PaddlePaddle
diff --git a/‎paddle/fluid/ir_adaptor/translator/op_translator.cc
Copy file name to clipboard
+60Lines changed: 60 additions & 0 deletions b/‎paddle/fluid/ir_adaptor/translator/op_translator.cc
Copy file name to clipboard
+60Lines changed: 60 additions & 0 deletions
diff --git a/‎paddle/fluid/pir/dialect/op_generator/op_verify_gen.py
Copy file name to clipboard
+9-4Lines changed: 9 additions & 4 deletions b/‎paddle/fluid/pir/dialect/op_generator/op_verify_gen.py
Copy file name to clipboard
+9-4Lines changed: 9 additions & 4 deletions
diff --git a/‎paddle/fluid/pir/dialect/op_generator/ops_api_gen.py
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions b/‎paddle/fluid/pir/dialect/op_generator/ops_api_gen.py
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions
diff --git a/‎paddle/fluid/pir/dialect/operator/ir/ops.yaml
Copy file name to clipboard
+12-1Lines changed: 12 additions & 1 deletion b/‎paddle/fluid/pir/dialect/operator/ir/ops.yaml
Copy file name to clipboard
+12-1Lines changed: 12 additions & 1 deletion
diff --git a/‎paddle/fluid/pir/dialect/operator/ir/ops_backward.yaml
Copy file name to clipboard
+20Lines changed: 20 additions & 0 deletions b/‎paddle/fluid/pir/dialect/operator/ir/ops_backward.yaml
Copy file name to clipboard
+20Lines changed: 20 additions & 0 deletions
diff --git a/‎paddle/phi/api/yaml/op_compat.yaml
Copy file name to clipboard
+33Lines changed: 33 additions & 0 deletions b/‎paddle/phi/api/yaml/op_compat.yaml
Copy file name to clipboard
+33Lines changed: 33 additions & 0 deletions
@@ -1808,6 +1808,65 @@ struct LegacySetValueDispatcher : public OpTranscriber {
   }
 };
 
+struct FusedFeedForwardOpTranscriber : public OpTranscriber {
+  void HandleNonexistentAttribute(pir::IrContext* ctx,
+                                  pir::AttributeMap* attribute_map,
+                                  const OpAttributeInfo& info) override {
+    if (info.name == "ln1_epsilon") {
+      (*attribute_map)[info.name] = pir::FloatAttribute::get(ctx, 1e-5f);
+    } else if (info.name == "ln2_epsilon") {
+      (*attribute_map)[info.name] = pir::FloatAttribute::get(ctx, 1e-5f);
+    } else if (info.name == "act_method") {
+      (*attribute_map)[info.name] = pir::StrAttribute::get(ctx, "gelu");
+    } else if (info.name == "dropout1_prob") {
+      (*attribute_map)[info.name] = pir::FloatAttribute::get(ctx, .5f);
+    } else if (info.name == "dropout2_prob") {
+      (*attribute_map)[info.name] = pir::FloatAttribute::get(ctx, .5f);
+    } else if (info.name == "dropout1_implementation") {
+      (*attribute_map)[info.name] =
+          pir::StrAttribute::get(ctx, "downgrade_in_infer");
+    } else if (info.name == "dropout2_implementation") {
+      (*attribute_map)[info.name] =
+          pir::StrAttribute::get(ctx, "downgrade_in_infer");
+    } else if (info.name == "is_test") {
+      (*attribute_map)[info.name] = pir::BoolAttribute::get(ctx, false);
+    } else if (info.name == "dropout1_fix_seed") {
+      (*attribute_map)[info.name] = pir::BoolAttribute::get(ctx, false);
+    } else if (info.name == "dropout2_fix_seed") {
+      (*attribute_map)[info.name] = pir::BoolAttribute::get(ctx, false);
+    } else if (info.name == "dropout1_seed_val") {
+      (*attribute_map)[info.name] = pir::Int32Attribute::get(ctx, false);
+    } else if (info.name == "dropout2_seed_val") {
+      (*attribute_map)[info.name] = pir::Int32Attribute::get(ctx, false);
+    } else if (info.name == "add_residual") {
+      (*attribute_map)[info.name] = pir::BoolAttribute::get(ctx, true);
+    } else if (info.name == "ring_id") {
+      (*attribute_map)[info.name] = pir::Int32Attribute::get(ctx, -1);
+    }
+  }
+
+  void RecordOpResultMapping(pir::IrContext* ctx,
+                             TranslationContext* param_map,
+                             const OpDesc& op_desc,
+                             pir::Operation* operation,
+                             const OpOutputMapping& arg_to_idx) override {
+    OpTranscriber::RecordOpResultMapping(
+        ctx, param_map, op_desc, operation, arg_to_idx);
+    if (op_desc.HasOutput("Out")) {
+      const auto& output_vars = op_desc.Output("Out");
+      IR_ENFORCE(output_vars.size() == 1,
+                 "Expected op[%s]'s Out has only 1 var but got %s",
+                 op_desc.Type(),
+                 output_vars.size());
+      auto output_var = output_vars[0];
+      auto fused_feedforward_op =
+          operation->dyn_cast<dialect::FusedFeedforwardOp>();
+      (*param_map)[output_var] =
+          VariableDefiningInfo{fused_feedforward_op.out()};
+    }
+  }
+};
+
 OpTranslator::OpTranslator() {
   pir::IrContext* ctx = pir::IrContext::Instance();
   ctx->GetOrRegisterDialect<paddle::dialect::OperatorDialect>();
@@ -1821,6 +1880,7 @@ OpTranslator::OpTranslator() {
   special_handlers["fetch"] = FetchOpTranscriber();
   special_handlers["fetch_v2"] = FetchOpTranscriber();
   special_handlers["fill_constant"] = FillConstantTranscriber();
+  special_handlers["fused_feedforward"] = FusedFeedForwardOpTranscriber();
   special_handlers["grad_add"] = GradAddOpTranscriber();
   special_handlers["increment"] = IncrementOpTranscriber();
   special_handlers["lookup_table_v2"] = EmbeddingOpTranscriber();
 
@@ -72,11 +72,16 @@
     }}
   }}"""
 ATTRIBUTE_CHECK_TEMPLATE = """
-  PADDLE_ENFORCE(attributes.count("{attribute_name}")>0 && attributes.at("{attribute_name}").isa<{standard}>(),
-                 phi::errors::PreconditionNotMet("Type of attribute: {attribute_name} is not right."));"""
+  PADDLE_ENFORCE(attributes.count("{attribute_name}")>0,
+                 phi::errors::PreconditionNotMet("{attribute_name} does not exist."));
+  PADDLE_ENFORCE(attributes.at("{attribute_name}").isa<{standard}>(),
+                 phi::errors::PreconditionNotMet("Type of attribute: {attribute_name} is not {standard}."));
+"""
 ATTRIBUTE_VECTOR_CHECK_TEMPLATE = """
-  PADDLE_ENFORCE(attributes.count("{attribute_name}")>0 && attributes.at("{attribute_name}").isa<pir::ArrayAttribute>(),
-                 phi::errors::PreconditionNotMet("Type of attribute: {attribute_name} is not right."));
+  PADDLE_ENFORCE(attributes.count("{attribute_name}")>0,
+                 phi::errors::PreconditionNotMet("{attribute_name} does not exist."));
+  PADDLE_ENFORCE(attributes.at("{attribute_name}").isa<pir::ArrayAttribute>(),
+                 phi::errors::PreconditionNotMet("Type of attribute: {attribute_name} is not pir::ArrayAttribute."));
   for (size_t i = 0; i < attributes.at("{attribute_name}").dyn_cast<pir::ArrayAttribute>().size(); i++) {{
     PADDLE_ENFORCE(attributes.at("{attribute_name}").dyn_cast<pir::ArrayAttribute>().at(i).isa<{standard}>(),
                    phi::errors::PreconditionNotMet("Type of attribute: {attribute_name} is not right."));
 
@@ -96,6 +96,7 @@
     'c_allgather',
     'seed',
     "fused_attention",
+    "fused_feedforward",
 ]
 
 
 
@@ -184,4 +184,15 @@
     func: fused_attention
   infer_meta:
     func: FusedAttentionInferMeta
-  optional: cache_kv, ln_scale, ln_bias, qkv_bias, src_mask, out_linear_bias, ln_scale_2, ln_bias_2
+  optional: cache_kv, ln_scale, ln_bias, qkv_bias, src_mask, out_linear_bias, ln_scale_2, ln_bias_2, ln_mean_2
+  backward: fused_attention_grad
+
+- op: fused_feedforward
+  args: (Tensor x, Tensor dropout1_seed, Tensor dropout2_seed, Tensor linear1_weight, Tensor linear1_bias, Tensor linear2_weight, Tensor linear2_bias, Tensor ln1_scale, Tensor ln1_bias, Tensor ln2_scale, Tensor ln2_bias, bool pre_layer_norm, float ln1_epsilon, float ln2_epsilon, str act_method, float dropout1_prob, float dropout2_prob, str dropout1_implementation, str dropout2_implementation, bool is_test, bool dropout1_fix_seed, bool dropout2_fix_seed, int dropout1_seed_val, int dropout2_seed_val, bool add_residual, int ring_id)
+  output: Tensor(out), Tensor(dropout1_mask), Tensor(dropout2_mask), Tensor(ln1_mean), Tensor(ln1_variance), Tensor(ln2_mean), Tensor(ln2_variance), Tensor(linear1_out), Tensor(ln1_out), Tensor(dropout1_out), Tensor(dropout2_out)
+  kernel:
+    func: fused_feedforward
+  infer_meta:
+    func: FusedFeedForwardInferMeta
+  optional: dropout1_seed, dropout2_seed, linear1_bias, linear2_bias, ln1_scale, ln1_bias, ln2_scale, ln2_bias, ln2_mean, ln2_variance, ln1_mean, ln1_variance, ln1_out
+  backward: fused_feedforward_grad
@@ -1,3 +1,23 @@
+- backward_op : fused_attention_grad
+  args : (Tensor out_grad, Tensor x, Tensor qkv_weight, Tensor qkv_bias, Tensor qkv_bias_out, Tensor src_mask, Tensor src_mask_out, Tensor out_linear_weight, Tensor out_linear_bias, Tensor ln_scale, Tensor ln_bias, Tensor ln_scale_2, Tensor ln_bias_2, Tensor ln_out, Tensor ln_mean, Tensor ln_var, Tensor ln_mean_2, Tensor ln_var_2, Tensor bias_dropout_residual_out, Tensor qkv_out, Tensor transpose_out_2, Tensor qk_out, Tensor qktv_out, Tensor softmax_out, Tensor attn_dropout_mask_out, Tensor attn_dropout_out, Tensor fmha_out, Tensor out_linear_out, Tensor dropout_mask_out, int num_heads, bool transpose_qkv_wb, bool pre_layer_norm, float epsilon, float attn_dropout_rate, bool is_test, bool attn_dropout_fix_seed, int attn_dropout_seed, str attn_dropout_implementation, float dropout_rate, bool dropout_fix_seed, int dropout_seed, str dropout_implementation, float ln_epsilon, bool add_residual, int ring_id)
+  output : Tensor(qkv_bias_grad), Tensor(qkv_bias_out_grad), Tensor(src_mask_out_grad), Tensor(out_linear_bias_grad), Tensor(ln_scale_grad), Tensor(ln_bias_grad), Tensor(ln_scale_2_grad), Tensor(ln_bias_2_grad), Tensor(x_grad), Tensor(qkv_weight_grad), Tensor(out_linear_weight_grad), Tensor(ln_out_grad), Tensor(bias_dropout_residual_out_grad), Tensor(qkv_out_grad), Tensor(qktv_out_grad), Tensor(transpose_out_2_grad), Tensor(qk_out_grad), Tensor(softmax_out_grad), Tensor(attn_dropout_out_grad), Tensor(fmha_out_grad), Tensor(out_linear_out_grad)
+  infer_meta:
+    func: FusedAttentionGradInferMeta
+  kernel:
+    func: fused_attention_grad
+    data_type : x
+  optional: ln_scale, ln_bias, qkv_bias, src_mask, out_linear_bias, ln_scale_2, ln_bias_2, qkv_bias_grad, qkv_bias_out_grad, src_mask_out_grad, out_linear_bias_grad, ln_scale_grad, ln_bias_grad, ln_scale_2_grad, ln_bias_2_grad, ln_out_grad, bias_dropout_residual_out_grad, ln_out, ln_mean, ln_var,  ln_mean_2, ln_var_2, bias_dropout_residual_out, qkv_bias, qkv_bias_out, src_mask, src_mask_out, out_linear_bias
+  no_need_buffer: qkv_bias_out, qkv_out, qk_out, qktv_out, out_linear_out, src_mask
+
+- backward_op : fused_feedforward_grad
+  args : (Tensor out_grad, Tensor x, Tensor linear1_weight, Tensor linear1_bias, Tensor linear2_weight, Tensor dropout1_mask, Tensor dropout2_mask, Tensor linear1_out, Tensor dropout1_out, Tensor dropout2_out, Tensor ln1_scale, Tensor ln1_bias, Tensor ln1_out, Tensor ln1_mean, Tensor ln1_variance, Tensor ln2_scale, Tensor ln2_bias, Tensor ln2_mean, Tensor ln2_variance, Tensor linear2_bias, bool pre_layer_norm, float ln1_epsilon, float ln2_epsilon, str act_method, float dropout1_prob, float dropout2_prob, str dropout1_implementation, str dropout2_implementation, bool is_test, bool dropout1_fix_seed, bool dropout2_fix_seed, int dropout1_seed_val, int dropout2_seed_val, bool add_residual, int ring_id)
+  output : Tensor(x_grad), Tensor(ln1_scale_grad), Tensor(ln1_bias_grad), Tensor(ln2_scale_grad), Tensor(ln2_bias_grad), Tensor(linear1_weight_grad), Tensor(linear1_bias_grad), Tensor(linear2_weight_grad), Tensor(linear2_bias_grad)
+  infer_meta:
+    func: FusedFeedForwardGradInferMeta
+  kernel:
+    func: fused_feedforward_grad
+  optional: linear1_bias, linear2_bias, ln1_scale, ln1_bias, ln1_out, ln1_mean, ln1_variance, ln2_scale, ln2_bias, ln2_mean, ln2_variance, dropout2_out, ln1_scale_grad, ln1_bias_grad, ln2_scale_grad, ln2_bias_grad,  linear2_bias_grad
+
 - backward_op : set_value_grad
   args : (Tensor out_grad, Tensor values, int64_t[] starts, int64_t[] ends, int64_t[] steps, int64_t[] axes, int64_t[] decrease_axes, int64_t[] none_axes)
   output : Tensor(x_grad), Tensor(values_grad)
 
@@ -1182,6 +1182,7 @@
       support_tensor : true
 
 - op : fused_attention
+  backward: fused_attention_grad
   inputs:
     x: X
     ln_scale: LnScale
@@ -1254,6 +1255,38 @@
     attrs : [bool use_cudnn = false, float fuse_alpha = 0.0f, float fuse_beta = 0.0f, float Scale_in = 1.0f,
              float Scale_out = 1.0f, float Scale_in_eltwise = 1.0f, 'float[] Scale_weights = {1.0f}']
 
+- op : fused_feedforward
+  backward: fused_feedforward_grad
+  inputs:
+    x: X
+    dropout1_seed: Dropout1Seed
+    dropout2_seed: Dropout2Seed
+    linear1_weight: Linear1Weight
+    linear1_bias: Linear1Bias
+    linear2_weight: Linear2Weight
+    linear2_bias: Linear2Bias
+    ln1_scale: Ln1Scale
+    ln1_bias: Ln1Bias
+    ln2_scale: Ln2Scale
+    ln2_bias: Ln2Bias
+  attrs:
+    dropout1_seed_val: dropout1_seed
+    dropout2_seed_val: dropout2_seed
+    dropout1_prob: dropout1_rate
+    dropout2_prob: dropout2_rate
+  outputs:
+    out: Out
+    dropout1_mask: Dropout1Mask
+    dropout2_mask: Dropout2Mask
+    ln1_mean: Ln1Mean
+    ln1_variance: Ln1Variance
+    ln2_mean: Ln2Mean
+    ln2_variance: Ln2Variance
+    linear1_out: Linear1Out
+    ln1_out: Ln1Out
+    dropout1_out: Dropout1Out
+    dropout2_out: Dropout2Out
+
 - op : fused_transpose
   extra :
     attrs : [str data_format = "AnyLayout"]
Original file line number	Diff line number	Diff line change
`@@ -96,6 +96,7 @@`
`96`	`96`	`'c_allgather',`
`97`	`97`	`'seed',`
`98`	`98`	`"fused_attention",`
	`99`	`+ "fused_feedforward",`
`99`	`100`	`]`
`100`	`101`
`101`	`102`