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

upgrade all other formats

Browse source
This commit is contained in:
Concedo 2023-05-17 16:28:20 +08:00
parent 00da2a5f4e
commit 57230b5196
3 changed files with 21 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

9
llama.cpp
View file

@ -56,6 +56,7 @@ static const size_t MB = 1024*1024;
static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0() static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
{ {
static std::map<e_model, size_t> k_sizes = { static std::map<e_model, size_t> k_sizes = {
{ MODEL_UNKNOWN, 512ull * MB },
{ MODEL_7B, 512ull * MB }, { MODEL_7B, 512ull * MB },
{ MODEL_13B, 512ull * MB }, { MODEL_13B, 512ull * MB },
{ MODEL_30B, 512ull * MB }, { MODEL_30B, 512ull * MB },
@ -67,6 +68,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1() static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
{ {
static std::map<e_model, size_t> k_sizes = { static std::map<e_model, size_t> k_sizes = {
{ MODEL_UNKNOWN, 512ull * MB },
{ MODEL_7B, 512ull * MB }, { MODEL_7B, 512ull * MB },
{ MODEL_13B, 512ull * MB }, { MODEL_13B, 512ull * MB },
{ MODEL_30B, 512ull * MB }, { MODEL_30B, 512ull * MB },
@ -79,6 +81,7 @@ static const std::map<e_model, size_t> & MEM_REQ_SCRATCH1()
static const std::map<e_model, size_t> & MEM_REQ_KV_SELF() static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
{ {
static std::map<e_model, size_t> k_sizes = { static std::map<e_model, size_t> k_sizes = {
{ MODEL_UNKNOWN, 1026ull * MB },
{ MODEL_7B, 1026ull * MB }, { MODEL_7B, 1026ull * MB },
{ MODEL_13B, 1608ull * MB }, { MODEL_13B, 1608ull * MB },
{ MODEL_30B, 3124ull * MB }, { MODEL_30B, 3124ull * MB },
@ -92,6 +95,7 @@ static const std::map<e_model, size_t> & MEM_REQ_KV_SELF()
static const std::map<e_model, size_t> & MEM_REQ_EVAL() static const std::map<e_model, size_t> & MEM_REQ_EVAL()
{ {
static std::map<e_model, size_t> k_sizes = { static std::map<e_model, size_t> k_sizes = {
{ MODEL_UNKNOWN, 800ull * MB },
{ MODEL_7B, 800ull * MB }, { MODEL_7B, 800ull * MB },
{ MODEL_13B, 1024ull * MB }, { MODEL_13B, 1024ull * MB },
{ MODEL_30B, 1280ull * MB }, { MODEL_30B, 1280ull * MB },
@ -887,7 +891,9 @@ static const char *llama_model_type_name(e_model type) {
case MODEL_13B: return "13B"; case MODEL_13B: return "13B";
case MODEL_30B: return "30B"; case MODEL_30B: return "30B";
case MODEL_65B: return "65B"; case MODEL_65B: return "65B";
default: LLAMA_ASSERT(false); default:
printf("\nWARNING: NON-STANDARD LLAMA FILE DETECTED. DEFAULT TO 7B SIZE.\n");
return "UNKNOWN";
} }
} }
@ -920,6 +926,7 @@ static void llama_model_load_internal(
case 40: model.type = e_model::MODEL_13B; break; case 40: model.type = e_model::MODEL_13B; break;
case 60: model.type = e_model::MODEL_30B; break; case 60: model.type = e_model::MODEL_30B; break;
case 80: model.type = e_model::MODEL_65B; break; case 80: model.type = e_model::MODEL_65B; break;
default: model.type = e_model::MODEL_UNKNOWN; break;
} }
hparams.n_ctx = n_ctx; hparams.n_ctx = n_ctx;

10
otherarch/gpt2_v2.cpp
View file

@ -49,7 +49,6 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
fin.read((char *) &hparams.ftype, sizeof(hparams.ftype)); fin.read((char *) &hparams.ftype, sizeof(hparams.ftype));
const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR; const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
hparams.ftype %= GGML_QNT_VERSION_FACTOR;
printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab); printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx); printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx);
@ -57,6 +56,9 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
printf("%s: n_head = %d\n", __func__, hparams.n_head); printf("%s: n_head = %d\n", __func__, hparams.n_head);
printf("%s: n_layer = %d\n", __func__, hparams.n_layer); printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
printf("%s: ftype = %d\n", __func__, hparams.ftype); printf("%s: ftype = %d\n", __func__, hparams.ftype);
printf("%s: qntvr = %d\n", __func__, qntvr);
hparams.ftype %= GGML_QNT_VERSION_FACTOR;
} }
// load vocab // load vocab
@ -134,8 +136,9 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
ctx_size += 1.5*(n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // memory_k ctx_size += 1.5*(n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // memory_k
ctx_size += 1.5*(n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // memory_v ctx_size += 1.5*(n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // memory_v
ctx_size += (6 + 12*n_layer)*256; // object overhead ctx_size += (6 + 12*n_layer)*512; // object overhead
printf("%s: ggml tensor size = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0)); printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
} }
@ -431,11 +434,10 @@ bool gpt2_eval(
{ {
struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd); struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd); struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
// store key and value to memory // store key and value to memory
if (N >= 1) { if (N >= 1) {
struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past)); struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past)); struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));

13
otherarch/gptj_v2.cpp
View file

@ -47,10 +47,9 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
fin.read((char *) &hparams.n_head, sizeof(hparams.n_head)); fin.read((char *) &hparams.n_head, sizeof(hparams.n_head));
fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer)); fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot)); fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot));
fin.read((char *) &hparams.ftype, sizeof(hparams.ftype)); fin.read((char *) &hparams.ftype, sizeof(hparams.ftype));
const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR; const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
hparams.ftype %= GGML_QNT_VERSION_FACTOR;
printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab); printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx); printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx);
@ -59,6 +58,9 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
printf("%s: n_layer = %d\n", __func__, hparams.n_layer); printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
printf("%s: n_rot = %d\n", __func__, hparams.n_rot); printf("%s: n_rot = %d\n", __func__, hparams.n_rot);
printf("%s: ftype = %d\n", __func__, hparams.ftype); printf("%s: ftype = %d\n", __func__, hparams.ftype);
printf("%s: qntvr = %d\n", __func__, qntvr);
hparams.ftype %= GGML_QNT_VERSION_FACTOR;
} }
// load vocab // load vocab
@ -134,7 +136,7 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_k ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_k
ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_v ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(memory_type); // memory_v
ctx_size += (5 + 10*n_layer)*256; // object overhead ctx_size += (5 + 10*n_layer)*512; // object overhead
printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0)); printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
} }
@ -160,7 +162,6 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
const int n_embd = hparams.n_embd; const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer; const int n_layer = hparams.n_layer;
const int n_ctx = hparams.n_ctx;
const int n_vocab = hparams.n_vocab; const int n_vocab = hparams.n_vocab;
model.layers.resize(n_layer); model.layers.resize(n_layer);
@ -358,8 +359,6 @@ bool gptj_eval(
const int n_vocab = hparams.n_vocab; const int n_vocab = hparams.n_vocab;
const int n_rot = hparams.n_rot; const int n_rot = hparams.n_rot;
const int d_key = n_embd/n_head;
static size_t buf_size = 256u*1024*1024; static size_t buf_size = 256u*1024*1024;
static void * buf = malloc(buf_size); static void * buf = malloc(buf_size);
@ -551,7 +550,7 @@ bool gptj_eval(
//if (n_past%100 == 0) { //if (n_past%100 == 0) {
// ggml_graph_print (&gf); // ggml_graph_print (&gf);
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot"); // ggml_graph_dump_dot(&gf, NULL, "gpt-j.dot");
//} //}
//embd_w.resize(n_vocab*N); //embd_w.resize(n_vocab*N);