vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-12 09:59:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

fossilize ggml library ver 3, to support ggjtv3

Browse source
This commit is contained in:
Concedo 2024-01-20 10:49:25 +08:00
parent 1804238e3f
commit db14de5c32
18 changed files with 44315 additions and 1591 deletions
Download patch file Download diff file Expand all files Collapse all files

326
otherarch/gpt2_v3.cpp
View file

@ -1,4 +1,4 @@
#include "ggml.h"
#include "ggml_v3.h"
#include "otherarch.h"
#include "utils.h"
@ -17,10 +17,10 @@
#include "model_adapter.h"
#ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h"
#include "ggml_v3-cuda.h"
#endif
#if defined(GGML_USE_CLBLAST)
#include "ggml-opencl.h"
#include "ggml_v3-opencl.h"
#endif
@ -57,7 +57,7 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
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_V3_QNT_VERSION_FACTOR;
printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
printf("%s: n_ctx = %d (%d)\n", __func__, hparams.n_ctx,origmaxctx);
@ -67,7 +67,7 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
printf("%s: ftype = %d\n", __func__, hparams.ftype);
printf("%s: qntvr = %d\n", __func__, qntvr);
hparams.ftype %= GGML_QNT_VERSION_FACTOR;
hparams.ftype %= GGML_V3_QNT_VERSION_FACTOR;
}
// load vocab
@ -113,8 +113,8 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
// for the big tensors, we have the option to store the data in 16-bit floats or quantized
// in order to save memory and also to speed up the computation
ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
if (wtype == GGML_TYPE_COUNT) {
ggml_v3_type wtype = ggml_v3_ftype_to_ggml_v3_type((ggml_v3_ftype) (model.hparams.ftype));
if (wtype == GGML_V3_TYPE_COUNT) {
fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
__func__, fname.c_str(), model.hparams.ftype);
return ModelLoadResult::FAIL;
@ -136,33 +136,33 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
const int kv_heads = hparams.n_head; // 1 if MQA else hparams.n_head
const int kv_dim = kv_heads * head_dim;
ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
ctx_size += n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F32); // ln_f_g
ctx_size += n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F32); // ln_f_b
ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype); // wte
ctx_size += n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype); // lm_head
ctx_size += n_vocab*n_embd*ggml_v3_type_sizef(wtype); // wte
ctx_size += n_ctx*n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F32); // wpe
ctx_size += n_vocab*n_embd*ggml_v3_type_sizef(wtype); // lm_head
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
ctx_size += n_layer*(n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F32)); // ln_1_g
ctx_size += n_layer*(n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F32)); // ln_1_b
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
ctx_size += n_layer*(n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F32)); // ln_2_g
ctx_size += n_layer*(n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F32)); // ln_2_b
ctx_size += n_layer*((n_embd + 2*kv_dim)*n_embd*ggml_type_sizef(wtype)); // c_attn_attn_w // TODO:
ctx_size += n_layer*( (n_embd + 2*kv_dim)*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
ctx_size += n_layer*((n_embd + 2*kv_dim)*n_embd*ggml_v3_type_sizef(wtype)); // c_attn_attn_w // TODO:
ctx_size += n_layer*( (n_embd + 2*kv_dim)*ggml_v3_type_sizef(GGML_V3_TYPE_F32)); // c_attn_attn_b
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
ctx_size += n_layer*( n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_proj_b
ctx_size += n_layer*(n_embd*n_embd*ggml_v3_type_sizef(wtype)); // c_attn_proj_w
ctx_size += n_layer*( n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F32)); // c_attn_proj_b
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_fc_w
ctx_size += n_layer*( 4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
ctx_size += n_layer*(4*n_embd*n_embd*ggml_v3_type_sizef(wtype)); // c_mlp_fc_w
ctx_size += n_layer*( 4*n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F32)); // c_mlp_fc_b
ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_proj_w
ctx_size += n_layer*( n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
ctx_size += n_layer*(4*n_embd*n_embd*ggml_v3_type_sizef(wtype)); // c_mlp_proj_w
ctx_size += n_layer*( n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F32)); // c_mlp_proj_b
ctx_size += std::max(origmaxctx,n_ctx)*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F16); // memory_k
ctx_size += std::max(origmaxctx,n_ctx)*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F16); // memory_v
ctx_size += std::max(origmaxctx,n_ctx)*n_layer*n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F16); // memory_k
ctx_size += std::max(origmaxctx,n_ctx)*n_layer*n_embd*ggml_v3_type_sizef(GGML_V3_TYPE_F16); // memory_v
ctx_size += (6 + 12*n_layer)*1024; // object overhead
@ -171,14 +171,14 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
// create the ggml context
{
struct ggml_init_params params;
struct ggml_v3_init_params params;
params.mem_size = ctx_size;
params.mem_buffer = NULL;
params.no_alloc = false;
model.ctx = ggml_init(params);
model.ctx = ggml_v3_init(params);
if (!model.ctx) {
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
fprintf(stderr, "%s: ggml_v3_init() failed\n", __func__);
return ModelLoadResult::FAIL;
}
}
@ -198,12 +198,12 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
model.layers.resize(n_layer);
model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
model.ln_f_g = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F32, n_embd);
model.ln_f_b = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F32, n_embd);
model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
model.lm_head = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.wte = ggml_v3_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.wpe = ggml_v3_new_tensor_2d(ctx, GGML_V3_TYPE_F32, n_embd, n_ctx);
model.lm_head = ggml_v3_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
// map by name
model.tensors["model/ln_f/g"] = model.ln_f_g;
@ -216,23 +216,23 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
for (int i = 0; i < n_layer; ++i) {
auto & layer = model.layers[i];
layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_1_g = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F32, n_embd);
layer.ln_1_b = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F32, n_embd);
layer.ln_2_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_2_g = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F32, n_embd);
layer.ln_2_b = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F32, n_embd);
layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd + 2*kv_dim);
layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd + 2*kv_dim);
layer.c_attn_attn_w = ggml_v3_new_tensor_2d(ctx, wtype, n_embd, n_embd + 2*kv_dim);
layer.c_attn_attn_b = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F32, n_embd + 2*kv_dim);
layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.c_attn_proj_w = ggml_v3_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_proj_b = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F32, n_embd);
layer.c_mlp_fc_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd); //TODO: 4*n_embd = config.n_inner
layer.c_mlp_fc_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
layer.c_mlp_fc_w = ggml_v3_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd); //TODO: 4*n_embd = config.n_inner
layer.c_mlp_fc_b = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F32, 4*n_embd);
layer.c_mlp_proj_w = ggml_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.c_mlp_proj_w = ggml_v3_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
layer.c_mlp_proj_b = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F32, n_embd);
// map by name
model.tensors["model/h" + std::to_string(i) + "/ln_1/g"] = layer.ln_1_g;
@ -266,10 +266,10 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
const int n_mem = n_layer*std::max(origmaxctx,n_ctx);
const int n_elements = n_embd*n_mem;
model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
model.memory_k = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F16, n_elements);
model.memory_v = ggml_v3_new_tensor_1d(ctx, GGML_V3_TYPE_F16, n_elements);
const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
const size_t memory_size = ggml_v3_nbytes(model.memory_k) + ggml_v3_nbytes(model.memory_v);
printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
}
@ -314,37 +314,37 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
__func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
return ModelLoadResult::FAIL;
}
if (ggml_nelements(tensor) != nelements) {
if (ggml_v3_nelements(tensor) != nelements) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file. got %d, expected %d\n",
__func__, name.data(), (int) ggml_nelements(tensor), nelements);
__func__, name.data(), (int) ggml_v3_nelements(tensor), nelements);
return ModelLoadResult::FAIL;
}
// for debugging
if (0) {
printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_v3_type_name(ggml_v3_type(ttype)), ggml_v3_nbytes(tensor)/1024.0/1024.0, ggml_v3_nbytes(tensor));
}
const size_t bpe = ggml_type_size(ggml_type(ttype));
const size_t bpe = ggml_v3_type_size(ggml_v3_type(ttype));
if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
if ((nelements*bpe)/ggml_v3_blck_size(tensor->type) != ggml_v3_nbytes(tensor)) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
__func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
__func__, name.data(), ggml_v3_nbytes(tensor), nelements*bpe);
return ModelLoadResult::FAIL;
}
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
fin.read(reinterpret_cast<char *>(tensor->data), ggml_v3_nbytes(tensor));
// GPT-2 models share the WTE tensor as the LM head
if (name == "model/wte" && has_lm_head == false) {
memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
memcpy(model.lm_head->data, tensor->data, ggml_v3_nbytes(tensor));
}
if (name == "model/lm_head") {
has_lm_head = true;
}
total_size += ggml_nbytes(tensor);
total_size += ggml_v3_nbytes(tensor);
}
printf("%s: model size = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
@ -366,20 +366,20 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g
#endif
for (int i = 0; i < n_gpu; ++i) {
const auto & layer = model.layers[i];
layer.c_attn_attn_w->backend = GGML_BACKEND_GPU;
layer.c_attn_proj_w->backend = GGML_BACKEND_GPU;
layer.c_mlp_fc_w->backend = GGML_BACKEND_GPU;
layer.c_mlp_proj_w->backend = GGML_BACKEND_GPU;
layer.c_attn_attn_w->backend = GGML_V3_BACKEND_GPU;
layer.c_attn_proj_w->backend = GGML_V3_BACKEND_GPU;
layer.c_mlp_fc_w->backend = GGML_V3_BACKEND_GPU;
layer.c_mlp_proj_w->backend = GGML_V3_BACKEND_GPU;
#if defined(GGML_USE_CLBLAST)
ggml_cl_transform_tensor(layer.c_attn_attn_w->data,layer.c_attn_attn_w); vram_total += ggml_nbytes(layer.c_attn_attn_w);
ggml_cl_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
ggml_cl_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
ggml_cl_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
ggml_v3_cl_transform_tensor(layer.c_attn_attn_w->data,layer.c_attn_attn_w); vram_total += ggml_v3_nbytes(layer.c_attn_attn_w);
ggml_v3_cl_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_v3_nbytes(layer.c_attn_proj_w);
ggml_v3_cl_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_v3_nbytes(layer.c_mlp_fc_w);
ggml_v3_cl_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_v3_nbytes(layer.c_mlp_proj_w);
#else
ggml_cuda_transform_tensor(layer.c_attn_attn_w->data,layer.c_attn_attn_w); vram_total += ggml_nbytes(layer.c_attn_attn_w);
ggml_cuda_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
ggml_cuda_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
ggml_cuda_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
ggml_v3_cuda_transform_tensor(layer.c_attn_attn_w->data,layer.c_attn_attn_w); vram_total += ggml_v3_nbytes(layer.c_attn_attn_w);
ggml_v3_cuda_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_v3_nbytes(layer.c_attn_proj_w);
ggml_v3_cuda_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_v3_nbytes(layer.c_mlp_fc_w);
ggml_v3_cuda_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_v3_nbytes(layer.c_mlp_proj_w);
#endif
}
#if defined(GGML_USE_CLBLAST)
@ -448,48 +448,48 @@ bool gpt2_eval(
}
}
struct ggml_init_params params;
struct ggml_v3_init_params params;
params.mem_size = buf_size;
params.mem_buffer = buf;
params.no_alloc = false;
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, 8192, false);
struct ggml_v3_context * ctx0 = ggml_v3_init(params);
struct ggml_v3_cgraph * gf = ggml_v3_new_graph_custom(ctx0, 8192, false);
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
struct ggml_v3_tensor * embd = ggml_v3_new_tensor_1d(ctx0, GGML_V3_TYPE_I32, N);
memcpy(embd->data, embd_inp.data(), N*ggml_v3_element_size(embd));
struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
struct ggml_v3_tensor * position = ggml_v3_new_tensor_1d(ctx0, GGML_V3_TYPE_I32, N);
for (int i = 0; i < N; ++i) {
((int32_t *) position->data)[i] = n_past + i;
}
// wte + wpe
struct ggml_tensor * inpL =
ggml_add(ctx0,
ggml_get_rows(ctx0, model.wte, embd),
ggml_get_rows(ctx0, model.wpe, position));
struct ggml_v3_tensor * inpL =
ggml_v3_add(ctx0,
ggml_v3_get_rows(ctx0, model.wte, embd),
ggml_v3_get_rows(ctx0, model.wpe, position));
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * cur;
struct ggml_v3_tensor * cur;
if(use_scratch){
ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
ggml_v3_set_scratch(ctx0, { 0, scr0_size, scr0, });
}
// norm
{
// [ 768, N]
cur = ggml_norm(ctx0, inpL, default_norm_eps);
cur = ggml_v3_norm(ctx0, inpL, default_norm_eps);
// cur = ln_1_g*cur + ln_1_b
// [ 768, N]
cur = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
cur = ggml_v3_add(ctx0,
ggml_v3_mul(ctx0,
ggml_v3_repeat(ctx0, model.layers[il].ln_1_g, cur),
cur),
ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
ggml_v3_repeat(ctx0, model.layers[il].ln_1_b, cur));
}
// attn
@ -501,104 +501,104 @@ bool gpt2_eval(
// cur = attn_w*cur + attn_b
// [2304, N]
{
cur = ggml_mul_mat(ctx0,
cur = ggml_v3_mul_mat(ctx0,
model.layers[il].c_attn_attn_w,
cur);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
cur = ggml_v3_add(ctx0,
ggml_v3_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
cur);
}
// self-attention
{
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 * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
struct ggml_v3_tensor * Qcur = ggml_v3_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
struct ggml_v3_tensor * Kcur = ggml_v3_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
struct ggml_v3_tensor * Vcur = ggml_v3_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
// store key and value to memory
if (N >= 1) {
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_v3_tensor * k = ggml_v3_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_v3_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
struct ggml_v3_tensor * v = ggml_v3_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_v3_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
ggml_v3_build_forward_expand(gf, ggml_v3_cpy(ctx0, Kcur, k));
ggml_v3_build_forward_expand(gf, ggml_v3_cpy(ctx0, Vcur, v));
}
// Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
// [64, N, 12]
struct ggml_tensor * Q =
ggml_permute(ctx0,
ggml_cpy(ctx0,
struct ggml_v3_tensor * Q =
ggml_v3_permute(ctx0,
ggml_v3_cpy(ctx0,
Qcur,
ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
ggml_v3_new_tensor_3d(ctx0, GGML_V3_TYPE_F32, n_embd/n_head, n_head, N)),
0, 2, 1, 3);
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
// [64, n_past + N, 12]
struct ggml_tensor * K =
ggml_permute(ctx0,
ggml_reshape_3d(ctx0,
ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
struct ggml_v3_tensor * K =
ggml_v3_permute(ctx0,
ggml_v3_reshape_3d(ctx0,
ggml_v3_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_v3_element_size(model.memory_k)*n_embd),
n_embd/n_head, n_head, n_past + N),
0, 2, 1, 3); //TODO: need to be tiled
// GG: flash attention
//struct ggml_tensor * V =
// ggml_cpy(ctx0,
// ggml_permute(ctx0,
// ggml_reshape_3d(ctx0,
// ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
//struct ggml_v3_tensor * V =
// ggml_v3_cpy(ctx0,
// ggml_v3_permute(ctx0,
// ggml_v3_reshape_3d(ctx0,
// ggml_v3_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_v3_element_size(model.memory_v)*n_embd),
// n_embd/n_head, n_head, n_past + N),
// 1, 2, 0, 3),
// ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
// ggml_v3_new_tensor_3d(ctx0, GGML_V3_TYPE_F32, n_past + N, n_embd/n_head, n_head));
//struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
//struct ggml_v3_tensor * KQV = ggml_v3_flash_attn(ctx0, Q, K, V, true);
// K * Q
// [n_past + N, N, 12]
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); //TODO: check if it broadcasts
struct ggml_v3_tensor * KQ = ggml_v3_mul_mat(ctx0, K, Q); //TODO: check if it broadcasts
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_scaled =
ggml_scale_inplace(ctx0,
struct ggml_v3_tensor * KQ_scaled =
ggml_v3_scale_inplace(ctx0,
KQ,
1.0f/sqrt(float(n_embd)/n_head)
);
// KQ_masked = mask_past(KQ_scaled)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
struct ggml_v3_tensor * KQ_masked = ggml_v3_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
// KQ = soft_max(KQ_masked)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
struct ggml_v3_tensor * KQ_soft_max = ggml_v3_soft_max_inplace(ctx0, KQ_masked);
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
// [n_past + N, 64, 12]
struct ggml_tensor * V_trans =
ggml_cpy(ctx0,
ggml_permute(ctx0,
ggml_reshape_3d(ctx0,
ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
struct ggml_v3_tensor * V_trans =
ggml_v3_cpy(ctx0,
ggml_v3_permute(ctx0,
ggml_v3_reshape_3d(ctx0,
ggml_v3_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_v3_element_size(model.memory_v)*n_embd),
n_embd/n_head, n_head, n_past + N),
1, 2, 0, 3),
ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
ggml_v3_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
// KQV = transpose(V) * KQ_soft_max
// [64, N, 12]
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
struct ggml_v3_tensor * KQV = ggml_v3_mul_mat(ctx0, V_trans, KQ_soft_max);
// KQV_merged = KQV.permute(0, 2, 1, 3)
// [64, 12, N]
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
struct ggml_v3_tensor * KQV_merged = ggml_v3_permute(ctx0, KQV, 0, 2, 1, 3);
// cur = KQV_merged.contiguous().view(n_embd, N)
// [768, N]
cur = ggml_cpy(ctx0,
cur = ggml_v3_cpy(ctx0,
KQV_merged,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
ggml_v3_new_tensor_2d(ctx0, GGML_V3_TYPE_F32, n_embd, N));
}
// projection
@ -610,37 +610,37 @@ bool gpt2_eval(
// cur = proj_w*cur + proj_b
// [768, N]
{
cur = ggml_mul_mat(ctx0,
cur = ggml_v3_mul_mat(ctx0,
model.layers[il].c_attn_proj_w,
cur);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
cur = ggml_v3_add(ctx0,
ggml_v3_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
cur);
}
// add the input
cur = ggml_add(ctx0, cur, inpL);
cur = ggml_v3_add(ctx0, cur, inpL);
struct ggml_tensor * inpFF = cur;
struct ggml_v3_tensor * inpFF = cur;
if(use_scratch){
ggml_set_scratch(ctx0, { 0, scr1_size, scr1, });
ggml_v3_set_scratch(ctx0, { 0, scr1_size, scr1, });
}
// feed-forward network
{
// norm
{
cur = ggml_norm(ctx0, inpFF, default_norm_eps);
cur = ggml_v3_norm(ctx0, inpFF, default_norm_eps);
// cur = ln_2_g*cur + ln_2_b
// [ 768, N]
cur = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
cur = ggml_v3_add(ctx0,
ggml_v3_mul(ctx0,
ggml_v3_repeat(ctx0, model.layers[il].ln_2_g, cur),
cur),
ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
ggml_v3_repeat(ctx0, model.layers[il].ln_2_b, cur));
}
// fully connected
@ -651,17 +651,17 @@ bool gpt2_eval(
//
// cur = fc_w*cur + fc_b
// [3072, N]
cur = ggml_mul_mat(ctx0,
cur = ggml_v3_mul_mat(ctx0,
model.layers[il].c_mlp_fc_w,
cur);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
cur = ggml_v3_add(ctx0,
ggml_v3_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
cur);
// GELU activation
// [3072, N]
cur = ggml_gelu(ctx0, cur);
cur = ggml_v3_gelu(ctx0, cur);
// projection
// [ 768, 3072] - model.layers[il].c_mlp_proj_w
@ -671,71 +671,71 @@ bool gpt2_eval(
//
// cur = proj_w*cur + proj_b
// [768, N]
cur = ggml_mul_mat(ctx0,
cur = ggml_v3_mul_mat(ctx0,
model.layers[il].c_mlp_proj_w,
cur);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
cur = ggml_v3_add(ctx0,
ggml_v3_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
cur);
}
// input for next layer
inpL = ggml_add(ctx0, cur, inpFF);
inpL = ggml_v3_add(ctx0, cur, inpFF);
}
if(use_scratch){
ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
ggml_v3_set_scratch(ctx0, { 0, scr0_size, scr0, });
}
// norm
{
// [ 768, N]
inpL = ggml_norm(ctx0, inpL, default_norm_eps);
inpL = ggml_v3_norm(ctx0, inpL, default_norm_eps);
// inpL = ln_f_g*inpL + ln_f_b
// [ 768, N]
inpL = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.ln_f_g, inpL),
inpL = ggml_v3_add(ctx0,
ggml_v3_mul(ctx0,
ggml_v3_repeat(ctx0, model.ln_f_g, inpL),
inpL),
ggml_repeat(ctx0, model.ln_f_b, inpL));
ggml_v3_repeat(ctx0, model.ln_f_b, inpL));
}
if(use_scratch){
ggml_set_scratch(ctx0, { 0, 0, nullptr, });
ggml_v3_set_scratch(ctx0, { 0, 0, nullptr, });
}
// inpL = WTE * inpL
// [ 768, 50257] - model.lm_head
// [ 768, N] - inpL
inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
inpL = ggml_v3_mul_mat(ctx0, model.lm_head, inpL);
// logits -> probs
//inpL = ggml_soft_max_inplace(ctx0, inpL);
//inpL = ggml_v3_soft_max_inplace(ctx0, inpL);
// run the computation
ggml_build_forward_expand(gf, inpL);
ggml_v3_build_forward_expand(gf, inpL);
kcpp_graph_compute_helper(gf, n_threads);
//if (n_past%100 == 0) {
// ggml_graph_print (&gf);
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
// ggml_v3_graph_print (&gf);
// ggml_v3_graph_dump_dot(&gf, NULL, "gpt-2.dot");
//}
//embd_w.resize(n_vocab*N);
//memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
//memcpy(embd_w.data(), ggml_v3_get_data(inpL), sizeof(float)*n_vocab*N);
// return result just for the last token
embd_w.resize(n_vocab);
memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
memcpy(embd_w.data(), (float *) ggml_v3_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
if (mem_per_token == 0) {
mem_per_token = ggml_used_mem(ctx0)/N;
mem_per_token = ggml_v3_used_mem(ctx0)/N;
}
//printf("used_mem = %zu MB\n", ggml_used_mem(ctx0)/(1024*1024));
//printf("used_mem = %zu MB\n", ggml_v3_used_mem(ctx0)/(1024*1024));
ggml_free(ctx0);
ggml_v3_free(ctx0);
return true;
}