vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 09:34:37 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'upstream' into concedo_experimental

Browse source 
# Conflicts:
#	.devops/full-cuda.Dockerfile
#	.devops/full-rocm.Dockerfile
#	.devops/full.Dockerfile
#	.devops/llama-cpp-clblast.srpm.spec
#	.devops/llama-cpp-cuda.srpm.spec
#	.devops/llama-cpp.srpm.spec
#	.devops/nix/package.nix
#	.devops/server-cuda.Dockerfile
#	.devops/server-intel.Dockerfile
#	.devops/server-rocm.Dockerfile
#	.devops/server-vulkan.Dockerfile
#	.devops/server.Dockerfile
#	.github/workflows/build.yml
#	.github/workflows/code-coverage.yml
#	.github/workflows/docker.yml
#	.github/workflows/editorconfig.yml
#	.github/workflows/gguf-publish.yml
#	.github/workflows/nix-ci-aarch64.yml
#	.github/workflows/nix-ci.yml
#	.github/workflows/python-check-requirements.yml
#	.github/workflows/python-lint.yml
#	.github/workflows/server.yml
#	.github/workflows/zig-build.yml
#	CMakeLists.txt
#	Makefile
#	README-sycl.md
#	README.md
#	ci/run.sh
#	examples/gguf-split/gguf-split.cpp
#	flake.lock
#	flake.nix
#	llama.cpp
#	scripts/compare-llama-bench.py
#	scripts/sync-ggml-am.sh
#	scripts/sync-ggml.last
#	scripts/sync-ggml.sh
#	tests/CMakeLists.txt
#	tests/test-backend-ops.cpp
#	tests/test-chat-template.cpp
This commit is contained in:
Concedo 2024-04-07 22:07:27 +08:00
commit 81ac0e5656
58 changed files with 32809 additions and 7635 deletions
Download patch file Download diff file Expand all files Collapse all files

43
examples/imatrix/imatrix.cpp
View file

@ -99,35 +99,38 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
const float * data = is_host ? (const float *) src1->data : m_src1_data.data();
// this has been adapted to the new format of storing merged experts in a single 3d tensor
// ref: https://github.com/ggerganov/llama.cpp/pull/6387
if (t->op == GGML_OP_MUL_MAT_ID) {
const int idx = ((int32_t *) t->op_params)[0];
const int n_as = ((int32_t *) t->op_params)[1];
const ggml_tensor * ids = t->src[2];
const int n_as = src0->ne[2];
// the top-k selected expert ids are stored in the src0 tensor
// for simplicity, always copy src0 to host, because it is small
// take into account that src0 is not contiguous!
GGML_ASSERT(src0->ne[1] == src1->ne[1]);
GGML_ASSERT(n_as*ggml_nrows(src0)*sizeof(int) == GGML_PAD(ggml_nbytes(src0), n_as*sizeof(int)));
m_ids.resize(ggml_nbytes(src0)/sizeof(int));
ggml_backend_tensor_get(src0, m_ids.data(), 0, ggml_nbytes(src0));
// the top-k selected expert ids are stored in the ids tensor
// for simplicity, always copy ids to host, because it is small
// take into account that ids is not contiguous!
GGML_ASSERT(ids->ne[1] == src1->ne[1]);
GGML_ASSERT(n_as*ggml_nrows(ids)*sizeof(int) == GGML_PAD(ggml_nbytes(ids), n_as*sizeof(int)));
m_ids.resize(ggml_nbytes(ids)/sizeof(int));
ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids));
auto & e = m_stats[wname];
++e.ncall;
// NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
// using the following line, we can correct for that if needed by replacing the line above with:
//if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
// loop over all possible experts, regardless if they are used or not in the batch
// this is necessary to guarantee equal number of "ncall" for each tensor
for (int ex = 0; ex < n_as; ++ex) {
src0 = t->src[2 + ex];
wname = filter_tensor_name(src0->name);
auto& e = m_stats[wname];
size_t e_start = ex*src1->ne[0];
if (e.values.empty()) {
e.values.resize(src1->ne[0], 0);
e.values.resize(src1->ne[0]*n_as, 0);
}
else if (e.values.size() != (size_t)src1->ne[0]) {
fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
exit(1); //GGML_ASSERT(false);
}
// NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
// using the following line, we can correct for that if needed
//if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
++e.ncall;
if (m_params.verbosity > 1) {
printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
}
@ -137,7 +140,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
if (excur != ex) continue;
const float * x = data + row * src1->ne[0];
for (int j = 0; j < (int)src1->ne[0]; ++j) {
e.values[j] += x[j]*x[j];
e.values[e_start + j] += x[j]*x[j];
}
}
if (e.ncall > m_last_call) {