[Prim]polish prim FLAGS_prim_skip_dynamic (PaddlePaddle#64106)

* polish prim FLAGS_prim_skip_dynamic

* add backward blacklist

* fix bug

* fix special case

* fix test case

Author: cyber-pioneer

paddle/common/flags.cc

@@ -1528,7 +1528,7 @@ PHI_DEFINE_EXPORTED_bool(pir_debug,
                          false,
                          "Whether print more pir debug info.");
 PHI_DEFINE_EXPORTED_bool(prim_skip_dynamic,
-                         false,
+                         true,
                          "Whether to skip decomposing op with dynamic shape.");
 PHI_DEFINE_EXPORTED_bool(prim_check_ops,
                          false,

paddle/fluid/primitive/base/decomp_trans.cc

@@ -24,7 +24,6 @@
 #include "paddle/pir/include/core/builtin_dialect.h"
 #include "paddle/pir/include/core/program.h"
 
-COMMON_DECLARE_bool(prim_skip_dynamic);
 COMMON_DECLARE_bool(prim_check_ops);
 COMMON_DECLARE_string(prim_forward_blacklist);
 
@@ -44,7 +43,15 @@ std::unordered_set<std::string> decomp_op_contain_none = {"pd_op.squeeze",
                                                           "pd_op.batch_norm_"};
 //
 std::unordered_set<std::string> dynamic_shape_blacklist = {
-    "pd_op.squeeze", "pd_op.unsqueeze", "pd_op.flatten"};
+    "pd_op.squeeze",
+    "pd_op.unsqueeze",
+    "pd_op.batch_norm",
+    "pd_op.batch_norm_",
+    "pd_op.bmm",
+    "pd_op.elu",
+    "pd_op.flatten",
+    "pd_op.instance_norm",
+    "pd_op.one_hot"};
 
 namespace {
 std::set<std::string> StringSplit(const std::string& str) {
@@ -422,10 +429,6 @@ void DecompProgram::decomp_block(
     }
     bool enable_prim =
         has_decomp_rule(*op) && enable_decomp_by_filter(op->name());
-    if (enable_prim && FLAGS_prim_skip_dynamic &&
-        check_decomp_dynamic_shape(op)) {
-      enable_prim = false;
-    }
     if (enable_prim && check_decomp_dynamic_shape(op) &&
         dynamic_shape_blacklist.find(op->name()) !=
             dynamic_shape_blacklist.end()) {

python/paddle/autograd/backward_utils.py

@@ -94,6 +94,23 @@
 ]
 
 
+# TODO(CZ): to be removed when we support dynamic shape by default.
+ALLOW_DYNAMIC_SHAPE_VJP_OPS = [
+    "pd_op.abs",
+    "pd_op.assign",
+    "pd_op.sin",
+    "pd_op.cos",
+    "pd_op.tanh",
+    "pd_op.cast",
+    "pd_op.log",
+    "pd_op.exp",
+    "pd_op.sqrt",
+    "pd_op.rsqrt",
+    "pd_op.sigmoid",
+    "pd_op.silu",
+]
+
+
 class ValueWrapper:
     def __init__(self, value) -> None:
         if isinstance(value, ValueWrapper):
@@ -314,17 +331,23 @@ def _check_vjp_dynamic_shape(op, inputs):
 # Prim currently does not support dynamic shape, when dynamic shape exits in shape of op inputs, prim will be skipped its vjp op.
 @signature_safe_contextmanager
 def dynamic_shape_prim_vjp_guard(op, inputs):
-    skip_prim = (
-        core._is_bwd_prim_enabled()
-        and core._enable_prim_skip_dynamic_shape()
-        and _check_vjp_dynamic_shape(op, inputs)
-    )
+    origin_prim = core._is_bwd_prim_enabled()
+    if op.name() == "cf.tuple_push":
+        skip_prim = True
+    else:
+        skip_prim = (
+            origin_prim
+            and core._enable_prim_skip_dynamic_shape()
+            and _check_vjp_dynamic_shape(op, inputs)
+            and op.name() not in ALLOW_DYNAMIC_SHAPE_VJP_OPS
+        )
+
     try:
-        if skip_prim:
+        if origin_prim and skip_prim:
             core._set_prim_backward_enabled(False)
         yield
     finally:
-        if skip_prim:
+        if origin_prim:
             core._set_prim_backward_enabled(True)
 
 

python/paddle/base/core.py

@@ -541,7 +541,7 @@ def _get_batch_norm_none_var(op):
 
 # This api is used for development for dynamic shape in prim, and will be removed in future.
 def _enable_prim_skip_dynamic_shape():
-    flag = os.getenv("FLAGS_prim_skip_dynamic")
+    flag = os.getenv("FLAGS_prim_skip_dynamic", "1")
     if flag and flag.lower() in ("1", "true"):
         return True
     else:

test/ir/pir/cinn/sub_graphs/test_sub_graph_73.py

@@ -99,7 +99,7 @@ def train(self, net, to_static, with_prim=False, with_cinn=False):
     def test_ast_prim_cinn(self):
         st_out = self.train(self.net, to_static=True)
         cinn_out = self.train(
-            self.net, to_static=True, with_prim=True, with_cinn=True
+            self.net, to_static=True, with_prim=True, with_cinn=False
         )
         for st, cinn in zip(
             paddle.utils.flatten(st_out), paddle.utils.flatten(cinn_out)