sleepy/llama.cpp
1
0
Fork 0
Code Issues 11 Pull Requests Actions 114 Packages Projects Releases Wiki Activity

ggml : add ggml_build_forward_select (#18550)

Browse Source 
* ggml : add ggml_build_forward_select

* cuda : adapt CUDA graph compat to new feature

* vulkan : update logic to handle command buffer closing

* ggml : check compute for fusion

* ggml : add comment
This commit is contained in:
Georgi Gerganov
2026-01-19 20:03:19 +02:00
committed by GitHub
parent 3d55846a5c
commit 365a3e8c31
17 changed files with 148 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/include/ggml.h
+39 -7
View File
@@ -630,10 +630,11 @@ extern "C" {
// this tensor...
enum ggml_tensor_flag {
GGML_TENSOR_FLAG_INPUT = 1, // ...is an input for the GGML compute graph
GGML_TENSOR_FLAG_OUTPUT = 2, // ...is an output for the GGML compute graph
GGML_TENSOR_FLAG_PARAM = 4, // ...contains trainable parameters
GGML_TENSOR_FLAG_LOSS = 8, // ...defines loss for numerical optimization (multiple loss tensors add up)
GGML_TENSOR_FLAG_INPUT = 1, // ...is an input for the GGML compute graph
GGML_TENSOR_FLAG_OUTPUT = 2, // ...is an output for the GGML compute graph
GGML_TENSOR_FLAG_PARAM = 4, // ...contains trainable parameters
GGML_TENSOR_FLAG_LOSS = 8, // ...defines loss for numerical optimization (multiple loss tensors add up)
GGML_TENSOR_FLAG_COMPUTE = 16, // ...must be computed
};
enum ggml_tri_type {
@@ -2577,11 +2578,42 @@ extern "C" {
struct ggml_tensor * grad,
struct ggml_tensor * sgd_params); // alpha, weight decay
// build forward mutiple tensors and select one of them for computing
// this is useful for creating graphs that have constant topology but compute different things based on the input
// ref: https://github.com/ggml-org/llama.cpp/pull/18550
//
// automatic differentiation
// nodes:
// | - build forward into the graph but do not compute
// c - build forward into the graph and compute
//
// | | ... c ... |
// | | ... c ... |
// | | ... c ... |
// [0 1 ... idx ... n-1] <-- ggml_build_forward_select(..., n, idx)
// c
// c
//
// example:
// struct ggml_tensor * curs[3];
//
// curs[0] = compute0(...);
// curs[1] = compute1(...);
// curs[2] = compute2(...);
//
// int idx = select_branch(some_input);
//
// struct ggml_tensor * out = ggml_build_forward_select(cgraph, curs, 3, idx);
//
GGML_API struct ggml_tensor * ggml_build_forward_select(
struct ggml_cgraph * cgraph,
struct ggml_tensor ** tensors,
int n_tensors,
int idx);
GGML_API void ggml_build_forward_expand(
struct ggml_cgraph * cgraph,
struct ggml_tensor * tensor);
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, // context for gradient computation
struct ggml_cgraph * cgraph,
@@ -2613,7 +2645,7 @@ extern "C" {
GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
// dump the graph into a file using the dot format
GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * cgraph, const char * filename);
// TODO these functions were sandwiched in the old optimization interface, is there a better place for them?
typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);