mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-04-16 11:36:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

CANN: Fix failed test cases (#12708)

Browse Source 
* CANN: Fix memory waste in aclnn_tensor

* CANN: fix backend ops fail

* CANN: fix acl_tensor memory alloc.

* CANN: format

* CANN: remove trailing whitespace
This commit is contained in:
hipudding 2025-04-03 08:49:51 +08:00 committed by GitHub
parent 97a20c012b
commit 2a0dc97e56
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
5 changed files with 30 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/ggml-cann/acl_tensor.cpp
View File

@ -54,9 +54,7 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
// added.
int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2];
int64_t acl_storage_len = 0;
if (ne == nullptr) {
acl_storage_len = ggml_nbytes(tensor);
for (int i = 0; i < GGML_MAX_DIMS; i++) {
acl_ne[i] = tensor->ne[i];
// The step size of acl is in elements.
@ -65,14 +63,18 @@ aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
} else {
// With bcast
for (int i = 0; i < dims; i++) {
acl_storage_len += (ne[i] - 1) * nb[i];
acl_ne[i] = ne[i];
acl_stride[i] = nb[i] / ggml_element_size(tensor);
}
}
// Reverse ne and stride.
int64_t final_dims = (dims == 0 ? GGML_MAX_DIMS : dims);
int64_t acl_storage_len = 1;
for (int i = 0; i < final_dims; i++) {
acl_storage_len += (acl_ne[i] - 1) * acl_stride[i];
}
// Reverse ne and stride.
std::reverse(acl_ne, acl_ne + final_dims);
std::reverse(acl_stride, acl_stride + final_dims);

10
ggml/src/ggml-cann/acl_tensor.h
View File

@ -101,14 +101,14 @@ aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
tmp_stride[i] = nb[i] / type_size;
}
int64_t acl_storage_len = 1;
for (int i = 0; i < dims; i++) {
acl_storage_len += (tmp_ne[i] - 1) * tmp_stride[i];
}
std::reverse(tmp_ne, tmp_ne + dims);
std::reverse(tmp_stride, tmp_stride + dims);
int64_t acl_storage_len = 0;
for (int i = 0; i < dims; i++) {
acl_storage_len += (ne[i] - 1) * nb[i];
}
aclTensor* acl_tensor =
aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
format, &acl_storage_len, 1, data_ptr);

6
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@ -358,8 +358,6 @@ void ggml_cann_sqr(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
GGML_ASSERT(src->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
float min;
float max;
@ -1090,8 +1088,6 @@ void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
float eps;
memcpy(&eps, dst->op_params, sizeof(float));
GGML_ASSERT(eps > 0.0f);
uint64_t workspaceSize = 0;
aclOpExecutor* executor;
void* workspaceAddr = nullptr;
@ -3152,7 +3148,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
// TODO: use ascendc
// Only test with LLAMA model.
ggml_tensor* src0 = dst->src[0]; // input
ggml_tensor* src2 = dst->src[2]; // freq_factors
// ggml_tensor* src2 = dst->src[2]; // freq_factors, not used now.
// param
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;

6
ggml/src/ggml-cann/aclnn_ops.h
View File

@ -535,9 +535,6 @@ template <aclnnStatus getWorkspaceSize(const aclTensor*, aclTensor*, uint64_t*,
void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
GGML_ASSERT(src->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
aclTensor* acl_src = ggml_cann_create_tensor(src);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);
@ -566,9 +563,6 @@ template <aclnnStatus getWorkspaceSize(const aclTensor*, const aclTensor*,
void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_tensor* src = dst->src[0];
GGML_ASSERT(src->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_F32);
aclTensor* acl_src = ggml_cann_create_tensor(src);
aclTensor* acl_dst = ggml_cann_create_tensor(dst);

28
ggml/src/ggml-cann/ggml-cann.cpp
View File

@ -1458,11 +1458,6 @@ static void ggml_backend_cann_free(ggml_backend_t backend) {
ACL_CHECK(aclrtSynchronizeDevice());
ACL_CHECK(aclrtResetDevice(cann_ctx->device));
// finalize when last backend freed.
if (cann_ctx->device == ggml_backend_cann_get_device_count() - 1) {
ACL_CHECK(aclFinalize());
}
delete cann_ctx;
delete backend;
}
@ -1688,11 +1683,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
}
case GGML_OP_MUL_MAT: {
switch (op->src[0]->type) {
case GGML_TYPE_Q8_0:
case GGML_TYPE_F16:
case GGML_TYPE_F32:
case GGML_TYPE_Q4_0:
return true;
case GGML_TYPE_Q8_0:
case GGML_TYPE_Q4_0:
// only support contiguous for quantized types.
return ggml_is_contiguous(op->src[0]) &&
ggml_is_contiguous(op->src[1]);
default:
return false;
}
@ -1738,13 +1736,14 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
}
case GGML_OP_ROPE: {
// TODO: with ops-test v == 1
float * ext_factor = (float*)((int32_t*)op->op_params + 7);
float ext_factor = 0.0f;
memcpy(&ext_factor, (const float *) op->op_params + 7, sizeof(float));
// TODO: n_dims <= ne0
if (op->src[0]->ne[0] != op->op_params[1]) {
return false;
}
// TODO: ext_factor != 0
if (*ext_factor != 0) {
if (ext_factor != 0) {
return false;
}
@ -1766,6 +1765,16 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
}
return true;
}
case GGML_OP_POOL_2D: {
const int32_t * opts = (const int32_t *) op->op_params;
const int k0 = opts[1];
const int k1 = opts[2];
const int p0 = opts[5];
const int p1 = opts[6];
// value of paddingH should be at most half of kernelH
// value of paddingW should be at most half of kernelW
return (p0 <= (k0 / 2)) && (p1 <= (k1 / 2));
}
case GGML_OP_DUP:
case GGML_OP_IM2COL:
case GGML_OP_CONCAT:
@ -1785,7 +1794,6 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev,
case GGML_OP_CLAMP:
case GGML_OP_DIAG_MASK_INF:
case GGML_OP_SOFT_MAX:
case GGML_OP_POOL_2D:
case GGML_OP_SUM_ROWS:
case GGML_OP_ARGSORT:
case GGML_OP_ACC: