl")
                           .rename("yr");
         auto func = graph->compile(
                 {{xr, {}}, {yr, {}}, make_callback_copy(z, host_z)});
@@ -144,12 +144,12 @@ TEST(TestOprIORemote, APlusB) {
         auto graph = ComputingGraph::make();
         auto x = opr::RemoteRecv::make("x", *graph.get(),
                                        client, {cns[1]}, host_x->shape(),
-                                       host_x->dtype()),
+                                       host_x->dtype(), "nccl"),
              y = opr::RemoteRecv::make("y", *graph.get(),
                                        client, {cns[1]}, host_y->shape(),
-                                       host_y->dtype()),
+                                       host_y->dtype(), "nccl"),
              z = x + y,
-             zr = opr::RemoteSend::make("z", z, client, false);
+             zr = opr::RemoteSend::make("z", z, client, false, "nccl");
         auto func = graph->compile({{zr, {}}});
         func->execute();
     };
@@ -178,10 +178,10 @@ TEST(TestOprIORemote, SendGrad) {
         sys::set_thread_name("sender");
         auto graph = ComputingGraph::make();
         auto x = opr::Host2DeviceCopy::make(*graph, host_x),
-             loss = opr::RemoteSend::make("loss", x, client, false);
+             loss = opr::RemoteSend::make("loss", x, client, false, "nccl");
         ASSERT_TRUE(!loss.shape().ndim &&
                     loss.node()->contain_flag(VarNode::Flag::VOLATILE_CONTENT));
-        loss = opr::RemoteSend::make("loss", x, client, true);
+        loss = opr::RemoteSend::make("loss", x, client, true, "nccl");
         auto gx = cg::grad(loss, x);
         set_priority(loss, 0);
         set_priority(gx, 1);
@@ -200,8 +200,8 @@ TEST(TestOprIORemote, SendGrad) {
         auto graph = ComputingGraph::make();
         auto x = opr::RemoteRecv::make("loss", *graph.get(),
                                        client, {cns[1]}, host_x->shape(),
-                                       host_x->dtype());
-        auto y = opr::RemoteSend::make("loss:grad", x + 1, client, false);
+                                       host_x->dtype(), "nccl");
+        auto y = opr::RemoteSend::make("loss:grad", x + 1, client, false, "nccl");
         auto func = graph->compile({{y, {}}});
         func->execute();
     };