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

Split model correctly even if tensor id is out-of-order

Browse source
This commit is contained in:
z5269887 2024-04-22 23:16:23 +08:00
parent 6d66e609b5
commit d6e453eb6c
1 changed files with 18 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

28
llama.cpp
View file

@ -13534,7 +13534,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
const size_t align = GGUF_DEFAULT_ALIGNMENT; const size_t align = GGUF_DEFAULT_ALIGNMENT;
struct gguf_context * ctx_out = gguf_init_empty(); struct gguf_context * ctx_out = gguf_init_empty();
std::vector<gguf_context*> ctx_outs = {ctx_out};
// copy the KV pairs from the input file // copy the KV pairs from the input file
gguf_set_kv (ctx_out, ml.meta); gguf_set_kv (ctx_out, ml.meta);
@ -13596,19 +13595,28 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
std::vector<no_init<uint8_t>> work; std::vector<no_init<uint8_t>> work;
std::vector<no_init<float>> f32_conv_buf; std::vector<no_init<float>> f32_conv_buf;
uint16_t n_split = 1;
// Assume split index is continuous
if (params->keep_split) {
for (int i = 0; i < ml.n_tensors; ++i) {
n_split = std::max(uint16_t(ml.get_weight(i)->idx+1), n_split);
}
}
std::vector<gguf_context*> ctx_outs(n_split, NULL);
ctx_outs[0] = ctx_out;
// populate the original tensors so we get an initial meta data // populate the original tensors so we get an initial meta data
for (int i = 0; i < ml.n_tensors; ++i) { for (int i = 0; i < ml.n_tensors; ++i) {
auto weight = ml.get_weight(i); auto weight = ml.get_weight(i);
uint16_t i_split = params->keep_split ? weight->idx : 0;
struct ggml_tensor * tensor = weight->tensor; struct ggml_tensor * tensor = weight->tensor;
if (weight->idx != (ctx_outs.size() - 1) && params->keep_split) { if (ctx_outs[i_split] == NULL) {
ctx_out = gguf_init_empty(); ctx_outs[i_split] = gguf_init_empty();
ctx_outs.push_back(ctx_out);
} }
gguf_add_tensor(ctx_out, tensor); gguf_add_tensor(ctx_outs[i_split], tensor);
} }
// Set split info if needed // Set split info if needed
uint16_t n_split = ctx_outs.size();
if (n_split > 1) { if (n_split > 1) {
for (int i = 0; i < ctx_outs.size(); ++i) { for (int i = 0; i < ctx_outs.size(); ++i) {
gguf_set_val_u16(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i); gguf_set_val_u16(ctx_outs[i], ml.llm_kv(LLM_KV_SPLIT_NO).c_str(), i);
@ -13629,8 +13637,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
fout.close(); fout.close();
} }
}; };
auto new_ofstream = [&]() { auto new_ofstream = [&](int index = 0) {
++cur_split; cur_split = index;
GGML_ASSERT(ctx_outs[cur_split] && "Find uninitialized gguf_context");
std::string fname = fname_out; std::string fname = fname_out;
if (params->keep_split) { if (params->keep_split) {
char split_path[PATH_MAX] = {0}; char split_path[PATH_MAX] = {0};
@ -13651,9 +13660,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
auto weight = ml.get_weight(i); auto weight = ml.get_weight(i);
struct ggml_tensor * tensor = weight->tensor; struct ggml_tensor * tensor = weight->tensor;
if (weight->idx != cur_split && params->keep_split) { if (weight->idx != cur_split && params->keep_split) {
GGML_ASSERT(cur_split == weight->idx-1 && "Invalid split index found in weight");
close_ofstream(); close_ofstream();
new_ofstream(); new_ofstream(weight->idx);
} }
const std::string name = ggml_get_name(tensor); const std::string name = ggml_get_name(tensor);