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

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	CMakeLists.txt
#	Makefile
This commit is contained in:
Concedo 2023-04-30 10:35:02 +08:00
commit 0061b90ec6
5 changed files with 114 additions and 62 deletions
Download patch file Download diff file Expand all files Collapse all files

80
examples/save-load-state/save-load-state.cpp
View file

@ -1,13 +1,10 @@
#include "common.h"
#include "llama.h"
#include <vector>
#include <cstdio>
#include <chrono>
#include "common.h"
#include "llama.h"
#include "llama.cpp"
using namespace std;
int main(int argc, char ** argv) {
gpt_params params;
params.model = "models/llama-7B/ggml-model.bin";
@ -20,21 +17,25 @@ int main(int argc, char ** argv) {
return 1;
}
if (params.n_predict < 0) {
params.n_predict = 16;
}
auto lparams = llama_context_default_params();
lparams.n_ctx = params.n_ctx;
lparams.n_parts = params.n_parts;
lparams.seed = params.seed;
lparams.f16_kv = params.memory_f16;
lparams.use_mmap = params.use_mmap;
lparams.use_mlock = params.use_mlock;
lparams.n_ctx = params.n_ctx;
lparams.n_parts = params.n_parts;
lparams.seed = params.seed;
lparams.f16_kv = params.memory_f16;
lparams.use_mmap = params.use_mmap;
lparams.use_mlock = params.use_mlock;
auto n_past = 0;
auto last_n_tokens_data = vector<llama_token>(params.repeat_last_n, 0);
auto last_n_tokens_data = std::vector<llama_token>(params.repeat_last_n, 0);
// init
auto ctx = llama_init_from_file(params.model.c_str(), lparams);
auto tokens = vector<llama_token>(params.n_ctx);
auto tokens = std::vector<llama_token>(params.n_ctx);
auto n_prompt_tokens = llama_tokenize(ctx, params.prompt.c_str(), tokens.data(), tokens.size(), true);
if (n_prompt_tokens < 1) {
@ -43,26 +44,29 @@ int main(int argc, char ** argv) {
}
// evaluate prompt
llama_eval(ctx, tokens.data(), n_prompt_tokens, n_past, params.n_threads);
last_n_tokens_data.insert(last_n_tokens_data.end(), tokens.data(), tokens.data() + n_prompt_tokens);
n_past += n_prompt_tokens;
const size_t state_size = llama_get_state_size(ctx);
uint8_t * state_mem = new uint8_t[state_size];
// Save state (rng, logits, embedding and kv_cache) to file
FILE *fp_write = fopen("dump_state.bin", "wb");
auto state_size = llama_get_state_size(ctx);
auto state_mem = new uint8_t[state_size];
llama_copy_state_data(ctx, state_mem); // could also copy directly to memory mapped file
fwrite(state_mem, 1, state_size, fp_write);
fclose(fp_write);
{
FILE *fp_write = fopen("dump_state.bin", "wb");
llama_copy_state_data(ctx, state_mem); // could also copy directly to memory mapped file
fwrite(state_mem, 1, state_size, fp_write);
fclose(fp_write);
}
// save state (last tokens)
auto last_n_tokens_data_saved = vector<llama_token>(last_n_tokens_data);
auto n_past_saved = n_past;
const auto last_n_tokens_data_saved = std::vector<llama_token>(last_n_tokens_data);
const auto n_past_saved = n_past;
// first run
printf("\n%s", params.prompt.c_str());
for (auto i = 0; i < params.n_predict; i++) {
auto logits = llama_get_logits(ctx);
auto n_vocab = llama_n_vocab(ctx);
@ -75,6 +79,7 @@ int main(int argc, char ** argv) {
auto next_token = llama_sample_token(ctx, &candidates_p);
auto next_token_str = llama_token_to_str(ctx, next_token);
last_n_tokens_data.push_back(next_token);
printf("%s", next_token_str);
if (llama_eval(ctx, &next_token, 1, n_past, params.n_threads)) {
fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
@ -82,24 +87,34 @@ int main(int argc, char ** argv) {
}
n_past += 1;
}
printf("\n\n");
// free old model
llama_free(ctx);
// load new model
auto ctx2 = llama_init_from_file(params.model.c_str(), lparams);
// Load state (rng, logits, embedding and kv_cache) from file
FILE *fp_read = fopen("dump_state.bin", "rb");
auto state_size2 = llama_get_state_size(ctx2);
if (state_size != state_size2) {
fprintf(stderr, "\n%s : failed to validate state size\n", __func__);
{
FILE *fp_read = fopen("dump_state.bin", "rb");
if (state_size != llama_get_state_size(ctx2)) {
fprintf(stderr, "\n%s : failed to validate state size\n", __func__);
return 1;
}
const size_t ret = fread(state_mem, 1, state_size, fp_read);
if (ret != state_size) {
fprintf(stderr, "\n%s : failed to read state\n", __func__);
return 1;
}
llama_set_state_data(ctx2, state_mem); // could also read directly from memory mapped file
fclose(fp_read);
}
fread(state_mem, 1, state_size, fp_read);
llama_set_state_data(ctx2, state_mem); // could also read directly from memory mapped file
fclose(fp_read);
delete[] state_mem;
// restore state (last tokens)
last_n_tokens_data = last_n_tokens_data_saved;
@ -118,6 +133,7 @@ int main(int argc, char ** argv) {
auto next_token = llama_sample_token(ctx2, &candidates_p);
auto next_token_str = llama_token_to_str(ctx2, next_token);
last_n_tokens_data.push_back(next_token);
printf("%s", next_token_str);
if (llama_eval(ctx2, &next_token, 1, n_past, params.n_threads)) {
fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
@ -125,6 +141,8 @@ int main(int argc, char ** argv) {
}
n_past += 1;
}
printf("\n\n");
return 0;
}

86
ggml.c
View file

@ -668,6 +668,33 @@ uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
return vget_high_u8(vcombine_u8(a, b));
}
int8x16_t vzip1q_s8(int8x16_t a, int8x16_t b) {
return vcombine_s8(vget_low_s8(a), vget_low_s8(b));
}
int8x16_t vzip2q_s8(int8x16_t a, int8x16_t b) {
return vcombine_s8(vget_high_s8(a), vget_high_s8(b));
}
uint8x16_t vzip1q_u8(uint8x16_t a, uint8x16_t b) {
return vcombine_u8(vget_low_u8(a), vget_low_u8(b));
}
uint8x16_t vzip2q_u8(uint8x16_t a, uint8x16_t b) {
return vcombine_u8(vget_high_u8(a), vget_high_u8(b));
}
int32x4_t vcvtnq_s32_f32(float32x4_t v) {
int32x4_t res;
res[0] = roundf(vgetq_lane_f32(v, 0));
res[1] = roundf(vgetq_lane_f32(v, 1));
res[2] = roundf(vgetq_lane_f32(v, 2));
res[3] = roundf(vgetq_lane_f32(v, 3));
return res;
}
#endif
#endif
@ -2748,35 +2775,35 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
// interleave
const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs);
const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs);
const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs);
const int8x16_t v0_1hz = vzip2q_s8(v0_1ls, v0_1hs);
// load y
const int8x16_t v1_0l = vld1q_s8(y0->qs);
const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
// interleave
const int8x16_t v1_0ls = vuzp1q_s8(v1_0l, v1_0h);
const int8x16_t v1_0hs = vuzp2q_s8(v1_0l, v1_0h);
const int8x16_t v1_1ls = vuzp1q_s8(v1_1l, v1_1h);
const int8x16_t v1_1hs = vuzp2q_s8(v1_1l, v1_1h);
#if defined(__ARM_FEATURE_DOTPROD)
// dot product into int32x4_t
const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls), v0_0hs, v1_0hs);
const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls), v0_1hs, v1_1hs);
const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h);
const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d);
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0hs));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0hs));
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l));
const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h));
const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1ls));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1ls));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1hs));
const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l));
const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l));
const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h));
const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h));
const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
@ -8470,8 +8497,6 @@ static void ggml_compute_forward_mul_mat_f16_f32(
ggml_fp16_t * d_X = ggml_cuda_pool_malloc(sizeof(float) * x_ne, &x_size);
ggml_fp16_t * d_Y = ggml_cuda_pool_malloc(sizeof(float) * y_ne, &y_size);
float * d_D = ggml_cuda_pool_malloc(sizeof(float) * d_ne, &d_size);
#else
float * const wdata = params->wdata;
#endif
for (int64_t i03 = 0; i03 < ne03; i03++) {
for (int64_t i02 = 0; i02 < ne02; i02++) {
@ -8488,8 +8513,11 @@ static void ggml_compute_forward_mul_mat_f16_f32(
wdata[id++] = GGML_FP32_TO_FP16(*(float *) ((char *) src1->data + i03*nb13 + i02*nb12 + i01*nb11 + i00*nb10));
}
}
assert(id*sizeof(ggml_fp16_t) <= params->wsize);
}
#else
float * const wdata = params->wdata;
{
size_t id = 0;
for (int64_t i01 = 0; i01 < ne01; ++i01) {
@ -8497,6 +8525,8 @@ static void ggml_compute_forward_mul_mat_f16_f32(
wdata[id++] = GGML_FP16_TO_FP32(*(ggml_fp16_t *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00));
}
}
assert(id*sizeof(float) <= params->wsize);
}
#endif
@ -8762,7 +8792,10 @@ static void ggml_compute_forward_mul_mat_q_f32(
dequantize_row_q((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01, wdata + id, ne00);
id += ne00;
}
assert(id*sizeof(float) <= params->wsize);
}
const float * x = wdata;
#endif
@ -9345,7 +9378,7 @@ static void ggml_compute_forward_alibi_f32(
//const int nb3 = src0->nb[3];
assert(nb0 == sizeof(float));
assert(ne1+n_past == ne0);
assert(ne1 + n_past == ne0); (void) n_past;
// add alibi to src0 (KQ_scaled)
const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
@ -9406,7 +9439,7 @@ static void ggml_compute_forward_alibi_f16(
//const int nb3 = src0->nb[3];
assert(nb0 == sizeof(ggml_fp16_t));
assert(ne1+n_past == ne0);
assert(ne1 + n_past == ne0); (void) n_past;
// add alibi to src0 (KQ_scaled)
const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
@ -11799,10 +11832,13 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
node->n_tasks = 1; // TODO: this actually is doing nothing
// the threads are still spinning
cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*MAX(ggml_nelements(node->src1), ggml_nelements(node->src0));
//printf("src0: ne0 = %d, ne1 = %d, ne = %d\n", node->src0->ne[0], node->src0->ne[1], node->src0->ne[0]*node->src0->ne[1]);
//printf("src1: ne0 = %d, ne1 = %d, ne = %d\n", node->src1->ne[0], node->src1->ne[1], node->src1->ne[0]*node->src1->ne[1]);
//printf("cur = %zu\n", cur);
#if defined(GGML_USE_CUBLAS)
// with cuBLAS, we need memory for the full 3D / 4D data of src1
cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1);
#else
// here we need memory just for single 2D matrix from src0
cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);
#endif
} else {
cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1);
}
@ -11811,7 +11847,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
#endif
} else if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) {
cur = 0;
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS)
#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
node->n_tasks = 1;
}

4
ggml.h
View file

@ -701,8 +701,8 @@ extern "C" {
struct ggml_tensor * c1);
// Mapping operations
GGML_API typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
GGML_API typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
GGML_API struct ggml_tensor * ggml_map_unary_f32(
struct ggml_context * ctx,

1
llama_util.h → llama-util.h Executable file → Normal file
View file

@ -430,5 +430,4 @@ struct llama_ctx_buffer {
typedef llama_buffer llama_ctx_buffer;
#endif
#endif

5
llama.cpp
View file

@ -5,7 +5,7 @@
#include <cstdio>
#endif
#include "llama_util.h"
#include "llama-util.h"
#include "llama.h"
#include "ggml.h"
@ -33,7 +33,6 @@
#define LLAMA_USE_SCRATCH
#define LLAMA_MAX_SCRATCH_BUFFERS 16
// available llama models
enum e_model {
MODEL_UNKNOWN,
@ -790,7 +789,7 @@ static bool kv_cache_init(
const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer;
const int64_t n_mem = (int64_t)n_layer*n_ctx;
const int64_t n_mem = n_layer*n_ctx;
const int64_t n_elements = n_embd*n_mem;
cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);