mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2025-09-11 01:24:36 +00:00
Merge branch 'upstream' into concedo_experimental
# Conflicts: # .devops/full-cuda.Dockerfile # .devops/full.Dockerfile # .devops/main-cuda.Dockerfile # .devops/main-rocm.Dockerfile # .devops/main-vulkan.Dockerfile # .devops/main.Dockerfile # .devops/server-cuda.Dockerfile # .devops/server.Dockerfile # README.md # common/CMakeLists.txt # grammars/README.md # tests/test-grammar-integration.cpp # tests/test-grammar-parser.cpp # tests/test-json-schema-to-grammar.cpp
This commit is contained in:
Concedo
commit
562d980140
25 changed files with 881 additions and 676 deletions
|
|
@ -201,19 +201,13 @@ void gpt_params_handle_model_default(gpt_params & params) {
|
|||
}
|
||||
params.hf_file = params.model;
|
||||
} else if (params.model.empty()) {
|
||||
std::string cache_directory = fs_get_cache_directory();
|
||||
const bool success = fs_create_directory_with_parents(cache_directory);
|
||||
if (!success) {
|
||||
throw std::runtime_error("failed to create cache directory: " + cache_directory);
|
||||
}
|
||||
params.model = cache_directory + string_split(params.hf_file, '/').back();
|
||||
params.model = fs_get_cache_file(string_split(params.hf_file, '/').back());
|
||||
}
|
||||
} else if (!params.model_url.empty()) {
|
||||
if (params.model.empty()) {
|
||||
auto f = string_split(params.model_url, '#').front();
|
||||
f = string_split(f, '?').front();
|
||||
f = string_split(f, '/').back();
|
||||
params.model = "models/" + f;
|
||||
params.model = fs_get_cache_file(string_split(f, '/').back());
|
||||
}
|
||||
} else if (params.model.empty()) {
|
||||
params.model = DEFAULT_MODEL_PATH;
|
||||
|
|
@ -274,6 +268,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
|
|||
}
|
||||
} catch (const std::invalid_argument & ex) {
|
||||
fprintf(stderr, "%s\n", ex.what());
|
||||
params = params_org;
|
||||
return false;
|
||||
}
|
||||
|
||||
|
|
@ -409,6 +404,20 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
|
|||
}
|
||||
return true;
|
||||
}
|
||||
if (arg == "--in-file") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
return true;
|
||||
}
|
||||
std::ifstream file(argv[i]);
|
||||
if (!file) {
|
||||
fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
|
||||
invalid_param = true;
|
||||
return true;
|
||||
}
|
||||
params.in_files.push_back(argv[i]);
|
||||
return true;
|
||||
}
|
||||
if (arg == "-n" || arg == "--predict" || arg == "--n-predict") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
|
|
@ -1082,7 +1091,15 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
|
|||
return true;
|
||||
}
|
||||
if (arg == "-v" || arg == "--verbose") {
|
||||
params.verbose = true;
|
||||
params.verbosity = 1;
|
||||
return true;
|
||||
}
|
||||
if (arg == "--verbosity") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
return true;
|
||||
}
|
||||
params.verbosity = std::stoi(argv[i]);
|
||||
return true;
|
||||
}
|
||||
if (arg == "--verbose-prompt") {
|
||||
|
|
@ -1392,6 +1409,14 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
|
|||
params.timeout_write = std::stoi(argv[i]);
|
||||
return true;
|
||||
}
|
||||
if (arg == "--threads-http") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
return true;
|
||||
}
|
||||
params.n_threads_http = std::stoi(argv[i]);
|
||||
return true;
|
||||
}
|
||||
if (arg == "-spf" || arg == "--system-prompt-file") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
|
|
@ -1461,6 +1486,14 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
|
|||
params.chat_template = argv[i];
|
||||
return true;
|
||||
}
|
||||
if (arg == "--slot-prompt-similarity" || arg == "-sps") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
return true;
|
||||
}
|
||||
params.slot_prompt_similarity = std::stof(argv[i]);
|
||||
return true;
|
||||
}
|
||||
if (arg == "-pps") {
|
||||
params.is_pp_shared = true;
|
||||
return true;
|
||||
|
|
@ -1538,6 +1571,46 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
|
|||
params.i_pos = std::stoi(argv[i]);
|
||||
return true;
|
||||
}
|
||||
if (arg == "-o" || arg == "--output" || arg == "--output-file") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
return true;
|
||||
}
|
||||
params.out_file = argv[i];
|
||||
return true;
|
||||
}
|
||||
if (arg == "-ofreq" || arg == "--output-frequency") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
return true;
|
||||
}
|
||||
params.n_out_freq = std::stoi(argv[i]);
|
||||
return true;
|
||||
}
|
||||
if (arg == "--save-frequency") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
return true;
|
||||
}
|
||||
params.n_save_freq = std::stoi(argv[i]);
|
||||
return true;
|
||||
}
|
||||
if (arg == "--process-output") {
|
||||
params.process_output = true;
|
||||
return true;
|
||||
}
|
||||
if (arg == "--no-ppl") {
|
||||
params.compute_ppl = false;
|
||||
return true;
|
||||
}
|
||||
if (arg == "--chunk" || arg == "--from-chunk") {
|
||||
if (++i >= argc) {
|
||||
invalid_param = true;
|
||||
return true;
|
||||
}
|
||||
params.i_chunk = std::stoi(argv[i]);
|
||||
return true;
|
||||
}
|
||||
#ifndef LOG_DISABLE_LOGS
|
||||
// Parse args for logging parameters
|
||||
if (log_param_single_parse(argv[i])) {
|
||||
|
|
@ -1613,6 +1686,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
|
|||
options.push_back({ "*", "-h, --help, --usage", "print usage and exit" });
|
||||
options.push_back({ "*", " --version", "show version and build info" });
|
||||
options.push_back({ "*", "-v, --verbose", "print verbose information" });
|
||||
options.push_back({ "*", " --verbosity N", "set specific verbosity level (default: %d)", params.verbosity });
|
||||
options.push_back({ "*", " --verbose-prompt", "print a verbose prompt before generation (default: %s)", params.verbose_prompt ? "true" : "false" });
|
||||
options.push_back({ "*", " --no-display-prompt", "don't print prompt at generation (default: %s)", !params.display_prompt ? "true" : "false" });
|
||||
options.push_back({ "*", "-co, --color", "colorise output to distinguish prompt and user input from generations (default: %s)", params.use_color ? "true" : "false" });
|
||||
|
|
@ -1638,6 +1712,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
|
|||
options.push_back({ "*", "-fa, --flash-attn", "enable Flash Attention (default: %s)", params.flash_attn ? "enabled" : "disabled" });
|
||||
options.push_back({ "*", "-p, --prompt PROMPT", "prompt to start generation with (default: '%s')", params.prompt.c_str() });
|
||||
options.push_back({ "*", "-f, --file FNAME", "a file containing the prompt (default: none)" });
|
||||
options.push_back({ "*", " --in-file FNAME", "an input file (repeat to specify multiple files)" });
|
||||
options.push_back({ "*", "-bf, --binary-file FNAME", "binary file containing the prompt (default: none)" });
|
||||
options.push_back({ "*", "-e, --escape", "process escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\) (default: %s)", params.escape ? "true" : "false" });
|
||||
options.push_back({ "*", " --no-escape", "do not process escape sequences" });
|
||||
|
|
@ -1805,6 +1880,14 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
|
|||
options.push_back({ "passkey", " --junk N", "number of times to repeat the junk text (default: %d)", params.n_junk });
|
||||
options.push_back({ "passkey", " --pos N", "position of the passkey in the junk text (default: %d)", params.i_pos });
|
||||
|
||||
options.push_back({ "imatrix" });
|
||||
options.push_back({ "imatrix", "-o, --output FNAME", "output file (default: '%s')", params.out_file.c_str() });
|
||||
options.push_back({ "imatrix", " --output-frequency N", "output the imatrix every N iterations (default: %d)", params.n_out_freq });
|
||||
options.push_back({ "imatrix", " --save-frequency N", "save an imatrix copy every N iterations (default: %d)", params.n_save_freq });
|
||||
options.push_back({ "imatrix", " --process-output", "collect data for the output tensor (default: %s)", params.process_output ? "true" : "false" });
|
||||
options.push_back({ "imatrix", " --no-ppl", "do not compute perplexity (default: %s)", params.compute_ppl ? "true" : "false" });
|
||||
options.push_back({ "imatrix", " --chunk N", "start processing the input from chunk N (default: %d)", params.i_chunk });
|
||||
|
||||
options.push_back({ "bench" });
|
||||
options.push_back({ "bench", "-pps", "is the prompt shared across parallel sequences (default: %s)", params.is_pp_shared ? "true" : "false" });
|
||||
options.push_back({ "bench", "-npp n0,n1,...", "number of prompt tokens" });
|
||||
|
|
@ -1821,6 +1904,7 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
|
|||
options.push_back({ "server", " --ssl-key-file FNAME", "path to file a PEM-encoded SSL private key" });
|
||||
options.push_back({ "server", " --ssl-cert-file FNAME", "path to file a PEM-encoded SSL certificate" });
|
||||
options.push_back({ "server", " --timeout N", "server read/write timeout in seconds (default: %d)", params.timeout_read });
|
||||
options.push_back({ "server", " --threads-http N", "number of threads used to process HTTP requests (default: %d)", params.n_threads_http });
|
||||
options.push_back({ "server", " --system-prompt-file FNAME",
|
||||
"set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications" });
|
||||
options.push_back({ "server", " --log-format {text,json}",
|
||||
|
|
@ -1832,6 +1916,8 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param
|
|||
"set custom jinja chat template (default: template taken from model's metadata)\n"
|
||||
"only commonly used templates are accepted:\n"
|
||||
"https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template" });
|
||||
options.push_back({ "server", "-sps, --slot-prompt-similarity SIMILARITY",
|
||||
"how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity });
|
||||
|
||||
#ifndef LOG_DISABLE_LOGS
|
||||
options.push_back({ "logging" });
|
||||
|
|
@ -2188,6 +2274,16 @@ std::string fs_get_cache_directory() {
|
|||
return ensure_trailing_slash(cache_directory);
|
||||
}
|
||||
|
||||
std::string fs_get_cache_file(const std::string & filename) {
|
||||
GGML_ASSERT(filename.find(DIRECTORY_SEPARATOR) == std::string::npos);
|
||||
std::string cache_directory = fs_get_cache_directory();
|
||||
const bool success = fs_create_directory_with_parents(cache_directory);
|
||||
if (!success) {
|
||||
throw std::runtime_error("failed to create cache directory: " + cache_directory);
|
||||
}
|
||||
return cache_directory + filename;
|
||||
}
|
||||
|
||||
|
||||
//
|
||||
// Model utils
|
||||
|
|
|
|||
|
|
@ -67,7 +67,6 @@ struct gpt_params {
|
|||
float p_split = 0.1f; // speculative decoding split probability
|
||||
int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default)
|
||||
int32_t n_gpu_layers_draft = -1; // number of layers to store in VRAM for the draft model (-1 - use default)
|
||||
llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
|
||||
int32_t main_gpu = 0; // the GPU that is used for scratch and small tensors
|
||||
float tensor_split[128] = {0}; // how split tensors should be distributed across GPUs
|
||||
int32_t n_beams = 0; // if non-zero then use beam search of given width.
|
||||
|
|
@ -82,13 +81,13 @@ struct gpt_params {
|
|||
float yarn_beta_slow = 1.0f; // YaRN high correction dim
|
||||
int32_t yarn_orig_ctx = 0; // YaRN original context length
|
||||
float defrag_thold = -1.0f; // KV cache defragmentation threshold
|
||||
std::string rpc_servers = ""; // comma separated list of RPC servers
|
||||
|
||||
ggml_backend_sched_eval_callback cb_eval = nullptr;
|
||||
void * cb_eval_user_data = nullptr;
|
||||
|
||||
ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
|
||||
|
||||
enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
|
||||
enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
|
||||
enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
|
||||
|
||||
|
|
@ -131,7 +130,9 @@ struct gpt_params {
|
|||
std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding
|
||||
std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
|
||||
std::string logits_file = ""; // file for saving *all* logits
|
||||
std::string rpc_servers = ""; // comma separated list of RPC servers
|
||||
|
||||
std::vector<std::string> in_files; // all input files
|
||||
std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
|
||||
std::vector<llama_model_kv_override> kv_overrides;
|
||||
|
||||
|
|
@ -141,6 +142,7 @@ struct gpt_params {
|
|||
|
||||
std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
|
||||
|
||||
int32_t verbosity = 0;
|
||||
int32_t control_vector_layer_start = -1; // layer range for control vector
|
||||
int32_t control_vector_layer_end = -1; // layer range for control vector
|
||||
|
||||
|
|
@ -180,7 +182,6 @@ struct gpt_params {
|
|||
bool logits_all = false; // return logits for all tokens in the batch
|
||||
bool use_mmap = true; // use mmap for faster loads
|
||||
bool use_mlock = false; // use mlock to keep model in memory
|
||||
bool verbose = false;
|
||||
bool verbose_prompt = false; // print prompt tokens before generation
|
||||
bool display_prompt = true; // print prompt before generation
|
||||
bool infill = false; // use infill mode
|
||||
|
|
@ -197,10 +198,10 @@ struct gpt_params {
|
|||
std::vector<std::string> image; // path to image file(s)
|
||||
|
||||
// server params
|
||||
int32_t port = 8080;
|
||||
int32_t timeout_read = 600;
|
||||
int32_t timeout_write = timeout_read;
|
||||
int32_t n_threads_http = -1;
|
||||
int32_t port = 8080; // server listens on this network port
|
||||
int32_t timeout_read = 600; // http read timeout in seconds
|
||||
int32_t timeout_write = timeout_read; // http write timeout in seconds
|
||||
int32_t n_threads_http = -1; // number of threads to process HTTP requests
|
||||
|
||||
std::string hostname = "127.0.0.1";
|
||||
std::string public_path = "";
|
||||
|
|
@ -219,6 +220,8 @@ struct gpt_params {
|
|||
|
||||
std::string slot_save_path;
|
||||
|
||||
float slot_prompt_similarity = 0.5f;
|
||||
|
||||
// batched-bench params
|
||||
bool is_pp_shared = false;
|
||||
|
||||
|
|
@ -236,6 +239,16 @@ struct gpt_params {
|
|||
// passkey params
|
||||
int32_t n_junk = 250; // number of times to repeat the junk text
|
||||
int32_t i_pos = -1; // position of the passkey in the junk text
|
||||
|
||||
// imatrix params
|
||||
std::string out_file = "imatrix.dat"; // save the resulting imatrix to this file
|
||||
|
||||
int32_t n_out_freq = 10; // output the imatrix every n_out_freq iterations
|
||||
int32_t n_save_freq = 0; // save the imatrix every n_save_freq iterations
|
||||
int32_t i_chunk = 0; // start processing from this chunk
|
||||
|
||||
bool process_output = false; // collect data for the output tensor
|
||||
bool compute_ppl = true; // whether to compute perplexity
|
||||
};
|
||||
|
||||
void gpt_params_handle_model_default(gpt_params & params);
|
||||
|
|
@ -281,6 +294,7 @@ bool fs_validate_filename(const std::string & filename);
|
|||
bool fs_create_directory_with_parents(const std::string & path);
|
||||
|
||||
std::string fs_get_cache_directory();
|
||||
std::string fs_get_cache_file(const std::string & filename);
|
||||
|
||||
//
|
||||
// Model utils
|
||||
|
|
|
|||
|
|
@ -46,8 +46,12 @@ namespace grammar_parser {
|
|||
state.rules[rule_id] = rule;
|
||||
}
|
||||
|
||||
static bool is_digit_char(char c) {
|
||||
return '0' <= c && c <= '9';
|
||||
}
|
||||
|
||||
static bool is_word_char(char c) {
|
||||
return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
|
||||
return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || is_digit_char(c);
|
||||
}
|
||||
|
||||
static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
|
||||
|
|
@ -99,6 +103,17 @@ namespace grammar_parser {
|
|||
return pos;
|
||||
}
|
||||
|
||||
static const char * parse_int(const char * src) {
|
||||
const char * pos = src;
|
||||
while (is_digit_char(*pos)) {
|
||||
pos++;
|
||||
}
|
||||
if (pos == src) {
|
||||
throw std::runtime_error(std::string("expecting integer at ") + src);
|
||||
}
|
||||
return pos;
|
||||
}
|
||||
|
||||
static std::pair<uint32_t, const char *> parse_char(const char * src) {
|
||||
if (*src == '\\') {
|
||||
switch (src[1]) {
|
||||
|
|
@ -137,6 +152,60 @@ namespace grammar_parser {
|
|||
bool is_nested) {
|
||||
size_t last_sym_start = out_elements.size();
|
||||
const char * pos = src;
|
||||
|
||||
auto handle_repetitions = [&](int min_times, int max_times) {
|
||||
|
||||
if (last_sym_start == out_elements.size()) {
|
||||
throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
|
||||
}
|
||||
|
||||
// apply transformation to previous symbol (last_sym_start to end) according to
|
||||
// the following rewrite rules:
|
||||
// S{m,n} --> S S S (m times) S'(n-m)
|
||||
// S'(x) ::= S S'(x-1) |
|
||||
// (... n-m definitions of these S' rules ...)
|
||||
// S'(1) ::= S |
|
||||
// S{m,} --> S S S (m times) S'
|
||||
// S' ::= S S' |
|
||||
// S* --> S{0,}
|
||||
// --> S' ::= S S' |
|
||||
// S+ --> S{1,}
|
||||
// --> S S'
|
||||
// S' ::= S S' |
|
||||
// S? --> S{0,1}
|
||||
// --> S'
|
||||
// S' ::= S |
|
||||
|
||||
std::vector<llama_grammar_element> previous_elements(out_elements.begin() + last_sym_start, out_elements.end());
|
||||
if (min_times == 0) {
|
||||
out_elements.resize(last_sym_start);
|
||||
} else {
|
||||
// Repeat the previous elements (min_times - 1) times
|
||||
for (int i = 1; i < min_times; i++) {
|
||||
out_elements.insert(out_elements.end(), previous_elements.begin(), previous_elements.end());
|
||||
}
|
||||
}
|
||||
|
||||
uint32_t last_rec_rule_id = 0;
|
||||
auto n_opt = max_times < 0 ? 1 : max_times - min_times;
|
||||
|
||||
std::vector<llama_grammar_element> rec_rule(previous_elements);
|
||||
for (int i = 0; i < n_opt; i++) {
|
||||
rec_rule.resize(previous_elements.size());
|
||||
uint32_t rec_rule_id = generate_symbol_id(state, rule_name);
|
||||
if (i > 0 || max_times < 0) {
|
||||
rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
|
||||
}
|
||||
rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
|
||||
rec_rule.push_back({LLAMA_GRETYPE_END, 0});
|
||||
add_rule(state, rec_rule_id, rec_rule);
|
||||
last_rec_rule_id = rec_rule_id;
|
||||
}
|
||||
if (n_opt > 0) {
|
||||
out_elements.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id});
|
||||
}
|
||||
};
|
||||
|
||||
while (*pos) {
|
||||
if (*pos == '"') { // literal string
|
||||
pos++;
|
||||
|
|
@ -197,40 +266,51 @@ namespace grammar_parser {
|
|||
throw std::runtime_error(std::string("expecting ')' at ") + pos);
|
||||
}
|
||||
pos = parse_space(pos + 1, is_nested);
|
||||
} else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
|
||||
if (last_sym_start == out_elements.size()) {
|
||||
throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
|
||||
}
|
||||
|
||||
// apply transformation to previous symbol (last_sym_start to end) according to
|
||||
// rewrite rules:
|
||||
// S* --> S' ::= S S' |
|
||||
// S+ --> S' ::= S S' | S
|
||||
// S? --> S' ::= S |
|
||||
uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
|
||||
std::vector<llama_grammar_element> sub_rule;
|
||||
// add preceding symbol to generated rule
|
||||
sub_rule.insert(
|
||||
sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
|
||||
if (*pos == '*' || *pos == '+') {
|
||||
// cause generated rule to recurse
|
||||
sub_rule.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
|
||||
}
|
||||
// mark start of alternate def
|
||||
sub_rule.push_back({LLAMA_GRETYPE_ALT, 0});
|
||||
if (*pos == '+') {
|
||||
// add preceding symbol as alternate only for '+' (otherwise empty)
|
||||
sub_rule.insert(
|
||||
sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
|
||||
}
|
||||
sub_rule.push_back({LLAMA_GRETYPE_END, 0});
|
||||
add_rule(state, sub_rule_id, sub_rule);
|
||||
|
||||
// in original rule, replace previous symbol with reference to generated rule
|
||||
out_elements.resize(last_sym_start);
|
||||
out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
|
||||
|
||||
} else if (*pos == '.') { // any char
|
||||
last_sym_start = out_elements.size();
|
||||
out_elements.push_back({LLAMA_GRETYPE_CHAR_ANY, 0});
|
||||
pos = parse_space(pos + 1, is_nested);
|
||||
} else if (*pos == '*') {
|
||||
pos = parse_space(pos + 1, is_nested);
|
||||
handle_repetitions(0, -1);
|
||||
} else if (*pos == '+') {
|
||||
pos = parse_space(pos + 1, is_nested);
|
||||
handle_repetitions(1, -1);
|
||||
} else if (*pos == '?') {
|
||||
pos = parse_space(pos + 1, is_nested);
|
||||
handle_repetitions(0, 1);
|
||||
} else if (*pos == '{') {
|
||||
pos = parse_space(pos + 1, is_nested);
|
||||
|
||||
if (!is_digit_char(*pos)) {
|
||||
throw std::runtime_error(std::string("expecting an int at ") + pos);
|
||||
}
|
||||
const char * int_end = parse_int(pos);
|
||||
int min_times = std::stoul(std::string(pos, int_end - pos));
|
||||
pos = parse_space(int_end, is_nested);
|
||||
|
||||
int max_times = -1;
|
||||
|
||||
if (*pos == '}') {
|
||||
max_times = min_times;
|
||||
pos = parse_space(pos + 1, is_nested);
|
||||
} else if (*pos == ',') {
|
||||
pos = parse_space(pos + 1, is_nested);
|
||||
|
||||
if (is_digit_char(*pos)) {
|
||||
const char * int_end = parse_int(pos);
|
||||
max_times = std::stoul(std::string(pos, int_end - pos));
|
||||
pos = parse_space(int_end, is_nested);
|
||||
}
|
||||
|
||||
if (*pos != '}') {
|
||||
throw std::runtime_error(std::string("expecting '}' at ") + pos);
|
||||
}
|
||||
pos = parse_space(pos + 1, is_nested);
|
||||
} else {
|
||||
throw std::runtime_error(std::string("expecting ',' at ") + pos);
|
||||
}
|
||||
handle_repetitions(min_times, max_times);
|
||||
} else {
|
||||
break;
|
||||
}
|
||||
|
|
@ -325,6 +405,7 @@ namespace grammar_parser {
|
|||
case LLAMA_GRETYPE_CHAR_NOT: return true;
|
||||
case LLAMA_GRETYPE_CHAR_ALT: return true;
|
||||
case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true;
|
||||
case LLAMA_GRETYPE_CHAR_ANY: return true;
|
||||
default: return false;
|
||||
}
|
||||
}
|
||||
|
|
@ -339,6 +420,7 @@ namespace grammar_parser {
|
|||
case LLAMA_GRETYPE_CHAR_NOT: fprintf(file, "CHAR_NOT"); break;
|
||||
case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
|
||||
case LLAMA_GRETYPE_CHAR_ALT: fprintf(file, "CHAR_ALT"); break;
|
||||
case LLAMA_GRETYPE_CHAR_ANY: fprintf(file, "CHAR_ANY"); break;
|
||||
}
|
||||
switch (elem.type) {
|
||||
case LLAMA_GRETYPE_END:
|
||||
|
|
@ -350,6 +432,7 @@ namespace grammar_parser {
|
|||
case LLAMA_GRETYPE_CHAR_NOT:
|
||||
case LLAMA_GRETYPE_CHAR_RNG_UPPER:
|
||||
case LLAMA_GRETYPE_CHAR_ALT:
|
||||
case LLAMA_GRETYPE_CHAR_ANY:
|
||||
fprintf(file, "(\"");
|
||||
print_grammar_char(file, elem.value);
|
||||
fprintf(file, "\") ");
|
||||
|
|
@ -407,11 +490,15 @@ namespace grammar_parser {
|
|||
}
|
||||
print_grammar_char(file, elem.value);
|
||||
break;
|
||||
case LLAMA_GRETYPE_CHAR_ANY:
|
||||
fprintf(file, ".");
|
||||
break;
|
||||
}
|
||||
if (is_char_element(elem)) {
|
||||
switch (rule[i + 1].type) {
|
||||
case LLAMA_GRETYPE_CHAR_ALT:
|
||||
case LLAMA_GRETYPE_CHAR_RNG_UPPER:
|
||||
case LLAMA_GRETYPE_CHAR_ANY:
|
||||
break;
|
||||
default:
|
||||
fprintf(file, "] ");
|
||||
|
|
|
|||
|
|
@ -16,58 +16,27 @@ static std::string join(Iterator begin, Iterator end, const std::string & separa
|
|||
|
||||
static std::string repeat(const std::string & str, size_t n);
|
||||
|
||||
static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "", bool item_rule_is_literal = false) {
|
||||
if (separator_rule.empty()) {
|
||||
static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
|
||||
auto has_max = max_items != std::numeric_limits<int>::max();
|
||||
|
||||
if (min_items == 0 && max_items == 1) {
|
||||
return item_rule + "?";
|
||||
} else if (min_items == 1 && max_items == std::numeric_limits<int>::max()) {
|
||||
}
|
||||
|
||||
if (separator_rule.empty()) {
|
||||
if (min_items == 1 && !has_max) {
|
||||
return item_rule + "+";
|
||||
}
|
||||
}
|
||||
|
||||
std::string result;
|
||||
if (min_items > 0) {
|
||||
if (item_rule_is_literal && separator_rule.empty()) {
|
||||
result = "\"" + repeat(std::string(item_rule.begin() + 1, item_rule.end() - 1), min_items) + "\"";
|
||||
} else if (min_items == 0 && !has_max) {
|
||||
return item_rule + "*";
|
||||
} else {
|
||||
std::vector<std::string> items(min_items, item_rule);
|
||||
result = join(items.begin(), items.end(), separator_rule.empty() ? " " : " " + separator_rule + " ");
|
||||
return item_rule + "{" + std::to_string(min_items) + "," + (has_max ? std::to_string(max_items) : "") + "}";
|
||||
}
|
||||
}
|
||||
|
||||
std::function<std::string(int, bool)> opt_repetitions = [&](int up_to_n, bool prefix_with_sep) -> std::string {
|
||||
auto content = prefix_with_sep && !separator_rule.empty() ? separator_rule + " " + item_rule : item_rule;
|
||||
|
||||
if (up_to_n == 0) {
|
||||
return "";
|
||||
} else if (up_to_n == 1) {
|
||||
return "(" + content + ")?";
|
||||
} else if (!separator_rule.empty() && !prefix_with_sep) {
|
||||
return "(" + content + " " + opt_repetitions(up_to_n - 1, true) + ")?";
|
||||
} else {
|
||||
std::string res = repeat("(" + content + " ", up_to_n);
|
||||
// strip trailing space
|
||||
res = res.substr(0, res.length() - 1);
|
||||
res += repeat(")?", up_to_n);
|
||||
return res;
|
||||
auto result = item_rule + " " + build_repetition("(" + separator_rule + " " + item_rule + ")", min_items == 0 ? 0 : min_items - 1, has_max ? max_items - 1 : max_items);
|
||||
if (min_items == 0) {
|
||||
result = "(" + result + ")?";
|
||||
}
|
||||
};
|
||||
|
||||
if (min_items > 0 && max_items != min_items) {
|
||||
result += " ";
|
||||
}
|
||||
|
||||
if (max_items != std::numeric_limits<int>::max()) {
|
||||
result += opt_repetitions(max_items - min_items, min_items > 0);
|
||||
} else {
|
||||
std::string item_operator = "(" + (separator_rule.empty() ? "" : separator_rule + " ") + item_rule + ")";
|
||||
if (min_items == 0 && !separator_rule.empty()) {
|
||||
result = "(" + item_rule + " " + item_operator + "*)?";
|
||||
} else {
|
||||
result += item_operator + "*";
|
||||
}
|
||||
}
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
|
|
@ -78,30 +47,24 @@ struct BuiltinRule {
|
|||
std::vector<std::string> deps;
|
||||
};
|
||||
|
||||
const std::string _up_to_15_digits = build_repetition("[0-9]", 0, 15);
|
||||
|
||||
std::unordered_map<std::string, BuiltinRule> PRIMITIVE_RULES = {
|
||||
{"boolean", {"(\"true\" | \"false\") space", {}}},
|
||||
{"decimal-part", {"[0-9] " + _up_to_15_digits, {}}},
|
||||
{"integral-part", {"[0-9] | [1-9] " + _up_to_15_digits, {}}},
|
||||
{"decimal-part", {"[0-9]{1,16}", {}}},
|
||||
{"integral-part", {"[0] | [1-9] [0-9]{0,15}", {}}},
|
||||
{"number", {"(\"-\"? integral-part) (\".\" decimal-part)? ([eE] [-+]? integral-part)? space", {"integral-part", "decimal-part"}}},
|
||||
{"integer", {"(\"-\"? integral-part) space", {"integral-part"}}},
|
||||
{"value", {"object | array | string | number | boolean | null", {"object", "array", "string", "number", "boolean", "null"}}},
|
||||
{"object", {"\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space", {"string", "value"}}},
|
||||
{"array", {"\"[\" space ( value (\",\" space value)* )? \"]\" space", {"value"}}},
|
||||
{"uuid", {"\"\\\"\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
|
||||
"\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
|
||||
"\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
|
||||
"\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
|
||||
"\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] \"\\\"\" space", {}}},
|
||||
{"char", {"[^\"\\\\] | \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])", {}}},
|
||||
{"uuid", {"\"\\\"\" [0-9a-fA-F]{8} \"-\" [0-9a-fA-F]{4} \"-\" [0-9a-fA-F]{4} \"-\" [0-9a-fA-F]{4} \"-\" [0-9a-fA-F]{12} \"\\\"\" space", {}}},
|
||||
{"char", {"[^\"\\\\] | \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F]{4})", {}}},
|
||||
{"string", {"\"\\\"\" char* \"\\\"\" space", {"char"}}},
|
||||
{"null", {"\"null\" space", {}}},
|
||||
};
|
||||
|
||||
std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
|
||||
{"date", {"[0-9] [0-9] [0-9] [0-9] \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )", {}}},
|
||||
{"time", {"([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9] [0-9] [0-9] )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )", {}}},
|
||||
{"date", {"[0-9]{4} \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )", {}}},
|
||||
{"time", {"([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9]{3} )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )", {}}},
|
||||
{"date-time", {"date \"T\" time", {"date", "time"}}},
|
||||
{"date-string", {"\"\\\"\" date \"\\\"\" space", {"date"}}},
|
||||
{"time-string", {"\"\\\"\" time \"\\\"\" space", {"time"}}},
|
||||
|
|
@ -385,8 +348,7 @@ private:
|
|||
sub_is_literal ? "\"" + sub + "\"" : sub,
|
||||
min_times,
|
||||
max_times,
|
||||
"",
|
||||
sub_is_literal
|
||||
""
|
||||
);
|
||||
seq.back().second = false;
|
||||
} else {
|
||||
|
|
|
|||
|
|
@ -83,6 +83,7 @@ models = [
|
|||
{"name": "jina-v2-es", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
|
||||
{"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
|
||||
{"name": "smaug-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct", },
|
||||
{"name": "jina-v2-code", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-code", },
|
||||
]
|
||||
|
||||
|
||||
|
|
|
|||
|
|
@ -47,11 +47,12 @@ class Model:
|
|||
_model_classes: dict[str, type[Model]] = {}
|
||||
|
||||
dir_model: Path
|
||||
ftype: int
|
||||
ftype: gguf.LlamaFileType
|
||||
is_big_endian: bool
|
||||
endianess: gguf.GGUFEndian
|
||||
use_temp_file: bool
|
||||
lazy: bool
|
||||
model_name: str | None
|
||||
part_names: list[str]
|
||||
is_safetensors: bool
|
||||
hparams: dict[str, Any]
|
||||
|
|
@ -64,7 +65,7 @@ class Model:
|
|||
# subclasses should define this!
|
||||
model_arch: gguf.MODEL_ARCH
|
||||
|
||||
def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool):
|
||||
def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool, use_temp_file: bool, eager: bool, model_name: str | None):
|
||||
if type(self) is Model:
|
||||
raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
|
||||
self.dir_model = dir_model
|
||||
|
|
@ -73,10 +74,11 @@ class Model:
|
|||
self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
|
||||
self.use_temp_file = use_temp_file
|
||||
self.lazy = not eager
|
||||
self.part_names = Model.get_model_part_names(self.dir_model, ".safetensors")
|
||||
self.model_name = model_name
|
||||
self.part_names = Model.get_model_part_names(self.dir_model, "model", ".safetensors")
|
||||
self.is_safetensors = len(self.part_names) > 0
|
||||
if not self.is_safetensors:
|
||||
self.part_names = Model.get_model_part_names(self.dir_model, ".bin")
|
||||
self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
|
||||
self.hparams = Model.load_hparams(self.dir_model)
|
||||
self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
|
||||
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
|
||||
|
|
@ -94,7 +96,7 @@ class Model:
|
|||
ftype_lw: str = ftype_up.lower()
|
||||
# allow templating the file name with the output ftype, useful with the "auto" ftype
|
||||
self.fname_out = fname_out.parent / fname_out.name.format(ftype_lw, outtype=ftype_lw, ftype=ftype_lw, OUTTYPE=ftype_up, FTYPE=ftype_up)
|
||||
self.gguf_writer = gguf.GGUFWriter(self.fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
|
||||
self.gguf_writer = gguf.GGUFWriter(path=None, arch=gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
|
||||
|
||||
@classmethod
|
||||
def __init_subclass__(cls):
|
||||
|
|
@ -182,7 +184,7 @@ class Model:
|
|||
return new_name
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
self.gguf_writer.add_name(self.dir_model.name)
|
||||
self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name)
|
||||
self.gguf_writer.add_block_count(self.block_count)
|
||||
|
||||
if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
|
||||
|
|
@ -324,21 +326,21 @@ class Model:
|
|||
|
||||
def write(self):
|
||||
self.write_tensors()
|
||||
self.gguf_writer.write_header_to_file()
|
||||
self.gguf_writer.write_header_to_file(self.fname_out)
|
||||
self.gguf_writer.write_kv_data_to_file()
|
||||
self.gguf_writer.write_tensors_to_file(progress=True)
|
||||
self.gguf_writer.close()
|
||||
|
||||
def write_vocab(self):
|
||||
self.gguf_writer.write_header_to_file()
|
||||
self.gguf_writer.write_header_to_file(self.fname_out)
|
||||
self.gguf_writer.write_kv_data_to_file()
|
||||
self.gguf_writer.close()
|
||||
|
||||
@staticmethod
|
||||
def get_model_part_names(dir_model: Path, suffix: str) -> list[str]:
|
||||
def get_model_part_names(dir_model: Path, prefix: str, suffix: str) -> list[str]:
|
||||
part_names: list[str] = []
|
||||
for filename in os.listdir(dir_model):
|
||||
if filename.endswith(suffix):
|
||||
if filename.startswith(prefix) and filename.endswith(suffix):
|
||||
part_names.append(filename)
|
||||
|
||||
part_names.sort()
|
||||
|
|
@ -475,6 +477,9 @@ class Model:
|
|||
if chkhsh == "c136ed14d01c2745d4f60a9596ae66800e2b61fa45643e72436041855ad4089d":
|
||||
# ref: https://huggingface.co/abacusai/Smaug-Llama-3-70B-Instruct
|
||||
res = "smaug-bpe"
|
||||
if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a":
|
||||
# ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code
|
||||
res = "jina-v2-code"
|
||||
|
||||
if res is None:
|
||||
logger.warning("\n")
|
||||
|
|
@ -662,7 +667,7 @@ class GPTNeoXModel(Model):
|
|||
def set_gguf_parameters(self):
|
||||
block_count = self.hparams["num_hidden_layers"]
|
||||
|
||||
self.gguf_writer.add_name(self.dir_model.name)
|
||||
self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name)
|
||||
self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
|
||||
self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
|
||||
self.gguf_writer.add_block_count(block_count)
|
||||
|
|
@ -795,7 +800,7 @@ class MPTModel(Model):
|
|||
|
||||
def set_gguf_parameters(self):
|
||||
block_count = self.hparams["n_layers"]
|
||||
self.gguf_writer.add_name(self.dir_model.name)
|
||||
self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name)
|
||||
self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
|
||||
self.gguf_writer.add_embedding_length(self.hparams["d_model"])
|
||||
self.gguf_writer.add_block_count(block_count)
|
||||
|
|
@ -847,7 +852,7 @@ class OrionModel(Model):
|
|||
raise ValueError("gguf: can not find ctx length parameter.")
|
||||
|
||||
self.gguf_writer.add_file_type(self.ftype)
|
||||
self.gguf_writer.add_name(self.dir_model.name)
|
||||
self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name)
|
||||
self.gguf_writer.add_source_hf_repo(hf_repo)
|
||||
self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
|
||||
self.gguf_writer.add_context_length(ctx_length)
|
||||
|
|
@ -884,7 +889,7 @@ class BaichuanModel(Model):
|
|||
else:
|
||||
raise ValueError("gguf: can not find ctx length parameter.")
|
||||
|
||||
self.gguf_writer.add_name(self.dir_model.name)
|
||||
self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name)
|
||||
self.gguf_writer.add_source_hf_repo(hf_repo)
|
||||
self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
|
||||
self.gguf_writer.add_context_length(ctx_length)
|
||||
|
|
@ -1007,7 +1012,7 @@ class XverseModel(Model):
|
|||
else:
|
||||
raise ValueError("gguf: can not find ctx length parameter.")
|
||||
|
||||
self.gguf_writer.add_name(self.dir_model.name)
|
||||
self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name)
|
||||
self.gguf_writer.add_source_hf_repo(hf_repo)
|
||||
self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
|
||||
self.gguf_writer.add_context_length(ctx_length)
|
||||
|
|
@ -1203,7 +1208,7 @@ class StableLMModel(Model):
|
|||
hparams = self.hparams
|
||||
block_count = hparams["num_hidden_layers"]
|
||||
|
||||
self.gguf_writer.add_name(self.dir_model.name)
|
||||
self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name)
|
||||
self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
|
||||
self.gguf_writer.add_embedding_length(hparams["hidden_size"])
|
||||
self.gguf_writer.add_block_count(block_count)
|
||||
|
|
@ -1678,7 +1683,7 @@ class GPT2Model(Model):
|
|||
model_arch = gguf.MODEL_ARCH.GPT2
|
||||
|
||||
def set_gguf_parameters(self):
|
||||
self.gguf_writer.add_name(self.dir_model.name)
|
||||
self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name)
|
||||
self.gguf_writer.add_block_count(self.hparams["n_layer"])
|
||||
self.gguf_writer.add_context_length(self.hparams["n_ctx"])
|
||||
self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
|
||||
|
|
@ -2245,7 +2250,7 @@ class GemmaModel(Model):
|
|||
hparams = self.hparams
|
||||
block_count = hparams["num_hidden_layers"]
|
||||
|
||||
self.gguf_writer.add_name(self.dir_model.name)
|
||||
self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name)
|
||||
self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
|
||||
self.gguf_writer.add_embedding_length(hparams["hidden_size"])
|
||||
self.gguf_writer.add_block_count(block_count)
|
||||
|
|
@ -2345,7 +2350,7 @@ class MambaModel(Model):
|
|||
# Fail early for models which don't have a block expansion factor of 2
|
||||
assert d_inner == 2 * d_model
|
||||
|
||||
self.gguf_writer.add_name(self.dir_model.name)
|
||||
self.gguf_writer.add_name(self.dir_model.name if self.model_name is None else self.model_name)
|
||||
self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
|
||||
self.gguf_writer.add_embedding_length(d_model)
|
||||
self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
|
||||
|
|
@ -2452,11 +2457,13 @@ class JinaBertV2Model(BertModel):
|
|||
|
||||
def get_tensors(self):
|
||||
for name, data in super().get_tensors():
|
||||
if 'gated_layers' in name:
|
||||
if 'gated_layer' in name:
|
||||
d1 = data[:self.intermediate_size, :]
|
||||
name1 = name.replace('gated_layers', 'gated_layers_w')
|
||||
name1 = name1.replace('up_gated_layer', 'gated_layers_v')
|
||||
d2 = data[self.intermediate_size:, :]
|
||||
name2 = name.replace('gated_layers', 'gated_layers_v')
|
||||
name2 = name2.replace('up_gated_layer', 'gated_layers_w')
|
||||
yield name1, d1
|
||||
yield name2, d2
|
||||
continue
|
||||
|
|
@ -2847,7 +2854,7 @@ def main() -> None:
|
|||
logger.error(f"Model {hparams['architectures'][0]} is not supported")
|
||||
sys.exit(1)
|
||||
|
||||
model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file, args.no_lazy)
|
||||
model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file, args.no_lazy, args.model_name)
|
||||
|
||||
logger.info("Set model parameters")
|
||||
model_instance.set_gguf_parameters()
|
||||
|
|
|
|||
|
|
@ -62,10 +62,10 @@ static size_t split_str_to_n_bytes(std::string str) {
|
|||
int n;
|
||||
if (str.back() == 'M') {
|
||||
sscanf(str.c_str(), "%d", &n);
|
||||
n_bytes = (size_t)n * 1024 * 1024; // megabytes
|
||||
n_bytes = (size_t)n * 1000 * 1000; // megabytes
|
||||
} else if (str.back() == 'G') {
|
||||
sscanf(str.c_str(), "%d", &n);
|
||||
n_bytes = (size_t)n * 1024 * 1024 * 1024; // gigabytes
|
||||
n_bytes = (size_t)n * 1000 * 1000 * 1000; // gigabytes
|
||||
} else {
|
||||
throw std::invalid_argument("error: supported units are M (megabytes) or G (gigabytes), but got: " + std::string(1, str.back()));
|
||||
}
|
||||
|
|
@ -285,7 +285,7 @@ struct split_strategy {
|
|||
struct ggml_tensor * t = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_out, i));
|
||||
total_size += ggml_nbytes(t);
|
||||
}
|
||||
total_size = total_size / 1024 / 1024; // convert to megabytes
|
||||
total_size = total_size / 1000 / 1000; // convert to megabytes
|
||||
printf("split %05d: n_tensors = %d, total_size = %ldM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
|
||||
i_split++;
|
||||
}
|
||||
|
|
|
|||
|
|
@ -6,16 +6,19 @@ More information is available here: https://github.com/ggerganov/llama.cpp/pull/
|
|||
## Usage
|
||||
|
||||
```
|
||||
./imatrix -m <some_fp_model> -f <some_training_data> [-o <output_file>] [--verbosity <verbosity_level>]
|
||||
[-ofreq num_chunks] [-ow <0 or 1>] [other common params]
|
||||
./imatrix \
|
||||
-m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \
|
||||
[--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \
|
||||
[--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]
|
||||
```
|
||||
|
||||
Here `-m` with a model name and `-f` with a file containing training data (such as e.g. `wiki.train.raw`) are mandatory.
|
||||
The parameters in square brackets are optional and have the following meaning:
|
||||
* `-o` (or `--output-file`) specifies the name of the file where the computed data will be stored. If missing `imatrix.dat` is used.
|
||||
* `--verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
|
||||
* `-ofreq` (or `--output-frequency`) specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
|
||||
* `-ow` (or `--output-weight`) specifies if data will be collected for the `output.weight` tensor. My experience is that it is better to not utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
|
||||
* `--output-frequency` specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
|
||||
* `--save-frequency` specifies how often to save a copy of the imatrix in a separate file. Default is 0 (i.e., never)
|
||||
* `--process-output` specifies if data will be collected for the `output.weight` tensor. My experience is that it is better to not utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
|
||||
|
||||
For faster computation, make sure to use GPU offloading via the `-ngl` argument
|
||||
|
||||
|
|
|
|||
|
|
@ -18,39 +18,37 @@
|
|||
#pragma warning(disable: 4244 4267) // possible loss of data
|
||||
#endif
|
||||
|
||||
static void print_usage(int argc, char ** argv, const gpt_params & params) {
|
||||
gpt_params_print_usage(argc, argv, params);
|
||||
|
||||
LOG_TEE("\nexample usage:\n");
|
||||
LOG_TEE("\n %s \\\n"
|
||||
" -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \\\n"
|
||||
" [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n"
|
||||
" [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]\n" , argv[0]);
|
||||
LOG_TEE("\n");
|
||||
}
|
||||
|
||||
struct Stats {
|
||||
std::vector<float> values;
|
||||
std::vector<int> counts;
|
||||
int ncall = 0;
|
||||
};
|
||||
|
||||
struct StatParams {
|
||||
std::string dataset;
|
||||
std::string ofile = "imatrix.dat";
|
||||
int n_output_frequency = 10;
|
||||
int verbosity = 1;
|
||||
int keep_every = 0;
|
||||
bool collect_output_weight = false;
|
||||
};
|
||||
|
||||
class IMatrixCollector {
|
||||
public:
|
||||
IMatrixCollector() = default;
|
||||
void set_parameters(StatParams&& params) { m_params = std::move(params); }
|
||||
void set_params(gpt_params params) { m_params = std::move(params); }
|
||||
bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
|
||||
void save_imatrix() const;
|
||||
bool load_imatrix(const char * file_name, bool add);
|
||||
static bool load_imatrix(const char * file_name, std::unordered_map<std::string, Stats>& imatrix);
|
||||
void save_imatrix(int ncall = -1) const;
|
||||
bool load_imatrix(const char * file_name);
|
||||
private:
|
||||
std::unordered_map<std::string, Stats> m_stats;
|
||||
StatParams m_params;
|
||||
gpt_params m_params;
|
||||
std::mutex m_mutex;
|
||||
int m_last_call = 0;
|
||||
std::vector<float> m_src1_data;
|
||||
std::vector<char> m_ids; // the expert ids from ggml_mul_mat_id
|
||||
//
|
||||
void save_imatrix(const char * file_name, const char * dataset) const;
|
||||
void keep_imatrix(int ncall) const;
|
||||
};
|
||||
|
||||
// remove any prefix and suffixes from the name
|
||||
|
|
@ -86,7 +84,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
|
|||
if (t->op != GGML_OP_MUL_MAT) return false;
|
||||
// why are small batches ignored (<16 tokens)?
|
||||
if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
|
||||
if (!(wname.substr(0, 4) == "blk." || (m_params.collect_output_weight && wname == "output.weight"))) return false;
|
||||
if (!(wname.substr(0, 4) == "blk." || (m_params.process_output && wname == "output.weight"))) return false;
|
||||
return true;
|
||||
}
|
||||
|
||||
|
|
@ -154,16 +152,20 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
|
|||
for (int j = 0; j < (int)src1->ne[0]; ++j) {
|
||||
e.values[e_start + j] += x[j]*x[j];
|
||||
e.counts[e_start + j]++;
|
||||
if (!std::isfinite(e.values[e_start + j])) {
|
||||
fprintf(stderr, "%f detected in %s\n", e.values[e_start + j], wname.c_str());
|
||||
exit(1);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
if (e.ncall > m_last_call) {
|
||||
m_last_call = e.ncall;
|
||||
if (m_last_call % m_params.n_output_frequency == 0) {
|
||||
if (m_last_call % m_params.n_out_freq == 0) {
|
||||
save_imatrix();
|
||||
}
|
||||
if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) {
|
||||
keep_imatrix(m_last_call);
|
||||
if (m_params.n_save_freq > 0 && m_last_call%m_params.n_save_freq == 0) {
|
||||
save_imatrix(m_last_call);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -186,15 +188,19 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
|
|||
for (int j = 0; j < (int)src1->ne[0]; ++j) {
|
||||
e.values[j] += x[j]*x[j];
|
||||
e.counts[j]++;
|
||||
if (!std::isfinite(e.values[j])) {
|
||||
fprintf(stderr, "%f detected in %s\n", e.values[j], wname.c_str());
|
||||
exit(1);
|
||||
}
|
||||
}
|
||||
}
|
||||
if (e.ncall > m_last_call) {
|
||||
m_last_call = e.ncall;
|
||||
if (m_last_call % m_params.n_output_frequency == 0) {
|
||||
if (m_last_call % m_params.n_out_freq == 0) {
|
||||
save_imatrix();
|
||||
}
|
||||
if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) {
|
||||
keep_imatrix(m_last_call);
|
||||
if (m_params.n_save_freq > 0 && m_last_call%m_params.n_save_freq == 0) {
|
||||
save_imatrix(m_last_call);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
|
@ -202,19 +208,17 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
|
|||
return true;
|
||||
}
|
||||
|
||||
void IMatrixCollector::save_imatrix() const {
|
||||
save_imatrix(m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str(), m_params.dataset.c_str());
|
||||
void IMatrixCollector::save_imatrix(int ncall) const {
|
||||
auto fname = m_params.out_file;
|
||||
if (fname.empty()) {
|
||||
fname = "imatrix.dat";
|
||||
}
|
||||
|
||||
void IMatrixCollector::keep_imatrix(int ncall) const {
|
||||
auto file_name = m_params.ofile;
|
||||
if (file_name.empty()) file_name = "imatrix.dat";
|
||||
file_name += ".at_";
|
||||
file_name += std::to_string(ncall);
|
||||
save_imatrix(file_name.c_str(), m_params.dataset.c_str());
|
||||
if (ncall > 0) {
|
||||
fname += ".at_";
|
||||
fname += std::to_string(ncall);
|
||||
}
|
||||
|
||||
void IMatrixCollector::save_imatrix(const char * fname, const char * dataset) const {
|
||||
std::ofstream out(fname, std::ios::binary);
|
||||
int n_entries = m_stats.size();
|
||||
out.write((const char *) &n_entries, sizeof(n_entries));
|
||||
|
|
@ -237,26 +241,28 @@ void IMatrixCollector::save_imatrix(const char * fname, const char * dataset) co
|
|||
// Write the number of call the matrix was computed with
|
||||
out.write((const char *) &m_last_call, sizeof(m_last_call));
|
||||
|
||||
// Write the dataset name at the end of the file to later on specify it in quantize
|
||||
int n_dataset = strlen(dataset);
|
||||
out.write((const char *) &n_dataset, sizeof(n_dataset));
|
||||
out.write(dataset, n_dataset);
|
||||
// Write the input filename at the end of the file to later on specify it in quantize
|
||||
{
|
||||
int len = m_params.prompt_file.size();
|
||||
out.write((const char *) &len, sizeof(len));
|
||||
out.write(m_params.prompt_file.c_str(), len);
|
||||
}
|
||||
|
||||
if (m_params.verbosity > 0) {
|
||||
fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname);
|
||||
fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname.c_str());
|
||||
}
|
||||
}
|
||||
|
||||
bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_map<std::string, Stats>& imatrix_data) {
|
||||
std::ifstream in(imatrix_file, std::ios::binary);
|
||||
bool IMatrixCollector::load_imatrix(const char * fname) {
|
||||
std::ifstream in(fname, std::ios::binary);
|
||||
if (!in) {
|
||||
printf("%s: failed to open %s\n",__func__,imatrix_file);
|
||||
printf("%s: failed to open %s\n",__func__, fname);
|
||||
return false;
|
||||
}
|
||||
int n_entries;
|
||||
in.read((char*)&n_entries, sizeof(n_entries));
|
||||
if (in.fail() || n_entries < 1) {
|
||||
printf("%s: no data in file %s\n", __func__, imatrix_file);
|
||||
printf("%s: no data in file %s\n", __func__, fname);
|
||||
return false;
|
||||
}
|
||||
for (int i = 0; i < n_entries; ++i) {
|
||||
|
|
@ -264,23 +270,22 @@ bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_ma
|
|||
std::vector<char> name_as_vec(len+1);
|
||||
in.read((char *)name_as_vec.data(), len);
|
||||
if (in.fail()) {
|
||||
printf("%s: failed reading name for entry %d from %s\n",__func__,i+1,imatrix_file);
|
||||
printf("%s: failed reading name for entry %d from %s\n",__func__,i+1, fname);
|
||||
return false;
|
||||
}
|
||||
name_as_vec[len] = 0;
|
||||
std::string name{name_as_vec.data()};
|
||||
auto& e = imatrix_data[std::move(name)];
|
||||
auto & e = m_stats[std::move(name)];
|
||||
int ncall;
|
||||
in.read((char*)&ncall, sizeof(ncall));
|
||||
int nval;
|
||||
in.read((char *)&nval, sizeof(nval));
|
||||
if (in.fail() || nval < 1) {
|
||||
printf("%s: failed reading number of values for entry %d\n",__func__,i);
|
||||
imatrix_data = {};
|
||||
m_stats = {};
|
||||
return false;
|
||||
}
|
||||
|
||||
// When re-called from load_imatrix() with add set, this will already be created.
|
||||
if (e.values.empty()) {
|
||||
e.values.resize(nval, 0);
|
||||
e.counts.resize(nval, 0);
|
||||
|
|
@ -290,7 +295,7 @@ bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_ma
|
|||
in.read((char*)tmp.data(), nval*sizeof(float));
|
||||
if (in.fail()) {
|
||||
printf("%s: failed reading data for entry %d\n",__func__,i);
|
||||
imatrix_data = {};
|
||||
m_stats = {};
|
||||
return false;
|
||||
}
|
||||
|
||||
|
|
@ -305,13 +310,6 @@ bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_ma
|
|||
return true;
|
||||
}
|
||||
|
||||
bool IMatrixCollector::load_imatrix(const char * file_name, bool add) {
|
||||
if (!add) {
|
||||
m_stats.clear();
|
||||
}
|
||||
return load_imatrix(file_name, m_stats);
|
||||
}
|
||||
|
||||
static IMatrixCollector g_collector;
|
||||
|
||||
static bool ik_collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
|
||||
|
|
@ -359,8 +357,7 @@ static results_log_softmax log_softmax(int n_vocab, const float * logits, int to
|
|||
|
||||
static void process_logits(
|
||||
int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
|
||||
double & nll, double & nll2, float * logit_history, float * prob_history
|
||||
) {
|
||||
double & nll, double & nll2, float * logit_history, float * prob_history) {
|
||||
std::mutex mutex;
|
||||
int counter = 0;
|
||||
auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
|
||||
|
|
@ -392,8 +389,7 @@ static void process_logits(
|
|||
}
|
||||
}
|
||||
|
||||
static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl, int from_chunk) {
|
||||
|
||||
static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
|
||||
const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
|
||||
GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
|
||||
const int n_ctx = llama_n_ctx(ctx);
|
||||
|
|
@ -406,13 +402,13 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
|
|||
auto tim2 = std::chrono::high_resolution_clock::now();
|
||||
fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
|
||||
|
||||
if (from_chunk > 0) {
|
||||
if (size_t((from_chunk + 2)*n_ctx) >= tokens.size()) {
|
||||
fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, from_chunk);
|
||||
if (params.i_chunk > 0) {
|
||||
if (size_t((params.i_chunk + 2)*n_ctx) >= tokens.size()) {
|
||||
fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
|
||||
return false;
|
||||
}
|
||||
fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, from_chunk, from_chunk*n_ctx);
|
||||
tokens.erase(tokens.begin(), tokens.begin() + from_chunk*n_ctx);
|
||||
fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
|
||||
tokens.erase(tokens.begin(), tokens.begin() + params.i_chunk*n_ctx);
|
||||
}
|
||||
|
||||
if (int(tokens.size()) < 2*n_ctx) {
|
||||
|
|
@ -425,7 +421,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
|
|||
std::vector<float> logit_history;
|
||||
std::vector<float> prob_history;
|
||||
|
||||
if (compute_ppl) {
|
||||
if (params.compute_ppl) {
|
||||
logit_history.resize(tokens.size());
|
||||
prob_history.resize(tokens.size());
|
||||
}
|
||||
|
|
@ -447,7 +443,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
|
|||
const int num_batches = (n_ctx + n_batch - 1) / n_batch;
|
||||
|
||||
std::vector<float> logits;
|
||||
if (compute_ppl && num_batches > 1) {
|
||||
if (params.compute_ppl && num_batches > 1) {
|
||||
logits.reserve((size_t)n_ctx * n_vocab);
|
||||
}
|
||||
|
||||
|
|
@ -483,7 +479,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
|
|||
// restore the original token in case it was set to BOS
|
||||
tokens[batch_start] = token_org;
|
||||
|
||||
if (compute_ppl && num_batches > 1) {
|
||||
if (params.compute_ppl && num_batches > 1) {
|
||||
const auto * batch_logits = llama_get_logits(ctx);
|
||||
logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
|
||||
}
|
||||
|
|
@ -502,7 +498,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
|
|||
fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
|
||||
}
|
||||
|
||||
if (compute_ppl) {
|
||||
if (params.compute_ppl) {
|
||||
const int first = n_ctx/2;
|
||||
const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
|
||||
process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
|
||||
|
|
@ -517,7 +513,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
|
|||
}
|
||||
printf("\n");
|
||||
|
||||
if (compute_ppl) {
|
||||
if (params.compute_ppl) {
|
||||
nll2 /= count;
|
||||
nll /= count;
|
||||
const double ppl = exp(nll);
|
||||
|
|
@ -534,109 +530,32 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
|
|||
}
|
||||
|
||||
int main(int argc, char ** argv) {
|
||||
StatParams sparams;
|
||||
std::string prev_result_file;
|
||||
std::string combine_files;
|
||||
bool compute_ppl = true;
|
||||
int from_chunk = 0;
|
||||
std::vector<char*> args;
|
||||
args.push_back(argv[0]);
|
||||
int iarg = 1;
|
||||
for (; iarg < argc-1; ++iarg) {
|
||||
std::string arg{argv[iarg]};
|
||||
if (arg == "-o" || arg == "--output-file") {
|
||||
sparams.ofile = argv[++iarg];
|
||||
}
|
||||
else if (arg == "-ofreq" || arg == "--output-frequency") {
|
||||
sparams.n_output_frequency = std::stoi(argv[++iarg]);
|
||||
}
|
||||
else if (arg == "-ow" || arg == "--output-weight") {
|
||||
sparams.collect_output_weight = std::stoi(argv[++iarg]);
|
||||
}
|
||||
else if (arg == "--verbosity") {
|
||||
sparams.verbosity = std::stoi(argv[++iarg]);
|
||||
} else if (arg == "--no-ppl") {
|
||||
compute_ppl = false;
|
||||
} else if (arg == "--keep-imatrix") {
|
||||
sparams.keep_every = std::stoi(argv[++iarg]);
|
||||
} else if (arg == "--continue-from") {
|
||||
prev_result_file = argv[++iarg];
|
||||
} else if (arg == "--combine") {
|
||||
combine_files = argv[++iarg];
|
||||
}
|
||||
else if (arg == "--from-chunk") {
|
||||
from_chunk = std::stoi(argv[++iarg]);
|
||||
} else {
|
||||
args.push_back(argv[iarg]);
|
||||
}
|
||||
}
|
||||
if (iarg < argc) {
|
||||
std::string arg{argv[iarg]};
|
||||
if (arg == "--no-ppl") {
|
||||
compute_ppl = false;
|
||||
} else {
|
||||
args.push_back(argv[iarg]);
|
||||
}
|
||||
}
|
||||
|
||||
gpt_params params;
|
||||
params.n_batch = 512;
|
||||
|
||||
params.n_ctx = 512;
|
||||
params.logits_all = true;
|
||||
params.verbosity = 1;
|
||||
|
||||
if (!gpt_params_parse(argc, argv, params)) {
|
||||
gpt_params_print_usage(argc, argv, params);
|
||||
print_usage(argc, argv, params);
|
||||
return 1;
|
||||
}
|
||||
|
||||
params.logits_all = true;
|
||||
params.n_batch = std::min(params.n_batch, params.n_ctx);
|
||||
|
||||
print_build_info();
|
||||
g_collector.set_params(params);
|
||||
|
||||
if (params.seed == LLAMA_DEFAULT_SEED) {
|
||||
params.seed = time(NULL);
|
||||
}
|
||||
|
||||
fprintf(stderr, "%s: seed = %u\n", __func__, params.seed);
|
||||
|
||||
std::mt19937 rng(params.seed);
|
||||
|
||||
sparams.dataset = params.prompt_file;
|
||||
g_collector.set_parameters(std::move(sparams));
|
||||
|
||||
if (!combine_files.empty()) {
|
||||
std::vector<std::string> files;
|
||||
size_t pos = 0;
|
||||
while (true) {
|
||||
auto new_pos = combine_files.find(',', pos);
|
||||
if (new_pos != std::string::npos) {
|
||||
files.emplace_back(combine_files.substr(pos, new_pos - pos));
|
||||
pos = new_pos + 1;
|
||||
} else {
|
||||
files.emplace_back(combine_files.substr(pos));
|
||||
break;
|
||||
}
|
||||
}
|
||||
if (files.size() < 2) {
|
||||
fprintf(stderr, "You must provide at least two comma separated files to use --combine\n");
|
||||
return 1;
|
||||
}
|
||||
printf("Combining the following %d files\n", int(files.size()));
|
||||
for (auto& file : files) {
|
||||
printf(" %s\n", file.c_str());
|
||||
if (!g_collector.load_imatrix(file.c_str(), true)) {
|
||||
fprintf(stderr, "Failed to load %s\n", file.c_str());
|
||||
for (const auto & in_file : params.in_files) {
|
||||
printf("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
|
||||
if (!g_collector.load_imatrix(in_file.c_str())) {
|
||||
fprintf(stderr, "%s : failed to load %s\n", __func__, in_file.c_str());
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
|
||||
if (params.in_files.size() > 1) {
|
||||
printf("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
|
||||
g_collector.save_imatrix();
|
||||
return 0;
|
||||
}
|
||||
|
||||
if (!prev_result_file.empty()) {
|
||||
if (!g_collector.load_imatrix(prev_result_file.c_str(), false)) {
|
||||
fprintf(stderr, "=============== Failed to load %s\n", prev_result_file.c_str());
|
||||
return 1;
|
||||
}
|
||||
}
|
||||
|
||||
llama_backend_init();
|
||||
|
|
@ -651,6 +570,7 @@ int main(int argc, char ** argv) {
|
|||
// init
|
||||
llama_model * model;
|
||||
llama_context * ctx;
|
||||
|
||||
std::tie(model, ctx) = llama_init_from_gpt_params(params);
|
||||
if (model == nullptr || ctx == nullptr) {
|
||||
fprintf(stderr, "%s : failed to init\n", __func__);
|
||||
|
|
@ -669,8 +589,7 @@ int main(int argc, char ** argv) {
|
|||
fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
|
||||
}
|
||||
|
||||
bool OK = compute_imatrix(ctx, params, compute_ppl, from_chunk);
|
||||
if (!OK) {
|
||||
if (!compute_imatrix(ctx, params)) {
|
||||
return 1;
|
||||
}
|
||||
|
||||
|
|
|
|||
|
|
@ -6,52 +6,22 @@ import re
|
|||
import sys
|
||||
from typing import Any, Dict, List, Set, Tuple, Union
|
||||
|
||||
def _build_repetition(item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False):
|
||||
if not separator_rule:
|
||||
|
||||
def _build_repetition(item_rule, min_items, max_items, separator_rule=None):
|
||||
|
||||
if min_items == 0 and max_items == 1:
|
||||
return f'{item_rule}?'
|
||||
elif min_items == 1 and max_items is None:
|
||||
|
||||
if not separator_rule:
|
||||
if min_items == 1 and max_items is None:
|
||||
return f'{item_rule}+'
|
||||
|
||||
result = ''
|
||||
|
||||
if min_items > 0:
|
||||
if item_rule_is_literal and separator_rule is None:
|
||||
result = '"' + (item_rule[1:-1] * min_items) + '"'
|
||||
elif min_items == 0 and max_items is None:
|
||||
return f'{item_rule}*'
|
||||
else:
|
||||
result = (f' {separator_rule} ' if separator_rule else ' ').join([item_rule] * min_items)
|
||||
return f'{item_rule}{{{min_items},{max_items if max_items is not None else ""}}}'
|
||||
|
||||
def opt_repetitions(up_to_n, prefix_with_sep=False):
|
||||
'''
|
||||
- n=4, no sep: '(a (a (a (a)?)?)?)?'
|
||||
- n=4, sep=',', prefix: '("," a ("," a ("," a ("," a)?)?)?)?'
|
||||
- n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?'
|
||||
'''
|
||||
|
||||
content = f'{separator_rule} {item_rule}' if prefix_with_sep and separator_rule else item_rule
|
||||
if up_to_n == 0:
|
||||
return ''
|
||||
elif up_to_n == 1:
|
||||
return f'({content})?'
|
||||
elif separator_rule and not prefix_with_sep:
|
||||
return f'({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?'
|
||||
else:
|
||||
return (f'({content} ' * up_to_n).rstrip() + (')?' * up_to_n)
|
||||
|
||||
if min_items > 0 and max_items != min_items:
|
||||
result += ' '
|
||||
|
||||
if max_items is not None:
|
||||
result += opt_repetitions(max_items - min_items, prefix_with_sep=min_items > 0)
|
||||
else:
|
||||
item_operator = f'({separator_rule + " " if separator_rule else ""}{item_rule})'
|
||||
|
||||
if min_items == 0 and separator_rule:
|
||||
result = f'({item_rule} {item_operator}*)?'
|
||||
else:
|
||||
result += f'{item_operator}*'
|
||||
|
||||
return result
|
||||
result = item_rule + ' ' + _build_repetition(f'({separator_rule} {item_rule})', min_items - 1 if min_items > 0 else 0, max_items - 1 if max_items is not None else None)
|
||||
return f'({result})?' if min_items == 0 else result
|
||||
|
||||
|
||||
class BuiltinRule:
|
||||
|
|
@ -59,31 +29,29 @@ class BuiltinRule:
|
|||
self.content = content
|
||||
self.deps = deps or []
|
||||
|
||||
_up_to_15_digits = _build_repetition('[0-9]', 0, 15)
|
||||
|
||||
# whitespace is constrained to a single space char to prevent model "running away" in
|
||||
# whitespace. Also maybe improves generation quality?
|
||||
SPACE_RULE = '" "?'
|
||||
|
||||
PRIMITIVE_RULES = {
|
||||
'boolean' : BuiltinRule('("true" | "false") space', []),
|
||||
'decimal-part' : BuiltinRule('[0-9] ' + _up_to_15_digits, []),
|
||||
'integral-part': BuiltinRule('[0-9] | [1-9] ' + _up_to_15_digits, []),
|
||||
'decimal-part' : BuiltinRule('[0-9]{1,16}', []),
|
||||
'integral-part': BuiltinRule('[0] | [1-9] [0-9]{0,15}', []),
|
||||
'number' : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']),
|
||||
'integer' : BuiltinRule('("-"? integral-part) space', ['integral-part']),
|
||||
'value' : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
|
||||
'object' : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
|
||||
'array' : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
|
||||
'uuid' : BuiltinRule(r'"\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + r' "\"" space', []),
|
||||
'char' : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', []),
|
||||
'uuid' : BuiltinRule(r'"\"" [0-9a-fA-F]{8} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{12} "\"" space', []),
|
||||
'char' : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F]{4})', []),
|
||||
'string' : BuiltinRule(r'"\"" char* "\"" space', ['char']),
|
||||
'null' : BuiltinRule('"null" space', []),
|
||||
}
|
||||
|
||||
# TODO: support "uri", "email" string formats
|
||||
STRING_FORMAT_RULES = {
|
||||
'date' : BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
|
||||
'time' : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
|
||||
'date' : BuiltinRule('[0-9]{4} "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
|
||||
'time' : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9]{3} )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
|
||||
'date-time' : BuiltinRule('date "T" time', ['date', 'time']),
|
||||
'date-string' : BuiltinRule('"\\"" date "\\"" space', ['date']),
|
||||
'time-string' : BuiltinRule('"\\"" time "\\"" space', ['time']),
|
||||
|
|
@ -333,7 +301,7 @@ class SchemaConverter:
|
|||
sub_rule_ids[sub] = id
|
||||
sub = id
|
||||
|
||||
seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times, item_rule_is_literal=sub_is_literal), False)
|
||||
seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times), False)
|
||||
else:
|
||||
literal = ''
|
||||
while i < length:
|
||||
|
|
|
|||
|
|
@ -624,7 +624,7 @@ string ::= "\"" (
|
|||
"\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
|
||||
)* "\"" ws
|
||||
ws ::= ([ \t\n] ws)?
|
||||
float ::= ("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws
|
||||
float ::= ("-"? ([0] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? ws
|
||||
|
||||
integer ::= [0-9]+"""
|
||||
|
||||
|
|
|
|||
|
|
@ -6,10 +6,6 @@
|
|||
#include "ggml-metal.h"
|
||||
#endif
|
||||
|
||||
#ifdef GGML_USE_SYCL
|
||||
#include "ggml-sycl.h"
|
||||
#endif
|
||||
|
||||
#include "ggml-rpc.h"
|
||||
#ifdef _WIN32
|
||||
# include <windows.h>
|
||||
|
|
@ -83,12 +79,6 @@ static ggml_backend_t create_backend() {
|
|||
if (!backend) {
|
||||
fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
|
||||
}
|
||||
#elif GGML_USE_SYCL
|
||||
fprintf(stderr, "%s: using SYCL backend\n", __func__);
|
||||
backend = ggml_backend_sycl_init(0); // init device 0
|
||||
if (!backend) {
|
||||
fprintf(stderr, "%s: ggml_backend_sycl_init() failed\n", __func__);
|
||||
}
|
||||
#endif
|
||||
|
||||
// if there aren't GPU Backends fallback to CPU backend
|
||||
|
|
|
|||
|
|
@ -279,7 +279,7 @@ node index.js
|
|||
|
||||
`id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot. Default: `-1`
|
||||
|
||||
`cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. Default: `false`
|
||||
`cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `false`
|
||||
|
||||
`system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
|
||||
|
||||
|
|
|
|||
|
|
@ -2,57 +2,26 @@
|
|||
const SPACE_RULE = '" "?';
|
||||
|
||||
function _buildRepetition(itemRule, minItems, maxItems, opts={}) {
|
||||
if (minItems === 0 && maxItems === 1) {
|
||||
return `${itemRule}?`;
|
||||
}
|
||||
|
||||
|
||||
const separatorRule = opts.separatorRule ?? '';
|
||||
const itemRuleIsLiteral = opts.itemRuleIsLiteral ?? false
|
||||
|
||||
if (separatorRule === '') {
|
||||
if (minItems === 0 && maxItems === 1) {
|
||||
return `${itemRule}?`;
|
||||
} else if (minItems === 1 && maxItems === undefined) {
|
||||
if (minItems === 1 && maxItems === undefined) {
|
||||
return `${itemRule}+`;
|
||||
}
|
||||
}
|
||||
|
||||
let result = '';
|
||||
if (minItems > 0) {
|
||||
if (itemRuleIsLiteral && separatorRule === '') {
|
||||
result = `"${itemRule.slice(1, -1).repeat(minItems)}"`;
|
||||
} else if (minItems === 0 && maxItems === undefined) {
|
||||
return `${itemRule}*`;
|
||||
} else {
|
||||
result = Array.from({ length: minItems }, () => itemRule)
|
||||
.join(separatorRule !== '' ? ` ${separatorRule} ` : ' ');
|
||||
return `${itemRule}{${minItems},${maxItems !== undefined ? maxItems : ''}}`;
|
||||
}
|
||||
}
|
||||
|
||||
const optRepetitions = (upToN, prefixWithSep=false) => {
|
||||
const content = separatorRule !== '' && prefixWithSep ? `${separatorRule} ${itemRule}` : itemRule;
|
||||
if (upToN === 0) {
|
||||
return '';
|
||||
} else if (upToN === 1) {
|
||||
return `(${content})?`;
|
||||
} else if (separatorRule !== '' && !prefixWithSep) {
|
||||
return `(${content} ${optRepetitions(upToN - 1, true)})?`;
|
||||
} else {
|
||||
return Array.from({ length: upToN }, () => `(${content}`).join(' ').trim() + Array.from({ length: upToN }, () => ')?').join('');
|
||||
}
|
||||
};
|
||||
|
||||
if (minItems > 0 && maxItems !== minItems) {
|
||||
result += ' ';
|
||||
}
|
||||
|
||||
if (maxItems !== undefined) {
|
||||
result += optRepetitions(maxItems - minItems, minItems > 0);
|
||||
} else {
|
||||
const itemOperator = `(${separatorRule !== '' ? separatorRule + ' ' : ''}${itemRule})`;
|
||||
|
||||
if (minItems === 0 && separatorRule !== '') {
|
||||
result = `(${itemRule} ${itemOperator}*)?`;
|
||||
} else {
|
||||
result += `${itemOperator}*`;
|
||||
}
|
||||
}
|
||||
|
||||
return result;
|
||||
const result = itemRule + ' ' + _buildRepetition(`(${separatorRule} ${itemRule})`, minItems > 0 ? minItems - 1 : 0, maxItems !== undefined ? maxItems - 1 : undefined);
|
||||
return minItems === 0 ? `(${result})?` : result;
|
||||
}
|
||||
|
||||
class BuiltinRule {
|
||||
|
|
@ -62,27 +31,25 @@ class BuiltinRule {
|
|||
}
|
||||
}
|
||||
|
||||
const UP_TO_15_DIGITS = _buildRepetition('[0-9]', 0, 15);
|
||||
|
||||
const PRIMITIVE_RULES = {
|
||||
boolean : new BuiltinRule('("true" | "false") space', []),
|
||||
'decimal-part' : new BuiltinRule('[0-9] ' + UP_TO_15_DIGITS, []),
|
||||
'integral-part': new BuiltinRule('[0-9] | [1-9] ' + UP_TO_15_DIGITS, []),
|
||||
'decimal-part' : new BuiltinRule('[0-9]{1,16}', []),
|
||||
'integral-part': new BuiltinRule('[0] | [1-9] [0-9]{0,15}', []),
|
||||
number : new BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']),
|
||||
integer : new BuiltinRule('("-"? integral-part) space', ['integral-part']),
|
||||
value : new BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
|
||||
object : new BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
|
||||
array : new BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
|
||||
uuid : new BuiltinRule('"\\"" ' + [8, 4, 4, 4, 12].map(n => [...new Array(n)].map(_ => '[0-9a-fA-F]').join('')).join(' "-" ') + ' "\\"" space', []),
|
||||
char : new BuiltinRule(`[^"\\\\] | "\\\\" (["\\\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])`, []),
|
||||
uuid : new BuiltinRule('"\\"" [0-9a-fA-F]{8} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{4} "-" [0-9a-fA-F]{12} "\\"" space', []),
|
||||
char : new BuiltinRule(`[^"\\\\] | "\\\\" (["\\\\/bfnrt] | "u" [0-9a-fA-F]{4})`, []),
|
||||
string : new BuiltinRule(`"\\"" char* "\\"" space`, ['char']),
|
||||
null : new BuiltinRule('"null" space', []),
|
||||
};
|
||||
|
||||
// TODO: support "uri", "email" string formats
|
||||
const STRING_FORMAT_RULES = {
|
||||
'date' : new BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
|
||||
'time' : new BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
|
||||
'date' : new BuiltinRule('[0-9]{4} "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
|
||||
'time' : new BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9]{3} )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
|
||||
'date-time' : new BuiltinRule('date "T" time', ['date', 'time']),
|
||||
'date-string' : new BuiltinRule('"\\"" date "\\"" space', ['date']),
|
||||
'time-string' : new BuiltinRule('"\\"" time "\\"" space', ['time']),
|
||||
|
|
|
|||
|
|
@ -648,6 +648,9 @@ struct server_context {
|
|||
|
||||
server_metrics metrics;
|
||||
|
||||
// Necessary similarity of prompt for slot selection
|
||||
float slot_prompt_similarity = 0.0f;
|
||||
|
||||
~server_context() {
|
||||
if (ctx) {
|
||||
llama_free(ctx);
|
||||
|
|
@ -796,24 +799,88 @@ struct server_context {
|
|||
return prompt_tokens;
|
||||
}
|
||||
|
||||
server_slot * get_slot(int id) {
|
||||
int64_t t_last = ggml_time_us();
|
||||
|
||||
server_slot * last_used = nullptr;
|
||||
|
||||
server_slot * get_slot_by_id(int id) {
|
||||
for (server_slot & slot : slots) {
|
||||
if (slot.id == id && slot.available()) {
|
||||
if (slot.id == id) {
|
||||
return &slot;
|
||||
}
|
||||
}
|
||||
|
||||
// among all available slots, find the one that has been least recently used
|
||||
if (slot.available() && slot.t_last_used < t_last) {
|
||||
last_used = &slot;
|
||||
return nullptr;
|
||||
}
|
||||
|
||||
server_slot * get_available_slot(const std::string & prompt) {
|
||||
server_slot * ret = nullptr;
|
||||
|
||||
// find the slot that has at least n% prompt similarity
|
||||
if (ret == nullptr && slot_prompt_similarity != 0.0f && !prompt.empty()) {
|
||||
int max_lcp_len = 0;
|
||||
float similarity = 0;
|
||||
|
||||
for (server_slot & slot : slots) {
|
||||
// skip the slot if it is not available
|
||||
if (!slot.available()) {
|
||||
continue;
|
||||
}
|
||||
|
||||
// skip the slot if it does not contains prompt
|
||||
if (!slot.prompt.is_string()) {
|
||||
continue;
|
||||
}
|
||||
|
||||
// current slot's prompt
|
||||
std::string slot_prompt = slot.prompt.get<std::string>();
|
||||
|
||||
// length of the current slot's prompt
|
||||
int slot_prompt_len = slot_prompt.size();
|
||||
|
||||
// length of the Longest Common Prefix between the current slot's prompt and the input prompt
|
||||
int lcp_len = common_part(slot_prompt, prompt);
|
||||
|
||||
// fraction of the common substring length compared to the current slot's prompt length
|
||||
similarity = static_cast<float>(lcp_len) / slot_prompt_len;
|
||||
|
||||
// select the current slot if the criteria match
|
||||
if (lcp_len > max_lcp_len && similarity > slot_prompt_similarity) {
|
||||
max_lcp_len = lcp_len;
|
||||
ret = &slot;
|
||||
}
|
||||
}
|
||||
|
||||
if (ret != nullptr) {
|
||||
LOG_VERBOSE("selected slot by lcp similarity", {
|
||||
{"id_slot", ret->id},
|
||||
{"max_lcp_len", max_lcp_len},
|
||||
{"similarity", similarity},
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
// find the slot that has been least recently used
|
||||
if (ret == nullptr) {
|
||||
int64_t t_last = ggml_time_us();
|
||||
for (server_slot & slot : slots) {
|
||||
// skip the slot if it is not available
|
||||
if (!slot.available()) {
|
||||
continue;
|
||||
}
|
||||
|
||||
// select the current slot if the criteria match
|
||||
if (slot.t_last_used < t_last) {
|
||||
t_last = slot.t_last_used;
|
||||
ret = &slot;
|
||||
}
|
||||
}
|
||||
|
||||
return last_used;
|
||||
if (ret != nullptr) {
|
||||
LOG_VERBOSE("selected slot by lru", {
|
||||
{"id_slot", ret->id},
|
||||
{"t_last", t_last},
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
return ret;
|
||||
}
|
||||
|
||||
bool launch_slot_with_task(server_slot & slot, const server_task & task) {
|
||||
|
|
@ -889,7 +956,7 @@ struct server_context {
|
|||
slot.params.input_suffix = json_value(data, "input_suffix", default_params.input_suffix);
|
||||
|
||||
// get prompt
|
||||
{
|
||||
if (!task.infill) {
|
||||
const auto & prompt = data.find("prompt");
|
||||
if (prompt == data.end()) {
|
||||
send_error(task, "Either \"prompt\" or \"messages\" must be provided", ERROR_TYPE_INVALID_REQUEST);
|
||||
|
|
@ -1516,13 +1583,29 @@ struct server_context {
|
|||
switch (task.type) {
|
||||
case SERVER_TASK_TYPE_COMPLETION:
|
||||
{
|
||||
server_slot * slot = get_slot(json_value(task.data, "id_slot", -1));
|
||||
int id_slot = json_value(task.data, "id_slot", -1);
|
||||
std::string prompt = json_value(task.data, "prompt", std::string());
|
||||
|
||||
server_slot * slot;
|
||||
|
||||
if (id_slot != -1) {
|
||||
slot = get_slot_by_id(id_slot);
|
||||
} else {
|
||||
slot = get_available_slot(prompt);
|
||||
}
|
||||
|
||||
if (slot == nullptr) {
|
||||
// if no slot is available, we defer this task for processing later
|
||||
LOG_VERBOSE("no slot is available", {{"id_task", task.id}});
|
||||
queue_tasks.defer(task);
|
||||
break;
|
||||
}
|
||||
if (!slot->available()) {
|
||||
// if requested slot is unavailable, we defer this task for processing later
|
||||
LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
|
||||
queue_tasks.defer(task);
|
||||
break;
|
||||
}
|
||||
|
||||
if (task.data.contains("system_prompt")) {
|
||||
std::string sys_prompt = json_value(task.data, "system_prompt", std::string());
|
||||
|
|
@ -1639,11 +1722,17 @@ struct server_context {
|
|||
case SERVER_TASK_TYPE_SLOT_SAVE:
|
||||
{
|
||||
int id_slot = task.data.at("id_slot");
|
||||
server_slot * slot = get_slot(id_slot);
|
||||
server_slot * slot = get_slot_by_id(id_slot);
|
||||
if (slot == nullptr) {
|
||||
send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
|
||||
break;
|
||||
}
|
||||
if (!slot->available()) {
|
||||
// if requested slot is unavailable, we defer this task for processing later
|
||||
LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
|
||||
queue_tasks.defer(task);
|
||||
break;
|
||||
}
|
||||
|
||||
const size_t token_count = slot->cache_tokens.size();
|
||||
const int64_t t_start = ggml_time_us();
|
||||
|
|
@ -1674,11 +1763,17 @@ struct server_context {
|
|||
case SERVER_TASK_TYPE_SLOT_RESTORE:
|
||||
{
|
||||
int id_slot = task.data.at("id_slot");
|
||||
server_slot * slot = get_slot(id_slot);
|
||||
server_slot * slot = get_slot_by_id(id_slot);
|
||||
if (slot == nullptr) {
|
||||
send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
|
||||
break;
|
||||
}
|
||||
if (!slot->available()) {
|
||||
// if requested slot is unavailable, we defer this task for processing later
|
||||
LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
|
||||
queue_tasks.defer(task);
|
||||
break;
|
||||
}
|
||||
|
||||
const int64_t t_start = ggml_time_us();
|
||||
|
||||
|
|
@ -1716,11 +1811,17 @@ struct server_context {
|
|||
case SERVER_TASK_TYPE_SLOT_ERASE:
|
||||
{
|
||||
int id_slot = task.data.at("id_slot");
|
||||
server_slot * slot = get_slot(id_slot);
|
||||
server_slot * slot = get_slot_by_id(id_slot);
|
||||
if (slot == nullptr) {
|
||||
send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
|
||||
break;
|
||||
}
|
||||
if (!slot->available()) {
|
||||
// if requested slot is unavailable, we defer this task for processing later
|
||||
LOG_VERBOSE("requested slot is unavailable", {{"id_task", task.id}});
|
||||
queue_tasks.defer(task);
|
||||
break;
|
||||
}
|
||||
|
||||
// Erase token cache
|
||||
const size_t n_erased = slot->cache_tokens.size();
|
||||
|
|
@ -2361,7 +2462,7 @@ int main(int argc, char ** argv) {
|
|||
|
||||
// TODO: not great to use extern vars
|
||||
server_log_json = params.log_json;
|
||||
server_verbose = params.verbose;
|
||||
server_verbose = params.verbosity > 0;
|
||||
|
||||
// struct that contains llama context and inference
|
||||
server_context ctx_server;
|
||||
|
|
@ -2468,6 +2569,9 @@ int main(int argc, char ** argv) {
|
|||
log_data["api_key"] = "api_key: " + std::to_string(params.api_keys.size()) + " keys loaded";
|
||||
}
|
||||
|
||||
// Necessary similarity of prompt for slot selection
|
||||
ctx_server.slot_prompt_similarity = params.slot_prompt_similarity;
|
||||
|
||||
// load the model
|
||||
if (!ctx_server.load_model(params)) {
|
||||
state.store(SERVER_STATE_ERROR);
|
||||
|
|
|
|||
|
|
@ -253,6 +253,13 @@ static size_t common_part(const std::vector<llama_token> & a, const std::vector<
|
|||
return i;
|
||||
}
|
||||
|
||||
static size_t common_part(const std::string & a, const std::string & b) {
|
||||
size_t i;
|
||||
for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
|
||||
|
||||
return i;
|
||||
}
|
||||
|
||||
static bool ends_with(const std::string & str, const std::string & suffix) {
|
||||
return str.size() >= suffix.size() && 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
|
||||
}
|
||||
|
|
|
|||
|
|
@ -9108,6 +9108,7 @@ static void soft_max_f32(const float * x, const float * mask, float * dst, const
|
|||
// find the sum of exps in the block
|
||||
tmp = warp_reduce_sum(tmp, item_ct1);
|
||||
if (block_size > WARP_SIZE) {
|
||||
item_ct1.barrier(sycl::access::fence_space::local_space);
|
||||
if (warp_id == 0) {
|
||||
buf[lane_id] = 0.f;
|
||||
}
|
||||
|
|
|
|||
124
ggml-vulkan.cpp
124
ggml-vulkan.cpp
|
|
@ -345,15 +345,12 @@ struct vk_context {
|
|||
};
|
||||
|
||||
struct ggml_tensor_extra_gpu {
|
||||
bool ready;
|
||||
|
||||
size_t ctx_idx;
|
||||
|
||||
vk_buffer_ref buffer_gpu;
|
||||
uint64_t offset;
|
||||
|
||||
void reset() {
|
||||
ready = false;
|
||||
ctx_idx = 0;
|
||||
buffer_gpu.reset();
|
||||
offset = 0;
|
||||
|
|
@ -2949,7 +2946,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
|
|||
const uint64_t d_sz = sizeof(float) * d_ne;
|
||||
|
||||
vk_buffer d_D = extra->buffer_gpu.lock();
|
||||
const uint64_t d_buf_offset = extra->offset;
|
||||
const uint64_t d_buf_offset = extra->offset + dst->view_offs;
|
||||
GGML_ASSERT(d_D != nullptr);
|
||||
GGML_ASSERT(d_D->size >= d_buf_offset + d_sz * ne02 * ne03);
|
||||
vk_buffer d_X;
|
||||
|
|
@ -2958,12 +2955,12 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context * su
|
|||
uint64_t y_buf_offset = 0;
|
||||
if (!src0_uma) {
|
||||
d_Qx = extra_src0->buffer_gpu.lock();
|
||||
qx_buf_offset = extra_src0->offset;
|
||||
qx_buf_offset = extra_src0->offset + src0->view_offs;
|
||||
GGML_ASSERT(d_Qx != nullptr);
|
||||
}
|
||||
if (!src1_uma) {
|
||||
d_Qy = extra_src1->buffer_gpu.lock();
|
||||
qy_buf_offset = extra_src1->offset;
|
||||
qy_buf_offset = extra_src1->offset + src1->view_offs;
|
||||
GGML_ASSERT(d_Qy != nullptr);
|
||||
}
|
||||
if (qx_needs_dequant) {
|
||||
|
|
@ -3114,7 +3111,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
|
|||
const uint64_t d_sz = sizeof(float) * d_ne;
|
||||
|
||||
vk_buffer d_D = extra->buffer_gpu.lock();
|
||||
const uint64_t d_buf_offset = extra->offset;
|
||||
const uint64_t d_buf_offset = extra->offset + dst->view_offs;
|
||||
GGML_ASSERT(d_D != nullptr);
|
||||
vk_buffer d_X;
|
||||
uint64_t x_buf_offset = 0;
|
||||
|
|
@ -3122,12 +3119,12 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context
|
|||
uint64_t y_buf_offset = 0;
|
||||
if(!src0_uma) {
|
||||
d_Qx = extra_src0->buffer_gpu.lock();
|
||||
qx_buf_offset = extra_src0->offset;
|
||||
qx_buf_offset = extra_src0->offset + src0->view_offs;
|
||||
GGML_ASSERT(d_Qx != nullptr);
|
||||
}
|
||||
if(!src1_uma) {
|
||||
d_Qy = extra_src1->buffer_gpu.lock();
|
||||
qy_buf_offset = extra_src1->offset;
|
||||
qy_buf_offset = extra_src1->offset + src1->view_offs;
|
||||
GGML_ASSERT(d_Qy != nullptr);
|
||||
}
|
||||
if (qx_needs_dequant) {
|
||||
|
|
@ -3246,14 +3243,14 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
|
|||
const uint64_t d_sz = sizeof(float) * d_ne;
|
||||
|
||||
vk_buffer d_D = extra->buffer_gpu.lock();
|
||||
const uint64_t d_buf_offset = extra->offset;
|
||||
const uint64_t d_buf_offset = extra->offset + dst->view_offs;
|
||||
GGML_ASSERT(d_D != nullptr);
|
||||
vk_buffer d_Qx = extra_src0->buffer_gpu.lock();
|
||||
const uint64_t qx_buf_offset = extra_src0->offset;
|
||||
const uint64_t qx_buf_offset = extra_src0->offset + src0->view_offs;
|
||||
GGML_ASSERT(d_Qx != nullptr);
|
||||
if (!src1_uma) {
|
||||
d_Qy = extra_src1->buffer_gpu.lock();
|
||||
qy_buf_offset = extra_src1->offset;
|
||||
qy_buf_offset = extra_src1->offset + src1->view_offs;
|
||||
GGML_ASSERT(d_Qx != nullptr);
|
||||
}
|
||||
|
||||
|
|
@ -3323,14 +3320,14 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
|
|||
const uint64_t d_sz = sizeof(float) * d_ne;
|
||||
|
||||
vk_buffer d_D = extra->buffer_gpu.lock();
|
||||
const uint64_t d_buf_offset = extra->offset;
|
||||
const uint64_t d_buf_offset = extra->offset + dst->view_offs;
|
||||
GGML_ASSERT(d_D != nullptr);
|
||||
vk_buffer d_Qx = extra_src0->buffer_gpu.lock();
|
||||
const uint64_t qx_buf_offset = extra_src0->offset;
|
||||
const uint64_t qx_buf_offset = extra_src0->offset + src0->view_offs;
|
||||
GGML_ASSERT(d_Qx != nullptr);
|
||||
if (!src1_uma) {
|
||||
d_Qy = extra_src1->buffer_gpu.lock();
|
||||
qy_buf_offset = extra_src1->offset;
|
||||
qy_buf_offset = extra_src1->offset + src1->view_offs;
|
||||
GGML_ASSERT(d_Qx != nullptr);
|
||||
}
|
||||
|
||||
|
|
@ -3459,7 +3456,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context *
|
|||
const uint64_t d_sz = sizeof(float) * d_ne;
|
||||
|
||||
vk_buffer d_D = extra->buffer_gpu.lock();
|
||||
const uint64_t d_buf_offset = extra->offset;
|
||||
const uint64_t d_buf_offset = extra->offset + dst->view_offs;
|
||||
GGML_ASSERT(d_D != nullptr);
|
||||
vk_buffer d_X;
|
||||
uint64_t x_buf_offset = 0;
|
||||
|
|
@ -3467,17 +3464,17 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context *
|
|||
uint64_t y_buf_offset = 0;
|
||||
if (!src0_uma) {
|
||||
d_Qx = extra_src0->buffer_gpu.lock();
|
||||
qx_buf_offset = extra_src0->offset;
|
||||
qx_buf_offset = extra_src0->offset + src0->view_offs;
|
||||
GGML_ASSERT(d_Qx != nullptr);
|
||||
}
|
||||
if (!src1_uma) {
|
||||
d_Qy = extra_src1->buffer_gpu.lock();
|
||||
qy_buf_offset = extra_src1->offset;
|
||||
qy_buf_offset = extra_src1->offset + src1->view_offs;
|
||||
GGML_ASSERT(d_Qy != nullptr);
|
||||
}
|
||||
if (!ids_uma) {
|
||||
d_ids = extra_ids->buffer_gpu.lock();
|
||||
ids_buf_offset = extra_ids->offset;
|
||||
ids_buf_offset = extra_ids->offset + ids->view_offs;
|
||||
GGML_ASSERT(d_ids != nullptr);
|
||||
}
|
||||
if (qx_needs_dequant) {
|
||||
|
|
@ -3636,7 +3633,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
|
|||
const uint64_t d_sz = sizeof(float) * d_ne;
|
||||
|
||||
vk_buffer d_D = extra->buffer_gpu.lock();
|
||||
const uint64_t d_buf_offset = extra->offset;
|
||||
const uint64_t d_buf_offset = extra->offset + dst->view_offs;
|
||||
GGML_ASSERT(d_D != nullptr);
|
||||
vk_buffer d_X;
|
||||
uint64_t x_buf_offset = 0;
|
||||
|
|
@ -3644,17 +3641,17 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
|
|||
uint64_t y_buf_offset = 0;
|
||||
if(!src0_uma) {
|
||||
d_Qx = extra_src0->buffer_gpu.lock();
|
||||
qx_buf_offset = extra_src0->offset;
|
||||
qx_buf_offset = extra_src0->offset + src0->view_offs;
|
||||
GGML_ASSERT(d_Qx != nullptr);
|
||||
}
|
||||
if(!src1_uma) {
|
||||
d_Qy = extra_src1->buffer_gpu.lock();
|
||||
qy_buf_offset = extra_src1->offset;
|
||||
qy_buf_offset = extra_src1->offset + src1->view_offs;
|
||||
GGML_ASSERT(d_Qy != nullptr);
|
||||
}
|
||||
if(!ids_uma) {
|
||||
d_ids = extra_ids->buffer_gpu.lock();
|
||||
ids_buf_offset = extra_ids->offset;
|
||||
ids_buf_offset = extra_ids->offset + ids->view_offs;
|
||||
GGML_ASSERT(d_ids != nullptr);
|
||||
}
|
||||
if (qx_needs_dequant) {
|
||||
|
|
@ -3769,9 +3766,9 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx
|
|||
ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
|
||||
|
||||
const vk_buffer src_buf = extra_src0->buffer_gpu.lock();
|
||||
const uint64_t src_offset = extra_src0->offset;
|
||||
const uint64_t src_offset = extra_src0->offset + src0->view_offs;
|
||||
vk_buffer dst_buf = extra->buffer_gpu.lock();
|
||||
const uint64_t dst_offset = extra->offset;
|
||||
const uint64_t dst_offset = extra->offset + dst->view_offs;
|
||||
|
||||
std::vector<vk::BufferCopy> copies;
|
||||
|
||||
|
|
@ -4062,21 +4059,21 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
|
|||
}
|
||||
|
||||
GGML_ASSERT(d_D != nullptr);
|
||||
uint64_t d_buf_offset = (extra->offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
|
||||
uint64_t d_buf_offset = ((extra->offset + dst->view_offs) / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
|
||||
GGML_ASSERT(d_buf_offset == extra->offset || op == GGML_OP_CPY); // NOLINT
|
||||
if(!src0_uma) {
|
||||
d_X = extra_src0->buffer_gpu.lock();
|
||||
x_buf_offset = extra_src0->offset;
|
||||
x_buf_offset = extra_src0->offset + src0->view_offs;
|
||||
GGML_ASSERT(d_X != nullptr);
|
||||
}
|
||||
if (use_src1 && !src1_uma) {
|
||||
d_Y = extra_src1->buffer_gpu.lock();
|
||||
y_buf_offset = extra_src1->offset;
|
||||
y_buf_offset = extra_src1->offset + src1->view_offs;
|
||||
GGML_ASSERT(d_Y != nullptr);
|
||||
}
|
||||
if (use_src2 && !src2_uma) {
|
||||
d_Z = extra_src2->buffer_gpu.lock();
|
||||
z_buf_offset = extra_src2->offset;
|
||||
z_buf_offset = extra_src2->offset + src2->view_offs;
|
||||
GGML_ASSERT(d_Z != nullptr);
|
||||
}
|
||||
|
||||
|
|
@ -4336,7 +4333,7 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, cons
|
|||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
|
||||
const uint32_t src0_type_size = ggml_type_size(src0->type);
|
||||
const uint32_t dst_type_size = ggml_type_size(dst->type);
|
||||
const uint32_t d_offset = (extra->offset % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
|
||||
const uint32_t d_offset = ((extra->offset + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
|
||||
|
||||
ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, {
|
||||
(uint32_t)ggml_nelements(src0),
|
||||
|
|
@ -5569,6 +5566,13 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
|
|||
const ggml_tensor * src2 = node->src[2];
|
||||
|
||||
switch (node->op) {
|
||||
// Return on empty ops to avoid generating a compute_ctx and setting exit_tensor
|
||||
case GGML_OP_RESHAPE:
|
||||
case GGML_OP_VIEW:
|
||||
case GGML_OP_PERMUTE:
|
||||
case GGML_OP_TRANSPOSE:
|
||||
case GGML_OP_NONE:
|
||||
return;
|
||||
case GGML_OP_UNARY:
|
||||
switch (ggml_get_unary_op(node)) {
|
||||
case GGML_UNARY_OP_SILU:
|
||||
|
|
@ -5590,10 +5594,6 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
|
|||
case GGML_OP_CPY:
|
||||
case GGML_OP_CONT:
|
||||
case GGML_OP_DUP:
|
||||
case GGML_OP_RESHAPE:
|
||||
case GGML_OP_VIEW:
|
||||
case GGML_OP_PERMUTE:
|
||||
case GGML_OP_TRANSPOSE:
|
||||
case GGML_OP_NORM:
|
||||
case GGML_OP_RMS_NORM:
|
||||
case GGML_OP_DIAG_MASK_INF:
|
||||
|
|
@ -5601,7 +5601,6 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
|
|||
case GGML_OP_ROPE:
|
||||
case GGML_OP_MUL_MAT:
|
||||
case GGML_OP_MUL_MAT_ID:
|
||||
case GGML_OP_NONE:
|
||||
case GGML_OP_ARGSORT:
|
||||
case GGML_OP_SUM_ROWS:
|
||||
break;
|
||||
|
|
@ -5654,12 +5653,6 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
|
|||
case GGML_OP_DUP:
|
||||
ggml_vk_cpy(ctx, ctx->compute_ctx, src0, node);
|
||||
|
||||
break;
|
||||
case GGML_OP_RESHAPE:
|
||||
case GGML_OP_VIEW:
|
||||
case GGML_OP_PERMUTE:
|
||||
case GGML_OP_TRANSPOSE:
|
||||
case GGML_OP_NONE:
|
||||
break;
|
||||
case GGML_OP_NORM:
|
||||
ggml_vk_norm(ctx, ctx->compute_ctx, src0, node);
|
||||
|
|
@ -5712,7 +5705,6 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
|
|||
return;
|
||||
}
|
||||
|
||||
extra->ready = true;
|
||||
extra->ctx_idx = ctx->compute_ctx->idx;
|
||||
|
||||
#ifdef GGML_VULKAN_CHECK_RESULTS
|
||||
|
|
@ -5796,8 +5788,6 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
|
|||
ggml_vk_check_results_0(ctx, params, tensor);
|
||||
#endif
|
||||
|
||||
GGML_ASSERT(extra->ready);
|
||||
|
||||
vk_context& subctx = ctx->gc.contexts[extra->ctx_idx];
|
||||
|
||||
// Only run if ctx hasn't been submitted yet
|
||||
|
|
@ -5822,8 +5812,6 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
|
|||
subctx.out_memcpys.clear();
|
||||
}
|
||||
|
||||
extra->ready = false;
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
|
|
@ -5943,8 +5931,10 @@ struct ggml_backend_vk_buffer_context {
|
|||
|
||||
~ggml_backend_vk_buffer_context() {
|
||||
ggml_vk_destroy_buffer(dev_buffer);
|
||||
if (temp_tensor_extras != nullptr) {
|
||||
delete[] temp_tensor_extras;
|
||||
}
|
||||
}
|
||||
|
||||
ggml_tensor_extra_gpu * ggml_vk_alloc_temp_tensor_extra() {
|
||||
if (temp_tensor_extras == nullptr) {
|
||||
|
|
@ -5990,19 +5980,17 @@ GGML_CALL static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t b
|
|||
#endif
|
||||
ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
|
||||
|
||||
ggml_tensor_extra_gpu * extra = ctx->ggml_vk_alloc_temp_tensor_extra();
|
||||
if (tensor->view_src != nullptr && tensor->view_src->extra != nullptr) {
|
||||
if (tensor->view_src != nullptr) {
|
||||
GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
|
||||
ggml_tensor_extra_gpu * extra_view = (ggml_tensor_extra_gpu *) tensor->view_src->extra;
|
||||
extra->buffer_gpu = extra_view->buffer_gpu;
|
||||
extra->offset = extra_view->offset + tensor->view_offs;
|
||||
GGML_ASSERT(tensor->view_src->extra != nullptr);
|
||||
tensor->extra = tensor->view_src->extra;
|
||||
} else {
|
||||
ggml_tensor_extra_gpu * extra = ctx->ggml_vk_alloc_temp_tensor_extra();
|
||||
extra->buffer_gpu = ctx->dev_buffer;
|
||||
extra->offset = (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base;
|
||||
}
|
||||
|
||||
tensor->extra = extra;
|
||||
}
|
||||
}
|
||||
|
||||
GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
|
||||
#ifdef GGML_VULKAN_DEBUG
|
||||
|
|
@ -6014,7 +6002,7 @@ GGML_CALL static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t bu
|
|||
|
||||
vk_buffer buf = extra->buffer_gpu.lock();
|
||||
|
||||
ggml_vk_buffer_write(ctx->ctx, buf, extra->offset + offset, data, size);
|
||||
ggml_vk_buffer_write(ctx->ctx, buf, extra->offset + tensor->view_offs + offset, data, size);
|
||||
}
|
||||
|
||||
GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
|
||||
|
|
@ -6027,7 +6015,7 @@ GGML_CALL static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t bu
|
|||
|
||||
vk_buffer buf = extra->buffer_gpu.lock();
|
||||
|
||||
ggml_vk_buffer_read(ctx->ctx, buf, extra->offset + offset, data, size);
|
||||
ggml_vk_buffer_read(ctx->ctx, buf, extra->offset + tensor->view_offs + offset, data, size);
|
||||
}
|
||||
|
||||
GGML_CALL static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
|
||||
|
|
@ -6038,7 +6026,7 @@ GGML_CALL static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t bu
|
|||
vk_buffer src_buf = src_extra->buffer_gpu.lock();
|
||||
vk_buffer dst_buf = dst_extra->buffer_gpu.lock();
|
||||
|
||||
ggml_vk_buffer_copy(dst_buf, dst_extra->offset, src_buf, src_extra->offset, ggml_nbytes(src));
|
||||
ggml_vk_buffer_copy(dst_buf, dst_extra->offset + dst->view_offs, src_buf, src_extra->offset + src->view_offs, ggml_nbytes(src));
|
||||
|
||||
return true;
|
||||
}
|
||||
|
|
@ -6264,7 +6252,7 @@ GGML_CALL static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, g
|
|||
|
||||
vk_buffer buf = extra->buffer_gpu.lock();
|
||||
|
||||
ggml_vk_buffer_write_async(ctx, ctx->transfer_ctx, buf, extra->offset + offset, data, size);
|
||||
ggml_vk_buffer_write_async(ctx, ctx->transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size);
|
||||
}
|
||||
|
||||
GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
|
||||
|
|
@ -6284,7 +6272,7 @@ GGML_CALL static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, c
|
|||
|
||||
vk_buffer buf = extra->buffer_gpu.lock();
|
||||
|
||||
ggml_vk_buffer_read_async(ctx, ctx->transfer_ctx, buf, extra->offset + offset, data, size);
|
||||
ggml_vk_buffer_read_async(ctx, ctx->transfer_ctx, buf, extra->offset + tensor->view_offs + offset, data, size);
|
||||
}
|
||||
|
||||
GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
|
||||
|
|
@ -6305,7 +6293,7 @@ GGML_CALL static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, c
|
|||
vk_buffer src_buf = src_extra->buffer_gpu.lock();
|
||||
vk_buffer dst_buf = dst_extra->buffer_gpu.lock();
|
||||
|
||||
ggml_vk_buffer_copy_async(ctx->transfer_ctx, dst_buf, dst_extra->offset, src_buf, src_extra->offset, ggml_nbytes(src));
|
||||
ggml_vk_buffer_copy_async(ctx->transfer_ctx, dst_buf, dst_extra->offset + dst->view_offs, src_buf, src_extra->offset + src->view_offs, ggml_nbytes(src));
|
||||
return true;
|
||||
}
|
||||
|
||||
|
|
@ -6478,11 +6466,7 @@ GGML_CALL static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const
|
|||
// return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
|
||||
// } break;
|
||||
case GGML_OP_ROPE:
|
||||
{
|
||||
const int mode = ((const int32_t *) op->op_params)[2];
|
||||
|
||||
return true;
|
||||
} break;
|
||||
case GGML_OP_NONE:
|
||||
case GGML_OP_RESHAPE:
|
||||
case GGML_OP_VIEW:
|
||||
|
|
@ -6725,7 +6709,7 @@ static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tenso
|
|||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
||||
|
||||
vk_buffer buffer_gpu = extra->buffer_gpu.lock();
|
||||
ggml_vk_buffer_read(ctx, buffer_gpu, extra->offset, tensor_data, tensor_size);
|
||||
ggml_vk_buffer_read(ctx, buffer_gpu, extra->offset + tensor->view_offs, tensor_data, tensor_size);
|
||||
}
|
||||
|
||||
std::cerr << "TENSOR CHECK " << name << " (" << tensor->name << "): " << ggml_op_name(tensor->op) << std::endl;
|
||||
|
|
@ -6809,7 +6793,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
|
|||
} else if (ggml_backend_buffer_is_vk(src0->buffer)) {
|
||||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src0->extra;
|
||||
vk_buffer buffer_gpu = extra->buffer_gpu.lock();
|
||||
uint64_t offset = extra->offset;
|
||||
uint64_t offset = extra->offset + src0->view_offs;
|
||||
if (!ggml_is_contiguous(src0) && ggml_vk_dim01_contiguous(src0)) {
|
||||
for (int i3 = 0; i3 < src0->ne[3]; i3++) {
|
||||
for (int i2 = 0; i2 < src0->ne[2]; i2++) {
|
||||
|
|
@ -6851,7 +6835,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
|
|||
} else if (ggml_backend_buffer_is_vk(src1->buffer)) {
|
||||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src1->extra;
|
||||
vk_buffer buffer_gpu = extra->buffer_gpu.lock();
|
||||
uint64_t offset = extra->offset;
|
||||
uint64_t offset = extra->offset + src1->view_offs;
|
||||
if (!ggml_is_contiguous(src1) && ggml_vk_dim01_contiguous(src1)) {
|
||||
for (int i3 = 0; i3 < src1->ne[3]; i3++) {
|
||||
for (int i2 = 0; i2 < src1->ne[2]; i2++) {
|
||||
|
|
@ -6909,7 +6893,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
|
|||
} else if (ggml_backend_buffer_is_vk(src2->buffer)) {
|
||||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src2->extra;
|
||||
vk_buffer buffer_gpu = extra->buffer_gpu.lock();
|
||||
uint64_t offset = extra->offset;
|
||||
uint64_t offset = extra->offset + src2->view_offs;
|
||||
if (!ggml_is_contiguous(src2) && ggml_vk_dim01_contiguous(src2)) {
|
||||
for (int i3 = 0; i3 < src2->ne[3]; i3++) {
|
||||
for (int i2 = 0; i2 < src2->ne[2]; i2++) {
|
||||
|
|
@ -7092,11 +7076,11 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
|
|||
ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
|
||||
|
||||
vk_buffer buffer_gpu = extra->buffer_gpu.lock();
|
||||
if (extra->offset + tensor_size >= buffer_gpu->size) {
|
||||
tensor_size = buffer_gpu->size - (extra->offset);
|
||||
if (extra->offset + tensor->view_offs + tensor_size >= buffer_gpu->size) {
|
||||
tensor_size = buffer_gpu->size - (extra->offset + tensor->view_offs);
|
||||
}
|
||||
|
||||
ggml_vk_buffer_read(ctx, buffer_gpu, extra->offset, tensor_data, tensor_size);
|
||||
ggml_vk_buffer_read(ctx, buffer_gpu, extra->offset + tensor->view_offs, tensor_data, tensor_size);
|
||||
}
|
||||
|
||||
float first_error_result = -1.0f;
|
||||
|
|
|
|||
|
|
@ -415,6 +415,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
|
|||
MODEL_TENSOR.TOKEN_EMBD,
|
||||
MODEL_TENSOR.TOKEN_EMBD_NORM,
|
||||
MODEL_TENSOR.TOKEN_TYPES,
|
||||
MODEL_TENSOR.ATTN_NORM_2,
|
||||
MODEL_TENSOR.ATTN_OUT_NORM,
|
||||
MODEL_TENSOR.ATTN_Q,
|
||||
MODEL_TENSOR.ATTN_Q_NORM,
|
||||
|
|
|
|||
|
|
@ -5,6 +5,7 @@ import os
|
|||
import shutil
|
||||
import struct
|
||||
import tempfile
|
||||
from dataclasses import dataclass
|
||||
from enum import Enum, auto
|
||||
from io import BufferedWriter
|
||||
from typing import IO, Any, Sequence, Mapping
|
||||
|
|
@ -30,17 +31,36 @@ from .quants import quant_shape_from_byte_shape
|
|||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
@dataclass
|
||||
class TensorInfo:
|
||||
shape: Sequence[int]
|
||||
dtype: GGMLQuantizationType
|
||||
nbytes: int
|
||||
tensor: np.ndarray[Any, Any] | None = None
|
||||
|
||||
|
||||
@dataclass
|
||||
class GGUFValue:
|
||||
value: Any
|
||||
type: GGUFValueType
|
||||
|
||||
|
||||
class WriterState(Enum):
|
||||
NO_FILE = auto()
|
||||
EMPTY = auto()
|
||||
HEADER = auto()
|
||||
KV_DATA = auto()
|
||||
TI_DATA = auto()
|
||||
WEIGHTS = auto()
|
||||
|
||||
|
||||
class GGUFWriter:
|
||||
fout: BufferedWriter
|
||||
fout: BufferedWriter | None
|
||||
path: os.PathLike[str] | str | None
|
||||
temp_file: tempfile.SpooledTemporaryFile[bytes] | None
|
||||
tensors: list[np.ndarray[Any, Any]]
|
||||
tensors: dict[str, TensorInfo]
|
||||
kv_data: dict[str, GGUFValue]
|
||||
state: WriterState
|
||||
_simple_value_packing = {
|
||||
GGUFValueType.UINT8: "B",
|
||||
GGUFValueType.INT8: "b",
|
||||
|
|
@ -56,141 +76,140 @@ class GGUFWriter:
|
|||
}
|
||||
|
||||
def __init__(
|
||||
self, path: os.PathLike[str] | str, arch: str, use_temp_file: bool = True,
|
||||
self, path: os.PathLike[str] | str | None, arch: str, use_temp_file: bool = False,
|
||||
endianess: GGUFEndian = GGUFEndian.LITTLE,
|
||||
):
|
||||
self.fout = open(path, "wb")
|
||||
self.fout = None
|
||||
self.path = path
|
||||
self.arch = arch
|
||||
self.endianess = endianess
|
||||
self.offset_tensor = 0
|
||||
self.data_alignment = GGUF_DEFAULT_ALIGNMENT
|
||||
self.kv_data = bytearray()
|
||||
self.kv_data_count = 0
|
||||
self.ti_data = bytearray()
|
||||
self.ti_data_count = 0
|
||||
self.ti_names = set()
|
||||
self.use_temp_file = use_temp_file
|
||||
self.temp_file = None
|
||||
self.tensors = []
|
||||
self.tensors = dict()
|
||||
self.kv_data = dict()
|
||||
logger.info("gguf: This GGUF file is for {0} Endian only".format(
|
||||
"Big" if self.endianess == GGUFEndian.BIG else "Little",
|
||||
))
|
||||
self.state = WriterState.EMPTY
|
||||
self.state = WriterState.NO_FILE
|
||||
|
||||
self.add_architecture()
|
||||
|
||||
def write_header_to_file(self) -> None:
|
||||
def open_output_file(self, path: os.PathLike[str] | str | None = None) -> None:
|
||||
if self.state is WriterState.EMPTY and self.fout is not None and (path is None or path == self.path):
|
||||
# allow calling this multiple times as long as the path is the same
|
||||
return
|
||||
if self.state is not WriterState.NO_FILE:
|
||||
raise ValueError(f'Expected output file to be not yet opened, got {self.state}')
|
||||
|
||||
if path is not None:
|
||||
self.path = path
|
||||
|
||||
if self.path is not None:
|
||||
if self.fout is not None:
|
||||
self.fout.close()
|
||||
self.fout = open(self.path, "wb")
|
||||
self.state = WriterState.EMPTY
|
||||
|
||||
def write_header_to_file(self, path: os.PathLike[str] | str | None = None) -> None:
|
||||
self.open_output_file(path)
|
||||
|
||||
if self.state is not WriterState.EMPTY:
|
||||
raise ValueError(f'Expected output file to be empty, got {self.state}')
|
||||
|
||||
self._write_packed("<I", GGUF_MAGIC, skip_pack_prefix = True)
|
||||
self._write_packed("I", GGUF_VERSION)
|
||||
self._write_packed("Q", self.ti_data_count)
|
||||
self._write_packed("Q", self.kv_data_count)
|
||||
self._write_packed("Q", len(self.tensors))
|
||||
self._write_packed("Q", len(self.kv_data))
|
||||
self.flush()
|
||||
self.state = WriterState.HEADER
|
||||
|
||||
def write_kv_data_to_file(self) -> None:
|
||||
if self.state is not WriterState.HEADER:
|
||||
raise ValueError(f'Expected output file to contain the header, got {self.state}')
|
||||
assert self.fout is not None
|
||||
|
||||
self.fout.write(self.kv_data)
|
||||
kv_data = bytearray()
|
||||
|
||||
for key, val in self.kv_data.items():
|
||||
kv_data += self._pack_val(key, GGUFValueType.STRING, add_vtype=False)
|
||||
kv_data += self._pack_val(val.value, val.type, add_vtype=True)
|
||||
|
||||
self.fout.write(kv_data)
|
||||
self.flush()
|
||||
self.state = WriterState.KV_DATA
|
||||
|
||||
def write_ti_data_to_file(self) -> None:
|
||||
if self.state is not WriterState.KV_DATA:
|
||||
raise ValueError(f'Expected output file to contain KV data, got {self.state}')
|
||||
assert self.fout is not None
|
||||
|
||||
self.fout.write(self.ti_data)
|
||||
ti_data = bytearray()
|
||||
offset_tensor = 0
|
||||
|
||||
for name, ti in self.tensors.items():
|
||||
ti_data += self._pack_val(name, GGUFValueType.STRING, add_vtype=False)
|
||||
n_dims = len(ti.shape)
|
||||
ti_data += self._pack("I", n_dims)
|
||||
for i in range(n_dims):
|
||||
ti_data += self._pack("Q", ti.shape[n_dims - 1 - i])
|
||||
ti_data += self._pack("I", ti.dtype)
|
||||
ti_data += self._pack("Q", offset_tensor)
|
||||
offset_tensor += GGUFWriter.ggml_pad(ti.nbytes, self.data_alignment)
|
||||
|
||||
self.fout.write(ti_data)
|
||||
self.flush()
|
||||
self.state = WriterState.TI_DATA
|
||||
|
||||
def add_key(self, key: str) -> None:
|
||||
self.add_val(key, GGUFValueType.STRING, add_vtype=False)
|
||||
def add_key_value(self, key: str, val: Any, vtype: GGUFValueType) -> None:
|
||||
if key in self.kv_data:
|
||||
raise ValueError(f'Duplicated key name {key!r}')
|
||||
|
||||
self.kv_data[key] = GGUFValue(value=val, type=vtype)
|
||||
|
||||
def add_uint8(self, key: str, val: int) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.UINT8)
|
||||
self.add_key_value(key,val, GGUFValueType.UINT8)
|
||||
|
||||
def add_int8(self, key: str, val: int) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.INT8)
|
||||
self.add_key_value(key, val, GGUFValueType.INT8)
|
||||
|
||||
def add_uint16(self, key: str, val: int) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.UINT16)
|
||||
self.add_key_value(key, val, GGUFValueType.UINT16)
|
||||
|
||||
def add_int16(self, key: str, val: int) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.INT16)
|
||||
self.add_key_value(key, val, GGUFValueType.INT16)
|
||||
|
||||
def add_uint32(self, key: str, val: int) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.UINT32)
|
||||
self.add_key_value(key, val, GGUFValueType.UINT32)
|
||||
|
||||
def add_int32(self, key: str, val: int) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.INT32)
|
||||
self.add_key_value(key, val, GGUFValueType.INT32)
|
||||
|
||||
def add_float32(self, key: str, val: float) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.FLOAT32)
|
||||
self.add_key_value(key, val, GGUFValueType.FLOAT32)
|
||||
|
||||
def add_uint64(self, key: str, val: int) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.UINT64)
|
||||
self.add_key_value(key, val, GGUFValueType.UINT64)
|
||||
|
||||
def add_int64(self, key: str, val: int) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.INT64)
|
||||
self.add_key_value(key, val, GGUFValueType.INT64)
|
||||
|
||||
def add_float64(self, key: str, val: float) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.FLOAT64)
|
||||
self.add_key_value(key, val, GGUFValueType.FLOAT64)
|
||||
|
||||
def add_bool(self, key: str, val: bool) -> None:
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.BOOL)
|
||||
self.add_key_value(key, val, GGUFValueType.BOOL)
|
||||
|
||||
def add_string(self, key: str, val: str) -> None:
|
||||
if not val:
|
||||
return
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.STRING)
|
||||
self.add_key_value(key, val, GGUFValueType.STRING)
|
||||
|
||||
def add_array(self, key: str, val: Sequence[Any]) -> None:
|
||||
if not isinstance(val, Sequence):
|
||||
raise ValueError("Value must be a sequence for array type")
|
||||
|
||||
self.add_key(key)
|
||||
self.add_val(val, GGUFValueType.ARRAY)
|
||||
|
||||
def add_val(self, val: Any, vtype: GGUFValueType | None = None, add_vtype: bool = True) -> None:
|
||||
if vtype is None:
|
||||
vtype = GGUFValueType.get_type(val)
|
||||
|
||||
if add_vtype:
|
||||
self.kv_data += self._pack("I", vtype)
|
||||
self.kv_data_count += 1
|
||||
|
||||
pack_fmt = self._simple_value_packing.get(vtype)
|
||||
if pack_fmt is not None:
|
||||
self.kv_data += self._pack(pack_fmt, val, skip_pack_prefix = vtype == GGUFValueType.BOOL)
|
||||
elif vtype == GGUFValueType.STRING:
|
||||
encoded_val = val.encode("utf-8") if isinstance(val, str) else val
|
||||
self.kv_data += self._pack("Q", len(encoded_val))
|
||||
self.kv_data += encoded_val
|
||||
elif vtype == GGUFValueType.ARRAY and isinstance(val, Sequence) and val:
|
||||
ltype = GGUFValueType.get_type(val[0])
|
||||
if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]):
|
||||
raise ValueError("All items in a GGUF array should be of the same type")
|
||||
self.kv_data += self._pack("I", ltype)
|
||||
self.kv_data += self._pack("Q", len(val))
|
||||
for item in val:
|
||||
self.add_val(item, add_vtype=False)
|
||||
else:
|
||||
raise ValueError("Invalid GGUF metadata value type or value")
|
||||
self.add_key_value(key, val, GGUFValueType.ARRAY)
|
||||
|
||||
@staticmethod
|
||||
def ggml_pad(x: int, n: int) -> int:
|
||||
|
|
@ -200,16 +219,12 @@ class GGUFWriter:
|
|||
self, name: str, tensor_shape: Sequence[int], tensor_dtype: np.dtype,
|
||||
tensor_nbytes: int, raw_dtype: GGMLQuantizationType | None = None,
|
||||
) -> None:
|
||||
if self.state is not WriterState.EMPTY:
|
||||
raise ValueError(f'Expected output file to be empty, got {self.state}')
|
||||
if self.state is not WriterState.NO_FILE:
|
||||
raise ValueError(f'Expected output file to be not yet opened, got {self.state}')
|
||||
|
||||
if name in self.ti_names:
|
||||
raise ValueError(f'Duplicated tensor name {name}')
|
||||
self.ti_names.add(name)
|
||||
if name in self.tensors:
|
||||
raise ValueError(f'Duplicated tensor name {name!r}')
|
||||
|
||||
encoded_name = name.encode("utf-8")
|
||||
self.ti_data += self._pack("Q", len(encoded_name))
|
||||
self.ti_data += encoded_name
|
||||
if raw_dtype is None:
|
||||
if tensor_dtype == np.float16:
|
||||
dtype = GGMLQuantizationType.F16
|
||||
|
|
@ -231,14 +246,8 @@ class GGUFWriter:
|
|||
dtype = raw_dtype
|
||||
if tensor_dtype == np.uint8:
|
||||
tensor_shape = quant_shape_from_byte_shape(tensor_shape, raw_dtype)
|
||||
n_dims = len(tensor_shape)
|
||||
self.ti_data += self._pack("I", n_dims)
|
||||
for i in range(n_dims):
|
||||
self.ti_data += self._pack("Q", tensor_shape[n_dims - 1 - i])
|
||||
self.ti_data += self._pack("I", dtype)
|
||||
self.ti_data += self._pack("Q", self.offset_tensor)
|
||||
self.offset_tensor += GGUFWriter.ggml_pad(tensor_nbytes, self.data_alignment)
|
||||
self.ti_data_count += 1
|
||||
|
||||
self.tensors[name] = TensorInfo(shape=tensor_shape, dtype=dtype, nbytes=tensor_nbytes)
|
||||
|
||||
def add_tensor(
|
||||
self, name: str, tensor: np.ndarray[Any, Any], raw_shape: Sequence[int] | None = None,
|
||||
|
|
@ -255,7 +264,7 @@ class GGUFWriter:
|
|||
self.add_tensor_info(name, shape, tensor.dtype, tensor.nbytes, raw_dtype=raw_dtype)
|
||||
|
||||
if self.temp_file is None:
|
||||
self.tensors.append(tensor)
|
||||
self.tensors[name].tensor = tensor
|
||||
return
|
||||
|
||||
tensor.tofile(self.temp_file)
|
||||
|
|
@ -267,8 +276,9 @@ class GGUFWriter:
|
|||
fp.write(bytes([0] * pad))
|
||||
|
||||
def write_tensor_data(self, tensor: np.ndarray[Any, Any]) -> None:
|
||||
if self.state is not WriterState.TI_DATA:
|
||||
raise ValueError(f'Expected output file to contain tensor info, got {self.state}')
|
||||
if self.state is not WriterState.TI_DATA and self.state is not WriterState.WEIGHTS:
|
||||
raise ValueError(f'Expected output file to contain tensor info or weights, got {self.state}')
|
||||
assert self.fout is not None
|
||||
|
||||
if self.endianess == GGUFEndian.BIG:
|
||||
tensor.byteswap(inplace=True)
|
||||
|
|
@ -276,50 +286,51 @@ class GGUFWriter:
|
|||
tensor.tofile(self.fout)
|
||||
self.write_padding(self.fout, tensor.nbytes)
|
||||
|
||||
self.state = WriterState.WEIGHTS
|
||||
|
||||
def write_tensors_to_file(self, *, progress: bool = False) -> None:
|
||||
self.write_ti_data_to_file()
|
||||
|
||||
assert self.fout is not None
|
||||
|
||||
self.write_padding(self.fout, self.fout.tell())
|
||||
|
||||
if self.temp_file is None:
|
||||
self.tensors.reverse() # to pop from the "beginning" in constant time
|
||||
bar = None
|
||||
|
||||
if progress:
|
||||
from tqdm import tqdm
|
||||
|
||||
total_bytes = sum(t.nbytes for t in self.tensors)
|
||||
total_bytes = sum(t.nbytes for t in self.tensors.values())
|
||||
|
||||
bar = tqdm(desc="Writing", total=total_bytes, unit="byte", unit_scale=True)
|
||||
|
||||
while True:
|
||||
try:
|
||||
tensor = self.tensors.pop()
|
||||
except IndexError:
|
||||
break
|
||||
tensor.tofile(self.fout)
|
||||
bar.update(tensor.nbytes)
|
||||
self.write_padding(self.fout, tensor.nbytes)
|
||||
return
|
||||
while True:
|
||||
try:
|
||||
tensor = self.tensors.pop()
|
||||
except IndexError:
|
||||
break
|
||||
tensor.tofile(self.fout)
|
||||
self.write_padding(self.fout, tensor.nbytes)
|
||||
return
|
||||
|
||||
# relying on the fact that Python dicts preserve insertion order (since 3.7)
|
||||
for ti in self.tensors.values():
|
||||
assert ti.tensor is not None # can only iterate once over the tensors
|
||||
assert ti.tensor.nbytes == ti.nbytes
|
||||
ti.tensor.tofile(self.fout)
|
||||
if bar is not None:
|
||||
bar.update(ti.nbytes)
|
||||
self.write_padding(self.fout, ti.nbytes)
|
||||
ti.tensor = None
|
||||
else:
|
||||
self.temp_file.seek(0)
|
||||
|
||||
shutil.copyfileobj(self.temp_file, self.fout)
|
||||
self.flush()
|
||||
self.temp_file.close()
|
||||
|
||||
self.state = WriterState.WEIGHTS
|
||||
|
||||
def flush(self) -> None:
|
||||
assert self.fout is not None
|
||||
self.fout.flush()
|
||||
|
||||
def close(self) -> None:
|
||||
if self.fout is not None:
|
||||
self.fout.close()
|
||||
self.fout = None
|
||||
|
||||
def add_architecture(self) -> None:
|
||||
self.add_string(Keys.General.ARCHITECTURE, self.arch)
|
||||
|
|
@ -449,7 +460,7 @@ class GGUFWriter:
|
|||
def add_rope_scaling_factor(self, value: float) -> None:
|
||||
self.add_float32(Keys.Rope.SCALING_FACTOR.format(arch=self.arch), value)
|
||||
|
||||
def add_rope_scaling_attn_factors(self, value: Sequence[float]) -> None:
|
||||
def add_rope_scaling_attn_factors(self, value: float) -> None:
|
||||
self.add_float32(Keys.Rope.SCALING_ATTN_FACTOR.format(arch=self.arch), value)
|
||||
|
||||
def add_rope_scaling_orig_ctx_len(self, value: int) -> None:
|
||||
|
|
@ -571,5 +582,32 @@ class GGUFWriter:
|
|||
pack_prefix = '<' if self.endianess == GGUFEndian.LITTLE else '>'
|
||||
return struct.pack(f'{pack_prefix}{fmt}', value)
|
||||
|
||||
def _pack_val(self, val: Any, vtype: GGUFValueType, add_vtype: bool) -> bytes:
|
||||
kv_data = bytearray()
|
||||
|
||||
if add_vtype:
|
||||
kv_data += self._pack("I", vtype)
|
||||
|
||||
pack_fmt = self._simple_value_packing.get(vtype)
|
||||
if pack_fmt is not None:
|
||||
kv_data += self._pack(pack_fmt, val, skip_pack_prefix = vtype == GGUFValueType.BOOL)
|
||||
elif vtype == GGUFValueType.STRING:
|
||||
encoded_val = val.encode("utf-8") if isinstance(val, str) else val
|
||||
kv_data += self._pack("Q", len(encoded_val))
|
||||
kv_data += encoded_val
|
||||
elif vtype == GGUFValueType.ARRAY and isinstance(val, Sequence) and val:
|
||||
ltype = GGUFValueType.get_type(val[0])
|
||||
if not all(GGUFValueType.get_type(i) is ltype for i in val[1:]):
|
||||
raise ValueError("All items in a GGUF array should be of the same type")
|
||||
kv_data += self._pack("I", ltype)
|
||||
kv_data += self._pack("Q", len(val))
|
||||
for item in val:
|
||||
kv_data += self._pack_val(item, ltype, add_vtype=False)
|
||||
else:
|
||||
raise ValueError("Invalid GGUF metadata value type or value")
|
||||
|
||||
return kv_data
|
||||
|
||||
def _write_packed(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> None:
|
||||
assert self.fout is not None
|
||||
self.fout.write(self._pack(fmt, value, skip_pack_prefix))
|
||||
|
|
|
|||
|
|
@ -102,6 +102,7 @@ class TensorNameMap:
|
|||
# Attention norm 2
|
||||
MODEL_TENSOR.ATTN_NORM_2: (
|
||||
"transformer.h.{bid}.ln_attn", # falcon40b
|
||||
"encoder.layer.{bid}.layer_norm_1", # jina-v2-code
|
||||
),
|
||||
|
||||
# Attention query-key-value
|
||||
|
|
@ -311,6 +312,7 @@ class TensorNameMap:
|
|||
"model.layers.{bid}.mlp.c_proj", # starcoder2
|
||||
"encoder.layer.{bid}.mlp.wo", # jina-bert-v2
|
||||
"model.layers.{bid}.residual_mlp.w2", # arctic
|
||||
"encoder.layer.{bid}.mlp.down_layer", # jina-bert-v2
|
||||
),
|
||||
|
||||
MODEL_TENSOR.FFN_DOWN_EXP: (
|
||||
|
|
@ -350,6 +352,7 @@ class TensorNameMap:
|
|||
"encoder.layers.{bid}.norm2", # nomic-bert
|
||||
"transformer.decoder_layer.{bid}.rms_norm_3", # Grok
|
||||
"encoder.layer.{bid}.mlp.layernorm", # jina-bert-v2
|
||||
"encoder.layer.{bid}.layer_norm_2" # jina-v2-code
|
||||
),
|
||||
|
||||
MODEL_TENSOR.SSM_IN: (
|
||||
|
|
|
|||
|
|
@ -101,8 +101,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
|
|||
logger.debug(f'Copying {field.name}')
|
||||
|
||||
if val.value is not None:
|
||||
writer.add_key(field.name)
|
||||
writer.add_val(val.value, val.type)
|
||||
writer.add_key_value(field.name, val.value, val.type)
|
||||
|
||||
if gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata:
|
||||
logger.debug('Adding chat template(s)')
|
||||
|
|
@ -111,8 +110,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
|
|||
|
||||
for key, val in new_metadata.items():
|
||||
logger.debug(f'Adding {key}: "{val.value}" {val.description}')
|
||||
writer.add_key(key)
|
||||
writer.add_val(val.value, val.type)
|
||||
writer.add_key_value(key, val.value, val.type)
|
||||
|
||||
total_bytes = 0
|
||||
|
||||
|
|
|
|||
37
klite.embd
37
klite.embd
|
|
@ -3671,7 +3671,7 @@ Current version: 145
|
|||
const XTTS_ID = 1000;
|
||||
const ALLTALK_ID = 1001;
|
||||
const HD_RES_PX = 512;
|
||||
const NO_HD_RES_PX = 256;
|
||||
const NO_HD_RES_PX = 320;
|
||||
|
||||
//all configurable globals
|
||||
var perfdata = null; //if it's null, we are not connected
|
||||
|
|
@ -12691,6 +12691,29 @@ Current version: 145
|
|||
//original source https://github.com/kdavis-mozilla/vad.js Copyright (c) 2015, Kelly Daviss
|
||||
let VAD=function(t){for(var e in this.options={fftSize:512,bufferLen:512,voice_stop:function(){},voice_start:function(){},smoothingTimeConstant:.99,energy_offset:1e-8,energy_threshold_ratio_pos:2,energy_threshold_ratio_neg:.5,energy_integration:1,filter:[{f:200,v:0},{f:2e3,v:1}],source:null,context:null},t)t.hasOwnProperty(e)&&(this.options[e]=t[e]);if(!this.options.source)throw Error("The options must specify a MediaStreamAudioSourceNode.");this.options.context=this.options.source.context,this.hertzPerBin=this.options.context.sampleRate/this.options.fftSize,this.iterationFrequency=this.options.context.sampleRate/this.options.bufferLen,this.iterationPeriod=1/this.iterationFrequency,this.setFilter=function(t){this.filter=[];for(var e=0,i=this.options.fftSize/2;e<i;e++){this.filter[e]=0;for(var s=0,o=t.length;s<o;s++)if(e*this.hertzPerBin<t[s].f){this.filter[e]=t[s].v;break}}},this.setFilter(this.options.filter),this.ready={},this.vadState=!1,this.energy_offset=this.options.energy_offset,this.energy_threshold_pos=this.energy_offset*this.options.energy_threshold_ratio_pos,this.energy_threshold_neg=this.energy_offset*this.options.energy_threshold_ratio_neg,this.voiceTrend=0,this.voiceTrendMax=10,this.voiceTrendMin=-10,this.voiceTrendStart=5,this.voiceTrendEnd=-5,this.analyser=this.options.context.createAnalyser(),this.analyser.smoothingTimeConstant=this.options.smoothingTimeConstant,this.analyser.fftSize=this.options.fftSize,this.floatFrequencyData=new Float32Array(this.analyser.frequencyBinCount),this.floatFrequencyDataLinear=new Float32Array(this.floatFrequencyData.length),this.options.source.connect(this.analyser),this.scriptProcessorNode=this.options.context.createScriptProcessor(this.options.bufferLen,1,1),this.scriptProcessorNode.connect(this.options.context.destination);var i=this;this.scriptProcessorNode.onaudioprocess=function(t){i.analyser.getFloatFrequencyData(i.floatFrequencyData),i.update(),i.monitor()},this.options.source.connect(this.scriptProcessorNode),this.update=function(){for(var t=this.floatFrequencyData,e=0,i=t.length;e<i;e++)this.floatFrequencyDataLinear[e]=Math.pow(10,t[e]/10);this.ready={}},this.getEnergy=function(){if(this.ready.energy)return this.energy;for(var t=0,e=this.floatFrequencyDataLinear,i=0,s=e.length;i<s;i++)t+=this.filter[i]*e[i]*e[i];return this.energy=t,this.ready.energy=!0,t},this.monitor=function(){var t=this.getEnergy()-this.energy_offset;t>this.energy_threshold_pos?this.voiceTrend=this.voiceTrend+1>this.voiceTrendMax?this.voiceTrendMax:this.voiceTrend+1:t<-this.energy_threshold_neg?this.voiceTrend=this.voiceTrend-1<this.voiceTrendMin?this.voiceTrendMin:this.voiceTrend-1:this.voiceTrend>0?this.voiceTrend--:this.voiceTrend<0&&this.voiceTrend++;var e=!1,i=!1;this.voiceTrend>this.voiceTrendStart?e=!0:this.voiceTrend<this.voiceTrendEnd&&(i=!0);var s=t*this.iterationPeriod*this.options.energy_integration;return s>0||!i?this.energy_offset+=s:this.energy_offset+=10*s,this.energy_offset=this.energy_offset<0?0:this.energy_offset,this.energy_threshold_pos=this.energy_offset*this.options.energy_threshold_ratio_pos,this.energy_threshold_neg=this.energy_offset*this.options.energy_threshold_ratio_neg,e&&!this.vadState&&(this.vadState=!0,this.options.voice_start()),i&&this.vadState&&(this.vadState=!1,this.options.voice_stop()),t}};
|
||||
|
||||
//resample audio freq (to 16khz)
|
||||
function resampleAudioBuffer(buffer, resampledRate, callback) {
|
||||
const originalSampleRate = buffer.sampleRate;
|
||||
const channels = buffer.numberOfChannels;
|
||||
const length = buffer.length;
|
||||
const ratio = originalSampleRate / resampledRate;
|
||||
const newLength = Math.round(length / ratio);
|
||||
const audioContext = new (window.AudioContext || window.webkitAudioContext)();
|
||||
const resampledBuffer = audioContext.createBuffer(channels, newLength, resampledRate);
|
||||
const offlineContext = new OfflineAudioContext(channels, newLength, resampledRate);
|
||||
const source = offlineContext.createBufferSource();
|
||||
source.buffer = buffer;
|
||||
source.connect(offlineContext.destination);
|
||||
const startRendering = offlineContext.startRendering();
|
||||
startRendering.then((renderedBuffer) => {
|
||||
callback(renderedBuffer);
|
||||
}).catch((error) => {
|
||||
console.error('Resample Err:', error);
|
||||
});
|
||||
source.start();
|
||||
return resampledBuffer;
|
||||
}
|
||||
|
||||
let audioBufferToWavBlob = function (audioBuffer) {
|
||||
let writeWavString = function (view, offset, string) {
|
||||
for (let i = 0; i < string.length; i++) {
|
||||
|
|
@ -12738,7 +12761,7 @@ Current version: 145
|
|||
reader.onloadend = function() {
|
||||
let arrayBuffer = reader.result;
|
||||
window.AudioContext = window.AudioContext || window.webkitAudioContext;
|
||||
let audioContext = new AudioContext({ sampleRate: 16000 });
|
||||
let audioContext = new AudioContext();
|
||||
audioContext.decodeAudioData(arrayBuffer, (buffer)=>{
|
||||
onDone(buffer);
|
||||
},(err)=>
|
||||
|
|
@ -12778,20 +12801,24 @@ Current version: 145
|
|||
if(preaudioblobs.length<2)
|
||||
{
|
||||
audioBlobToDecodedAudioBuffer(completeRecording,(buffer)=>{
|
||||
let wavblob = audioBufferToWavBlob(buffer);
|
||||
resampleAudioBuffer(buffer,16000,(rsBuffer)=>{
|
||||
let wavblob = audioBufferToWavBlob(rsBuffer);
|
||||
audiodatareader.readAsDataURL(wavblob);
|
||||
});
|
||||
});
|
||||
} else {
|
||||
audioBlobToDecodedAudioBuffer(completeRecording,(buffer)=>{
|
||||
audioBlobToDecodedAudioBuffer(preaudioblobs[0],(buffer2)=>{
|
||||
audioBlobToDecodedAudioBuffer(preaudioblobs[1],(buffer3)=>{
|
||||
let prefix = concatenateAudioBuffers(buffer2,buffer3);
|
||||
let finalbuf = concatenateAudioBuffers(prefix,buffer);
|
||||
let wavblob = audioBufferToWavBlob(finalbuf);
|
||||
resampleAudioBuffer(finalbuf,16000,(rsBuffer)=>{
|
||||
let wavblob = audioBufferToWavBlob(rsBuffer);
|
||||
audiodatareader.readAsDataURL(wavblob);
|
||||
});
|
||||
});
|
||||
});
|
||||
});
|
||||
}
|
||||
|
||||
audiodatareader.onloadend = function () {
|
||||
|
|
@ -12825,7 +12852,7 @@ Current version: 145
|
|||
voicerecorder = new MediaRecorder(stream);
|
||||
voicerecorder.addEventListener('dataavailable', onRecordingReady);
|
||||
window.AudioContext = window.AudioContext || window.webkitAudioContext;
|
||||
let audioContext = new AudioContext({ sampleRate: 16000 });
|
||||
let audioContext = new AudioContext();
|
||||
let source = audioContext.createMediaStreamSource(stream);
|
||||
let options = {
|
||||
source: source,
|
||||
|
|
|
|||
35
llama.cpp
35
llama.cpp
|
|
@ -728,6 +728,7 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
|
|||
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
|
||||
{ LLM_TENSOR_TOKEN_EMBD_NORM, "token_embd_norm" },
|
||||
{ LLM_TENSOR_TOKEN_TYPES, "token_types" },
|
||||
{ LLM_TENSOR_ATTN_NORM_2, "blk.%d.attn_norm_2" },
|
||||
{ LLM_TENSOR_ATTN_OUT_NORM, "blk.%d.attn_output_norm" },
|
||||
{ LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
|
||||
{ LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" },
|
||||
|
|
@ -4715,8 +4716,7 @@ static void llm_load_vocab(
|
|||
LLAMA_LOG_WARN("%s: ************************************ \n", __func__);
|
||||
LLAMA_LOG_WARN("%s: \n", __func__);
|
||||
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
|
||||
} else if (
|
||||
tokenizer_pre == "default") {
|
||||
} else if (tokenizer_pre == "default") {
|
||||
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
|
||||
} else if (
|
||||
tokenizer_pre == "llama3" ||
|
||||
|
|
@ -4743,7 +4743,8 @@ static void llm_load_vocab(
|
|||
tokenizer_pre == "jina-es" ||
|
||||
tokenizer_pre == "jina-de" ||
|
||||
tokenizer_pre == "jina-v2-es" ||
|
||||
tokenizer_pre == "jina-v2-de") {
|
||||
tokenizer_pre == "jina-v2-de" ||
|
||||
tokenizer_pre == "jina-v2-code") {
|
||||
vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_GPT2;
|
||||
} else if (
|
||||
tokenizer_pre == "refact") {
|
||||
|
|
@ -5634,6 +5635,9 @@ static bool llm_load_tensors(
|
|||
layer.attn_out_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}); //output_norm
|
||||
layer.attn_out_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i), {n_embd});
|
||||
|
||||
layer.attn_norm_2 = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
|
||||
layer.attn_norm_2_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM_2, "bias", i), {n_embd}, llama_model_loader::TENSOR_NOT_REQUIRED);
|
||||
|
||||
layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
|
||||
layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
|
||||
|
||||
|
|
@ -8597,6 +8601,11 @@ struct llm_build_context {
|
|||
// attention layer norm
|
||||
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, cb, il);
|
||||
|
||||
if (model.layers[il].attn_norm_2 != nullptr) {
|
||||
cur = ggml_add(ctx0, cur, inpL); // re-add the layer input
|
||||
cur = llm_build_norm(ctx0, cur, hparams, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, cb, il);
|
||||
}
|
||||
|
||||
struct ggml_tensor * ffn_inp = cur;
|
||||
cb(ffn_inp, "ffn_inp", il);
|
||||
|
||||
|
|
@ -13940,7 +13949,7 @@ static std::pair<bool, const llama_grammar_element *> llama_grammar_match_char(
|
|||
const uint32_t chr) {
|
||||
|
||||
bool found = false;
|
||||
bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR;
|
||||
bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR || pos->type == LLAMA_GRETYPE_CHAR_ANY;
|
||||
|
||||
GGML_ASSERT(is_positive_char || pos->type == LLAMA_GRETYPE_CHAR_NOT); // NOLINT
|
||||
|
||||
|
|
@ -13949,6 +13958,10 @@ static std::pair<bool, const llama_grammar_element *> llama_grammar_match_char(
|
|||
// inclusive range, e.g. [a-z]
|
||||
found = found || (pos->value <= chr && chr <= pos[1].value);
|
||||
pos += 2;
|
||||
} else if (pos->type == LLAMA_GRETYPE_CHAR_ANY) {
|
||||
// Any character matches "."
|
||||
found = true;
|
||||
pos += 1;
|
||||
} else {
|
||||
// exact char match, e.g. [a] or "a"
|
||||
found = found || pos->value == chr;
|
||||
|
|
@ -13966,7 +13979,7 @@ static bool llama_grammar_match_partial_char(
|
|||
const llama_grammar_element * pos,
|
||||
const llama_partial_utf8 partial_utf8) {
|
||||
|
||||
bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR;
|
||||
bool is_positive_char = pos->type == LLAMA_GRETYPE_CHAR || pos->type == LLAMA_GRETYPE_CHAR_ANY;
|
||||
GGML_ASSERT(is_positive_char || pos->type == LLAMA_GRETYPE_CHAR_NOT);
|
||||
|
||||
uint32_t partial_value = partial_utf8.value;
|
||||
|
|
@ -13996,6 +14009,9 @@ static bool llama_grammar_match_partial_char(
|
|||
return is_positive_char;
|
||||
}
|
||||
pos += 2;
|
||||
} else if (pos->type == LLAMA_GRETYPE_CHAR_ANY) {
|
||||
// Any character matches "."
|
||||
return true;
|
||||
} else {
|
||||
// exact char match, e.g. [a] or "a"
|
||||
if (low <= pos->value && pos->value <= high) {
|
||||
|
|
@ -14056,6 +14072,7 @@ static void llama_grammar_advance_stack(
|
|||
}
|
||||
case LLAMA_GRETYPE_CHAR:
|
||||
case LLAMA_GRETYPE_CHAR_NOT:
|
||||
case LLAMA_GRETYPE_CHAR_ANY:
|
||||
if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
|
||||
// only add the stack if it's not a duplicate of one we already have
|
||||
new_stacks.emplace_back(stack);
|
||||
|
|
@ -15543,6 +15560,14 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
|
|||
if (imatrix_data) {
|
||||
LLAMA_LOG_INFO("================================ Have weights data with %d entries\n",int(imatrix_data->size()));
|
||||
qs.has_imatrix = true;
|
||||
// check imatrix for nans or infs
|
||||
for (const auto & kv : *imatrix_data) {
|
||||
for (float f : kv.second) {
|
||||
if (!std::isfinite(f)) {
|
||||
throw std::runtime_error(format("imatrix contains non-finite value %f\n", f));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
|
|
|
|||
3
llama.h
3
llama.h
|
|
@ -365,6 +365,9 @@ extern "C" {
|
|||
// modifies a preceding LLAMA_GRETYPE_CHAR or
|
||||
// LLAMA_GRETYPE_CHAR_RNG_UPPER to add an alternate char to match ([ab], [a-zA])
|
||||
LLAMA_GRETYPE_CHAR_ALT = 6,
|
||||
|
||||
// any character (.)
|
||||
LLAMA_GRETYPE_CHAR_ANY = 7,
|
||||
};
|
||||
|
||||
typedef struct llama_grammar_element {
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue