feat(mge/imperative): add sublinear options

GitOrigin-RevId: f0e917f716

imperative/python/megengine/jit/__init__.py

@@ -1 +1,2 @@
+from .sublinear_memory_config import SublinearMemoryConfig
 from .tracing import exclude_from_trace, trace

imperative/python/megengine/jit/sublinear_memory_config.py

@@ -0,0 +1,56 @@
+# -*- coding: utf-8 -*-
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+
+from ..device import get_device_count
+
+
+class SublinearMemoryConfig:
+    r"""
+    Configuration for sublinear memory optimization.
+
+    :param thresh_nr_try: number of samples both for searching in linear space
+        and around current thresh in sublinear memory optimization. Default: 10.
+        It can also be set through the environmental variable 'MGB_SUBLINEAR_MEMORY_THRESH_NR_TRY'.
+    :param genetic_nr_iter: number of iterations to find the best checkpoints in genetic algorithm.
+        Default: 0.
+        It can also be set through the environmental variable 'MGB_SUBLINEAR_MEMORY_GENETIC_NR_ITER'.
+    :param genetic_pool_size: number of samples for the crossover random selection
+        during genetic optimization. Default: 20.
+        It can also be set through the environmental variable 'MGB_SUBLINEAR_MEMORY_GENETIC_POOL_SIZE'.
+    :param lb_memory: memory lower bound of bottleneck size in MB for sublinear memory optimization.
+        It can be used to perform manual tradeoff between memory and speed. Default: 0.
+        It can also be set through the environmental variable 'MGB_SUBLINEAR_MEMORY_LOWER_BOUND_MB'.
+    :param num_worker: number of thread workers to search the optimum checkpoints
+        in sublinear memory optimization. Default: half of cpu number in the system.
+        Note: the value must be greater or equal to one.
+        It can also be set through the environmental variable 'MGB_SUBLINEAR_MEMORY_WORKERS'.
+
+    Note that the environmental variable MGB_COMP_GRAPH_OPT must be set to 'enable_sublinear_memory_opt=1'
+    in order for the above environmental variable to be effective.
+    """
+
+    def __init__(
+        self,
+        thresh_nr_try: int = 10,
+        genetic_nr_iter: int = 0,
+        genetic_pool_size: int = 20,
+        lb_memory: int = 0,
+        num_worker: int = max(1, get_device_count("cpu") // 2),
+    ):
+        assert thresh_nr_try >= 0, "thresh_nr_try must be greater or equal to zero"
+        self.thresh_nr_try = thresh_nr_try
+        assert genetic_nr_iter >= 0, "genetic_nr_iter must be greater or equal to zero"
+        self.genetic_nr_iter = genetic_nr_iter
+        assert (
+            genetic_pool_size >= 0
+        ), "genetic_pool_size must be greater or equal to zero"
+        self.genetic_pool_size = genetic_pool_size
+        self.lb_memory = lb_memory
+        assert num_worker > 0, "num_worker must be greater or equal to one"
+        self.num_worker = num_worker

imperative/python/megengine/jit/tracing.py

@@ -7,6 +7,7 @@ from ..core.ops.special import Const
 from ..core.tensor import megbrain_graph as G
 from ..core.tensor.core import OpBase, apply
 from ..core.tensor.raw_tensor import OpDef, RawTensor, as_raw_tensor
+from .sublinear_memory_config import SublinearMemoryConfig
 
 
 class TraceMismatchError(RuntimeError):
@@ -72,11 +73,18 @@ class trace:
         self.__init__(*args, **kwargs)
         return self
 
-    def __init__(self, function, symbolic=False, capture_as_const=False):
+    def __init__(
+        self,
+        function,
+        symbolic=False,
+        capture_as_const=False,
+        sublinear_memory_config: SublinearMemoryConfig = None,
+    ):
         self.__wrapped__ = function
         self._symbolic = symbolic
         self._capture_as_const = capture_as_const
         self._capture_static_shape = False
+        self._sublinear_memory_config = sublinear_memory_config
 
         self._untraced = True
         self._tinfo = []  # handle -> TensorInfo
@@ -227,6 +235,7 @@ class trace:
                         G.OutputNode(x._LazyEvalTensor__varnode).outputs[0]
                         for x in lazy_eval_tensors
                     ]
+                    self._apply_graph_options(self._lazy_eval_graph)
                     self._lazy_eval_graph.compile(*readers)
                     self._lazy_eval_graph()
                     for r, x in zip(readers, lazy_eval_tensors):
@@ -259,9 +268,26 @@ class trace:
                 info.exported = True
                 info.data_read = True
 
+    def _apply_graph_options(self, graph):
+
+        # sublinear
+        if self._sublinear_memory_config is not None:
+            graph.options.enable_sublinear_memory_opt = True
+            sublinear_config = graph.options.sublinear_mem_config
+            sublinear_config.lb_memory = self._sublinear_memory_config.lb_memory
+            sublinear_config.genetic_nr_iter = (
+                self._sublinear_memory_config.genetic_nr_iter
+            )
+            sublinear_config.genetic_pool_size = (
+                self._sublinear_memory_config.genetic_pool_size
+            )
+            sublinear_config.thresh_nr_try = self._sublinear_memory_config.thresh_nr_try
+            sublinear_config.num_worker = self._sublinear_memory_config.num_worker
+
     def _compile(self):
         graph = self._graph = G.Graph()
         graph.options.no_force_inplace = True
+        self._apply_graph_options(graph)
         # graph.options.graph_opt_level = 0
         need_reset_nodes = self._need_reset_nodes = []
         # links enforce ordering of I/O nodes

imperative/python/src/graph_rt.cpp

@@ -119,6 +119,7 @@ void init_graph_rt(py::module m) {
         DEF_READWRITE(enable_memory_swap)
         DEF_READWRITE(comp_node_seq_record_level)
         DEF_READWRITE(no_force_inplace)
+        DEF_READWRITE(sublinear_mem_config)
         // DEF_READWRITE(eager_evaluation)
         // DEF_READWRITE(imperative_proxy_graph)
         // DEF_READWRITE(extra_vardeps)
@@ -142,6 +143,16 @@ void init_graph_rt(py::module m) {
 
 #undef CURRENT_CLASS
 
+#define CURRENT_CLASS cg::ComputingGraph::Options::SublinearMemConfig
+
+    py::class_<cg::ComputingGraph::Options::SublinearMemConfig>(PyComputingGraphOptions, "SublinearMemConfig")
+        DEF_READWRITE(thresh_nr_try)
+        DEF_READWRITE(genetic_nr_iter)
+        DEF_READWRITE(genetic_pool_size)
+        DEF_READWRITE(lb_memory)
+        DEF_READWRITE(num_worker);
+
+#undef CURRENT_CLASS
     auto common = rel_import("common", m, 1);
 
     common.def("invoke_op", [](const OpDef& def, const std::vector<cg::VarNode*> inputs, cg::ComputingGraph* graph) {

imperative/python/test/integration/test_correctness.py

@@ -19,6 +19,7 @@ import megengine.functional as F
 from megengine import jit
 from megengine.core._trace_option import set_tensor_shape
 from megengine.functional.debug_param import set_conv_execution_strategy
+from megengine.jit import SublinearMemoryConfig
 from megengine.module import AvgPool2d, BatchNorm2d, Conv2d, Linear, Module
 from megengine.optimizer import SGD
 from megengine.tensor import Tensor
@@ -217,14 +218,14 @@ def test_correctness():
     set_conv_execution_strategy("HEURISTIC_REPRODUCIBLE")
 
     run_train(model_path, False, False, max_err=1e-5)
-    # run_test(model_path, True, False)
-    # run_test(model_path, True, True)
+    run_train(model_path, True, False, max_err=1e-5)
+    run_train(model_path, True, True, max_err=1e-5)
 
     # sublinear
-    # config = SublinearMemoryConfig(genetic_nr_iter=10)
-    # run_test(
-    #     model_path, True, True, sublinear_memory_config=config, max_err=1e-5,
-    # )
+    config = SublinearMemoryConfig(genetic_nr_iter=10)
+    run_train(
+        model_path, True, True, sublinear_memory_config=config, max_err=1e-5,
+    )
 
     run_eval(model_path, False, max_err=1e-7)
-    # run_eval(model_path, True, max_err=1e-7) # XXX: fix me
+    run_eval(model_path, True, max_err=1e-7)

python_module/megengine/jit/__init__.py

@@ -298,23 +298,23 @@ class trace:
         if self._sublinear_memory_config is not None:
             cg.set_option("enable_sublinear_memory_opt", True)
             cg.set_option(
-                "sublinear_mem_cofig.lb_memory",
+                "sublinear_mem_config.lb_memory",
                 self._sublinear_memory_config.lb_memory,
             )
             cg.set_option(
-                "sublinear_mem_cofig.genetic_nr_iter",
+                "sublinear_mem_config.genetic_nr_iter",
                 self._sublinear_memory_config.genetic_nr_iter,
             )
             cg.set_option(
-                "sublinear_mem_cofig.genetic_pool_size",
+                "sublinear_mem_config.genetic_pool_size",
                 self._sublinear_memory_config.genetic_pool_size,
             )
             cg.set_option(
-                "sublinear_mem_cofig.thresh_nr_try",
+                "sublinear_mem_config.thresh_nr_try",
                 self._sublinear_memory_config.thresh_nr_try,
             )
             cg.set_option(
-                "sublinear_mem_cofig.num_worker",
+                "sublinear_mem_config.num_worker",
                 self._sublinear_memory_config.num_worker,
             )
         # pack allreduce

python_module/src/cpp/megbrain_config.cpp

@@ -116,11 +116,11 @@ bool _config::set_comp_graph_option(
     SET_CG_OPTION(allocate_static_mem_after_graph_compile);
     SET_CG_OPTION(log_level);
     SET_CG_OPTION(enable_sublinear_memory_opt);
-    SET_CG_OPTION(sublinear_mem_cofig.lb_memory);
-    SET_CG_OPTION(sublinear_mem_cofig.genetic_nr_iter);
-    SET_CG_OPTION(sublinear_mem_cofig.genetic_pool_size);
-    SET_CG_OPTION(sublinear_mem_cofig.thresh_nr_try);
-    SET_CG_OPTION(sublinear_mem_cofig.num_worker);
+    SET_CG_OPTION(sublinear_mem_config.lb_memory);
+    SET_CG_OPTION(sublinear_mem_config.genetic_nr_iter);
+    SET_CG_OPTION(sublinear_mem_config.genetic_pool_size);
+    SET_CG_OPTION(sublinear_mem_config.thresh_nr_try);
+    SET_CG_OPTION(sublinear_mem_config.num_worker);
     SET_CG_OPTION(enable_var_mem_defragment);
     SET_CG_OPTION(eager_evaluation);
     SET_CG_OPTION(enable_memory_swap);

src/core/impl/graph/cg_impl.cpp

@@ -219,7 +219,7 @@ ComputingGraphImpl::Components::Components(ComputingGraphImpl* owner)
           grad_manager{owner},
 #if MGB_ENABLE_SUBLINEAR
           seq_modifier_for_sublinear_memory{owner,
-              &(owner->options().sublinear_mem_cofig)},
+              &(owner->options().sublinear_mem_config)},
 #endif
 #if MGB_ENABLE_MEMORY_SWAP
           memory_swap_support{owner},

src/core/include/megbrain/graph/cg.h

@@ -409,7 +409,7 @@ class ComputingGraph : public std::enable_shared_from_this<ComputingGraph>,
                 int genetic_pool_size = 20;
                 int lb_memory = 0;
                 int num_worker = sys::get_cpu_count() / 2;
-            } sublinear_mem_cofig;
+            } sublinear_mem_config;
 
             //! do not re-profile to select best impl algo when input shape
             //! changes (use previous algo)

src/core/test/sublinear_memory.cpp

@@ -522,7 +522,7 @@ TEST(TestSublinearMemory, BadOpr) {
        set_priority(z, 3);
        graph->options().graph_opt_level = 0;
        graph->options().enable_sublinear_memory_opt = 1;
-       graph->options().sublinear_mem_cofig.genetic_nr_iter = 50;
+       graph->options().sublinear_mem_config.genetic_nr_iter = 50;
        auto func = graph->compile({{y, {}}, {z, {}}});
        auto&& results = static_cast<cg::ComputingGraphImpl*>(graph.get())
            ->seq_modifier_for_sublinear_memory().prev_min_bottleneck();