vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-11 01:24:36 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

still merging in process

Browse source
This commit is contained in:
Concedo 2023-05-20 15:58:33 +08:00
parent a8958f6b76
commit a0cfed1e30
19 changed files with 6903 additions and 5644 deletions
Download patch file Download diff file Expand all files Collapse all files

270
otherarch/gptj_v2.cpp
View file

@ -1,4 +1,4 @@
#include "ggml.h"
#include "ggml_v2.h"
#include "otherarch.h"
#include "utils.h"
@ -49,7 +49,7 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot));
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_V2_QNT_VERSION_FACTOR;
printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx);
@ -60,7 +60,7 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_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_V2_QNT_VERSION_FACTOR;
}
// load vocab
@ -89,8 +89,8 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_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_v2_type wtype = ggml_v2_ftype_to_ggml_v2_type((ggml_v2_ftype) (model.hparams.ftype));
if (wtype == GGML_V2_TYPE_COUNT) {
fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
__func__, fname.c_str(), model.hparams.ftype);
return ModelLoadResult::FAIL;
@ -98,7 +98,7 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
auto & ctx = model.ctx;
auto memory_type = GGML_TYPE_F16;
auto memory_type = GGML_V2_TYPE_F16;
size_t ctx_size = 0;
@ -110,31 +110,31 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
const int n_ctx = hparams.n_ctx;
const int n_vocab = hparams.n_vocab;
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_v2_type_sizef(GGML_V2_TYPE_F32); // ln_f_g
ctx_size += n_embd*ggml_v2_type_sizef(GGML_V2_TYPE_F32); // ln_f_b
ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
ctx_size += n_embd*n_vocab*ggml_v2_type_sizef(wtype); // wte
ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // lmh_g
ctx_size += n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
ctx_size += n_embd*n_vocab*ggml_v2_type_sizef(wtype); // lmh_g
ctx_size += n_vocab*ggml_v2_type_sizef(GGML_V2_TYPE_F32); // lmh_b
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_v2_type_sizef(GGML_V2_TYPE_F32)); // ln_1_g
ctx_size += n_layer*(n_embd*ggml_v2_type_sizef(GGML_V2_TYPE_F32)); // ln_1_b
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_v2_type_sizef(wtype)); // c_attn_q_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_v2_type_sizef(wtype)); // c_attn_k_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_v2_type_sizef(wtype)); // c_attn_v_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
ctx_size += n_layer*(n_embd*n_embd*ggml_v2_type_sizef(wtype)); // c_attn_proj_w
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_v2_type_sizef(wtype)); // c_mlp_fc_w
ctx_size += n_layer*( 4*n_embd*ggml_v2_type_sizef(GGML_V2_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_v2_type_sizef(wtype)); // c_mlp_proj_w
ctx_size += n_layer*( n_embd*ggml_v2_type_sizef(GGML_V2_TYPE_F32)); // c_mlp_proj_b
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_v2_type_sizef(memory_type); // memory_k
ctx_size += n_ctx*n_layer*n_embd*ggml_v2_type_sizef(memory_type); // memory_v
ctx_size += (5 + 10*n_layer)*512; // object overhead
@ -143,15 +143,15 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
// create the ggml context
{
struct ggml_init_params params;
struct ggml_v2_init_params params;
params.mem_size = ctx_size;
params.mem_buffer = NULL;
params.no_alloc = false;
model.ctx = ggml_init(params);
model.ctx = ggml_v2_init(params);
if (!model.ctx) {
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
fprintf(stderr, "%s: ggml_v2_init() failed\n", __func__);
return ModelLoadResult::FAIL;
}
}
@ -166,13 +166,13 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
model.layers.resize(n_layer);
model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.wte = ggml_v2_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
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_v2_new_tensor_1d(ctx, GGML_V2_TYPE_F32, n_embd);
model.ln_f_b = ggml_v2_new_tensor_1d(ctx, GGML_V2_TYPE_F32, n_embd);
model.lmh_g = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.lmh_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
model.lmh_g = ggml_v2_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.lmh_b = ggml_v2_new_tensor_1d(ctx, GGML_V2_TYPE_F32, n_vocab);
// map by name
model.tensors["transformer.wte.weight"] = model.wte;
@ -186,20 +186,20 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_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_v2_new_tensor_1d(ctx, GGML_V2_TYPE_F32, n_embd);
layer.ln_1_b = ggml_v2_new_tensor_1d(ctx, GGML_V2_TYPE_F32, n_embd);
layer.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_q_proj_w = ggml_v2_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_k_proj_w = ggml_v2_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_v_proj_w = ggml_v2_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_attn_proj_w = ggml_v2_new_tensor_2d(ctx, wtype, n_embd, n_embd);
layer.c_mlp_fc_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd);
layer.c_mlp_fc_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
layer.c_mlp_fc_w = ggml_v2_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd);
layer.c_mlp_fc_b = ggml_v2_new_tensor_1d(ctx, GGML_V2_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_v2_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
layer.c_mlp_proj_b = ggml_v2_new_tensor_1d(ctx, GGML_V2_TYPE_F32, n_embd);
// map by name
model.tensors["transformer.h." + std::to_string(i) + ".ln_1.weight"] = layer.ln_1_g;
@ -230,10 +230,10 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
const int n_mem = n_layer*n_ctx;
const int n_elements = n_embd*n_mem;
model.memory_k = ggml_new_tensor_1d(ctx, memory_type, n_elements);
model.memory_v = ggml_new_tensor_1d(ctx, memory_type, n_elements);
model.memory_k = ggml_v2_new_tensor_1d(ctx, memory_type, n_elements);
model.memory_v = ggml_v2_new_tensor_1d(ctx, memory_type, n_elements);
const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
const size_t memory_size = ggml_v2_nbytes(model.memory_k) + ggml_v2_nbytes(model.memory_v);
printf("%s: memory_size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
}
@ -274,7 +274,7 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
}
auto tensor = model.tensors[name.data()];
if (ggml_nelements(tensor) != nelements) {
if (ggml_v2_nelements(tensor) != nelements) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
return ModelLoadResult::FAIL;
}
@ -286,7 +286,7 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
if(tensor->ne[0]==ne[1] && tensor->ne[1]==ne[0] && should_transpose_layer(name))
{
printf("\nFound a transposed tensor. This could be an older or newer model. Retrying load...");
ggml_free(ctx);
ggml_v2_free(ctx);
return ModelLoadResult::RETRY_LOAD;
}
else
@ -300,21 +300,21 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
// 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_v2_type_name(ggml_v2_type(ttype)), ggml_v2_nbytes(tensor)/1024.0/1024.0, ggml_v2_nbytes(tensor));
}
const size_t bpe = ggml_type_size(ggml_type(ttype));
const size_t bpe = ggml_v2_type_size(ggml_v2_type(ttype));
if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
if ((nelements*bpe)/ggml_v2_blck_size(tensor->type) != ggml_v2_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_v2_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_v2_nbytes(tensor));
//printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ttype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
total_size += ggml_nbytes(tensor);
//printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ttype == 0 ? "float" : "f16", ggml_v2_nbytes(tensor)/1024.0/1024.0);
total_size += ggml_v2_nbytes(tensor);
if (++n_tensors % 8 == 0) {
printf(".");
fflush(stdout);
@ -344,16 +344,16 @@ ModelLoadResult gptj_model_load(const std::string & fname, gptj_model & model, g
// for (int i = 0; i < n_gpu; ++i) {
// const auto & layer = model.layers[i];
// ggml_cl_transform_tensor(layer.ln_1_g); vram_total += ggml_nbytes(layer.ln_1_g);
// ggml_cl_transform_tensor(layer.ln_1_b); vram_total += ggml_nbytes(layer.ln_1_b);
// ggml_cl_transform_tensor(layer.c_attn_q_proj_w); vram_total += ggml_nbytes(layer.c_attn_q_proj_w);
// ggml_cl_transform_tensor(layer.c_attn_k_proj_w); vram_total += ggml_nbytes(layer.c_attn_k_proj_w);
// ggml_cl_transform_tensor(layer.c_attn_v_proj_w); vram_total += ggml_nbytes(layer.c_attn_v_proj_w);
// ggml_cl_transform_tensor(layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
// ggml_cl_transform_tensor(layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
// ggml_cl_transform_tensor(layer.c_mlp_fc_b); vram_total += ggml_nbytes(layer.c_mlp_fc_b);
// ggml_cl_transform_tensor(layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
// ggml_cl_transform_tensor(layer.c_mlp_proj_b); vram_total += ggml_nbytes(layer.c_mlp_proj_b);
// ggml_v2_cl_transform_tensor(layer.ln_1_g); vram_total += ggml_v2_nbytes(layer.ln_1_g);
// ggml_v2_cl_transform_tensor(layer.ln_1_b); vram_total += ggml_v2_nbytes(layer.ln_1_b);
// ggml_v2_cl_transform_tensor(layer.c_attn_q_proj_w); vram_total += ggml_v2_nbytes(layer.c_attn_q_proj_w);
// ggml_v2_cl_transform_tensor(layer.c_attn_k_proj_w); vram_total += ggml_v2_nbytes(layer.c_attn_k_proj_w);
// ggml_v2_cl_transform_tensor(layer.c_attn_v_proj_w); vram_total += ggml_v2_nbytes(layer.c_attn_v_proj_w);
// ggml_v2_cl_transform_tensor(layer.c_attn_proj_w); vram_total += ggml_v2_nbytes(layer.c_attn_proj_w);
// ggml_v2_cl_transform_tensor(layer.c_mlp_fc_w); vram_total += ggml_v2_nbytes(layer.c_mlp_fc_w);
// ggml_v2_cl_transform_tensor(layer.c_mlp_fc_b); vram_total += ggml_v2_nbytes(layer.c_mlp_fc_b);
// ggml_v2_cl_transform_tensor(layer.c_mlp_proj_w); vram_total += ggml_v2_nbytes(layer.c_mlp_proj_w);
// ggml_v2_cl_transform_tensor(layer.c_mlp_proj_b); vram_total += ggml_v2_nbytes(layer.c_mlp_proj_b);
// }
// fprintf(stderr, "%s: [opencl] total VRAM used: %zu MB\n", __func__, vram_total / 1024 / 1024);
@ -420,193 +420,193 @@ bool gptj_eval(
}
}
struct ggml_init_params params;
struct ggml_v2_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 = {};
struct ggml_v2_context * ctx0 = ggml_v2_init(params);
struct ggml_v2_cgraph gf = {};
gf.n_threads = n_threads;
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_v2_tensor * embd = ggml_v2_new_tensor_1d(ctx0, GGML_V2_TYPE_I32, N);
memcpy(embd->data, embd_inp.data(), N*ggml_v2_element_size(embd));
// wte
struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
struct ggml_v2_tensor * inpL = ggml_v2_get_rows(ctx0, model.wte, embd);
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * cur;
struct ggml_v2_tensor * cur;
// norm
{
cur = ggml_norm(ctx0, inpL);
cur = ggml_v2_norm(ctx0, inpL);
// cur = ln_1_g*cur + ln_1_b
cur = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
cur = ggml_v2_add(ctx0,
ggml_v2_mul(ctx0,
ggml_v2_repeat(ctx0, model.layers[il].ln_1_g, cur),
cur),
ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
ggml_v2_repeat(ctx0, model.layers[il].ln_1_b, cur));
}
struct ggml_tensor * inpSA = cur;
struct ggml_v2_tensor * inpSA = cur;
// self-attention
{
struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
struct ggml_v2_tensor * Qcur = ggml_v2_rope_inplace(ctx0, ggml_v2_reshape_3d(ctx0, ggml_v2_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
struct ggml_v2_tensor * Kcur = ggml_v2_rope_inplace(ctx0, ggml_v2_reshape_3d(ctx0, ggml_v2_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
// store key and value to memory
{
struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur));
struct ggml_v2_tensor * Vcur = ggml_v2_transpose(ctx0, ggml_v2_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur));
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_2d(ctx0, model.memory_v, N, n_embd,
( n_ctx)*ggml_element_size(model.memory_v),
(il*n_ctx)*ggml_element_size(model.memory_v)*n_embd + n_past*ggml_element_size(model.memory_v));
struct ggml_v2_tensor * k = ggml_v2_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_v2_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
struct ggml_v2_tensor * v = ggml_v2_view_2d(ctx0, model.memory_v, N, n_embd,
( n_ctx)*ggml_v2_element_size(model.memory_v),
(il*n_ctx)*ggml_v2_element_size(model.memory_v)*n_embd + n_past*ggml_v2_element_size(model.memory_v));
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
ggml_v2_build_forward_expand(&gf, ggml_v2_cpy(ctx0, Kcur, k));
ggml_v2_build_forward_expand(&gf, ggml_v2_cpy(ctx0, Vcur, v));
}
// Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
struct ggml_tensor * Q =
ggml_permute(ctx0,
struct ggml_v2_tensor * Q =
ggml_v2_permute(ctx0,
Qcur,
0, 2, 1, 3);
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
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_v2_tensor * K =
ggml_v2_permute(ctx0,
ggml_v2_reshape_3d(ctx0,
ggml_v2_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_v2_element_size(model.memory_k)*n_embd),
n_embd/n_head, n_head, n_past + N),
0, 2, 1, 3);
// K * Q
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
struct ggml_v2_tensor * KQ = ggml_v2_mul_mat(ctx0, K, Q);
// KQ_scaled = KQ / sqrt(n_embd/n_head)
struct ggml_tensor * KQ_scaled =
ggml_scale_inplace(ctx0,
struct ggml_v2_tensor * KQ_scaled =
ggml_v2_scale_inplace(ctx0,
KQ,
ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
ggml_v2_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
);
// KQ_masked = mask_past(KQ_scaled)
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
struct ggml_v2_tensor * KQ_masked = ggml_v2_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
// KQ = soft_max(KQ_masked)
struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
struct ggml_v2_tensor * KQ_soft_max = ggml_v2_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()
struct ggml_tensor * V =
ggml_view_3d(ctx0, model.memory_v,
struct ggml_v2_tensor * V =
ggml_v2_view_3d(ctx0, model.memory_v,
n_past + N, n_embd/n_head, n_head,
n_ctx*ggml_element_size(model.memory_v),
n_ctx*ggml_element_size(model.memory_v)*n_embd/n_head,
il*n_ctx*ggml_element_size(model.memory_v)*n_embd);
n_ctx*ggml_v2_element_size(model.memory_v),
n_ctx*ggml_v2_element_size(model.memory_v)*n_embd/n_head,
il*n_ctx*ggml_v2_element_size(model.memory_v)*n_embd);
// KQV = transpose(V) * KQ_soft_max
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
struct ggml_v2_tensor * KQV = ggml_v2_mul_mat(ctx0, V, KQ_soft_max);
// KQV_merged = KQV.permute(0, 2, 1, 3)
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
struct ggml_v2_tensor * KQV_merged = ggml_v2_permute(ctx0, KQV, 0, 2, 1, 3);
// cur = KQV_merged.contiguous().view(n_embd, N)
cur = ggml_cpy(ctx0,
cur = ggml_v2_cpy(ctx0,
KQV_merged,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
ggml_v2_new_tensor_2d(ctx0, GGML_V2_TYPE_F32, n_embd, N));
// projection (no bias)
cur = ggml_mul_mat(ctx0,
cur = ggml_v2_mul_mat(ctx0,
model.layers[il].c_attn_proj_w,
cur);
}
struct ggml_tensor * inpFF = cur;
struct ggml_v2_tensor * inpFF = cur;
// feed-forward network
// this is independent of the self-attention result, so it could be done in parallel to the self-attention
{
// note here we pass inpSA instead of cur
cur = ggml_mul_mat(ctx0,
cur = ggml_v2_mul_mat(ctx0,
model.layers[il].c_mlp_fc_w,
inpSA);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
cur = ggml_v2_add(ctx0,
ggml_v2_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
cur);
// GELU activation
cur = ggml_gelu(ctx0, cur);
cur = ggml_v2_gelu(ctx0, cur);
// projection
// cur = proj_w*cur + proj_b
cur = ggml_mul_mat(ctx0,
cur = ggml_v2_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_v2_add(ctx0,
ggml_v2_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
cur);
}
// self-attention + FF
cur = ggml_add(ctx0, cur, inpFF);
cur = ggml_v2_add(ctx0, cur, inpFF);
// input for next layer
inpL = ggml_add(ctx0, cur, inpL);
inpL = ggml_v2_add(ctx0, cur, inpL);
}
// norm
{
inpL = ggml_norm(ctx0, inpL);
inpL = ggml_v2_norm(ctx0, inpL);
// inpL = ln_f_g*inpL + ln_f_b
inpL = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.ln_f_g, inpL),
inpL = ggml_v2_add(ctx0,
ggml_v2_mul(ctx0,
ggml_v2_repeat(ctx0, model.ln_f_g, inpL),
inpL),
ggml_repeat(ctx0, model.ln_f_b, inpL));
ggml_v2_repeat(ctx0, model.ln_f_b, inpL));
}
// lm_head
{
inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
inpL = ggml_v2_mul_mat(ctx0, model.lmh_g, inpL);
inpL = ggml_add(ctx0,
ggml_repeat(ctx0, model.lmh_b, inpL),
inpL = ggml_v2_add(ctx0,
ggml_v2_repeat(ctx0, model.lmh_b, inpL),
inpL);
}
// logits -> probs
//inpL = ggml_soft_max_inplace(ctx0, inpL);
//inpL = ggml_v2_soft_max_inplace(ctx0, inpL);
// run the computation
ggml_build_forward_expand(&gf, inpL);
ggml_graph_compute (ctx0, &gf);
ggml_v2_build_forward_expand(&gf, inpL);
ggml_v2_graph_compute (ctx0, &gf);
//if (n_past%100 == 0) {
// ggml_graph_print (&gf);
// ggml_graph_dump_dot(&gf, NULL, "gpt-j.dot");
// ggml_v2_graph_print (&gf);
// ggml_v2_graph_dump_dot(&gf, NULL, "gpt-j.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_v2_get_data(inpL), sizeof(float)*n_vocab*N);
// return result for just 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_v2_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_v2_used_mem(ctx0)/N;
}
//printf("used_mem = %zu\n", ggml_used_mem(ctx0));
//printf("used_mem = %zu\n", ggml_v2_used_mem(ctx0));
ggml_free(ctx0);
ggml_v2_free(ctx0);
return true;
}