Merge pull request #6 from xymyeah/paddlebox

fix fused_seqpool_cvm_op_xpu op bug
jack603047588 · Sep 13, 2023 · de6dd8f · de6dd8f
2 parents 495198b + 59b3ab2
commit de6dd8f
Showing 1 changed file with 3 additions and 1 deletion.
diff --git a/paddle/fluid/operators/fused/fused_seqpool_cvm_op_xpu.cc b/paddle/fluid/operators/fused/fused_seqpool_cvm_op_xpu.cc
@@ -74,6 +74,7 @@ class FusedSeqpoolCVMOpXPUKernel : public framework::OpKernel<T> {
     auto clk_coeff = ctx.Attr<float>("clk_coeff");
     auto threshold = ctx.Attr<float>("threshold");
     auto cvm_offset = ctx.Attr<int>("cvm_offset");
+    auto embed_thres_size = ctx.Attr<int>("embed_thres_size");
 
     auto x0_lod = ins[0]->lod();
     auto x0_dims = ins[0]->dims();
@@ -98,12 +99,13 @@ class FusedSeqpoolCVMOpXPUKernel : public framework::OpKernel<T> {
     phi::Place l3_place = ctx.template device_context<DeviceContext>().GetL3Place();
     int w = ins[0]->numel() / x0_dims[0];
     if(use_cvm) {
+      if(clk_filter) w = w - 1;
       PADDLE_ENFORCE_EQ(y_dims[1] % w, 0,
                         paddle::platform::errors::InvalidArgument(
                             "The output of dims[1] should be dividable of w"));
     }
     else{
-      PADDLE_ENFORCE_EQ(y_dims[1] % (w-2), 0,
+      PADDLE_ENFORCE_EQ(y_dims[1] % (w - cvm_offset - embed_thres_size), 0,
                   paddle::platform::errors::InvalidArgument(
                       "The output of dims[1] should be dividable of (w-2)"));
     }