extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-19 17:28:09 +02:00
Code Issues Packages Projects Releases Wiki Activity

CANN: Support MOE Model MUL_MAT_ID (llama/13042)

Browse Source 
Signed-off-by: noemotiovon <757486878@qq.com>
This commit is contained in:
Chenguang Li 2025-05-19 14:21:17 +08:00 committed by Georgi Gerganov
parent 2c13651e08
commit 9da3fc27be
3 changed files with 183 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

147
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@ -65,6 +65,7 @@
#include <aclnnop/aclnn_eq_tensor.h>
#include <aclnnop/aclnn_gt_scalar.h>
#include <aclnnop/aclnn_pow.h>
#include <aclnnop/aclnn_grouped_matmul_v2.h>
#include <float.h>
#include <cmath>
@ -2587,3 +2588,149 @@ void ggml_cann_step(ggml_backend_cann_context& ctx, ggml_tensor* dst){
ggml_cann_release_resources(ctx, acl_src, acl_dst, alpha);
}
/**
* @brief Performs expert-specific matrix multiplication (MoE) with
* floating-point precision using the CANN backend.
*
* This function executes a matrix multiplication operation tailored for
* Mixture of Experts (MoE) models, where the input tensor is multiplied
* with expert-specific weight matrices. It uses the CANN backend for
* efficient computation and stores the result in the destination tensor `dst`.
* The operation may leverage identity-based optimizations or routing masks
* as part of sparse expert selection.
*
* @param ctx The context for executing CANN backend operations.
* @param dst The destination tensor where the MoE multiplication result
* will be stored.
*
* @note This function assumes floating-point data types and is designed for
* MoE architectures, possibly involving sparse expert routing.
*/
static void ggml_cann_mul_mat_id_fp(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
//dst [M, K, N, 1]
ggml_tensor * src0 = dst->src[0]; //src0 [D, M, A, 1]
ggml_tensor * src1 = dst->src[1]; //src1 [D, B, N, 1], B = K or B = 1
ggml_tensor * ids = dst->src[2]; //ids [K, N]
GGML_TENSOR_BINARY_OP_LOCALS
// copy index from npu to cpu
int64_t n_as = ne02; // A
int64_t n_ids = ids->ne[0]; // K
std::vector<char> ids_host(ggml_nbytes(ids));
ggml_cann_async_memcpy(ctx, ids_host.data(), ids->data, ggml_nbytes(ids),
ACL_MEMCPY_DEVICE_TO_HOST);
ACL_CHECK(aclrtSynchronizeStream(ctx.stream()));
char * src0_original = (char *) src0->data;
char * src1_original = (char *) src1->data;
char * dst_original = (char *) dst->data;
size_t ori_src0_nb[4] = {nb00, nb01, nb02, nb03};
// src0 is F16, src1 is F32, dst is F32
ggml_cann_pool_alloc src0_cast_allocator;
if (src0->type == GGML_TYPE_F16) {
src0_cast_allocator.alloc(ctx.pool(), sizeof(float) * ggml_nelements(src0));
void* src0_cast_buf = src0_cast_allocator.get();
size_t cast_nb[GGML_MAX_DIMS];
cast_nb[0] = sizeof(float_t);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
cast_nb[i] = cast_nb[i - 1] * src0->ne[i - 1];
}
aclTensor* acl_src0_f16 = ggml_cann_create_tensor(src0);
aclTensor* acl_cast = ggml_cann_create_tensor(src0_cast_buf,
ACL_FLOAT, sizeof(float), src0->ne, cast_nb, 4);
GGML_CANN_CALL_ACLNN_OP(ctx, Cast, acl_src0_f16, ACL_FLOAT, acl_cast);
ggml_cann_release_resources(ctx, acl_cast, acl_src0_f16);
src0_original = (char *) src0_cast_buf;
memcpy(ori_src0_nb, cast_nb, sizeof(ori_src0_nb));
}
std::vector<aclTensor*> src0_tensor_vec;
std::vector<aclTensor*> src1_tensor_vec;
std::vector<aclTensor*> dst_tensor_vec;
for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
for (int64_t id = 0; id < n_ids; id++) {
// src0_row [M, D] -> weight && permute
int64_t src0_ne[2] = {ne01, ne00};
size_t src0_nb[2] = {ori_src0_nb[1], ori_src0_nb[0]};
// src1_row [D, 1] -> input
int64_t src1_ne[2] = {ne10, 1};
size_t src1_nb[2] = {nb10, nb11};
// dst_row [M, 1] -> out
int64_t dst_ne[2] = {ne0, 1};
size_t dst_nb[2] = {nb0, nb1};
// expert index
int32_t i02 = *(int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
GGML_ASSERT(i02 >= 0 && i02 < n_as);
// If B = 1 (broadcast), always use 0; otherwise, use id.
int64_t i11 = (ne11 == 1 ? 0 : id);
int64_t i12 = iid1;
int64_t i1 = id;
int64_t i2 = i12;
void* src0_tmp_ptr = src0_original + i02*ori_src0_nb[2];
void* src1_tmp_ptr = src1_original + i11*nb11 + i12*nb12;
void* dst_tmp_ptr = dst_original + i1*nb1 + i2*nb2;
aclTensor* acl_src0 = ggml_cann_create_tensor(src0_tmp_ptr,
ACL_FLOAT, sizeof(float),
src0_ne, src0_nb, 2);
aclTensor* acl_src1 = ggml_cann_create_tensor(src1_tmp_ptr,
ACL_FLOAT, sizeof(float),
src1_ne, src1_nb, 2);
aclTensor* acl_dst = ggml_cann_create_tensor(dst_tmp_ptr,
ACL_FLOAT, sizeof(float),
dst_ne, dst_nb, 2);
src0_tensor_vec.push_back(acl_src0);
src1_tensor_vec.push_back(acl_src1);
dst_tensor_vec.push_back(acl_dst);
}
}
// GroupedMatmulV2 required tensor_list.size < 128
size_t GROUP_SIZE = 128;
std::vector<std::vector<aclTensor*>> src0_tensor_vec_vec;
std::vector<std::vector<aclTensor*>> src1_tensor_vec_vec;
std::vector<std::vector<aclTensor*>> dst_tensor_vec_vec;
// split and call GroupedMatmulV2
for (size_t i = 0; i < src0_tensor_vec.size(); i += GROUP_SIZE) {
size_t end = std::min(i + GROUP_SIZE, src0_tensor_vec.size());
std::vector<aclTensor*> src0_tensor_vec_split(src0_tensor_vec.begin() + i, src0_tensor_vec.begin() + end);
std::vector<aclTensor*> src1_tensor_vec_split(src1_tensor_vec.begin() + i, src1_tensor_vec.begin() + end);
std::vector<aclTensor*> dst_tensor_vec_split(dst_tensor_vec.begin() + i, dst_tensor_vec.begin() + end);
aclTensorList* src0_tensor_list = aclCreateTensorList(src0_tensor_vec_split.data(), src0_tensor_vec_split.size());
aclTensorList* src1_tensor_list = aclCreateTensorList(src1_tensor_vec_split.data(), src1_tensor_vec_split.size());
aclTensorList* dst_tensor_list = aclCreateTensorList(dst_tensor_vec_split.data(), dst_tensor_vec_split.size());
GGML_CANN_CALL_ACLNN_OP(ctx, GroupedMatmulV2, src1_tensor_list, src0_tensor_list,
nullptr, nullptr, nullptr, nullptr, nullptr, nullptr, 0, -1, dst_tensor_list);
ggml_cann_release_resources(ctx, src0_tensor_list, src1_tensor_list, dst_tensor_list);
}
return;
}
void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
const enum ggml_type type = dst->src[0]->type;
switch (type) {
case GGML_TYPE_F32:
case GGML_TYPE_F16:
ggml_cann_mul_mat_id_fp(ctx, dst);
break;
default:
GGML_ABORT("Unsupported type for mul_mat_id");
break;
}
}

27
ggml/src/ggml-cann/aclnn_ops.h
View File

@ -978,6 +978,33 @@ inline void ggml_cann_async_memset(ggml_backend_cann_context & ctx, void * buffe
}
}
/**
* @brief Performs sparse expert-based matrix multiplication using the CANN backend.
*
* @details This function implements a MoE-style batched matrix multiplication, where each input token
* is routed to one or more experts, and each expert corresponds to a specific [D, M] weight matrix
* in the source tensor `src0`. The routing indices are provided via the `ids` tensor.
*
* For each token (from `src1`), the function selects the corresponding expert(s) as specified by `ids`,
* performs the matrix multiplication with the selected expert's weight submatrix (from `src0`),
* and stores the results in `dst`. This operation is optimized and executed on the CANN backend.
*
* Dimensions:
* - src0: [D, M, A, 1], where A is the number of experts
* - src1: [D, B, N, 1], where N is batch size and B is the slot count per sample
* - ids : [K, N], where K is the number of experts each token is routed to
* - dst : [M, K, N, 1], output tensor storing the result of expert × token multiplication
*
* The function handles two main modes:
* - If `ne12 == 1`, a simpler per-token loop is used.
* - TODO: If `ne12 > 1`, grouped multiplication and memory copying is used for efficiency.
*
* @param ctx The CANN context used for operations.
* @param dst The destination tensor where the expert-weighted token outputs are stored.
* Expected to be of shape [M, K, N, 1].
*/
void ggml_cann_mul_mat_id(ggml_backend_cann_context& ctx, ggml_tensor* dst);
/**
* @brief Applies a element-wise operation to two input tensors using the CANN
* backend.

11
ggml/src/ggml-cann/ggml-cann.cpp
View File

@ -1672,7 +1672,8 @@ static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
ggml_cann_mul_mat(ctx, dst);
break;
case GGML_OP_MUL_MAT_ID:
return false;
ggml_cann_mul_mat_id(ctx, dst);
break;
case GGML_OP_SCALE:
ggml_cann_scale(ctx, dst);
break;
@ -2030,7 +2031,13 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
}
}
case GGML_OP_MUL_MAT_ID:
return false;
switch (op->src[0]->type) {
case GGML_TYPE_F16:
case GGML_TYPE_F32:
return true;
default:
return false;
}
// embedding
case GGML_OP_GET_ROWS: {
switch (op->src[0]->type) {