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

llama/ggml: add LLM training support

Browse source 
more compact progress bar

refactor: llama_prepare_sbatch/ubatch

llama_save_model_to_file

gqa_mode arg for repeat_back

llama_opt_param_filter

ggml_graph_dup force_grads

refactor ggml_opt, fix test-opt
This commit is contained in:
Johannes Gäßler 2024-11-17 14:58:51 +01:00
parent a5203b4465
commit c25557362a
26 changed files with 1294 additions and 339 deletions
Download patch file Download diff file Expand all files Collapse all files

13
ggml/include/ggml.h
View file

@ -763,7 +763,7 @@ extern "C" {
// Tensor flags
GGML_API void ggml_set_input(struct ggml_tensor * tensor);
GGML_API void ggml_set_output(struct ggml_tensor * tensor);
GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
GGML_API void ggml_set_param(struct ggml_tensor * tensor);
GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
//
@ -933,7 +933,7 @@ extern "C" {
GGML_API struct ggml_tensor * ggml_repeat_back(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
struct ggml_tensor * b); // sum up values that are adjacent in dims > 0 instead of repeated with same stride
// concat a and b along dim
// used in stable-diffusion
@ -2054,15 +2054,14 @@ extern "C" {
GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
GGML_API void ggml_build_backward_expand(
struct ggml_context * ctx_static, // context for static gradients (loss + gradient accumulation)
struct ggml_context * ctx_compute, // context for gradient computation
struct ggml_cgraph * cgraph,
bool accumulate); // whether or not gradients should be accumulated, requires static allocation of tensors in ctx_static
struct ggml_context * ctx, // context for gradient computation
struct ggml_cgraph * cgraph,
struct ggml_tensor ** grad_accs);
// graph allocation in a context
GGML_API struct ggml_cgraph * ggml_new_graph (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
GGML_API struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads);
GGML_API struct ggml_cgraph * ggml_graph_dup (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
GGML_API struct ggml_cgraph * ggml_graph_dup (struct ggml_context * ctx, struct ggml_cgraph * cgraph, bool force_grads);
GGML_API void ggml_graph_cpy (struct ggml_cgraph * src, struct ggml_cgraph * dst);
GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); // set regular grads + optimizer momenta to 0, set loss grad to 1
GGML_API void ggml_graph_clear (struct ggml_cgraph * cgraph);