Skip to content

修复paddle.incubate.nn.functional.fused_rotary_position_embedding的非法地址访问问题 #74347

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 4 commits into from
Aug 2, 2025
Merged

修复paddle.incubate.nn.functional.fused_rotary_position_embedding的非法地址访问问题 #74347

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
b3d0010
error 700:fused_rotary_position_embedding test=develop
LCStayingdullCircuit Jul 31, 2025
8e24b32
error 700:fused_rotary_position_embedding test=develop
LCStayingdullCircuit Jul 31, 2025
8a824c5
error 700:fused_rotary_position_embedding test=develop
LCStayingdullCircuit Aug 1, 2025
bdbf895
bugfix:fused_rotary_position_embedding test=develop
LCStayingdullCircuit Aug 1, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
27 changes: 25 additions & 2 deletions paddle/phi/kernels/fusion/gpu/fused_rope_kernel.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,13 +75,36 @@ void FusedRopeKernel(const Context& dev_ctx,
outs_data[0] = out_q->data<T>();
int num_inputs = 1;

if (out_k && out_k->numel() > 0) {
if (k) {
auto k_dims = k->dims();
auto k_batch_size = time_major ? k_dims[1] : k_dims[0];
PADDLE_ENFORCE_LE(
batch_size,
k_batch_size,
common::errors::InvalidArgument("The batch_size of q (%d) must be less "
"than or equal to k's (%d).",
batch_size,
k_batch_size));

dev_ctx.template Alloc<T>(out_k);
ins_data[num_inputs] = k->data<T>();
outs_data[num_inputs] = out_k->data<T>();
inputs_num_heads[num_inputs] = k->dims()[2];
num_inputs++;
}
if (out_v && out_v->numel() > 0) {

if (v) {
auto v_dims = v->dims();
auto v_batch_size = time_major ? v_dims[1] : v_dims[0];
PADDLE_ENFORCE_LE(
batch_size,
v_batch_size,
common::errors::InvalidArgument("The batch_size of q (%d) must be less "
"than or equal to v's (%d).",
batch_size,
v_batch_size));

dev_ctx.template Alloc<T>(out_v);
ins_data[num_inputs] = v->data<T>();
outs_data[num_inputs] = out_v->data<T>();
inputs_num_heads[num_inputs] = v->dims()[2];
Expand Down
Loading