Skip to content
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.

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

[Ops] RoPE kernel support theta input #8440

Merged
merged 3 commits into from
May 23, 2024
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
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
12 changes: 9 additions & 3 deletions csrc/generation/fused_get_rope.cu
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,7 @@ __global__ __launch_bounds__(kBlockSize) void fused_get_rotary_embedding_neox(co
const int32_t prompt_num,
const float inv_head_dim,
const int32_t elem_cnt,
const float theta,
float* rope_embedding) {
/*
In Naive implementation, it will stacks [freqs, freqs]
Expand All @@ -63,7 +64,7 @@ __global__ __launch_bounds__(kBlockSize) void fused_get_rotary_embedding_neox(co
const int64_t position_offset = bsz_idx * max_position_seq_length + seq_idx + prompt_num;
const int32_t half_head_idx = (idx % half_head_dim) * PackSize;
const float exponent_factor = -static_cast<float>(half_head_idx) * inv_head_dim; // * inv_head_dim equals to / head_dim.
const float inv_freq_val = powf(10000.0f, exponent_factor);
const float inv_freq_val = powf(theta, exponent_factor);
const float freqs_val = static_cast<float>(position_ids[position_offset]) * inv_freq_val;
const float cos_embedding_val = cos(freqs_val);
const float sin_embedding_val = sin(freqs_val);
Expand Down Expand Up @@ -100,6 +101,7 @@ __global__ __launch_bounds__(kBlockSize) void fused_get_rotary_embedding(const i
const int32_t prompt_num,
const float inv_head_dim,
const int32_t elem_cnt,
const float theta,
float* rope_embedding) {
/*
In Naive implementation, it will stacks [freqs, freqs]
Expand All @@ -119,7 +121,7 @@ __global__ __launch_bounds__(kBlockSize) void fused_get_rotary_embedding(const i
const int64_t position_offset = bsz_idx * max_position_seq_length + seq_idx + prompt_num;
const int32_t half_head_idx = (idx % half_head_dim) * PackSize;
const float exponent_factor = -static_cast<float>(half_head_idx) * inv_head_dim; // * inv_head_dim equals to / head_dim.
const float inv_freq_val = powf(10000.0f, exponent_factor);
const float inv_freq_val = powf(theta, exponent_factor);
const float freqs_val = static_cast<float>(position_ids[position_offset]) * inv_freq_val;
const float cos_embedding_val = cos(freqs_val);
const float sin_embedding_val = sin(freqs_val);
Expand All @@ -145,6 +147,7 @@ std::vector<paddle::Tensor> GetRoPE(const paddle::Tensor& input_ids,
const paddle::Tensor& position_ids,
const paddle::Tensor& head_dim_shape_tensor,
int prompt_num,
float theta,
bool use_neox) {
const int64_t batch_size = input_ids.shape()[0];
const int64_t max_seq_length = input_ids.shape()[1];
Expand All @@ -170,6 +173,7 @@ std::vector<paddle::Tensor> GetRoPE(const paddle::Tensor& input_ids,
prompt_num,
inv_head_dim,
elem_cnt,
theta,
reinterpret_cast<float*>(rotary_embedding.data<float>()));
} else {
fused_get_rotary_embedding<<<grid_size, kBlockSize, 0, cu_stream>>> (
Expand All @@ -181,6 +185,7 @@ std::vector<paddle::Tensor> GetRoPE(const paddle::Tensor& input_ids,
prompt_num,
inv_head_dim,
elem_cnt,
theta,
reinterpret_cast<float*>(rotary_embedding.data<float>()));
}
return {rotary_embedding};
Expand Down Expand Up @@ -209,7 +214,8 @@ PD_BUILD_OP(fused_get_rotary_embedding)
.Inputs({"input_ids", "position_ids", "head_dim_shape_tensor"})
.Outputs({"rotary_embedding"})
.Attrs({"prompt_num: int",
"theta: float",
"use_neox: bool"})
.SetKernelFn(PD_KERNEL(GetRoPE))
.SetInferShapeFn(PD_INFER_SHAPE(GetRoPEInferShape))
.SetInferDtypeFn(PD_INFER_DTYPE(GetRoPEInferDtype));
.SetInferDtypeFn(PD_INFER_DTYPE(GetRoPEInferDtype));
3 changes: 2 additions & 1 deletion paddlenlp/experimental/transformers/llama/modeling.py
Original file line number Diff line number Diff line change
Expand Up @@ -432,10 +432,11 @@
seq_lens = seq_len_decoder if is_decoder else seq_len_encoder

position_offset = 0
theta = 10000.0

Check warning on line 435 in paddlenlp/experimental/transformers/llama/modeling.py

View check run for this annotation

Codecov / codecov/patch

paddlenlp/experimental/transformers/llama/modeling.py#L435

Added line #L435 was not covered by tests
if not is_decoder and pre_caches is not None:
position_offset = 128
new_rope = fused_get_rotary_embedding(
input_ids, position_ids, self.head_dim_shape_tensor, position_offset, True
input_ids, position_ids, self.head_dim_shape_tensor, position_offset, theta, True
)

with dy2st_nocheck_guard_context():
Expand Down
3 changes: 2 additions & 1 deletion paddlenlp/experimental/transformers/qwen/modeling.py
Original file line number Diff line number Diff line change
Expand Up @@ -320,11 +320,12 @@
seq_lens = seq_len_decoder if is_decoder else seq_len_encoder

position_offset = 0
theta = 10000.0

Check warning on line 323 in paddlenlp/experimental/transformers/qwen/modeling.py

View check run for this annotation

Codecov / codecov/patch

paddlenlp/experimental/transformers/qwen/modeling.py#L323

Added line #L323 was not covered by tests
if not is_decoder and pre_caches is not None:
position_offset = 128

new_rope = fused_get_rotary_embedding(
input_ids, position_ids, self.head_dim_shape_tensor, position_offset, True
input_ids, position_ids, self.head_dim_shape_tensor, position_offset, theta, True
)

with dy2st_nocheck_guard_context():
Expand Down
Loading