Cxx prim custom vjp (#8)

* [CINN]Enhance CacheKey hash logic by considering input dtypes (PaddlePaddle#50557)

---------

Co-authored-by: jiangcheng <[email protected]>

* [prim] enable dygraph_to_static to support custom_vjp

* Pr 50885 (#7)

* [CINN]Enhance CacheKey hash logic by considering input dtypes (PaddlePaddle#50557)

* [CINN]Enhance CacheKey hash logic by considering input dtypes

---------

Co-authored-by: jiangcheng <[email protected]>

* [prim] enable dygraph_to_static to support custom_vjp

* fix code in a dy2static-friendly way.

* [dystatic] add hooker for prim

---------

Co-authored-by: Aurelius84 <[email protected]>
Co-authored-by: jiangcheng <[email protected]>
Co-authored-by: cxxly <[email protected]>

* [prim] enable dygraph_to_static to support custom_vjp

* fix cast prim and vjp dtype mapping error bug

* [dy2static-ci] fix dy2static ci errors.

---------

Co-authored-by: Aurelius84 <[email protected]>
Co-authored-by: jiangcheng <[email protected]>
Co-authored-by: cxxly <[email protected]>

Author: 4 people

python/paddle/fluid/framework.py

@@ -3751,7 +3751,6 @@ def __init__(self, program, idx):
         self.vars = collections.OrderedDict()  # var_name --> var
         self.ops = list()  # operator list
         self.program = program
-        self.removed_vars = collections.OrderedDict()
 
     def __str__(self):
         return self._to_readable_code()

python/paddle/fluid/tests/unittests/dygraph_to_static/test_cinn_prim.py

@@ -77,7 +77,12 @@ def train(self, use_prim):
     def check_prim(self, net, use_prim):
         if not use_prim:
             return
-        fwd_ops = [op.type for op in net.forward.main_program.block(0).ops]
+        fwd_ops = [
+            op.type
+            for op in net.forward.get_concrete_program(self.x)[1]
+            .train_program.block(0)
+            .ops
+        ]
         # Ensure that softmax is splitted into small ops
         self.assertTrue('softmax' not in fwd_ops)
 
@@ -128,7 +133,12 @@ def train(self, use_prim):
     def check_prim(self, net, use_prim):
         if not use_prim:
             return
-        fwd_ops = [op.type for op in net.forward.main_program.block(0).ops]
+        fwd_ops = [
+            op.type
+            for op in net.forward.get_concrete_program(self.x)[1]
+            .train_program.block(0)
+            .ops
+        ]
         all_ops = [
             op.type
             for op in net.forward.program_cache.last()[-1][-1]

python/paddle/fluid/tests/unittests/dygraph_to_static/test_cinn_prim_gelu.py

@@ -77,6 +77,7 @@ def _train(self, use_prim, approximate, data):
             net = apply_to_static(net, use_prim)
 
         res = []
+        self.x = data
         for _ in range(10):
             out = net(data)
             loss = paddle.mean(out)
@@ -92,7 +93,12 @@ def _train(self, use_prim, approximate, data):
     def check_prim(self, net, use_prim):
         if not use_prim:
             return
-        fwd_ops = [op.type for op in net.forward.main_program.block(0).ops]
+        fwd_ops = [
+            op.type
+            for op in net.forward.get_concrete_program(self.x)[1]
+            .train_program.block(0)
+            .ops
+        ]
         # Ensure that gelu is splitted into small ops
         self.assertTrue('gelu' not in fwd_ops)
 

python/paddle/fluid/tests/unittests/dygraph_to_static/test_cinn_prim_layer_norm.py

@@ -89,7 +89,14 @@ def train(self, use_prim):
     def check_prim(self, net, use_prim):
         if not use_prim:
             return
-        fwd_ops = [op.type for op in net.forward.main_program.block(0).ops]
+        fwd_ops = [
+            op.type
+            for op in net.forward.get_concrete_program(self.x, self.w, self.b)[
+                1
+            ]
+            .train_program.block(0)
+            .ops
+        ]
         # Ensure that layer_norm is splitted into small ops
         self.assertTrue('layer_norm' not in fwd_ops)
 
@@ -150,7 +157,14 @@ def train(self, use_prim):
     def check_prim(self, net, use_prim):
         if not use_prim:
             return
-        fwd_ops = [op.type for op in net.forward.main_program.block(0).ops]
+        fwd_ops = [
+            op.type
+            for op in net.forward.get_concrete_program(self.x, self.w, self.b)[
+                1
+            ]
+            .train_program.block(0)
+            .ops
+        ]
         # Ensure that layer_norm is splitted into small ops
         self.assertTrue('layer_norm' not in fwd_ops)
 

python/paddle/fluid/tests/unittests/dygraph_to_static/test_cinn_prim_mean.py

@@ -83,6 +83,7 @@ def _train(self, use_prim, data, axis, keep_dim):
             net = apply_to_static(net, use_prim)
 
         res = []
+        self.x = data
         for _ in range(10):
             out = net(data)
             loss = paddle.mean(out, axis, keep_dim)
@@ -99,7 +100,12 @@ def _train(self, use_prim, data, axis, keep_dim):
     def check_prim(self, net, use_prim):
         if not use_prim:
             return
-        fwd_ops = [op.type for op in net.forward.main_program.block(0).ops]
+        fwd_ops = [
+            op.type
+            for op in net.forward.get_concrete_program(self.x)[1]
+            .train_program.block(0)
+            .ops
+        ]
         # Ensure that reduce_mean is splitted into small ops
         self.assertTrue('reduce_mean' not in fwd_ops)
 
@@ -150,6 +156,7 @@ def _train(self, use_prim, data, axis, keep_dim):
             net = apply_to_static(net, use_prim)
 
         res = []
+        self.x = data
         for _ in range(10):
             out = net(data)
             loss = paddle.mean(out, axis, keep_dim)
@@ -166,7 +173,12 @@ def _train(self, use_prim, data, axis, keep_dim):
     def check_prim(self, net, use_prim):
         if not use_prim:
             return
-        fwd_ops = [op.type for op in net.forward.main_program.block(0).ops]
+        fwd_ops = [
+            op.type
+            for op in net.forward.get_concrete_program(self.x)[1]
+            .train_program.block(0)
+            .ops
+        ]
         # Ensure that reduce_mean is splitted into small ops
         self.assertTrue('reduce_mean' not in fwd_ops)
 

python/paddle/jit/dy2static/partial_program.py

@@ -315,9 +315,7 @@ def _create_pure_fp16_program(self, is_infer_mode=False):
     def _create_forward_backward_train_program(self):
         whole_program = self._train_program
         # _, forward_end_op_index = self._infer_info('fp32', self._create_program)
-        forward_end_op_index = self._forward_end_index_map[
-            _hash_with_id(whole_program, self)
-        ]
+        forward_end_op_index = self.get_forward_end_op_idx(whole_program)
         assert forward_end_op_index >= 0
 
         return self._get_forward_backward_program_form(
@@ -438,11 +436,14 @@ def _infer_pure_fp16_program_id(self):
     def _param_grad_names(self):
         return _param_grad_names(self._train_program.desc, self._params)
 
+    def get_forward_end_op_idx(self, program):
+        return self._forward_end_index_map[_hash_with_id(program, self)]
+
     @LazyInitialized
     def _out_grad_names(self):
         return _out_grad_names(
             self._train_program.desc,
-            self._create_program(is_infer_mode=True).desc.block(0).op_size(),
+            self.get_forward_end_op_idx(self._train_program),
             len(self._outputs.var_ids),
         )
 
@@ -642,6 +643,7 @@ def _append_backward_desc(self, main_program):
             if isinstance(out, framework.Variable):
                 targets.append(program.global_block().var(out.name))
 
+        start_idx = len(program.block(0).ops) + len(self._outputs.tolist())
         if targets:
             # TODO(CZ): later when use cinn, set_prim_all_enabled and check_and_set_prim_all_enabled will be set at else branch.
             core.check_and_set_prim_all_enabled()
@@ -652,12 +654,11 @@ def _append_backward_desc(self, main_program):
                 program, start_idx = self._hooker.after_append_backward(
                     self, program, start_idx
                 )
-            self._forward_end_index_map[
-                _hash_with_id(program, self)
-            ] = start_idx - len(self._outputs.tolist())
-            # TODO: prim make this complicate
             self.prepare_gradient_aggregation(start_idx, main_program, program)
 
+        self._forward_end_index_map[
+            _hash_with_id(program, self)
+        ] = start_idx - len(self._outputs.tolist())
         return program
 
     def _prune_unused_params(self, program):
@@ -1155,5 +1156,8 @@ def add_build_strategy_for(
         if hasattr(compiled_program._program, 'lr_sheduler'):
             builded_program.lr_sheduler = compiled_program._program.lr_sheduler
     else:
+        # can't just create a new program, we need copy the vardesc.
         builded_program = paddle.static.Program()
+        for var in program.block(0).vars.values():
+            builded_program.block(0)._clone_variable(var, False)
     return builded_program

python/paddle/jit/dy2static/program_translator.py

@@ -1226,7 +1226,6 @@ def after_infer(self, partial_program_layer, infer_program):
             partial_program.set_hooker(PrimHooker())
         return concrete_program, partial_program
 
-
     def __getitem__(self, item):
         if not isinstance(item, CacheKey):
             raise ValueError(

python/paddle/jit/dy2static/utils.py

@@ -1568,7 +1568,7 @@ def _out_grad_names(program_desc, fwd_end_op_index, out_size):
         min(fwd_end_op_index + out_size, program_desc.block(0).op_size()),
     ):
         op = program_desc.block(0).op(i)
-        if op.type() == 'fill_any_like':
+        if op.type() in ['fill_any_like', "fill_constant"]:
             var_name = op.output('Out')[0]
             names.append(var_name)
     return names