ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

llama/ggml: add LLM training support (#10544)

Browse source 
* llama/ggml: add LLM training support

more compact progress bar

llama_save_model_to_file

llama_opt_param_filter

ggml_graph_dup force_grads

refactor ggml_opt, fix test-opt

* remove logits_all

* refactor CUDA implementation for ACC

* reset graph at beginning of opt period
This commit is contained in:
Johannes Gäßler 2025-05-12 14:44:49 +02:00 • committed by GitHub
parent 064cc596ac
commit 10d2af0eaa
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
31 changed files with 1415 additions and 359 deletions
Download patch file Download diff file Expand all files Collapse all files

17
common/common.cpp
View file

@ -1565,3 +1565,20 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
return result;
}
ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride) {
const int64_t ne_datapoint = llama_n_ctx(ctx);
const int64_t ndata = (tokens.size() - ne_datapoint - 1) / stride;
ggml_opt_dataset_t result = ggml_opt_dataset_init(
GGML_TYPE_I32, GGML_TYPE_I32, ne_datapoint, ne_datapoint, ndata, /*ndata_shard =*/ 1);
llama_token * data = (llama_token *) ggml_opt_dataset_data(result)->data;
llama_token * labels = (llama_token *) ggml_opt_dataset_labels(result)->data;
for (int64_t idata = 0; idata < ndata; ++idata) {
memcpy(data + idata*ne_datapoint, tokens.data() + idata*stride + 0, ne_datapoint*sizeof(llama_token));
memcpy(labels + idata*ne_datapoint, tokens.data() + idata*stride + 1, ne_datapoint*sizeof(llama_token));
}
return result;
}

6
common/common.h
View file

@ -666,3 +666,9 @@ const char * const LLM_KV_SPLIT_COUNT = "split.count";
const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
}
//
// training utils
//
ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);