Skip to content

Commit

Permalink
feat: sync llama.cpp (#45)
Browse files Browse the repository at this point in the history 
* feat: sync llama.cpp

* fix(ci): bump cocoapods version to 1.15.2
  • Loading branch information
@jhen0409
jhen0409 authored Feb 22, 2024
1 parent b1fb538 commit ae30692
Show file tree
Hide file tree
Showing 27 changed files with 9,510 additions and 2,748 deletions.
5 changes: 5 additions & 0 deletions .github/workflows/ci.yml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,11 @@ jobs:
restore-keys: |
${{ runner.os }}-pods-
- name: Upgrade CocoaPods to version 1.15.2
run: |
gem uninstall cocoapods --ignore-dependencies
gem install cocoapods -v 1.15.2
- name: Install cocoapods
run: |
yarn example pods
Expand Down
309 changes: 225 additions & 84 deletions cpp/common.cpp
View file

Large diffs are not rendered by default.

82 changes: 45 additions & 37 deletions cpp/common.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -32,40 +32,40 @@
int32_t get_num_physical_cores();

struct gpt_params {
uint32_t seed = -1; // RNG seed

int32_t n_threads = get_num_physical_cores();
int32_t n_threads_draft = -1;
int32_t n_threads_batch = -1; // number of threads to use for batch processing (-1 = use n_threads)
int32_t n_threads_batch_draft = -1;
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 512; // context size
int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_draft = 8; // number of tokens to draft during speculative decoding
int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited)
int32_t n_parallel = 1; // number of parallel sequences to decode
int32_t n_sequences = 1; // number of sequences to decode
float p_accept = 0.5f; // speculative decoding accept probability
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_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[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
int32_t n_beams = 0; // if non-zero then use beam search of given width.
int32_t grp_attn_n = 1; // group-attention factor
int32_t grp_attn_w = 512; // group-attention width
int32_t n_print = -1; // print token count every n tokens (-1 = disabled)
float rope_freq_base = 0.0f; // RoPE base frequency
float rope_freq_scale = 0.0f; // RoPE frequency scaling factor
float yarn_ext_factor = -1.0f; // YaRN extrapolation mix factor
float yarn_attn_factor = 1.0f; // YaRN magnitude scaling factor
float yarn_beta_fast = 32.0f; // YaRN low correction dim
float yarn_beta_slow = 1.0f; // YaRN high correction dim
int32_t yarn_orig_ctx = 0; // YaRN original context length
int8_t rope_scaling_type = LLAMA_ROPE_SCALING_UNSPECIFIED; // TODO: better to be int32_t for alignment
// pinging @cebtenzzre
uint32_t seed = -1; // RNG seed

int32_t n_threads = get_num_physical_cores();
int32_t n_threads_draft = -1;
int32_t n_threads_batch = -1; // number of threads to use for batch processing (-1 = use n_threads)
int32_t n_threads_batch_draft = -1;
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 512; // context size
int32_t n_batch = 512; // batch size for prompt processing (must be >=32 to use BLAS)
int32_t n_keep = 0; // number of tokens to keep from initial prompt
int32_t n_draft = 8; // number of tokens to draft during speculative decoding
int32_t n_chunks = -1; // max number of chunks to process (-1 = unlimited)
int32_t n_parallel = 1; // number of parallel sequences to decode
int32_t n_sequences = 1; // number of sequences to decode
float p_accept = 0.5f; // speculative decoding accept probability
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_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.
int32_t grp_attn_n = 1; // group-attention factor
int32_t grp_attn_w = 512; // group-attention width
int32_t n_print = -1; // print token count every n tokens (-1 = disabled)
float rope_freq_base = 0.0f; // RoPE base frequency
float rope_freq_scale = 0.0f; // RoPE frequency scaling factor
float yarn_ext_factor = -1.0f; // YaRN extrapolation mix factor
float yarn_attn_factor = 1.0f; // YaRN magnitude scaling factor
float yarn_beta_fast = 32.0f; // YaRN low correction dim
float yarn_beta_slow = 1.0f; // YaRN high correction dim
int32_t yarn_orig_ctx = 0; // YaRN original context length
int32_t rope_scaling_type = LLAMA_ROPE_SCALING_UNSPECIFIED;
lm_ggml_numa_strategy numa = LM_GGML_NUMA_STRATEGY_DISABLED;

// // sampling parameters
struct llama_sampling_params sparams;
Expand All @@ -80,6 +80,7 @@ struct gpt_params {
std::string input_suffix = ""; // string to suffix user inputs with
std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
std::string logdir = ""; // directory in which to save YAML log files
std::string logits_file = ""; // file for saving *all* logits

std::vector<llama_model_kv_override> kv_overrides;

Expand All @@ -97,6 +98,11 @@ struct gpt_params {
bool winogrande = false; // compute Winogrande score over random tasks from datafile supplied in prompt
size_t winogrande_tasks= 0; // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed

bool multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt
size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed

bool kl_divergence = false; // compute KL-divergence

bool mul_mat_q = true; // if true, use mul_mat_q kernels instead of cuBLAS
bool random_prompt = false; // do not randomize prompt if none provided
bool use_color = false; // use color to distinguish generations and inputs
Expand All @@ -118,7 +124,6 @@ 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 numa = false; // attempt optimizations that help on some NUMA systems
bool verbose_prompt = false; // print prompt tokens before generation
bool display_prompt = true; // print prompt before generation
bool infill = false; // use infill mode
Expand Down Expand Up @@ -146,10 +151,13 @@ std::string gpt_random_prompt(std::mt19937 & rng);
void process_escapes(std::string& input);

//
// String parsing
// String utils
//

std::string parse_samplers_input(std::string input);
std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
std::vector<std::string> string_split(std::string input, char separator);
std::string sampler_type_to_name_string(llama_sampler_type sampler_type);

//
// Model utils
Expand Down
Loading

0 comments on commit ae30692

Please sign in to comment.