Skip to content

Commit

Permalink
Merge branch 'upstream' into concedo_experimental
Browse files Browse the repository at this point in the history 
# Conflicts:
#	.github/workflows/build.yml
#	README.md
#	examples/CMakeLists.txt
#	examples/batched/batched.cpp
#	examples/gritlm/gritlm.cpp
#	examples/llama.android/llama/build.gradle.kts
#	examples/main/README.md
#	examples/retrieval/retrieval.cpp
#	examples/server/CMakeLists.txt
#	examples/server/README.md
#	ggml/CMakeLists.txt
#	ggml/src/ggml-cpu/CMakeLists.txt
#	ggml/src/ggml.c
#	scripts/compare-commits.sh
#	scripts/sync-ggml.last
#	tests/CMakeLists.txt
#	tests/test-backend-ops.cpp
#	tests/test-chat-template.cpp
#	tests/test-sampling.cpp
  • Loading branch information
@LostRuins
LostRuins committed Dec 19, 2024
2 parents 6343604 + 9177484 commit ee486ba
Show file tree
Hide file tree
Showing 59 changed files with 20,502 additions and 13,156 deletions.
73 changes: 50 additions & 23 deletions common/arg.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -120,29 +120,33 @@ std::string common_arg::to_string() {
// utils
//

static void common_params_handle_model_default(common_params & params) {
if (!params.hf_repo.empty()) {
static void common_params_handle_model_default(
std::string & model,
std::string & model_url,
std::string & hf_repo,
std::string & hf_file) {
if (!hf_repo.empty()) {
// short-hand to avoid specifying --hf-file -> default it to --model
if (params.hf_file.empty()) {
if (params.model.empty()) {
if (hf_file.empty()) {
if (model.empty()) {
throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n");
}
params.hf_file = params.model;
} else if (params.model.empty()) {
hf_file = model;
} else if (model.empty()) {
// this is to avoid different repo having same file name, or same file name in different subdirs
std::string filename = params.hf_repo + "_" + params.hf_file;
std::string filename = hf_repo + "_" + hf_file;
// to make sure we don't have any slashes in the filename
string_replace_all(filename, "/", "_");
params.model = fs_get_cache_file(filename);
model = fs_get_cache_file(filename);
}
} else if (!params.model_url.empty()) {
if (params.model.empty()) {
auto f = string_split<std::string>(params.model_url, '#').front();
} else if (!model_url.empty()) {
if (model.empty()) {
auto f = string_split<std::string>(model_url, '#').front();
f = string_split<std::string>(f, '?').front();
params.model = fs_get_cache_file(string_split<std::string>(f, '/').back());
model = fs_get_cache_file(string_split<std::string>(f, '/').back());
}
} else if (params.model.empty()) {
params.model = DEFAULT_MODEL_PATH;
} else if (model.empty()) {
model = DEFAULT_MODEL_PATH;
}
}

Expand Down Expand Up @@ -277,7 +281,9 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
}

common_params_handle_model_default(params);
// TODO: refactor model params in a common struct
common_params_handle_model_default(params.model, params.model_url, params.hf_repo, params.hf_file);
common_params_handle_model_default(params.vocoder.model, params.vocoder.model_url, params.vocoder.hf_repo, params.vocoder.hf_file);

if (params.escape) {
string_process_escapes(params.prompt);
Expand Down Expand Up @@ -843,7 +849,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_sparam());
add_opt(common_arg(
{"--sampling-seq"}, "SEQUENCE",
{"--sampling-seq", "--sampler-seq"}, "SEQUENCE",
string_format("simplified sequence for samplers that will be used (default: %s)", sampler_type_chars.c_str()),
[](common_params & params, const std::string & value) {
params.sampling.samplers = common_sampler_types_from_chars(value);
Expand All @@ -856,13 +862,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.sampling.ignore_eos = true;
}
).set_sparam());
add_opt(common_arg(
{"--penalize-nl"},
string_format("penalize newline tokens (default: %s)", params.sampling.penalize_nl ? "true" : "false"),
[](common_params & params) {
params.sampling.penalize_nl = true;
}
).set_sparam());
add_opt(common_arg(
{"--temp"}, "N",
string_format("temperature (default: %.1f)", (double)params.sampling.temp),
Expand Down Expand Up @@ -917,6 +916,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--repeat-last-n"}, "N",
string_format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sampling.penalty_last_n),
[](common_params & params, int value) {
if (value < -1) {
throw std::runtime_error(string_format("error: invalid repeat-last-n = %d\n", value));
}
params.sampling.penalty_last_n = value;
params.sampling.n_prev = std::max(params.sampling.n_prev, params.sampling.penalty_last_n);
}
Expand Down Expand Up @@ -971,6 +973,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--dry-penalty-last-n"}, "N",
string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sampling.dry_penalty_last_n),
[](common_params & params, int value) {
if (value < -1) {
throw std::runtime_error(string_format("error: invalid dry-penalty-last-n = %d\n", value));
}
params.sampling.dry_penalty_last_n = value;
}
).set_sparam());
Expand Down Expand Up @@ -1583,6 +1588,20 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.hf_file = value;
}
).set_env("LLAMA_ARG_HF_FILE"));
add_opt(common_arg(
{"-hfrv", "--hf-repo-v"}, "REPO",
"Hugging Face model repository for the vocoder model (default: unused)",
[](common_params & params, const std::string & value) {
params.vocoder.hf_repo = value;
}
).set_env("LLAMA_ARG_HF_REPO_V"));
add_opt(common_arg(
{"-hffv", "--hf-file-v"}, "FILE",
"Hugging Face model file for the vocoder model (default: unused)",
[](common_params & params, const std::string & value) {
params.vocoder.hf_file = value;
}
).set_env("LLAMA_ARG_HF_FILE_V"));
add_opt(common_arg(
{"-hft", "--hf-token"}, "TOKEN",
"Hugging Face access token (default: value from HF_TOKEN environment variable)",
Expand Down Expand Up @@ -2180,5 +2199,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));

add_opt(common_arg(
{"-mv", "--model-vocoder"}, "FNAME",
"vocoder model for audio generation (default: unused)",
[](common_params & params, const std::string & value) {
params.vocoder.model = value;
}
).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));

return ctx_arg;
}
30 changes: 26 additions & 4 deletions common/common.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -942,6 +942,25 @@ struct common_init_result common_init_from_params(common_params & params) {
params.sampling.ignore_eos = false;
}

if (params.sampling.ignore_eos) {
for (llama_token i = 0; i < llama_n_vocab(model); i++) {
if (llama_token_is_eog(model, i)) {
LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
params.sampling.logit_bias.push_back({i, -INFINITY});
}
}
}

if (params.sampling.penalty_last_n == -1) {
LOG_INF("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
params.sampling.penalty_last_n = llama_n_ctx(lctx);
}

if (params.sampling.dry_penalty_last_n == -1) {
LOG_INF("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
}

if (params.warmup) {
LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);

Expand Down Expand Up @@ -1078,7 +1097,7 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
#define CURL_MAX_RETRY 3
#define CURL_RETRY_DELAY_SECONDS 2

static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_attempts, int retry_delay_seconds) {
static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
int remaining_attempts = max_attempts;

while (remaining_attempts > 0) {
Expand All @@ -1102,7 +1121,6 @@ static bool curl_perform_with_retry(const std::string& url, CURL* curl, int max_
}

static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {

// Initialize libcurl
std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup);
if (!curl) {
Expand Down Expand Up @@ -1175,11 +1193,13 @@ static bool common_download_file(const std::string & url, const std::string & pa
std::string etag;
std::string last_modified;
};

common_load_model_from_url_headers headers;

{
typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
common_load_model_from_url_headers *headers = (common_load_model_from_url_headers *) userdata;
common_load_model_from_url_headers * headers = (common_load_model_from_url_headers *) userdata;

static std::regex header_regex("([^:]+): (.*)\r\n");
static std::regex etag_regex("ETag", std::regex_constants::icase);
Expand Down Expand Up @@ -1763,7 +1783,9 @@ void common_embd_normalize(const float * inp, float * out, int n, int embd_norm)
break;
case 0: // max absolute
for (int i = 0; i < n; i++) {
if (sum < std::abs(inp[i])) sum = std::abs(inp[i]);
if (sum < std::abs(inp[i])) {
sum = std::abs(inp[i]);
}
}
sum /= 32760.0; // make an int16 range
break;
Expand Down
31 changes: 23 additions & 8 deletions common/common.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -76,6 +76,7 @@ enum llama_example {
LLAMA_EXAMPLE_LLAVA,
LLAMA_EXAMPLE_LOOKUP,
LLAMA_EXAMPLE_PARALLEL,
LLAMA_EXAMPLE_TTS,

LLAMA_EXAMPLE_COUNT,
};
Expand All @@ -91,6 +92,7 @@ enum common_sampler_type {
COMMON_SAMPLER_TYPE_TEMPERATURE = 7,
COMMON_SAMPLER_TYPE_XTC = 8,
COMMON_SAMPLER_TYPE_INFILL = 9,
COMMON_SAMPLER_TYPE_PENALTIES = 10,
};

// dimensionality reduction methods, used by cvector-generator
Expand Down Expand Up @@ -126,7 +128,6 @@ struct common_params_sampling {
int32_t mirostat = 0; // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
float mirostat_tau = 5.00f; // target entropy
float mirostat_eta = 0.10f; // learning rate
bool penalize_nl = false; // consider newlines as a repeatable token
bool ignore_eos = false;
bool no_perf = false; // disable performance metrics
bool timing_per_token = false;
Expand All @@ -135,6 +136,7 @@ struct common_params_sampling {


std::vector<enum common_sampler_type> samplers = {
COMMON_SAMPLER_TYPE_PENALTIES,
COMMON_SAMPLER_TYPE_DRY,
COMMON_SAMPLER_TYPE_TOP_K,
COMMON_SAMPLER_TYPE_TYPICAL_P,
Expand All @@ -154,6 +156,7 @@ struct common_params_sampling {

struct common_params_speculative {
std::vector<ggml_backend_dev_t> devices; // devices to use for offloading

int32_t n_ctx = 0; // draft context size
int32_t n_max = 16; // maximum number of tokens to draft during speculative decoding
int32_t n_min = 5; // minimum number of draft tokens to use for speculative decoding
Expand All @@ -167,6 +170,14 @@ struct common_params_speculative {
std::string model = ""; // draft model for speculative decoding // NOLINT
};

struct common_params_vocoder {
std::string hf_repo = ""; // HF repo // NOLINT
std::string hf_file = ""; // HF file // NOLINT

std::string model = ""; // model path // NOLINT
std::string model_url = ""; // model url to download // NOLINT
};

struct common_params {
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 4096; // context size
Expand All @@ -189,11 +200,13 @@ struct common_params {
float defrag_thold = 0.1f; // KV cache defragmentation threshold

// offload params
std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default)
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
enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
std::vector<ggml_backend_dev_t> devices; // devices to use for offloading

int32_t n_gpu_layers = -1; // number of layers to store in VRAM (-1 - use default)
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

enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs

struct cpu_params cpuparams;
struct cpu_params cpuparams_batch;
Expand All @@ -207,8 +220,9 @@ struct common_params {
enum llama_pooling_type pooling_type = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
enum llama_attention_type attention_type = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings

struct common_params_sampling sampling;
struct common_params_sampling sampling;
struct common_params_speculative speculative;
struct common_params_vocoder vocoder;

std::string model = ""; // model path // NOLINT
std::string model_alias = ""; // model alias // NOLINT
Expand Down Expand Up @@ -589,7 +603,8 @@ void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_si
// Embedding utils
//

void common_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
// TODO: repace embd_norm with an enum
void common_embd_normalize(const float * inp, float * out, int n, int embd_norm);

float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);

Expand Down
27 changes: 11 additions & 16 deletions common/sampling.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -161,32 +161,20 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
params.logit_bias.size(),
params.logit_bias.data()));

llama_sampler_chain_add(result->chain,
llama_sampler_init_penalties(
llama_n_vocab (model),
llama_token_eos(model),
llama_token_nl (model),
params.penalty_last_n,
params.penalty_repeat,
params.penalty_freq,
params.penalty_present,
params.penalize_nl,
params.ignore_eos));

if (params.mirostat == 0) {
for (const auto & cnstr : params.samplers) {
switch (cnstr) {
case COMMON_SAMPLER_TYPE_DRY:
case COMMON_SAMPLER_TYPE_DRY:
{
std::vector<const char*> c_breakers;
std::vector<const char *> c_breakers;
c_breakers.reserve(params.dry_sequence_breakers.size());
for (const auto& str : params.dry_sequence_breakers) {
for (const auto & str : params.dry_sequence_breakers) {
c_breakers.push_back(str.c_str());
}

llama_sampler_chain_add(result->chain, llama_sampler_init_dry (model, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
}
break;
break;
case COMMON_SAMPLER_TYPE_TOP_K:
llama_sampler_chain_add(result->chain, llama_sampler_init_top_k (params.top_k));
break;
Expand All @@ -208,6 +196,9 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
case COMMON_SAMPLER_TYPE_INFILL:
llama_sampler_chain_add(result->chain, llama_sampler_init_infill (model));
break;
case COMMON_SAMPLER_TYPE_PENALTIES:
llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
break;
default:
GGML_ASSERT(false && "unknown sampler type");
}
Expand Down Expand Up @@ -415,6 +406,7 @@ char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
case COMMON_SAMPLER_TYPE_XTC: return 'x';
case COMMON_SAMPLER_TYPE_INFILL: return 'i';
case COMMON_SAMPLER_TYPE_PENALTIES: return 'e';
default : return '?';
}
}
Expand All @@ -429,6 +421,7 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
case COMMON_SAMPLER_TYPE_XTC: return "xtc";
case COMMON_SAMPLER_TYPE_INFILL: return "infill";
case COMMON_SAMPLER_TYPE_PENALTIES: return "penalties";
default : return "";
}
}
Expand All @@ -443,6 +436,7 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
{ "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
{ "xtc", COMMON_SAMPLER_TYPE_XTC },
{ "infill", COMMON_SAMPLER_TYPE_INFILL },
{ "penalties", COMMON_SAMPLER_TYPE_PENALTIES },
};

// since samplers names are written multiple ways
Expand Down Expand Up @@ -489,6 +483,7 @@ std::vector<common_sampler_type> common_sampler_types_from_chars(const std::stri
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC), COMMON_SAMPLER_TYPE_XTC },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_INFILL), COMMON_SAMPLER_TYPE_INFILL },
{ common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_PENALTIES), COMMON_SAMPLER_TYPE_PENALTIES },
};

std::vector<common_sampler_type> samplers;
Expand Down
Loading

0 comments on commit ee486ba

Please sign in to comment.