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

imatrix : offload to GPU support (#4957)

Browse source 
* backend : add eval callback

ggml-ci

* backend : group nodes in a single compute when user don't need them

* backend : clean-up the implementation

ggml-ci

* simple : do not perform tensor data copy if not needed

* simple : fix

* imatrix : offload to GPU support

* imatrix : fix ggml_mul_mat_id hanlding

ggml-ci

* ci : add imatrix test

ggml-ci

* ci : rearrange output

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-01-17 18:46:30 +02:00 • committed by GitHub
parent 44a1a4a41a
commit ba69bbc84c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 129 additions and 54 deletions
Download patch file Download diff file Expand all files Collapse all files

152
examples/imatrix/imatrix.cpp
View file

@ -33,43 +33,120 @@ class IMatrixCollector {
public:
IMatrixCollector() = default;
void set_parameters(StatParams&& params) { m_params = std::move(params); }
void collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1);
bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
void save_imatrix() const;
private:
std::unordered_map<std::string, Stats> m_stats;
StatParams m_params;
std::mutex m_mutex;
int m_last_call = 0;
std::vector<float> m_src1_data;
std::vector<int> m_ids; // the expert ids from ggml_mul_mat_id
};
void IMatrixCollector::collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) {
if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return;
if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return;
bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
GGML_UNUSED(user_data);
const struct ggml_tensor * src0 = t->src[0];
const struct ggml_tensor * src1 = t->src[1];
// when ask is true, the scheduler wants to know if we are interested in data from this tensor
// if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
if (ask) {
if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications
if (t->op != GGML_OP_MUL_MAT) return false;
if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return false;
return true;
}
std::lock_guard<std::mutex> lock(m_mutex);
auto& e = m_stats[src0->name];
if (e.values.empty()) {
e.values.resize(src1->ne[0], 0);
// copy the data from the GPU memory if needed
const bool is_host = ggml_backend_buffer_is_host(src1->buffer);
if (!is_host) {
m_src1_data.resize(ggml_nelements(src1));
ggml_backend_tensor_get(src1, m_src1_data.data(), 0, ggml_nbytes(src1));
}
else if (e.values.size() != (size_t)src1->ne[0]) {
fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
exit(1); //GGML_ASSERT(false);
}
++e.ncall;
if (m_params.verbosity > 1) {
printf("%s[%d]: %s, %d x %d, %d\n",__func__,m_last_call,src0->name,(int)src1->ne[0],(int)src1->ne[1],(int)src1->type);
}
for (int row = 0; row < (int)src1->ne[1]; ++row) {
const float * x = (const float *)src1->data + row * src1->ne[0];
for (int j = 0; j < (int)src1->ne[0]; ++j) {
e.values[j] += x[j]*x[j];
}
}
if (e.ncall > m_last_call) {
m_last_call = e.ncall;
if (m_last_call % m_params.n_output_frequency == 0) {
save_imatrix();
const float * data = is_host ? (const float *) src1->data : m_src1_data.data();
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];
// 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));
m_ids.resize(ggml_nbytes(src0)/sizeof(int));
ggml_backend_tensor_get(src0, m_ids.data(), 0, ggml_nbytes(src0));
// 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];
auto& e = m_stats[src0->name];
if (e.values.empty()) {
e.values.resize(src1->ne[0], 0);
}
else if (e.values.size() != (size_t)src1->ne[0]) {
fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
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, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
}
for (int row = 0; row < (int)src1->ne[1]; ++row) {
const int excur = m_ids[row*n_as + idx];
GGML_ASSERT(excur >= 0 && excur < n_as); // sanity check
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];
}
}
if (e.ncall > m_last_call) {
m_last_call = e.ncall;
if (m_last_call % m_params.n_output_frequency == 0) {
save_imatrix();
}
}
}
} else {
auto& e = m_stats[src0->name];
if (e.values.empty()) {
e.values.resize(src1->ne[0], 0);
}
else if (e.values.size() != (size_t)src1->ne[0]) {
fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
exit(1); //GGML_ASSERT(false);
}
++e.ncall;
if (m_params.verbosity > 1) {
printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
}
for (int row = 0; row < (int)src1->ne[1]; ++row) {
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];
}
}
if (e.ncall > m_last_call) {
m_last_call = e.ncall;
if (m_last_call % m_params.n_output_frequency == 0) {
save_imatrix();
}
}
}
return true;
}
void IMatrixCollector::save_imatrix() const {
@ -93,8 +170,8 @@ void IMatrixCollector::save_imatrix() const {
static IMatrixCollector g_collector;
static void ik_collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) {
g_collector.collect_imatrix(src0, src1);
static bool ik_collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
return g_collector.collect_imatrix(t, ask, user_data);
}
@ -320,8 +397,6 @@ int main(int argc, char ** argv) {
g_collector.set_parameters(std::move(sparams));
ggml_set_imatrix_collection(ik_collect_imatrix);
params.logits_all = true;
params.n_batch = std::min(params.n_batch, params.n_ctx);
@ -340,16 +415,27 @@ int main(int argc, char ** argv) {
llama_backend_init(params.numa);
llama_model * model;
llama_context * ctx;
llama_model_params mparams = llama_model_params_from_gpt_params(params);
// load the model and apply lora adapter, if any
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
return 1;
}
llama_context_params cparams = llama_context_params_from_gpt_params(params);
// pass the callback to the backend scheduler
// it will be executed for each node during the graph computation
cparams.cb_eval = ik_collect_imatrix;
cparams.cb_eval_user_data = NULL;
llama_context * ctx = llama_new_context_with_model(model, cparams);
if (ctx == NULL) {
fprintf(stderr, "%s: error: unable to create context\n", __func__);
return 1;
}
const int n_ctx_train = llama_n_ctx_train(model);
if (params.n_ctx > n_ctx_train) {
fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",