CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-13 10:57:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

CANN: Implement GLU ops (#14884)

Browse Source 
Implement REGLU, GEGLU, SWIGLU ops according to #14158
This commit is contained in:
hipudding
2025-07-26 17:56:18 +08:00
committed by GitHub
parent 9b8f3c6c77
commit 11dd5a44eb
4 changed files with 194 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

120
ggml/src/ggml-cann/aclnn_ops.h
View File

@@ -1098,7 +1098,7 @@ void ggml_cann_binary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
* @param dst The destination tensor. Its src[0] is treated as the input tensor.
*/
template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
void ggml_cann_unary_op(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
void ggml_cann_op_unary(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
aclTensor* acl_src = ggml_cann_create_tensor(src);
@@ -1109,49 +1109,125 @@ template <void unary_op(ggml_backend_cann_context&, aclTensor*, aclTensor*)>
}
/**
* @brief Applies a unary operation to a ggml tensor using the CANN backend.
* @brief Applies a unary operation to a ggml tensor using the CANN backend.
*
* @details This function performs a unary operation on the input tensor using
* a user-provided lambda or callable object `unary_op`, which accepts the CANN
* context and two ACL tensors (source and destination). Internally, this function
* creates ACL representations of the ggml tensors and invokes the unary operation.
* The result is stored in the destination tensor `dst`. This utility abstracts the
* common boilerplate of tensor conversion and cleanup when implementing unary ops.
* @details This function applies a unary operation to the input tensor using
* a user-provided lambda or callable `unary_op`. The lambda receives the
* CANN backend context and two ACL tensors: the source and the destination.
*
* @param unary_op A callable that performs the unary operation using CANN APIs.
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the result will be stored.
* The source tensor is retrieved from `dst->src[0]`.
* Internally, this function handles the conversion from GGML tensors to ACL tensors,
* calls the provided unary op, and manages resource cleanup. The input is assumed
* to be `dst->src[0]`, and the result is written to `dst`.
*
* This utility simplifies writing unary op wrappers by abstracting tensor preparation.
*
* @param unary_op A callable that performs the unary operation using CANN ACL APIs.
* @param ctx The CANN context for operation execution.
* @param dst The destination ggml_tensor where the result will be stored.
* The input tensor is assumed to be `dst->src[0]`.
*
* @see GGML_CANN_CALL_OP_UNARY
*/
void ggml_cann_unary_op(
void ggml_cann_op_unary(
std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Helper macro to invoke a unary ACL operation using ggml_cann_unary_op.
* @brief Applies a gated (GLU-style) unary operation using the CANN backend.
*
* This macro defines an inline lambda wrapping a specific ACL operation name,
* and passes it to the templated ggml_cann_unary_op function. It simplifies
* calling unary ops by hiding the lambda boilerplate.
* @details This function performs a gated activation such as GEGLU or ReGLU.
* It supports two input modes:
*
* 1. **Dual input mode**: `dst->src[0]` and `dst->src[1]` are both valid tensors.
* These are used directly as the value and gate tensors.
*
* 2. **Packed input mode**: Only `dst->src[0]` is valid, and it is assumed to
* contain a concatenation of value and gate along the first dimension. This tensor
* will be split into two equal halves to form the value and gate inputs.
*
* The function applies a user-provided unary operation (e.g., GELU) to the value tensor,
* then multiplies the result in-place with the gate tensor:
*
* Internally, the lambda will call:
* @code
* GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
* dst = unary_op(value) * gate;
* @endcode
*
* The `swapped` parameter (from `dst->op_params[1]`) allows flipping the
* order of value/gate in the packed input case.
*
* @param unary_op A callable that performs the unary operation using CANN ACL APIs.
* It receives (ctx, acl_value_tensor, acl_output_tensor).
* @param ctx The CANN context used for execution.
* @param dst The destination ggml_tensor. Source tensors are in `dst->src[0]` and optionally `src[1]`.
*
* @see GGML_CANN_CALL_OP_UNARY_GATED
*/
void ggml_cann_op_unary_gated(
std::function<void(ggml_backend_cann_context&, aclTensor*, aclTensor*)> unary_op,
ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Helper macro to call a unary ACL operator via ggml_cann_op_unary.
*
* This macro wraps the specified ACLNN unary operator name into a lambda expression,
* and passes it to `ggml_cann_op_unary`, which handles the common logic for executing
* unary ops in the CANN backend.
*
* Internally, this macro expands to a lambda like:
* @code
* [](ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) {
* GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
* };
* @endcode
*
* This lambda is then passed to `ggml_cann_op_unary`, which applies the operation.
*
* @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
*
* @see ggml_cann_unary_op
* @see ggml_cann_op_unary
* @see GGML_CANN_CALL_ACLNN_OP
*/
#define GGML_CANN_CALL_UNARY_OP(OP_NAME) \
#define GGML_CANN_CALL_OP_UNARY(OP_NAME) \
do { \
auto lambda = [](ggml_backend_cann_context& ctx, \
aclTensor* acl_src, \
aclTensor* acl_dst) { \
GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst); \
}; \
ggml_cann_unary_op(lambda, ctx, dst); \
ggml_cann_op_unary(lambda, ctx, dst); \
} \
while (0)
/**
* @brief Helper macro to call a gated unary ACL operator via ggml_cann_op_unary_gated.
*
* This macro wraps the specified ACLNN unary operator name into a lambda expression,
* and passes it to `ggml_cann_op_unary_gated`, which handles the common logic for
* executing gated unary ops in the CANN backend.
*
* Internally, this macro expands to a lambda like:
* @code
* [](ggml_backend_cann_context& ctx, aclTensor* acl_src, aclTensor* acl_dst) {
* GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst);
* };
* @endcode
*
* This lambda is then passed to `ggml_cann_op_unary_gated`, which applies the operation.
*
* @param OP_NAME The name of the ACL unary operator to invoke via GGML_CANN_CALL_ACLNN_OP.
*
* @see ggml_cann_op_unary_gated
* @see GGML_CANN_CALL_ACLNN_OP
*/
#define GGML_CANN_CALL_OP_UNARY_GATED(OP_NAME) \
do { \
auto lambda = [](ggml_backend_cann_context& ctx, \
aclTensor* acl_src, \
aclTensor* acl_dst) { \
GGML_CANN_CALL_ACLNN_OP(ctx, OP_NAME, acl_src, acl_dst); \
}; \
ggml_cann_op_unary_gated(lambda, ctx, dst); \
} \
while (0)
#endif // CANN_ACLNN_OPS