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

sync : ggml (backend v2) (#3912)

Browse source 
* sync : ggml (backend v2) (wip)

* sync : migrate examples and llama.cpp to dynamic graphs (wip)

* sync : update tests + fix max op params to 64

ggml-ci

* sync : ggml-cuda

ggml-ci

* llama : fix save/load state context size

ggml-ci

* sync : try to fix build on tvOS

* sync : pass custom graph sizes in training examples

* sync : update graph copies to new ggml API

* sync : update sync-ggml.sh with new files

* scripts : fix header in sync script

* train : fix context size calculations

* llama : increase inference graph size up to 4096 nodes

* train : allocate grads for backward graphs

* train : allocate grads for gb_tmp
This commit is contained in:
Georgi Gerganov 2023-11-13 14:16:23 +02:00 • committed by GitHub
parent bb50a792ec
commit 4760e7cc0b
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
22 changed files with 1994 additions and 864 deletions
Download patch file Download diff file Expand all files Collapse all files

40
llama.cpp
View file

@ -91,6 +91,8 @@
#define LLAMA_ATTRIBUTE_FORMAT(...)
#endif
#define LLAMA_MAX_NODES 4096
//
// logging
//
@ -3618,7 +3620,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_llama() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
GGML_ASSERT(n_embd_head == hparams.n_rot);
@ -3730,7 +3732,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_baichuan() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -3850,7 +3852,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_falcon() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -3972,7 +3974,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_starcoder() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * pos;
@ -4071,7 +4073,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_persimmon() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
const int64_t n_rot = n_embd_head / 2;
@ -4281,7 +4283,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_refact() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -4372,7 +4374,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_bloom() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -4466,7 +4468,7 @@ struct llm_build_context {
}
struct ggml_cgraph * build_mpt() {
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
struct ggml_tensor * cur;
struct ggml_tensor * inpL;
@ -8208,7 +8210,7 @@ struct llama_context * llama_new_context_with_model(
{
static const size_t tensor_alignment = 32;
// the compute buffer is used to store the tensor and graph structs, while the allocator buffer is used for the tensor data
ctx->buf_compute.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead());
ctx->buf_compute.resize(ggml_tensor_overhead()*LLAMA_MAX_NODES + ggml_graph_overhead());
// create measure allocator
ctx->alloc = ggml_allocr_new_measure(tensor_alignment);
@ -8597,8 +8599,8 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
if (kv_buf_size) {
const size_t elt_size = ggml_element_size(kv_self.k);
ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
ggml_cgraph gf{};
ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true });
ggml_cgraph * gf = ggml_new_graph(cpy_ctx);
ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
std::vector<uint8_t> kout3d_data(ggml_nbytes(kout3d), 0);
@ -8616,9 +8618,9 @@ static void llama_copy_state_data_internal(struct llama_context * ctx, llama_dat
kv_head, n_embd, n_layer,
elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, k3d, kout3d));
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, v3d, vout3d));
ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1);
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, k3d, kout3d));
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, v3d, vout3d));
ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1);
ggml_free(cpy_ctx);
@ -8725,8 +8727,8 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
const size_t elt_size = ggml_element_size(kv_self.k);
ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
ggml_cgraph gf{};
ggml_context * cpy_ctx = ggml_init({ 6*ggml_tensor_overhead() + ggml_graph_overhead(), NULL, /* no_alloc */ true });
ggml_cgraph * gf = ggml_new_graph(cpy_ctx);
ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_head, n_layer);
kin3d->data = (void *) inp;
@ -8744,9 +8746,9 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
kv_head, n_embd, n_layer,
elt_size*n_ctx, elt_size*n_ctx*n_embd, 0);
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, kin3d, k3d));
ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, vin3d, v3d));
ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1);
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, kin3d, k3d));
ggml_build_forward_expand(gf, ggml_cpy(cpy_ctx, vin3d, v3d));
ggml_graph_compute_helper(ctx->work_buffer, gf, /*n_threads*/ 1);
ggml_free(cpy_ctx);
}