Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

… fix_roll_op
PaddlePaddle · Oct 26, 2021 · 928eda2 · 928eda2 · paddle-bot-old · Oct 26, 2021
2 parents 6f23fce + 81e0c1b
commit 928eda2
Show file tree

Hide file tree

Showing 36 changed files with 3,428 additions and 1,833 deletions.
diff --git a/paddle/fluid/framework/fleet/gloo_wrapper.h b/paddle/fluid/framework/fleet/gloo_wrapper.h
@@ -27,6 +27,7 @@ limitations under the License. */
 #include <vector>
 #ifdef PADDLE_WITH_GLOO
 #include <gloo/allgather.h>
+#include <gloo/allgatherv.h>
 #include <gloo/allreduce.h>
 #include <gloo/barrier.h>
 #include <gloo/rendezvous/context.h>
@@ -238,10 +239,25 @@ class GlooWrapper {
     return ret;
   }
 
-  // TODO(xiongkun03): support all gather array of
+  // NOTE(@xiongkun03): support all gather array of
   //                   numbers with different length
-  //                   can use AllgathervOptions, may be work in different
-  //                   occasion. Need some survey.
+  //                   if the third argument is int, use allgather,
+  //                   if it is vector, use AllgathervOptions,
+  //                   which works in different length occasion.
+  template <typename T>
+  void AllGatherVector(T* input_ptr, T* output_ptr,
+                       std::vector<size_t>& element_nums) {  // NOLINT
+    CHECK_EQ(is_initialized_, true);
+#ifdef PADDLE_WITH_GLOO
+    gloo::AllgathervOptions opts(context_);
+    opts.setInput(input_ptr, element_nums[rank_]);
+    opts.setOutput(output_ptr, element_nums);
+    gloo::allgatherv(opts);
+#else
+    LOG(WARNING) << "AllGather does nothing when WITH_GLOO=OFF";
+#endif
+  }
+
   template <typename T>
   void AllGatherVector(T* input_ptr, T* output_ptr,
                        size_t element_num) {  // NOLINT

diff --git a/paddle/fluid/framework/ir/delete_quant_dequant_filter_op_pass.cc b/paddle/fluid/framework/ir/delete_quant_dequant_filter_op_pass.cc
@@ -181,7 +181,7 @@ void DeleteQuantDequantFilterOpPass::ApplyImpl(ir::Graph* graph) const {
                               "Weight scale should be nonzero, but get zero."));
         weight_scale[i] = weight_scale[i] / range;
       }
-    } else {
+    } else if (dequant_type == "fake_quantize_dequantize_abs_max") {
       // Implement quantize_dequantize_abs_max quantization algorithm
       float abs_max_weight = 0.;
       for (int j = 0; j < weight_tensor->numel(); j++) {
@@ -192,6 +192,9 @@ void DeleteQuantDequantFilterOpPass::ApplyImpl(ir::Graph* graph) const {
                         platform::errors::InvalidArgument(
                             "Weight scale should be nonzero, but get zero"));
       weight_scale.push_back(abs_max_weight / range);
+    } else {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "Unsupported quantize_dequantize op type: %s", dequant_type));
     }
 
     nodes2rm.insert(quant_dequant_op_outscale);

diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc
@@ -1606,6 +1606,7 @@ PDNode *patterns::Matmul::operator()() {
                          ->assert_is_op_input("matmul", "X");
   auto matmul_in_y = pattern->NewNode(matmul_in_y_repr())
                          ->AsInput()
+                         ->assert_is_persistable_var()
                          ->assert_is_op_input("matmul", "Y");
   auto matmul_out = pattern->NewNode(matmul_out_repr())
                         ->AsOutput()
@@ -1615,23 +1616,45 @@ PDNode *patterns::Matmul::operator()() {
   return matmul_out;
 }
 
+// MatmulV2: tensor * weight
+PDNode *patterns::MatmulV2Weight::operator()() {
+  auto matmul_v2_op =
+      pattern->NewNode(matmul_v2_op_repr())->assert_is_op("matmul_v2");
+
+  auto matmul_v2_in_x = pattern->NewNode(matmul_v2_in_x_repr())
+                            ->AsInput()
+                            ->assert_is_op_input("matmul_v2", "X");
+  auto matmul_v2_in_y = pattern->NewNode(matmul_v2_in_y_repr())
+                            ->AsInput()
+                            ->assert_is_persistable_var()  // Y is weight
+                            ->assert_is_op_input("matmul_v2", "Y");
+  auto matmul_v2_out = pattern->NewNode(matmul_v2_out_repr())
+                           ->AsOutput()
+                           ->assert_is_op_output("matmul_v2", "Out");
+
+  matmul_v2_op->LinksFrom({matmul_v2_in_x, matmul_v2_in_y})
+      .LinksTo({matmul_v2_out});
+  return matmul_v2_out;
+}
+
+// MatmulV2: tensor * tensor or tensor * weight
 PDNode *patterns::MatmulV2::operator()() {
-  auto matmul_op =
-      pattern->NewNode(matmul_op_repr())->assert_is_op("matmul_v2");
+  auto matmul_v2_op =
+      pattern->NewNode(matmul_v2_op_repr())->assert_is_op("matmul_v2");
 
-  auto matmul_in_x = pattern->NewNode(matmul_in_x_repr())
-                         ->AsInput()
-                         ->assert_is_op_input("matmul_v2", "X");
-  auto matmul_in_y = pattern->NewNode(matmul_in_y_repr())
-                         ->assert_is_persistable_var()
-                         ->AsInput()
-                         ->assert_is_op_input("matmul_v2", "Y");
-  auto matmul_out = pattern->NewNode(matmul_out_repr())
-                        ->AsOutput()
-                        ->assert_is_op_output("matmul_v2", "Out");
+  auto matmul_v2_in_x = pattern->NewNode(matmul_v2_in_x_repr())
+                            ->AsInput()
+                            ->assert_is_op_input("matmul_v2", "X");
+  auto matmul_v2_in_y = pattern->NewNode(matmul_v2_in_y_repr())
+                            ->AsInput()
+                            ->assert_is_op_input("matmul_v2", "Y");
+  auto matmul_v2_out = pattern->NewNode(matmul_v2_out_repr())
+                           ->AsOutput()
+                           ->assert_is_op_output("matmul_v2", "Out");
 
-  matmul_op->LinksFrom({matmul_in_x, matmul_in_y}).LinksTo({matmul_out});
-  return matmul_out;
+  matmul_v2_op->LinksFrom({matmul_v2_in_x, matmul_v2_in_y})
+      .LinksTo({matmul_v2_out});
+  return matmul_v2_out;
 }
 
 PDNode *patterns::Squeeze2Matmul::operator()() {

diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.h b/paddle/fluid/framework/ir/graph_pattern_detector.h
@@ -976,17 +976,28 @@ struct Matmul : public PatternBase {
   PATTERN_DECL_NODE(matmul_out);
 };
 
-// Matmul_v2 op
-// Forward pass for matmul_v2.
+// MatmulV2: tensor * weight
+struct MatmulV2Weight : public PatternBase {
+  MatmulV2Weight(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "matmul_v2_weight") {}
+
+  PDNode* operator()();
+  PATTERN_DECL_NODE(matmul_v2_in_x);
+  PATTERN_DECL_NODE(matmul_v2_in_y);
+  PATTERN_DECL_NODE(matmul_v2_op);
+  PATTERN_DECL_NODE(matmul_v2_out);
+};
+
+// MatmulV2: tensor * tensor or tensor * weight
 struct MatmulV2 : public PatternBase {
   MatmulV2(PDPattern* pattern, const std::string& name_scope)
       : PatternBase(pattern, name_scope, "matmul_v2") {}
 
   PDNode* operator()();
-  PATTERN_DECL_NODE(matmul_in_x);
-  PATTERN_DECL_NODE(matmul_in_y);
-  PATTERN_DECL_NODE(matmul_op);
-  PATTERN_DECL_NODE(matmul_out);
+  PATTERN_DECL_NODE(matmul_v2_in_x);
+  PATTERN_DECL_NODE(matmul_v2_in_y);
+  PATTERN_DECL_NODE(matmul_v2_op);
+  PATTERN_DECL_NODE(matmul_v2_out);
 };
 
 // Squeeze2 + Matmul