happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

more clean up

This commit is contained in:
bssrdf 2025-10-29 21:57:12 -04:00
parent a3b4d8d31e
commit 70132278cb
3 changed files with 9 additions and 35 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ggml/src/ggml-cuda/cpy.cu
View File

@ -138,7 +138,7 @@ void ggml_cuda_cpy_dest_ptrs_copy(ggml_cuda_graph * cuda_graph, char ** host_des
#endif
}
template<typename src_t, typename dst_t, bool transpose = false>
template<typename src_t, typename dst_t>
static void ggml_cpy_flt_cuda(
const char * cx, char * cdst, const int ne,
const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
@ -332,7 +332,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
}
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
ggml_cpy_flt_cuda<float, float, false> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
ggml_cpy_flt_cuda<float, float> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_BF16) {
ggml_cpy_flt_cuda<float, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
} else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
@ -363,7 +363,7 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
} else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_F32) {
ggml_cpy_q5_1_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
ggml_cpy_flt_cuda<half, half, false> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
ggml_cpy_flt_cuda<half, half> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_BF16) {
ggml_cpy_flt_cuda<half, nv_bfloat16> (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream, dest_ptrs_d, graph_cpynode_index);
} else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {

2
ggml/src/ggml.c
View File

@ -4570,7 +4570,7 @@ struct ggml_tensor * ggml_conv_2d_direct(
return result;
}
// ggml_conv_3d
// ggml_conv_3d_direct
struct ggml_tensor * ggml_conv_3d_direct(
struct ggml_context * ctx,

36
tests/test-backend-ops.cpp
View File

@ -2414,7 +2414,6 @@ struct test_cpy : public test_case {
const std::array<int64_t, 4> permute_dst;
bool _src_use_permute;
bool _dst_use_permute;
bool is_transpose;
std::string vars() override {
return VARS_TO_STR5(type_src, type_dst, ne, permute_src, permute_dst);
@ -2431,12 +2430,10 @@ 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},
std::array<int64_t, 4> permute_src = {0, 0, 0, 0},
std::array<int64_t, 4> permute_dst = {0, 0, 0, 0},
bool transpose = false)
std::array<int64_t, 4> permute_dst = {0, 0, 0, 0})
: type_src(type_src), type_dst(type_dst), ne(ne), permute_src(permute_src), permute_dst(permute_dst),
_src_use_permute(permute_src[0] + permute_src[1] + permute_src[2] + permute_src[3] > 0),
_dst_use_permute(permute_dst[0] + permute_dst[1] + permute_dst[2] + permute_dst[3] > 0),
is_transpose(transpose) {}
_dst_use_permute(permute_dst[0] + permute_dst[1] + permute_dst[2] + permute_dst[3] > 0) {}
ggml_tensor * build_graph(ggml_context * ctx) override {
ggml_tensor * src = ggml_new_tensor(ctx, type_src, 4, ne.data());
@ -2457,8 +2454,6 @@ struct test_cpy : public test_case {
}
ggml_tensor * out = ggml_cpy(ctx, src, dst);
if(is_transpose)
src->op_params[10] = 999;
ggml_set_name(out, "out");
return out;
@ -6024,7 +6019,6 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_I32, {256, 2, 3, 4}, {1, 0, 2, 3}));
test_cases.emplace_back(new test_cpy(GGML_TYPE_I32, GGML_TYPE_F32, {256, 2, 3, 4}));
test_cases.emplace_back(new test_cpy(GGML_TYPE_I32, GGML_TYPE_F32, {256, 2, 3, 4}, {1, 0, 2, 3}));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {48, 48, 1, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}, true));
test_cases.emplace_back(new test_cont());
test_cases.emplace_back(new test_cont(GGML_TYPE_F32, {2, 1, 1 ,1}));
@ -6685,32 +6679,12 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
GGML_TYPE_F32, 1, 1, p0, p1, 1, 1, false));
}
// for (auto act_case : cases_sd) {
// GGML_ASSERT(act_case[idx_sd["kw"]] == 3 || act_case[idx_sd["kw"]] == 1);
// GGML_ASSERT(act_case[idx_sd["kh"]] == 3 || act_case[idx_sd["kh"]] == 1);
// uint32_t p0 = act_case[idx_sd["kw"]] == 3 ? 1 : 0;
// uint32_t p1 = act_case[idx_sd["kh"]] == 3 ? 1 : 0;
// test_cases.emplace_back(new test_conv_2d_implicit(
// { act_case[idx_sd["iw"]], act_case[idx_sd["ih"]], act_case[idx_sd["Cin"]], act_case[idx_sd["B"]] },
// { act_case[idx_sd["kw"]], act_case[idx_sd["kh"]], act_case[idx_sd["Cin"]], act_case[idx_sd["Cout"]] },
// GGML_TYPE_F16, 1, 1, p0, p1, 1, 1, false));
// }
test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1, 1, 1, 1}));
test_cases.emplace_back(new test_bin_bcast(ggml_add, GGML_TYPE_F32, {4096, 1, 1, 1}, {1, 512, 1, 1}));
// test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F16, {512, 3072, 1, 1}));
// test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {8192, 512, 2, 1}, {0, 2, 1, 3}));
// test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {3072, 512, 2, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {768*1024, 256, 1, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}, true));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {768*1024, 256, 1, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}, false));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {768*1024, 256, 1, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}, true));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {768*1024, 256, 1, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}, false));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {768, 1024, 256, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}, true));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F16, GGML_TYPE_F16, {768, 1024, 256, 1}, {1, 0, 2, 3}, {0, 0, 0, 0}, false));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F16, {512, 3072, 1, 1}));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {8192, 512, 2, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_cpy(GGML_TYPE_F32, GGML_TYPE_F32, {3072, 512, 2, 1}, {0, 2, 1, 3}));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {4096, 4096, 5, 1}, false, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));
test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {12888, 256, 5, 1}, false, false, GGML_TYPE_F32, {1, 1}, 1.0f, 0.0f));