CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-10 10:27:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/mamba2

Browse Source
This commit is contained in:
Francis Couture-Harpin
2025-06-10 19:22:15 -04:00
parent e94f3932f2 dad5c44398
commit 9864bfcd01
581 changed files with 159374 additions and 37809 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
ggml/include/ggml.h
View File

@@ -536,6 +536,7 @@ extern "C" {
GGML_UNARY_OP_HARDSWISH,
GGML_UNARY_OP_HARDSIGMOID,
GGML_UNARY_OP_EXP,
GGML_UNARY_OP_GELU_ERF,
GGML_UNARY_OP_COUNT,
};
@@ -673,11 +674,15 @@ extern "C" {
GGML_API bool ggml_is_3d (const struct ggml_tensor * tensor);
GGML_API int ggml_n_dims (const struct ggml_tensor * tensor); // returns 1 for scalars
// returns whether the tensor elements can be iterated over with a flattened index (no gaps, no permutation)
GGML_API bool ggml_is_contiguous (const struct ggml_tensor * tensor);
GGML_API bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous()
GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
// returns whether the tensor elements are allocated as one contiguous block of memory (no gaps, but permutation ok)
GGML_API bool ggml_is_contiguously_allocated(const struct ggml_tensor * tensor);
// true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
@@ -764,7 +769,7 @@ 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);
//
@@ -930,11 +935,20 @@ extern "C" {
struct ggml_tensor * a,
struct ggml_tensor * b);
// repeat a to the specified shape
GGML_API struct ggml_tensor * ggml_repeat_4d(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3);
// sums repetitions in a into shape of b
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
@@ -1020,6 +1034,16 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// GELU using erf (error function) when possible
// some backends may fallback to approximation based on Abramowitz and Stegun formula
GGML_API struct ggml_tensor * ggml_gelu_erf(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_gelu_erf_inplace(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_gelu_quick(
struct ggml_context * ctx,
struct ggml_tensor * a);
@@ -2046,15 +2070,14 @@ 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);
@@ -2073,9 +2096,6 @@ extern "C" {
GGML_API struct ggml_tensor * ggml_graph_get_grad (const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
GGML_API struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
GGML_API void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
// print info and performance information for the graph
GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
@@ -2159,6 +2179,7 @@ extern "C" {
// scheduling priorities
enum ggml_sched_priority {
GGML_SCHED_PRIO_LOW = -1,
GGML_SCHED_PRIO_NORMAL,
GGML_SCHED_PRIO_MEDIUM,
GGML_SCHED_PRIO_HIGH,