提交 ff112813 编写于 作者: P phlrain

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

...@@ -78,12 +78,6 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> { ...@@ -78,12 +78,6 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
"The numel of 'pad_value' can only be 1 or be equal to the " "The numel of 'pad_value' can only be 1 or be equal to the "
"'step_width'."); "'step_width'.");
if (!norm_by_times && seq_num == 1UL && pad_seq_len == max_seq_len) {
TensorCopy(seq_tensor, context.GetPlace(), context, pad_tensor);
pad_tensor->Resize(pad_tensor_dims);
return;
}
const int kBlockSize = 512; const int kBlockSize = 512;
/* At least use 32 threads to copy sequence_width elements, /* At least use 32 threads to copy sequence_width elements,
...@@ -129,12 +123,13 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> { ...@@ -129,12 +123,13 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len, CheckDims(seq_tensor_dims, pad_tensor_dims, seq_offsets, pad_seq_len,
step_width, layout); step_width, layout);
/*
if (!norm_by_times && seq_num == 1UL && pad_seq_len == max_seq_len) { if (!norm_by_times && seq_num == 1UL && pad_seq_len == max_seq_len) {
TensorCopy(pad_tensor, context.GetPlace(), context, seq_tensor); TensorCopy(pad_tensor, context.GetPlace(), context, seq_tensor);
seq_tensor->Resize(seq_tensor_dims); seq_tensor->Resize(seq_tensor_dims);
return; return;
} }
*/
const int kBlockSize = 512; const int kBlockSize = 512;
......
...@@ -290,8 +290,10 @@ class MatMulOp : public framework::OperatorWithKernel { ...@@ -290,8 +290,10 @@ class MatMulOp : public framework::OperatorWithKernel {
context->Attrs().Get<bool>("transpose_Y")); context->Attrs().Get<bool>("transpose_Y"));
PADDLE_ENFORCE_EQ(mat_dim_x.width_, mat_dim_y.height_); PADDLE_ENFORCE_EQ(mat_dim_x.width_, mat_dim_y.height_);
PADDLE_ENFORCE(mat_dim_x.batch_size_ == mat_dim_y.batch_size_ || if (context->IsRuntime()) {
mat_dim_x.batch_size_ == 0 || mat_dim_y.batch_size_ == 0); PADDLE_ENFORCE(mat_dim_x.batch_size_ == mat_dim_y.batch_size_ ||
mat_dim_x.batch_size_ == 0 || mat_dim_y.batch_size_ == 0);
}
std::vector<int64_t> dim_out; std::vector<int64_t> dim_out;
if (mat_dim_x.batch_size_ != 0) { if (mat_dim_x.batch_size_ != 0) {
dim_out = framework::vectorize(dim_x); dim_out = framework::vectorize(dim_x);
......
...@@ -4901,6 +4901,9 @@ def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None): ...@@ -4901,6 +4901,9 @@ def matmul(x, y, transpose_x=False, transpose_y=False, alpha=1.0, name=None):
if len(y_shape) > 2 and len(x_shape) > 2: if len(y_shape) > 2 and len(x_shape) > 2:
for i, dim_x in enumerate(x_shape[:-2]): for i, dim_x in enumerate(x_shape[:-2]):
# don't check neg shape
if dim_x < 0 or y_shape[i] < 0:
continue
if dim_x != y_shape[i]: if dim_x != y_shape[i]:
raise ValueError("Invalid inputs for matmul. x(%s), y(%s)" % raise ValueError("Invalid inputs for matmul. x(%s), y(%s)" %
(x.shape, y.shape)) (x.shape, y.shape))
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册