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

Merge branch 'ggerganov:master' into patch-4

Browse source
This commit is contained in:
FirstTimeEZ 2024-11-16 19:08:13 +13:00 • committed by GitHub
commit aa51d1dd43
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 123 additions and 56 deletions
Download patch file Download diff file Expand all files Collapse all files

34
CMakePresets.json
View file

@ -24,11 +24,12 @@
"CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
}
},
{ "name": "debug", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
{ "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
{ "name": "reldbg", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
{ "name": "static", "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
{ "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
{ "name": "debug", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
{ "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
{ "name": "reldbg", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
{ "name": "static", "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
{ "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
{ "name": "vulkan", "hidden": true, "cacheVariables": { "GGML_VULKAN": "ON" } },
{
"name": "arm64-windows-msvc", "hidden": true,
@ -57,25 +58,28 @@
}
},
{ "name": "arm64-windows-llvm-debug" , "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
{ "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },
{ "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] },
{ "name": "arm64-windows-llvm-debug", "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
{ "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },
{ "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] },
{ "name": "arm64-apple-clang-debug" , "inherits": [ "base", "arm64-apple-clang", "debug" ] },
{ "name": "arm64-apple-clang-release" , "inherits": [ "base", "arm64-apple-clang", "reldbg" ] },
{ "name": "arm64-apple-clang+static-release" , "inherits": [ "base", "arm64-apple-clang", "reldbg", "static" ] },
{ "name": "arm64-apple-clang-debug", "inherits": [ "base", "arm64-apple-clang", "debug" ] },
{ "name": "arm64-apple-clang-release", "inherits": [ "base", "arm64-apple-clang", "reldbg" ] },
{ "name": "arm64-apple-clang+static-release", "inherits": [ "base", "arm64-apple-clang", "reldbg", "static" ] },
{ "name": "arm64-windows-msvc-debug" , "inherits": [ "base", "arm64-windows-msvc", "debug" ] },
{ "name": "arm64-windows-msvc-debug", "inherits": [ "base", "arm64-windows-msvc", "debug" ] },
{ "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg" ] },
{ "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc", "reldbg", "static" ] },
{ "name": "x64-windows-msvc-debug" , "inherits": [ "base", "debug" ] },
{ "name": "x64-windows-msvc-debug", "inherits": [ "base", "debug" ] },
{ "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] },
{ "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
{ "name": "x64-windows-sycl-debug" , "inherits": [ "sycl-base", "debug" ] },
{ "name": "x64-windows-sycl-debug", "inherits": [ "sycl-base", "debug" ] },
{ "name": "x64-windows-sycl-debug-f16", "inherits": [ "sycl-base", "debug", "sycl_f16" ] },
{ "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] },
{ "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] }
{ "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] },
{ "name": "x64-windows-vulkan-debug", "inherits": [ "base", "vulkan", "debug" ] },
{ "name": "x64-windows-vulkan-release", "inherits": [ "base", "vulkan", "release" ] }
]
}

57
ggml/src/ggml-aarch64.c
View file

@ -8,19 +8,42 @@
#define UNUSED GGML_UNUSED
static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
block_q4_0x4 out;
for (int i = 0; i < 4; i++) {
out.d[i] = in[i].d;
}
for (int i = 0; i < QK4_0 * 2; i++) {
int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (i % blck_size_interleave);
const int end = QK4_0 * 2 / blck_size_interleave;
out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
if (blck_size_interleave == 8) {
const uint64_t xor_mask = 0x8888888888888888ULL;
for (int i = 0; i < end; ++i) {
int src_id = i % 4;
int src_offset = (i / 4) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint64_t elems;
// Using memcpy to avoid unaligned memory accesses
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
}
} else if (blck_size_interleave == 4) {
const uint32_t xor_mask = 0x88888888;
for (int i = 0; i < end; ++i) {
int src_id = i % 4;
int src_offset = (i / 4) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint32_t elems;
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
}
} else {
GGML_ASSERT(false);
}
return out;
@ -30,19 +53,25 @@ static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_in
// returns an interleaved block_q4_0x8
// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
block_q4_0x8 out;
for (int i = 0; i < 8; i++) {
out.d[i] = in[i].d;
}
for (int i = 0; i < QK4_0 * 4; i++) {
int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
src_offset += (i % blck_size_interleave);
const int end = QK4_0 * 4 / blck_size_interleave;
const uint64_t xor_mask = 0x8888888888888888ULL;
out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
for (int i = 0; i < end; ++i) {
int src_id = i % 8;
int src_offset = (i / 8) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint64_t elems;
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
}
return out;
@ -71,11 +100,11 @@ static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict ds
}
if (nrows_interleaved == 8) {
*(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88);
*(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave);
out_ptr = (block_q4_0x8 *) out_ptr + 1;
}
else if (nrows_interleaved == 4) {
*(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88);
*(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave);
out_ptr = (block_q4_0x4 *) out_ptr + 1;
}
}

57
ggml/src/ggml-cpu/ggml-cpu-aarch64.c
View file

@ -3387,19 +3387,42 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
}
// FIXME: this code is duplicated from ggml-aarch64.c
static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
block_q4_0x4 out;
for (int i = 0; i < 4; i++) {
out.d[i] = in[i].d;
}
for (int i = 0; i < QK4_0 * 2; i++) {
int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (i % blck_size_interleave);
const int end = QK4_0 * 2 / blck_size_interleave;
out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
if (blck_size_interleave == 8) {
const uint64_t xor_mask = 0x8888888888888888ULL;
for (int i = 0; i < end; ++i) {
int src_id = i % 4;
int src_offset = (i / 4) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint64_t elems;
// Using memcpy to avoid unaligned memory accesses
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
}
} else if (blck_size_interleave == 4) {
const uint32_t xor_mask = 0x88888888;
for (int i = 0; i < end; ++i) {
int src_id = i % 4;
int src_offset = (i / 4) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint32_t elems;
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
}
} else {
GGML_ASSERT(false);
}
return out;
@ -3409,19 +3432,25 @@ static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_in
// returns an interleaved block_q4_0x8
// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
block_q4_0x8 out;
for (int i = 0; i < 8; i++) {
out.d[i] = in[i].d;
}
for (int i = 0; i < QK4_0 * 4; i++) {
int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
src_offset += (i % blck_size_interleave);
const int end = QK4_0 * 4 / blck_size_interleave;
const uint64_t xor_mask = 0x8888888888888888ULL;
out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
for (int i = 0; i < end; ++i) {
int src_id = i % 8;
int src_offset = (i / 8) * blck_size_interleave;
int dst_offset = i * blck_size_interleave;
uint64_t elems;
memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
elems ^= xor_mask;
memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
}
return out;
@ -3449,7 +3478,7 @@ static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block
for (int i = 0; i < nrows_interleaved; i++) {
dst_tmp[i] = src[x + i * nblocks];
}
*dst++ = make_block_q4_0x4(dst_tmp, interleave_block, 0x88);
*dst++ = make_block_q4_0x4(dst_tmp, interleave_block);
}
src += nrows_interleaved * nblocks;
}
@ -3480,7 +3509,7 @@ static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor *t, int interleave_block,
for (int i = 0; i < nrows_interleaved; i++ ) {
dst_tmp[i] = src[x + i * nblocks];
}
*dst++ = make_block_q4_0x8(dst_tmp, interleave_block, 0x88);
*dst++ = make_block_q4_0x8(dst_tmp, interleave_block);
}
src += nrows_interleaved * nblocks;
}

31
src/llama.cpp
View file

@ -2907,9 +2907,15 @@ struct llama_model {
// for quantize-stats only
std::vector<std::pair<std::string, struct ggml_tensor *>> tensors_by_name;
int64_t t_load_us = 0;
int64_t t_load_us = 0;
int64_t t_start_us = 0;
// total number of parameters in the model
uint64_t n_elements = 0;
// total size of all the tensors in the model in bytes
size_t n_bytes = 0;
// keep track of loaded lora adapters
std::set<struct llama_lora_adapter *> lora_adapters;
@ -4275,8 +4281,8 @@ struct llama_model_loader {
int n_tensors = 0;
int n_created = 0;
int64_t n_elements = 0;
size_t n_bytes = 0;
uint64_t n_elements = 0;
size_t n_bytes = 0;
bool use_mmap = false;
bool check_tensors;
@ -5344,6 +5350,11 @@ static const char * llama_model_vocab_type_name(enum llama_vocab_type type){
}
}
static void llm_load_stats(llama_model_loader & ml, llama_model & model) {
model.n_elements = ml.n_elements;
model.n_bytes = ml.n_bytes;
}
static void llm_load_arch(llama_model_loader & ml, llama_model & model) {
model.arch = ml.get_arch();
if (model.arch == LLM_ARCH_UNKNOWN) {
@ -9256,6 +9267,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
throw std::runtime_error("error loading model vocabulary: " + std::string(e.what()));
}
llm_load_stats(ml, model);
llm_load_print_meta(ml, model);
if (model.vocab.type != LLAMA_VOCAB_TYPE_NONE &&
@ -18601,6 +18613,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
llama_model model;
llm_load_arch(ml, model);
llm_load_hparams(ml, model);
llm_load_stats(ml, model);
struct quantize_state_internal qs(model, params);
@ -19953,19 +19966,11 @@ int32_t llama_model_desc(const struct llama_model * model, char * buf, size_t bu
}
uint64_t llama_model_size(const struct llama_model * model) {
uint64_t size = 0;
for (const auto & it : model->tensors_by_name) {
size += ggml_nbytes(it.second);
}
return size;
return model->n_bytes;
}
uint64_t llama_model_n_params(const struct llama_model * model) {
uint64_t nparams = 0;
for (const auto & it : model->tensors_by_name) {
nparams += ggml_nelements(it.second);
}
return nparams;
return model->n_elements;
}
struct ggml_tensor * llama_get_model_tensor(struct llama_model * model, const char * name) {