root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-12 19:50:17 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

Merge 'origin/master' into steering

Browse Source
This commit is contained in:
Henri Vasserman 2023-05-19 23:19:57 +03:00
commit 95dc4d7270
No known key found for this signature in database
GPG Key ID: 2995FC0F58B1A986
18 changed files with 333 additions and 153 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Makefile
View File

@ -115,7 +115,7 @@ ifndef LLAMA_NO_ACCELERATE
endif
endif
ifdef LLAMA_OPENBLAS
CFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
CFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
ifneq ($(shell grep -e "Arch Linux" -e "ID_LIKE=arch" /etc/os-release 2>/dev/null),)
LDFLAGS += -lopenblas -lcblas
else

22
README.md
View File

@ -9,6 +9,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
**Hot topics:**
- Quantization formats `Q4` and `Q8` have changed again (19 May) - [(info)](https://github.com/ggerganov/llama.cpp/pull/1508)
- Quantization formats `Q4` and `Q5` have changed - requantize any old models [(info)](https://github.com/ggerganov/llama.cpp/pull/1405)
- [Roadmap May 2023](https://github.com/ggerganov/llama.cpp/discussions/1220)
@ -80,6 +81,7 @@ as the main playground for developing new features for the [ggml](https://github
- [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
- [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy)
- [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b)
- [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
**Bindings:**
@ -333,16 +335,16 @@ Several quantization methods are supported. They differ in the resulting model d
| Model | Measure | F16 | Q4_0 | Q4_1 | Q5_0 | Q5_1 | Q8_0 |
|------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
| 7B | perplexity | 5.9066 | 6.1565 | 6.0910 | 5.9862 | 5.9481 | 5.9069 |
| 7B | file size | 13.0G | 4.0G | 4.8G | 4.4G | 4.8G | 7.1G |
| 7B | ms/tok @ 4th | 128 | 50 | 54 | 75 | 83 | 75 |
| 7B | ms/tok @ 8th | 123 | 44 | 52 | 53 | 58 | 72 |
| 7B | bits/weight | 16.0 | 5.0 | 6.0 | 5.5 | 6.0 | 9.0 |
| 13B | perplexity | 5.2543 | 5.3860 | 5.3607 | 5.2856 | 5.2706 | 5.2548 |
| 13B | file size | 25.0G | 7.6G | 9.1G | 8.4G | 9.1G | 14G |
| 13B | ms/tok @ 4th | 239 | 93 | 101 | 150 | 164 | 141 |
| 13B | ms/tok @ 8th | 240 | 81 | 96 | 96 | 104 | 136 |
| 13B | bits/weight | 16.0 | 5.0 | 6.0 | 5.5 | 6.0 | 9.0 |
| 7B | perplexity | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
| 7B | file size | 13.0G | 3.5G | 3.9G | 4.3G | 4.7G | 6.7G |
| 7B | ms/tok @ 4th | 127 | 55 | 54 | 76 | 83 | 72 |
| 7B | ms/tok @ 8th | 122 | 43 | 45 | 52 | 56 | 67 |
| 7B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 |
| 13B | perplexity | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 |
| 13B | file size | 25.0G | 6.8G | 7.6G | 8.3G | 9.1G | 13G |
| 13B | ms/tok @ 4th | - | 103 | 105 | 148 | 160 | 131 |
| 13B | ms/tok @ 8th | - | 73 | 82 | 98 | 105 | 128 |
| 13B | bits/weight | 16.0 | 4.5 | 5.0 | 5.5 | 6.0 | 8.5 |
### Perplexity (measuring model quality)

3
convert.py
View File

@ -121,7 +121,6 @@ def make_tensors_list() -> List[str]:
f'layers.{i}.feed_forward.w1.weight',
f'layers.{i}.feed_forward.w2.weight',
f'layers.{i}.feed_forward.w3.weight',
f'layers.{i}.atttention_norm.weight',
f'layers.{i}.ffn_norm.weight',
]
return ret
@ -1055,7 +1054,7 @@ def load_some_model(path: Path) -> ModelPlus:
files = list(path.glob("model-00001-of-*.safetensors"))
if not files:
# Try the PyTorch patterns too, with lower priority
globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt"]
globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin" ]
files = [file for glob in globs for file in path.glob(glob)]
if not files:
# Try GGML too, but with lower priority, since if both a non-GGML

5
examples/benchmark/benchmark-matmult.cpp
View File

@ -211,6 +211,7 @@ int main(int argc, char ** argv) {
printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
printf("=====================================================================================\n");
double gflops_sum = 0;
for (int i=0;i<benchmark_params.n_iterations ;i++) {
long long int start = ggml_time_us();
@ -219,6 +220,7 @@ int main(int argc, char ** argv) {
long long int stop = ggml_time_us();
long long int usec = stop-start;
double gflops = (double)(flops_per_matrix)/usec/1000.0;
gflops_sum += gflops;
printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%10.2f\n",
i,
gf31.n_threads,
@ -248,4 +250,7 @@ int main(int argc, char ** argv) {
// Running a different graph computation to make sure we override the CPU cache lines
ggml_graph_compute(ctx, &gf32);
}
printf("\n");
printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));
printf("=====================================================================================\n");
}

151
examples/chat-persistent.sh Executable file
View File

@ -0,0 +1,151 @@
#!/bin/bash
set -euo pipefail
cd "$(dirname "$0")/.." || exit
if [[ -z "${PROMPT_CACHE_FILE+x}" || -z "${CHAT_SAVE_DIR+x}" ]]; then
echo >&2 "error: PROMPT_CACHE_FILE and CHAT_SAVE_DIR must be provided"
exit 1
fi
MODEL="${MODEL:-./models/13B/ggml-model-q4_0.bin}"
PROMPT_TEMPLATE="${PROMPT_TEMPLATE:-./prompts/chat.txt}"
USER_NAME="${USER_NAME:-User}"
AI_NAME="${AI_NAME:-ChatLLaMa}"
DATE_TIME="$(date +%H:%M)"
DATE_YEAR="$(date +%Y)"
LOG="${CHAT_SAVE_DIR}/main.log"
LOG_BG="${CHAT_SAVE_DIR}/main-bg.log"
CUR_PROMPT_FILE="${CHAT_SAVE_DIR}/current-prompt.txt"
CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
SESSION_SIZE_MSG_PATTERN='main: session file matches \d+ / \d+'
SAMPLE_TIME_MSG_PATTERN='sample time =\s+\d+.\d+ ms /\s+\d+'
SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
CTX_SIZE=2048
CTX_ROTATE_POINT=$((CTX_SIZE * 3 / 5)) # REVIEW
OPTS=(--model "$MODEL" --ctx_size "$CTX_SIZE" --repeat_last_n 256 "$@")
# An unbuffered `tail -c+N`
skip_bytes() {
LANG=C IFS= read -r -n "$1" -d '' c
while LANG=C IFS= read -r -n 1 -d '' c; do
printf '%s' "$c"
done
}
mkdir -p "$CHAT_SAVE_DIR"
echo >"$LOG"
trap "tail -n100 ${LOG}" EXIT
if [[ ! -e "$CUR_PROMPT_FILE" ]]; then
sed -e "s/\[\[USER_NAME\]\]/${USER_NAME}/g" \
-e "s/\[\[AI_NAME\]\]/${AI_NAME}/g" \
-e "s/\[\[DATE_TIME\]\]/${DATE_TIME}/g" \
-e "s/\[\[DATE_YEAR\]\]/${DATE_YEAR}/g" \
"$PROMPT_TEMPLATE" >"$CUR_PROMPT_FILE"
fi
if [[ ! -e "$NEXT_PROMPT_FILE" ]]; then
sed -r "$SED_DELETE_MESSAGES" "$CUR_PROMPT_FILE" >"$NEXT_PROMPT_FILE"
fi
if [[ "$(tail -c4 "$NEXT_PROMPT_FILE")" != "..." ]]; then
echo '...' >>"$NEXT_PROMPT_FILE"
fi
if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then
echo 'Prompt cache does not exist, building...'
# Default batch_size to 8 here for better user feedback during initial prompt processing
./main 2>>"$LOG" \
--batch_size 8 \
"${OPTS[@]}" \
--prompt-cache "$PROMPT_CACHE_FILE" \
--file "$CUR_PROMPT_FILE" \
--n_predict 1
echo
echo 'Done!'
fi
if [[ ! -e "$CUR_PROMPT_CACHE" ]]; then
cp "$PROMPT_CACHE_FILE" "$CUR_PROMPT_CACHE"
fi
if [[ ! -e "$NEXT_PROMPT_CACHE" ]]; then
cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE"
fi
printf '%s ' "$(< "$CUR_PROMPT_FILE")"
n_tokens=0
while read -e line; do
# Limit generation to remaining context, with a buffer and estimating 2 chars/token for input
n_predict=$((CTX_SIZE - n_tokens - ${#line} / 2 - 32))
# Swap prompts when we're about to run out of context
if ((n_predict <= 0)); then
wait # for background main (below) to finish with next prompt
mv "$NEXT_PROMPT_FILE" "$CUR_PROMPT_FILE"
mv "$NEXT_PROMPT_CACHE" "$CUR_PROMPT_CACHE"
sed -r "$SED_DELETE_MESSAGES" "$CUR_PROMPT_FILE" >"$NEXT_PROMPT_FILE"
echo '...' >>"$NEXT_PROMPT_FILE"
cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE"
n_tokens=0
n_predict=$((CTX_SIZE / 2))
fi
echo " ${line}" >>"$CUR_PROMPT_FILE"
if ((n_tokens > CTX_ROTATE_POINT)); then
echo " ${line}" >>"$NEXT_PROMPT_FILE"
fi
n_prompt_len_pre=$(($(wc -c <"$CUR_PROMPT_FILE")))
printf '%s: ' "$AI_NAME" >>"$CUR_PROMPT_FILE"
./main 2>>"$LOG" "${OPTS[@]}" \
--prompt-cache "$CUR_PROMPT_CACHE" \
--prompt-cache-all \
--file "$CUR_PROMPT_FILE" \
--reverse-prompt "${USER_NAME}:" \
--n_predict "$n_predict" |
skip_bytes 1 | # skip BOS token added by ./main
tee "$CUR_PROMPT_FILE.tmp" | # save prompt + generation to tmp file
skip_bytes "$n_prompt_len_pre" # print generation
mv "$CUR_PROMPT_FILE.tmp" "$CUR_PROMPT_FILE"
# if we hit n_predict instead of reverse-prompt, we need to add the prompt
if [[ "$(tail -n1 "$CUR_PROMPT_FILE")" != "${USER_NAME}:" ]]; then
printf '\n%s:' "$USER_NAME"
printf '\n%s:' "$USER_NAME" >> "$CUR_PROMPT_FILE"
fi
printf ' '
# HACK get num tokens from debug message
# TODO get both messages in one go
if ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
! sample_time_msg="$( tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
echo >&2 "Couldn't get number of tokens from ./main output!"
exit 1
fi
n_tokens=$(($(cut -d/ -f2 <<<"$session_size_msg") + $(cut -d/ -f2 <<<"$sample_time_msg")))
if ((n_tokens > CTX_ROTATE_POINT)); then
tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE"
fi
# Update cache for next prompt in background, ideally during user input
./main >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
--prompt-cache "$NEXT_PROMPT_CACHE" \
--file "$NEXT_PROMPT_FILE" \
--n_predict 1 &
done

18
examples/common.cpp
View File

@ -321,12 +321,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
invalid_param = true;
break;
}
} else if (arg == "--n-parts") {
if (++i >= argc) {
invalid_param = true;
break;
}
params.n_parts = std::stoi(argv[i]);
} else if (arg == "-h" || arg == "--help") {
gpt_print_usage(argc, argv, default_params);
exit(0);
@ -387,7 +381,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
}
if (params.prompt_cache_all &&
(params.interactive || params.interactive_first ||
params.instruct || params.antiprompt.size())) {
params.instruct)) {
fprintf(stderr, "error: --prompt-cache-all not supported in interactive mode yet\n");
gpt_print_usage(argc, argv, default_params);
exit(1);
@ -409,8 +403,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stderr, " -ins, --instruct run in instruction mode (use with Alpaca models)\n");
fprintf(stderr, " --multiline-input allows you to write or paste multiple lines without ending each in '\\'\n");
fprintf(stderr, " -r PROMPT, --reverse-prompt PROMPT\n");
fprintf(stderr, " run in interactive mode and poll user input upon seeing PROMPT (can be\n");
fprintf(stderr, " specified more than once for multiple prompts).\n");
fprintf(stderr, " halt generation at PROMPT, return control in interactive mode\n");
fprintf(stderr, " (can be specified more than once for multiple prompts).\n");
fprintf(stderr, " --color colorise output to distinguish prompt and user input from generations\n");
fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1, use random seed for < 0)\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
@ -448,7 +442,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stderr, " --no-penalize-nl do not penalize newline token\n");
fprintf(stderr, " --memory-f32 use f32 instead of f16 for memory key+value\n");
fprintf(stderr, " --temp N temperature (default: %.1f)\n", (double)params.temp);
fprintf(stderr, " --n-parts N number of model parts (default: -1 = determine from dimensions)\n");
fprintf(stderr, " -b N, --batch-size N batch size for prompt processing (default: %d)\n", params.n_batch);
fprintf(stderr, " --perplexity compute perplexity over the prompt\n");
fprintf(stderr, " --keep number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
@ -508,7 +501,6 @@ struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
auto lparams = llama_context_default_params();
lparams.n_ctx = params.n_ctx;
lparams.n_parts = params.n_parts;
lparams.n_gpu_layers = params.n_gpu_layers;
lparams.seed = params.seed;
lparams.f16_kv = params.memory_f16;
@ -792,7 +784,7 @@ bool console_readline(console_state & con_st, std::string & line) {
break;
}
if (input_char == WEOF || input_char == 0x04 /* Ctrl+D*/) {
if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
end_of_stream = true;
break;
}
@ -807,7 +799,7 @@ bool console_readline(console_state & con_st, std::string & line) {
char32_t code = getchar32();
if (code == '[' || code == 0x1B) {
// Discard the rest of the escape sequence
while ((code = getchar32()) != WEOF) {
while ((code = getchar32()) != (char32_t) WEOF) {
if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
break;
}

7
examples/common.h
View File

@ -24,7 +24,6 @@ struct gpt_params {
int32_t seed = -1; // RNG seed
int32_t n_threads = get_num_physical_cores();
int32_t n_predict = -1; // new tokens to predict
int32_t n_parts = -1; // amount of model parts (-1 = determine from model dimensions)
int32_t n_ctx = 512; // context size
int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
@ -45,15 +44,15 @@ struct gpt_params {
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
std::string model = "models/lamma-7B/ggml-model.bin"; // model path
std::string prompt = "";
std::string model = "models/7B/ggml-model.bin"; // model path
std::string prompt = "";
std::string path_prompt_cache = ""; // path to file for saving/loading prompt eval state
std::string input_prefix = ""; // string to prefix user inputs with
std::string input_suffix = ""; // string to suffix user inputs with
std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
std::string lora_adapter = ""; // lora adapter path
std::string lora_base = ""; // base model path for the lora adapter
std::string lora_base = ""; // base model path for the lora adapter
bool memory_f16 = true; // use f16 instead of f32 for memory kv
bool random_prompt = false; // do not randomize prompt if none provided

1
examples/embedding/embedding.cpp
View File

@ -6,7 +6,6 @@
int main(int argc, char ** argv) {
gpt_params params;
params.model = "models/llama-7B/ggml-model.bin";
if (gpt_params_parse(argc, argv, params) == false) {
return 1;

29
examples/main/main.cpp
View File

@ -50,7 +50,6 @@ void sigint_handler(int signo) {
int main(int argc, char ** argv) {
gpt_params params;
params.model = "models/llama-7B/ggml-model.bin";
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
@ -239,8 +238,8 @@ int main(int argc, char ** argv) {
params.antiprompt.push_back("### Instruction:\n\n");
}
// enable interactive mode if reverse prompt or interactive start is specified
if (params.antiprompt.size() != 0 || params.interactive_first) {
// enable interactive mode if interactive start is specified
if (params.interactive_first) {
params.interactive = true;
}
@ -272,7 +271,7 @@ int main(int argc, char ** argv) {
sigint_action.sa_flags = 0;
sigaction(SIGINT, &sigint_action, NULL);
#elif defined (_WIN32)
auto console_ctrl_handler = [](DWORD ctrl_type) -> BOOL {
auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
};
SetConsoleCtrlHandler(static_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
@ -336,7 +335,7 @@ int main(int argc, char ** argv) {
std::vector<llama_token> embd;
while (n_remain != 0 || params.interactive) {
while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
// predict
if (embd.size() > 0) {
// infinite text generation via context swapping
@ -536,9 +535,8 @@ int main(int argc, char ** argv) {
console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
}
// in interactive mode, and not currently processing queued inputs;
// check if we should prompt the user for more
if (params.interactive && (int) embd_inp.size() <= n_consumed) {
// if not currently processing queued inputs;
if ((int) embd_inp.size() <= n_consumed) {
// check for reverse prompt
if (params.antiprompt.size()) {
@ -549,10 +547,21 @@ int main(int argc, char ** argv) {
is_antiprompt = false;
// Check if each of the reverse prompts appears at the end of the output.
// If we're not running interactively, the reverse prompt might be tokenized with some following characters
// so we'll compensate for that by widening the search window a bit.
for (std::string & antiprompt : params.antiprompt) {
if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
is_interacting = true;
size_t extra_padding = params.interactive ? 0 : 2;
size_t search_start_pos = last_output.length() > static_cast<size_t>(antiprompt.length() + extra_padding)
? last_output.length() - static_cast<size_t>(antiprompt.length() + extra_padding)
: 0;
if (last_output.find(antiprompt.c_str(), search_start_pos) != std::string::npos) {
if (params.interactive) {
is_interacting = true;
console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
}
is_antiprompt = true;
fflush(stdout);
break;
}
}

1
examples/perplexity/perplexity.cpp
View File

@ -116,7 +116,6 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
int main(int argc, char ** argv) {
gpt_params params;
params.model = "models/llama-7B/ggml-model.bin";
params.n_batch = 512;
if (gpt_params_parse(argc, argv, params) == false) {

1
examples/quantize-stats/quantize-stats.cpp
View File

@ -321,7 +321,6 @@ int main(int argc, char ** argv) {
auto lparams = llama_context_default_params();
lparams.n_ctx = 256;
lparams.n_parts = 1;
lparams.seed = 1;
lparams.f16_kv = false;
lparams.use_mlock = false;

2
examples/save-load-state/save-load-state.cpp
View File

@ -8,7 +8,6 @@
int main(int argc, char ** argv) {
gpt_params params;
params.model = "models/llama-7B/ggml-model.bin";
params.seed = 42;
params.n_threads = 4;
params.repeat_last_n = 64;
@ -27,7 +26,6 @@ int main(int argc, char ** argv) {
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;

14
ggml-cuda.cu
View File

@ -42,19 +42,19 @@ typedef void (*dequantize_mul_mat_vec_cuda_t)(const void * vx, const float * y,
#define QK4_0 32
#define QR4_0 2
typedef struct {
float d; // delta
half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
#define QR4_1 2
typedef struct {
float d; // delta
float m; // min
half d; // delta
half m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK5_0 32
#define QR5_0 2
@ -78,10 +78,10 @@ static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) +
#define QK8_0 32
#define QR8_0 1
typedef struct {
float d; // delta
half d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
#define CUDA_DMMV_BLOCK_SIZE 32 // dmmv = dequantize_mul_mat_vec

193
ggml.c
View File

@ -543,12 +543,7 @@ static inline __m256 sum_i16_pairs_float(const __m256i x) {
return _mm256_cvtepi32_ps(summed_pairs);
}
// multiply int8_t, add results pairwise twice and return as float vector
static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
// Get absolute values of x vectors
const __m256i ax = _mm256_sign_epi8(x, x);
// Sign the values of the y vectors
const __m256i sy = _mm256_sign_epi8(y, x);
static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
#if __AVXVNNI__
const __m256i zero = _mm256_setzero_si256();
const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
@ -560,6 +555,21 @@ static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
#endif
}
// multiply int8_t, add results pairwise twice and return as float vector
static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
#if __AVXVNNIINT8__
const __m256i zero = _mm256_setzero_si256();
const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y);
return _mm256_cvtepi32_ps(summed_pairs);
#else
// Get absolute values of x vectors
const __m256i ax = _mm256_sign_epi8(x, x);
// Sign the values of the y vectors
const __m256i sy = _mm256_sign_epi8(y, x);
return mul_sum_us8_pairs_float(ax, sy);
#endif
}
static inline __m128i packNibbles( __m256i bytes )
{
// Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
@ -619,6 +629,17 @@ static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) {
return _mm256_cvtepi32_ps(summed_pairs);
}
static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
const __m128i axl = _mm256_castsi256_si128(ax);
const __m128i axh = _mm256_extractf128_si256(ax, 1);
const __m128i syl = _mm256_castsi256_si128(sy);
const __m128i syh = _mm256_extractf128_si256(sy, 1);
// Perform multiplication and create 16-bit values
const __m128i dotl = _mm_maddubs_epi16(axl, syl);
const __m128i doth = _mm_maddubs_epi16(axh, syh);
return sum_i16_pairs_float(doth, dotl);
}
// multiply int8_t, add results pairwise twice and return as float vector
static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
const __m128i xl = _mm256_castsi256_si128(x);
@ -748,18 +769,18 @@ int32x4_t vcvtnq_s32_f32(float32x4_t v) {
#define QK4_0 32
typedef struct {
float d; // delta
ggml_fp16_t d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
typedef struct {
float d; // delta
float m; // min
ggml_fp16_t d; // delta
ggml_fp16_t m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == 2 * sizeof(float) + QK4_1 / 2, "wrong q4_1 block size/padding");
static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK5_0 32
typedef struct {
@ -780,16 +801,16 @@ static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) +
#define QK8_0 32
typedef struct {
float d; // delta
int8_t qs[QK8_0]; // quants
ggml_fp16_t d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
#define QK8_1 32
typedef struct {
float d; // delta
float s; // d * sum(qs[i])
int8_t qs[QK8_1]; // quants
float d; // delta
float s; // d * sum(qs[i])
int8_t qs[QK8_1]; // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
@ -816,7 +837,7 @@ static void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * r
const float d = max / -8;
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
for (int j = 0; j < qk/2; ++j) {
const float x0 = x[i*qk + 0 + j]*id;
@ -856,8 +877,8 @@ static void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * r
const float d = (max - min) / ((1 << 4) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].m = min;
y[i].d = GGML_FP32_TO_FP16(d);
y[i].m = GGML_FP32_TO_FP16(min);
for (int j = 0; j < qk/2; ++j) {
const float x0 = (x[i*qk + 0 + j] - min)*id;
@ -988,7 +1009,7 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
for (int j = 0; j < QK8_0; ++j) {
const float x0 = x[i*QK8_0 + j]*id;
@ -1023,7 +1044,7 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
for (int j = 0; j < 8; j++) {
const float32x4_t v = vmulq_n_f32(srcv[j], id);
@ -1058,7 +1079,7 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
// Quantize these floats
const float d = maxScalar / 127.f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
const __m256 mul = _mm256_set1_ps( id );
@ -1157,7 +1178,7 @@ static void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * r
sum += y[i].qs[QK8_1/2 + j];
}
y[i].s = d * sum;
y[i].s = sum*d;
}
}
@ -1309,7 +1330,7 @@ static void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
const float d = x[i].d;
const float d = GGML_FP16_TO_FP32(x[i].d);
for (int j = 0; j < qk/2; ++j) {
const int x0 = (x[i].qs[j] & 0x0F) - 8;
@ -1329,8 +1350,8 @@ static void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict
const int nb = k / qk;
for (int i = 0; i < nb; i++) {
const float d = x[i].d;
const float m = x[i].m;
const float d = GGML_FP16_TO_FP32(x[i].d);
const float m = GGML_FP16_TO_FP32(x[i].m);
for (int j = 0; j < qk/2; ++j) {
const int x0 = (x[i].qs[j] & 0x0F);
@ -1405,7 +1426,7 @@ static void dequantize_row_q8_0(const void * restrict vx, float * restrict y, in
const block_q8_0 * restrict x = vx;
for (int i = 0; i < nb; i++) {
const float d = x[i].d;
const float d = GGML_FP16_TO_FP32(x[i].d);
for (int j = 0; j < qk; ++j) {
y[i*qk + j] = x[i].qs[j]*d;
@ -1669,8 +1690,9 @@ quantize_fns_t ggml_internal_get_quantize_fn(size_t i) {
static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
float tmp[8];
for (int i = 0; i < 8; i++)
for (int i = 0; i < 8; i++) {
tmp[i] = GGML_FP16_TO_FP32(x[i]);
}
return _mm256_loadu_ps(tmp);
}
@ -2090,8 +2112,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
const block_q8_0 * restrict y0 = &y[i + 0];
const block_q8_0 * restrict y1 = &y[i + 1];
const uint8x16_t m4b = vdupq_n_u8(0x0F);
const int8x16_t s8b = vdupq_n_s8(0x8);
const uint8x16_t m4b = vdupq_n_u8(0x0F);
const int8x16_t s8b = vdupq_n_s8(0x8);
const uint8x16_t v0_0 = vld1q_u8(x0->qs);
const uint8x16_t v0_1 = vld1q_u8(x1->qs);
@ -2119,8 +2141,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, 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);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0l));
@ -2137,8 +2159,8 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#endif
}
@ -2150,7 +2172,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
// Main loop
for (int i = 0; i < nb; ++i) {
/* Compute combined scale for the block */
const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );
const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
__m256i bx = bytes_from_nibbles_32(x[i].qs);
@ -2174,7 +2196,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
// Main loop
for (int i = 0; i < nb; ++i) {
// Compute combined scale for the block
const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );
const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
const __m128i lowMask = _mm_set1_epi8(0xF);
const __m128i off = _mm_set1_epi8(8);
@ -2216,7 +2238,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
_mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 0 and 1
const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[0].d ), _mm_set1_ps( y[0].d ) );
const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[0].d) * GGML_FP16_TO_FP32(y[0].d) );
const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs);
@ -2234,7 +2256,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
_mm_prefetch(&y[1] + sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 2 and 3
const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[1].d ), _mm_set1_ps( y[1].d ) );
const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[1].d) * GGML_FP16_TO_FP32(y[1].d) );
const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs);
@ -2267,7 +2289,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
_mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 0 and 1
const __m128 d_0_1 = _mm_mul_ps( _mm_set1_ps( x[i].d ), _mm_set1_ps( y[i].d ) );
const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs);
@ -2285,7 +2307,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
_mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
// Compute combined scale for the block 2 and 3
const __m128 d_2_3 = _mm_mul_ps( _mm_set1_ps( x[i + 1].d ), _mm_set1_ps( y[i + 1].d ) );
const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i + 1].d) * GGML_FP16_TO_FP32(y[i + 1].d) );
const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs);
@ -2333,7 +2355,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
}
sumf += (x[i].d*y[i].d)*sumi;
sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d);
}
*s = sumf;
@ -2363,7 +2385,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
const block_q8_1 * restrict y0 = &y[i + 0];
const block_q8_1 * restrict y1 = &y[i + 1];
summs += x0->m * y0->s + x1->m * y1->s;
summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s;
const uint8x16_t m4b = vdupq_n_u8(0x0F);
@ -2387,8 +2409,8 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, 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);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d);
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0l));
@ -2405,8 +2427,8 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0->d*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1->d*y1->d);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d);
#endif
}
@ -2419,13 +2441,13 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
// Main loop
for (int i = 0; i < nb; ++i) {
const float * d0 = &x[i].d;
const float * d1 = &y[i].d;
const float d0 = GGML_FP16_TO_FP32(x[i].d);
const float d1 = y[i].d;
summs += x[i].m * y[i].s;
summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
const __m256 d0v = _mm256_broadcast_ss( d0 );
const __m256 d1v = _mm256_broadcast_ss( d1 );
const __m256 d0v = _mm256_set1_ps( d0 );
const __m256 d1v = _mm256_set1_ps( d1 );
// Compute combined scales
const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
@ -2434,7 +2456,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
const __m256i bx = bytes_from_nibbles_32(x[i].qs);
const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs );
const __m256 xy = mul_sum_i8_pairs_float(bx, by);
const __m256 xy = mul_sum_us8_pairs_float(bx, by);
// Accumulate d0*d1*x*y
#if defined(__AVX2__)
@ -2459,7 +2481,7 @@ static void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void *
sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
}
sumf += (x[i].d*y[i].d)*sumi + x[i].m*y[i].s;
sumf += (GGML_FP16_TO_FP32(x[i]).d*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
}
*s = sumf;
@ -2535,16 +2557,13 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
const float x0d = GGML_FP16_TO_FP32(x0->d);
const float x1d = GGML_FP16_TO_FP32(x1->d);
#if defined(__ARM_FEATURE_DOTPROD)
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d);
vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d);
vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l));
@ -2561,8 +2580,8 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#endif
}
@ -2637,7 +2656,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
// Main loop
for (int i = 0; i < nb; i++) {
/* Compute combined scale for the block */
const __m256 d = _mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)), _mm256_broadcast_ss(&y[i].d));
const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
__m256i bx = bytes_from_nibbles_32(x[i].qs);
__m256i bxhi = bytes_from_bits_32(x[i].qh);
@ -2661,7 +2680,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
// Main loop
for (int i = 0; i < nb; i++) {
/* Compute combined scale for the block */
const __m256 d = _mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d)), _mm256_broadcast_ss(&y[i].d));
const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
__m256i bx = bytes_from_nibbles_32(x[i].qs);
const __m256i bxhi = bytes_from_bits_32(x[i].qh);
@ -2704,7 +2723,7 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
}
sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi;
sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi;
}
*s = sumf;
@ -2786,16 +2805,13 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
const int8x16_t v1_1l = vld1q_s8(y1->qs);
const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
const float x0d = GGML_FP16_TO_FP32(x0->d);
const float x1d = GGML_FP16_TO_FP32(x1->d);
#if defined(__ARM_FEATURE_DOTPROD)
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), x0d*y0->d);
vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), x1d*y1->d);
vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d);
#else
const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lf), vget_low_s8 (v1_0l));
const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lf), vget_high_s8(v1_0l));
@ -2812,8 +2828,8 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), x0d*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), x1d*y1->d);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(pl0, ph0)), GGML_FP16_TO_FP32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(pl1, ph1)), GGML_FP16_TO_FP32(x1->d)*y1->d);
#endif
}
@ -2873,15 +2889,14 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
const float x0d = GGML_FP16_TO_FP32(x0->d);
// dot product
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
wasm_i32x4_add(
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
sumv = wasm_f32x4_add(sumv,
wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add(
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d));
}
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@ -2903,10 +2918,10 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
bx = _mm256_or_si256(bx, bxhi);
const __m256 dy = _mm256_broadcast_ss(&y[i].d);
const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_i8_pairs_float(bx, by);
const __m256 q = mul_sum_us8_pairs_float(bx, by);
acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
}
@ -2937,10 +2952,10 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
bxh = _mm_or_si128(bxh, bxhih);
bx = _mm256_set_m128i(bxh, bxl);
const __m256 dy = _mm256_broadcast_ss(&y[i].d);
const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_i8_pairs_float(bx, by);
const __m256 q = mul_sum_us8_pairs_float(bx, by);
acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
}
@ -3007,11 +3022,11 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
#if defined(__ARM_FEATURE_DOTPROD)
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), x0->d*y0->d);
vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), x1->d*y1->d);
vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#else
const int16x8_t p0_0 = vmull_s8(vget_low_s8 (x0_0), vget_low_s8 (y0_0));
@ -3029,8 +3044,8 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
const int32x4_t p2 = vaddq_s32(vpaddlq_s16(p1_0), vpaddlq_s16(p1_1));
const int32x4_t p3 = vaddq_s32(vpaddlq_s16(p1_2), vpaddlq_s16(p1_3));
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), x0->d*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), x1->d*y1->d);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(p0, p1)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(p2, p3)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
#endif
}
@ -3042,7 +3057,7 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
// Main loop
for (int i = 0; i < nb; ++i) {
// Compute combined scale for the block
const __m256 d = _mm256_mul_ps( _mm256_broadcast_ss( &x[i].d ), _mm256_broadcast_ss( &y[i].d ) );
const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
__m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs);
__m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
@ -3068,7 +3083,7 @@ static void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void *
sumi += x[i].qs[j]*y[i].qs[j];
}
sumf += (x[i].d*y[i].d)*sumi;
sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d));
}
*s = sumf;

2
ggml.h
View File

@ -190,7 +190,7 @@
#define GGML_FILE_MAGIC 0x67676d6c // "ggml"
#define GGML_FILE_VERSION 1
#define GGML_QNT_VERSION 1 // bump this on quantization format changes
#define GGML_QNT_VERSION 2 // bump this on quantization format changes
#define GGML_QNT_VERSION_FACTOR 1000 // do not change this
#define GGML_MAX_DIMS 4

23
llama.cpp
View File

@ -434,6 +434,7 @@ enum llama_file_version {
LLAMA_FILE_VERSION_GGMF_V1, // added version field and scores in vocab
LLAMA_FILE_VERSION_GGJT_V1, // added padding
LLAMA_FILE_VERSION_GGJT_V2, // changed quantization format
LLAMA_FILE_VERSION_GGJT_V3, // changed Q4 and Q8 quantization format
};
struct llama_file_loader {
@ -466,6 +467,8 @@ struct llama_file_loader {
file_version = LLAMA_FILE_VERSION_GGJT_V1;
} else if (magic == 'ggjt' && version == 2) {
file_version = LLAMA_FILE_VERSION_GGJT_V2;
} else if (magic == 'ggjt' && version == 3) {
file_version = LLAMA_FILE_VERSION_GGJT_V3;
} else {
throw format("unknown (magic, version) combination: %08x, %08x; is this really a GGML file?",
magic, version);
@ -840,10 +843,9 @@ static bool kv_cache_init(
struct llama_context_params llama_context_default_params() {
struct llama_context_params result = {
/*.n_ctx =*/ 512,
/*.n_parts =*/ -1,
/*.gpu_layers =*/ 0,
/*.seed =*/ -1,
/*.f16_kv =*/ false,
/*.f16_kv =*/ true,
/*.logits_all =*/ false,
/*.vocab_only =*/ false,
/*.use_mmap =*/ true,
@ -873,7 +875,8 @@ static const char *llama_file_version_name(llama_file_version version) {
case LLAMA_FILE_VERSION_GGML: return "'ggml' (old version with low tokenizer quality and no mmap support)";
case LLAMA_FILE_VERSION_GGMF_V1: return "ggmf v1 (old version with no mmap support)";
case LLAMA_FILE_VERSION_GGJT_V1: return "ggjt v1 (pre #1405)";
case LLAMA_FILE_VERSION_GGJT_V2: return "ggjt v2 (latest)";
case LLAMA_FILE_VERSION_GGJT_V2: return "ggjt v2 (pre #1508)";
case LLAMA_FILE_VERSION_GGJT_V3: return "ggjt v3 (latest)";
}
return "unknown";
@ -953,11 +956,19 @@ static void llama_model_load_internal(
fprintf(stderr, "%s: model size = %s\n", __func__, llama_model_type_name(model.type));
}
if (file_version != LLAMA_FILE_VERSION_GGJT_V2) {
if (file_version < LLAMA_FILE_VERSION_GGJT_V2) {
if (hparams.ftype != LLAMA_FTYPE_ALL_F32 &&
hparams.ftype != LLAMA_FTYPE_MOSTLY_F16 &&
hparams.ftype != LLAMA_FTYPE_MOSTLY_Q8_0) {
throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1305)");
throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1405)");
}
}
if (file_version < LLAMA_FILE_VERSION_GGJT_V3) {
if (hparams.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ||
hparams.ftype == LLAMA_FTYPE_MOSTLY_Q4_1 ||
hparams.ftype == LLAMA_FTYPE_MOSTLY_Q8_0) {
throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1508)");
}
}
@ -970,7 +981,7 @@ static void llama_model_load_internal(
size_t ctx_size;
size_t mmapped_size;
ml->calc_sizes(&ctx_size, &mmapped_size);
fprintf(stderr, "%s: ggml ctx size = %6.2f KB\n", __func__, ctx_size/1024.0);
fprintf(stderr, "%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/1024.0/1024.0);
// print memory requirements
{

3
llama.h
View File

@ -19,7 +19,7 @@
# define LLAMA_API
#endif
#define LLAMA_FILE_VERSION 2
#define LLAMA_FILE_VERSION 3
#define LLAMA_FILE_MAGIC 'ggjt'
#define LLAMA_FILE_MAGIC_UNVERSIONED 'ggml'
#define LLAMA_SESSION_MAGIC 'ggsn'
@ -55,7 +55,6 @@ extern "C" {
struct llama_context_params {
int n_ctx; // text context
int n_parts; // -1 for default
int n_gpu_layers; // number of layers to store in VRAM
int seed; // RNG seed, -1 for random

9
tests/test-sampling.cpp
View File

@ -1,6 +1,10 @@
#include "llama.h"
#include "ggml.h"
#include <cassert>
#include "llama.h"
#ifdef NDEBUG
#undef NDEBUG
#endif
#include <cmath>
#include <numeric>
#include <cassert>
@ -8,7 +12,6 @@
#include <vector>
#include <algorithm>
void dump(const llama_token_data_array * candidates) {
for (size_t i = 0; i < candidates->size; i++) {
printf("%d: %f (%f)\n", candidates->data[i].id, candidates->data[i].p, candidates->data[i].logit);