Skip to content

CANN: optimize the rope ops #15335

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 19, 2025
Merged

CANN: optimize the rope ops #15335

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0ff18e0
optimize rope ops
YangShuai52 Aug 14, 2025
930ee57
amendment
YangShuai52 Aug 15, 2025
682600c
delete trailing whitespace
YangShuai52 Aug 15, 2025
170d40e
change the variable name
YangShuai52 Aug 18, 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
170 changes: 105 additions & 65 deletions ggml/src/ggml-cann/aclnn_ops.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2154,87 +2154,128 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,

GGML_TENSOR_BINARY_OP_LOCALS

// theta_scale arange, [0,1,...,ne00/2 - 1]
int64_t theta_scale_length = ne00 / 2;
ggml_cann_pool_alloc theta_scale_allocator(ctx.pool(),
theta_scale_length * sizeof(float_t));
void* theta_scale_buffer = theta_scale_allocator.get();
int64_t theta_scale_ne[] = {theta_scale_length, 1, 1, 1};
size_t theta_scale_nb[] = {sizeof(float_t), sizeof(float_t), sizeof(float_t),
theta_scale_length * sizeof(float_t)};

aclTensor* acl_theta_scale_tensor =
ggml_cann_create_tensor(theta_scale_buffer, ACL_FLOAT, sizeof(float_t),
theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
float start = 0;
float step = 1;
float stop = ne00 / 2;
float n_elements = ne00 / 2;
aclnn_arange(ctx, acl_theta_scale_tensor, start, stop, step, n_elements);

// power
aclScalar* acl_theta_scale = aclCreateScalar(&theta_scale, aclDataType::ACL_FLOAT);
GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, acl_theta_scale, acl_theta_scale_tensor,
acl_theta_scale_tensor);

// freq_scale
if (freq_scale != 1) {
aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
}

// freq_factors
if (src2) {
aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
src2->data, ggml_cann_type_mapping(src2->type),
ggml_type_size(src2->type), theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
aclnn_div(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor);
ggml_cann_release_resources(ctx, acl_freq_factors_tensor);
}

// position
GGML_ASSERT(src1->type == GGML_TYPE_I32);
int64_t position_length = src1->ne[0];
int64_t position_ne[] = {1, 1, position_length, 1};
size_t position_nb[] = {sizeof(int32_t), sizeof(int32_t), sizeof(int32_t),
sizeof(int32_t) * position_length};
aclTensor* acl_position_tensor = ggml_cann_create_tensor(
src1->data, ggml_cann_type_mapping(src1->type),
ggml_type_size(src1->type), position_ne, position_nb, GGML_MAX_DIMS);

// power * position
int64_t theta_length = theta_scale_length * position_length;
ggml_cann_pool_alloc theta_allocator(ctx.pool(),
theta_length * sizeof(float_t));
void* theta_buffer = theta_allocator.get();

int64_t theta_ne[] = {theta_scale_length, 1, position_length, 1};
size_t theta_nb[GGML_MAX_DIMS];
theta_nb[0] = sizeof(float_t);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
theta_nb[i] = theta_nb[i - 1] * theta_ne[i - 1];
}
aclTensor* acl_theta_tensor =
ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float_t),
theta_ne, theta_nb, GGML_MAX_DIMS);
aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
acl_theta_tensor);

// sin/cos
ggml_cann_pool_alloc sin_allocator(ctx.pool(),
theta_length * sizeof(float_t));
void* sin_buffer = sin_allocator.get();
aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
sin_buffer, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
GGML_MAX_DIMS, ACL_FORMAT_ND);
aclnn_sin(ctx, acl_theta_tensor, acl_sin_tensor);

ggml_cann_pool_alloc cos_allocator(ctx.pool(),
theta_length * sizeof(float_t));
void* cos_buffer = cos_allocator.get();
aclTensor* acl_cos_tensor = ggml_cann_create_tensor(
cos_buffer, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
GGML_MAX_DIMS, ACL_FORMAT_ND);
aclnn_cos(ctx, acl_theta_tensor, acl_cos_tensor);
bool is_q = (std::strncmp(dst->name, "Qcur-", 5) == 0);
bool is_k = (std::strncmp(dst->name, "Kcur-", 5) == 0);
bool is_attention = is_q || is_k;

if(ctx.init_ptr == nullptr || !is_attention) {
// theta_scale arange, [0,1,...,ne00/2 - 1]
if(ctx.init_ptr != nullptr){
ACL_CHECK(aclrtFree(ctx.init_ptr));
}
ACL_CHECK(aclrtMalloc(&ctx.init_ptr,theta_scale_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
Copy link
Collaborator

@noemotiovon noemotiovon Aug 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There is a missing space after &ctx.init_ptr,.

Copy link
Contributor Author

@YangShuai52 YangShuai52 Aug 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you for your suggestion; it has been revised.


aclTensor* acl_theta_scale_tensor =
ggml_cann_create_tensor(ctx.init_ptr, ACL_FLOAT, sizeof(float_t),
theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
float start = 0;
float step = 1;
float stop = ne00 / 2;
float n_elements = ne00 / 2;
aclnn_arange(ctx, acl_theta_scale_tensor, start, stop, step, n_elements);

// power
aclScalar* acl_theta_scale = aclCreateScalar(&theta_scale, aclDataType::ACL_FLOAT);
GGML_CANN_CALL_ACLNN_OP(ctx, PowScalarTensor, acl_theta_scale, acl_theta_scale_tensor,
acl_theta_scale_tensor);

// freq_scale
if (freq_scale != 1) {
aclnn_muls(ctx, acl_theta_scale_tensor, freq_scale, nullptr, true);
}

// freq_factors
if (src2) {
aclTensor* acl_freq_factors_tensor = ggml_cann_create_tensor(
src2->data, ggml_cann_type_mapping(src2->type),
ggml_type_size(src2->type), theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);
aclnn_div(ctx, acl_theta_scale_tensor, acl_freq_factors_tensor);
ggml_cann_release_resources(ctx, acl_freq_factors_tensor);
}
// release
ggml_cann_release_resources(ctx, acl_theta_scale_tensor,acl_theta_scale);
}

if(ctx.sin_ptr == nullptr) {
int64_t theta_length = theta_scale_length * ctx.max_position_length;
ACL_CHECK(aclrtMalloc(&ctx.sin_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMalloc(&ctx.cos_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
}
if(position_length > ctx.max_position_length) {
ctx.max_position_length = position_length;
int64_t theta_length = theta_scale_length * ctx.max_position_length;
ACL_CHECK(aclrtFree(ctx.sin_ptr));
ACL_CHECK(aclrtFree(ctx.cos_ptr));
ACL_CHECK(aclrtMalloc(&ctx.sin_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
ACL_CHECK(aclrtMalloc(&ctx.cos_ptr, theta_length * sizeof(float_t), ACL_MEM_MALLOC_HUGE_FIRST));
}

bool is_fisrt_layer = (std::strncmp(dst->name, "Qcur-0", GGML_MAX_NAME) == 0);

if(is_fisrt_layer || !is_attention) {

aclTensor* acl_theta_scale_tensor =
ggml_cann_create_tensor(ctx.init_ptr, ACL_FLOAT, sizeof(float_t),
theta_scale_ne, theta_scale_nb, GGML_MAX_DIMS);

// position
aclTensor* acl_position_tensor = ggml_cann_create_tensor(
src1->data, ggml_cann_type_mapping(src1->type),
ggml_type_size(src1->type), position_ne, position_nb, GGML_MAX_DIMS);

// power * position
int64_t theta_length = theta_scale_length * position_length;
ggml_cann_pool_alloc theta_allocator(ctx.pool(),
theta_length * sizeof(float_t));
void* theta_buffer = theta_allocator.get();

aclTensor* acl_theta_tensor =
ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float_t),
theta_ne, theta_nb, GGML_MAX_DIMS);
aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
acl_theta_tensor);

// sin/cos
aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
ctx.sin_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
GGML_MAX_DIMS, ACL_FORMAT_ND);
aclnn_sin(ctx, acl_theta_tensor, acl_sin_tensor);

aclTensor* acl_cos_tensor = ggml_cann_create_tensor(
ctx.cos_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
GGML_MAX_DIMS, ACL_FORMAT_ND);
aclnn_cos(ctx, acl_theta_tensor, acl_cos_tensor);

// release
ggml_cann_release_resources(ctx, acl_theta_scale_tensor, acl_position_tensor,
acl_theta_tensor, acl_sin_tensor, acl_cos_tensor);
}

aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
ctx.sin_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
GGML_MAX_DIMS, ACL_FORMAT_ND);
aclTensor* acl_cos_tensor = ggml_cann_create_tensor(
ctx.cos_ptr, ACL_FLOAT, sizeof(float_t), theta_ne, theta_nb,
GGML_MAX_DIMS, ACL_FORMAT_ND);

// attn_factor
if (attn_factor != 1) {
aclnn_muls(ctx, acl_sin_tensor, attn_factor, nullptr, true);
Expand All @@ -2257,8 +2298,7 @@ static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
}

// release
ggml_cann_release_resources(ctx, acl_theta_scale_tensor, acl_position_tensor,
acl_theta_tensor, acl_sin_tensor, acl_cos_tensor, acl_theta_scale);
ggml_cann_release_resources(ctx, acl_sin_tensor, acl_cos_tensor);
}

#ifdef __cplusplus
Expand Down
13 changes: 13 additions & 0 deletions ggml/src/ggml-cann/common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -368,6 +368,10 @@ struct ggml_backend_cann_context {
std::string name; /**< Name of the device. */
std::string description; /**< Description of the device. */
aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
void* init_ptr = nullptr;
void* sin_ptr = nullptr;
void* cos_ptr = nullptr;
int64_t max_position_length = 200000;
Copy link
Collaborator

@noemotiovon noemotiovon Aug 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

A maximum prompt length of 200,000 is a bit excessive; let's initialize it to 65,536 here. And rename it to max_prompt_length.

Copy link
Contributor Author

@YangShuai52 YangShuai52 Aug 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thank you for your suggestion; it has been revised.

#ifdef USE_ACL_GRAPH
/// Cached CANN ACL graph used for executing the current ggml computation graph.
std::unique_ptr<ggml_cann_graph> cann_graph;
Expand Down Expand Up @@ -414,6 +418,15 @@ struct ggml_backend_cann_context {
ACL_CHECK(aclrtDestroyStream(streams[i]));
}
}
if(init_ptr != nullptr) {
ACL_CHECK(aclrtFree(init_ptr));
}
if(sin_ptr != nullptr) {
ACL_CHECK(aclrtFree(sin_ptr));
}
if(cos_ptr != nullptr) {
ACL_CHECK(aclrtFree(cos_ptr));
}
}

/**
Expand Down
Loading