vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

[CANN] Add Ascend NPU backend (#6035)

Browse source 
* [CANN] Add Ascend NPU backend

Ascend is a full-stack AI computing infrastructure for industry
applications and services based on Huawei Ascend processors and
software.

CANN (Compute Architecture of Neural Networks), developped by
Huawei, is a heterogeneous computing architecture for AI.

Co-authored-by: wangshuai09 <391746016@qq.com>

* delete trailing whitespaces

* Modify the code based on review comment

* Rename LLAMA_CANN to GGML_CANN

* Make ggml-common.h private

* add ggml_cann prefix for acl funcs

* Add logging for CANN backend

* Delete Trailing whitespace

---------

Co-authored-by: wangshuai09 <391746016@qq.com>
This commit is contained in:
hipudding 2024-07-17 19:23:50 +08:00 • committed by GitHub
parent da3913d8f9
commit 1bdd8ae19f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
27 changed files with 10756 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

33
tests/test-backend-ops.cpp
View file

@ -759,7 +759,7 @@ struct test_dup : public test_case {
}
test_dup(ggml_type type = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {10, 10, 10, 1},
std::array<int64_t, 4> ne = {10, 10, 20, 1},
std::array<int64_t, 4> permute = {0, 0, 0, 0})
: type(type), ne(ne), permute(permute),
_use_permute(permute[0] + permute[1] + permute[2] + permute[3] > 0) {}
@ -779,9 +779,11 @@ struct test_cpy : public test_case {
const ggml_type type_src;
const ggml_type type_dst;
const std::array<int64_t, 4> ne;
const std::array<int64_t, 4> permute;
bool _src_use_permute;
std::string vars() override {
return VARS_TO_STR3(type_src, type_dst, ne);
return VARS_TO_STR4(type_src, type_dst, ne, permute);
}
double max_nmse_err() override {
@ -793,12 +795,18 @@ struct test_cpy : public test_case {
}
test_cpy(ggml_type type_src = GGML_TYPE_F32, ggml_type type_dst = GGML_TYPE_F32,
std::array<int64_t, 4> ne = {10, 10, 10, 1})
: type_src(type_src), type_dst(type_dst), ne(ne) {}
std::array<int64_t, 4> ne = {10, 10, 10, 1},
std::array<int64_t, 4> permute = {0, 0, 0, 0},
bool _dst_use_permute = false)
: type_src(type_src), type_dst(type_dst), ne(ne), permute(permute),
_src_use_permute(permute[0] + permute[1] + permute[2] + permute[3] > 0) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
ggml_tensor * dst = ggml_new_tensor(ctx, type_dst, 4, ne.data());
if (_src_use_permute) {
src = ggml_permute(ctx, src, permute[0], permute[1], permute[2], permute[3]);
}
ggml_tensor* dst = ggml_new_tensor(ctx, type_dst, 4, src->ne);
ggml_tensor * out = ggml_cpy(ctx, src, dst);
return out;
}
@ -1174,6 +1182,7 @@ struct test_soft_max : public test_case {
}
};
// GGML_OP_ROPE
struct test_rope : public test_case {
const ggml_type type;
@ -2146,12 +2155,22 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
test_cases.emplace_back(new test_dup(GGML_TYPE_F16));
test_cases.emplace_back(new test_dup(GGML_TYPE_I32));
test_cases.emplace_back(new test_dup(GGML_TYPE_I16));
test_cases.emplace_back(new test_dup(GGML_TYPE_F32, {10, 10, 5, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_dup(GGML_TYPE_F16, {10, 10, 5, 1}, {0, 2, 1, 3})); // dup by rows
test_cases.emplace_back(new test_dup(GGML_TYPE_F32, {10, 10, 5, 1}, {1, 0, 2, 3}));
test_cases.emplace_back(new test_dup(GGML_TYPE_F16, {10, 10, 5, 1}, {1, 0, 2, 3})); // dup dst not-contiguous
test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_dup(GGML_TYPE_I16, {10, 8, 3, 1}, {1, 2, 0, 3}));
for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_F32}) {
for (ggml_type type_dst : all_types) {
test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 4, 4, 4}));
test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {0, 2, 1, 3})); // cpy by rows
}
}
for (ggml_type type_src : {GGML_TYPE_F16, GGML_TYPE_F32}) {
for (ggml_type type_dst : {GGML_TYPE_F16, GGML_TYPE_F32}) {
test_cases.emplace_back(new test_cpy(type_src, type_dst, {256, 2, 3, 4}, {1, 0, 2, 3})); // cpy not-contiguous
}
}
@ -2283,7 +2302,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
for (int n = 0; n < 10; ++n) {
int64_t ne0 = dist_ne0(rng);
int64_t ne1 = dist_ne1(rng);
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}, n/2 == 0, 0.1f, ne0 < 1000 ? 4.0f : 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, GGML_TYPE_F32, {ne0, ne1, 1, 1}, n/2 == 0, 0.1f, ne0 < 1000 ? 4.0f : 0.0f));
}
exponent <<= 1;
@ -2302,7 +2321,7 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
}
}
}
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, true, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {16, 2, 32, 1}, false, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true, 0.1f, 8.0f));