Skip to content

[BIG tensor] fix cuda error of layer_norm #74404

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Aug 5, 2025
Merged

[BIG tensor] fix cuda error of layer_norm #74404

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion paddle/phi/kernels/funcs/layer_norm_impl.cu.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -550,7 +550,9 @@ __inline__ __device__ void cuLoadAddStridedInputs(const int64_t i1_block,
for (int k = 0; k < VPT; ++k) {
const int i2 = i2_off + k;
const int64_t load_idx = i1 * n2 + i2;
const int write_idx = thr_load_row_off * row_stride + thr_load_col_off + k;
const int64_t write_idx =
static_cast<int64_t>(thr_load_row_off) * row_stride + thr_load_col_off +
k;
if (i2 < n2) {
U curr_input = static_cast<U>(input[load_idx]);
U curr_dout = static_cast<U>(dout[load_idx]);
Expand Down
22 changes: 11 additions & 11 deletions paddle/phi/kernels/gpu/layer_norm_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -159,7 +159,7 @@ struct LayerNormDataWriter {
temp_dst[j] = static_cast<T>((buffer[i * VecSize + j] - row_mean) *
row_inv_var);
}
v_dst[threadIdx.x + blockDim.x * i] = temp_dst;
v_dst[threadIdx.x + static_cast<int64_t>(blockDim.x) * i] = temp_dst;
}
} else {
const VecScaleT *__restrict__ v_scale =
Expand All @@ -168,7 +168,7 @@ struct LayerNormDataWriter {
reinterpret_cast<const VecScaleT *__restrict__>(bias);
if (valid_scale && valid_bias) {
for (int i = 0; i < write_times; ++i) {
int idx = threadIdx.x + blockDim.x * i;
int64_t idx = threadIdx.x + static_cast<int64_t>(blockDim.x) * i;
VecT temp_dst;
VecScaleT temp_v_scale = v_scale[idx];
VecScaleT temp_v_bias = v_bias[idx];
Expand All @@ -184,7 +184,7 @@ struct LayerNormDataWriter {
} else {
if (valid_scale) {
for (int i = 0; i < write_times; ++i) {
int idx = threadIdx.x + blockDim.x * i;
int64_t idx = threadIdx.x + static_cast<int64_t>(blockDim.x) * i;
VecT temp_dst;
VecScaleT temp_v_scale = v_scale[idx];
#pragma unroll
Expand Down Expand Up @@ -232,27 +232,27 @@ struct LayerNormDataWriter<T, U, IsSameType, 1> {
if ((!valid_scale) && (!valid_bias)) {
if (threadIdx.x < last_tid_idx) {
for (int i = 0; i < cols_this_thread; ++i) {
row_dst[threadIdx.x + last_tid_idx * i] =
row_dst[threadIdx.x + static_cast<int64_t>(last_tid_idx) * i] =
(buffer[i] - row_mean) * row_inv_var;
}
} else {
for (int i = 0; i < cols_this_thread; ++i) {
row_dst[last_tid_idx * write_times + i] =
row_dst[static_cast<int64_t>(last_tid_idx) * write_times + i] =
(buffer[i] - row_mean) * row_inv_var;
}
}
} else if (valid_scale && valid_bias) {
if (threadIdx.x < last_tid_idx) {
for (int i = 0; i < cols_this_thread; ++i) {
int idx = threadIdx.x + last_tid_idx * i;
int64_t idx = threadIdx.x + static_cast<int64_t>(last_tid_idx) * i;
row_dst[idx] =
static_cast<T>(static_cast<U>(scale[idx]) *
(buffer[i] - row_mean) * row_inv_var +
static_cast<U>(bias[idx]));
}
} else {
for (int i = 0; i < cols_this_thread; ++i) {
int idx = last_tid_idx * write_times + i;
int64_t idx = static_cast<int64_t>(last_tid_idx) * write_times + i;
row_dst[idx] =
static_cast<T>(static_cast<U>(scale[idx]) *
(buffer[i] - row_mean) * row_inv_var +
Expand All @@ -263,27 +263,27 @@ struct LayerNormDataWriter<T, U, IsSameType, 1> {
if (valid_scale) {
if (threadIdx.x < last_tid_idx) {
for (int i = 0; i < cols_this_thread; ++i) {
int idx = threadIdx.x + last_tid_idx * i;
int64_t idx = threadIdx.x + static_cast<int64_t>(last_tid_idx) * i;
row_dst[idx] = static_cast<T>(static_cast<U>(scale[idx]) *
(buffer[i] - row_mean) * row_inv_var);
}
} else {
for (int i = 0; i < cols_this_thread; ++i) {
int idx = last_tid_idx * write_times + i;
int64_t idx = static_cast<int64_t>(last_tid_idx) * write_times + i;
row_dst[idx] = static_cast<T>(static_cast<U>(scale[idx]) *
(buffer[i] - row_mean) * row_inv_var);
}
}
} else {
if (threadIdx.x < last_tid_idx) {
for (int i = 0; i < cols_this_thread; ++i) {
int idx = threadIdx.x + last_tid_idx * i;
int64_t idx = threadIdx.x + static_cast<int64_t>(last_tid_idx) * i;
row_dst[idx] = static_cast<T>((buffer[i] - row_mean) * row_inv_var +
static_cast<U>(bias[idx]));
}
} else {
for (int i = 0; i < cols_this_thread; ++i) {
int idx = last_tid_idx * write_times + i;
int64_t idx = static_cast<int64_t>(last_tid_idx) * write_times + i;
row_dst[idx] = static_cast<T>((buffer[i] - row_mean) * row_inv_var +
static_cast<U>(bias[idx]));
}
Expand Down