Module: LLaMACpp

Defined in:
lib/llama_cpp.rb,
lib/llama_cpp/version.rb,
ext/llama_cpp/llama_cpp.cpp

Overview

llama_cpp.rb provides Ruby bindings for the llama.cpp.

Constant Summary collapse

Params =

Class alias to match interface of whispercpp gem.

ContextParams
VERSION =

The version of llama_cpp.rb you install.

'0.5.1'
LLAMA_CPP_VERSION =

The version of llama.cpp bundled with llama_cpp.rb.

'b1198'
LLAMA_MAX_DEVICES = 
INT2NUM(LLAMA_MAX_DEVICES)
LLAMA_LOG_LEVEL_ERROR = 
INT2NUM(LLAMA_LOG_LEVEL_ERROR)
LLAMA_LOG_LEVEL_WARN = 
INT2NUM(LLAMA_LOG_LEVEL_WARN)
LLAMA_LOG_LEVEL_INFO = 
INT2NUM(LLAMA_LOG_LEVEL_INFO)
LLAMA_VOCAB_TYPE_SPM = 
INT2NUM(LLAMA_VOCAB_TYPE_SPM)
LLAMA_VOCAB_TYPE_BPE = 
INT2NUM(LLAMA_VOCAB_TYPE_BPE)
LLAMA_TOKEN_TYPE_UNDEFINED = 
INT2NUM(LLAMA_TOKEN_TYPE_UNDEFINED)
LLAMA_TOKEN_TYPE_NORMAL = 
INT2NUM(LLAMA_TOKEN_TYPE_NORMAL)
LLAMA_TOKEN_TYPE_UNKNOWN = 
INT2NUM(LLAMA_TOKEN_TYPE_UNKNOWN)
LLAMA_TOKEN_TYPE_CONTROL = 
INT2NUM(LLAMA_TOKEN_TYPE_CONTROL)
LLAMA_TOKEN_TYPE_USER_DEFINED = 
INT2NUM(LLAMA_TOKEN_TYPE_USER_DEFINED)
LLAMA_TOKEN_TYPE_UNUSED = 
INT2NUM(LLAMA_TOKEN_TYPE_UNUSED)
LLAMA_TOKEN_TYPE_BYTE = 
INT2NUM(LLAMA_TOKEN_TYPE_BYTE)
LLAMA_FTYPE_ALL_F32 = 
INT2NUM(LLAMA_FTYPE_ALL_F32)
LLAMA_FTYPE_MOSTLY_F16 = 
INT2NUM(LLAMA_FTYPE_MOSTLY_F16)
LLAMA_FTYPE_MOSTLY_Q4_0 = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q4_0)
LLAMA_FTYPE_MOSTLY_Q4_1 = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q4_1)
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16)
LLAMA_FTYPE_MOSTLY_Q8_0 = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q8_0)
LLAMA_FTYPE_MOSTLY_Q5_0 = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q5_0)
LLAMA_FTYPE_MOSTLY_Q5_1 = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q5_1)
LLAMA_FTYPE_MOSTLY_Q2_K = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q2_K)
LLAMA_FTYPE_MOSTLY_Q3_K_S = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q3_K_S)
LLAMA_FTYPE_MOSTLY_Q3_K_M = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q3_K_M)
LLAMA_FTYPE_MOSTLY_Q3_K_L = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q3_K_L)
LLAMA_FTYPE_MOSTLY_Q4_K_S = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q4_K_S)
LLAMA_FTYPE_MOSTLY_Q4_K_M = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q4_K_M)
LLAMA_FTYPE_MOSTLY_Q5_K_S = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q5_K_S)
LLAMA_FTYPE_MOSTLY_Q5_K_M = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q5_K_M)
LLAMA_FTYPE_MOSTLY_Q6_K = 
INT2NUM(LLAMA_FTYPE_MOSTLY_Q6_K)
LLAMA_FTYPE_GUESSED = 
INT2NUM(LLAMA_FTYPE_GUESSED)
LLAMA_GRETYPE_END = 
INT2NUM(LLAMA_GRETYPE_END)
LLAMA_GRETYPE_ALT = 
INT2NUM(LLAMA_GRETYPE_ALT)
LLAMA_GRETYPE_RULE_REF = 
INT2NUM(LLAMA_GRETYPE_RULE_REF)
LLAMA_GRETYPE_CHAR = 
INT2NUM(LLAMA_GRETYPE_CHAR)
LLAMA_GRETYPE_CHAR_NOT = 
INT2NUM(LLAMA_GRETYPE_CHAR_NOT)
LLAMA_GRETYPE_CHAR_RNG_UPPER = 
INT2NUM(LLAMA_GRETYPE_CHAR_RNG_UPPER)
LLAMA_GRETYPE_CHAR_ALT = 
INT2NUM(LLAMA_GRETYPE_CHAR_ALT)
LLAMA_FILE_MAGIC_GGSN = 
rb_str_new2(ss_magic.str().c_str())
LLAMA_SESSION_MAGIC = 
rb_str_new2(ss_magic.str().c_str())
LLAMA_DEFAULT_SEED = 
rb_str_new2(ss_magic.str().c_str())
LLAMA_SESSION_VERSION = 
rb_str_new2(std::to_string(LLAMA_SESSION_VERSION).c_str())

Class Method Summary collapse

Class Method Details

.backend_free ⇒ Object 



2468
2469
2470
2471
2472
 
# File 'ext/llama_cpp/llama_cpp.cpp', line 2468

static VALUE rb_llama_llama_backend_free(VALUE self) {
  llama_backend_free();

  return Qnil;
}

.backend_init(*args) ⇒ Object

module functions



2455
2456
2457
2458
2459
2460
2461
2462
2463
2464
2465
2466
 
# File 'ext/llama_cpp/llama_cpp.cpp', line 2455

static VALUE rb_llama_llama_backend_init(int argc, VALUE* argv, VALUE self) {
  VALUE kw_args = Qnil;
  ID kw_table[1] = { rb_intern("numa") };
  VALUE kw_values[1] = { Qundef };
  rb_scan_args(argc, argv, ":", &kw_args);
  rb_get_kwargs(kw_args, kw_table, 0, 1, kw_values);

  const bool numa = kw_values[0] == Qundef ? false : (RTEST(kw_values[0]) ? true : false);
  llama_backend_init(numa);

  return Qnil;
}

.generate(context, prompt, n_predict: 128, n_threads: 1, n_keep: 10, n_batch: 512, repeat_last_n: 64, repeat_penalty: 1.1, frequency: 0.0, presence: 0.0, top_k: 40, top_p: 0.95, tfs_z: 1.0, typical_p: 1.0, temperature: 0.8) ⇒ String

Generates sentences following the given prompt for operation check.

Parameters:

  • context (LLaMACpp::Context) —

    The context to use.

  • prompt (String) —

    The prompt to start generation with.

  • n_predict (Integer) (defaults to: 128) —

    The number of tokens to predict.

  • n_threads (Integer) (defaults to: 1) —

    The number of threads.

  • n_keep (Integer) (defaults to: 10) —

    The number of tokens to keep in the context.

  • n_batch (Integer) (defaults to: 512) —

    The number of tokens to process in a batch.

  • repeat_last_n (Integer) (defaults to: 64) —

    The number of tokens to consider for repetition penalty.

  • repeat_penalty (Float) (defaults to: 1.1) —

    The repetition penalty.

  • frequency (Float) (defaults to: 0.0) —

    The frequency penalty.

  • presence (Float) (defaults to: 0.0) —

    The presence penalty.

  • top_k (Integer) (defaults to: 40) —

    The number of tokens to consider for top-k sampling.

  • top_p (Float) (defaults to: 0.95) —

    The probability threshold for nucleus sampling.

  • tfs_z (Float) (defaults to: 1.0) —

    The z parameter for tail-free sampling.

  • typical_p (Float) (defaults to: 1.0) —

    The probability for typical sampling.

  • temperature (Float) (defaults to: 0.8) —

    The temperature for temperature sampling.

Returns:

  • (String)

Raises:

  • (ArgumentError) 


31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
 
# File 'lib/llama_cpp.rb', line 31

def generate(context, prompt, # rubocop:disable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/ParameterLists, Metrics/PerceivedComplexity
             n_predict: 128, n_threads: 1, n_keep: 10, n_batch: 512, repeat_last_n: 64,
             repeat_penalty: 1.1, frequency: 0.0, presence: 0.0, top_k: 40,
             top_p: 0.95, tfs_z: 1.0, typical_p: 1.0, temperature: 0.8)
  raise ArgumentError, 'context must be an instance of LLaMACpp::Context' unless context.is_a?(LLaMACpp::Context)
  raise ArgumentError, 'prompt must be a String' unless prompt.is_a?(String)

  spaced_prompt = " #{prompt}"
  embd_input = context.tokenize(text: spaced_prompt, add_bos: true)

  n_ctx = context.n_ctx
  raise ArgumentError, "prompt is too long #{embd_input.size} tokens, maximum is #{n_ctx - 4}" if embd_input.size > n_ctx - 4

  last_n_tokens = [0] * n_ctx

  embd = []
  n_consumed = 0
  n_past = 0
  n_remain = n_predict
  n_vocab = context.n_vocab
  output = []

  while n_remain != 0
    unless embd.empty?
      if n_past + embd.size > n_ctx
        n_left = n_past - n_keep
        n_past = n_keep
        embd.insert(0, last_n_tokens[(n_ctx - (n_left / 2) - embd.size)...-embd.size])
      end

      context.eval(tokens: embd, n_past: n_past, n_threads: n_threads)
    end

    n_past += embd.size
    embd.clear

    if embd_input.size <= n_consumed
      logits = context.logits
      base_candidates = Array.new(n_vocab) { |i| LLaMACpp::TokenData.new(id: i, logit: logits[i], p: 0.0) }
      candidates = LLaMACpp::TokenDataArray.new(base_candidates)

      # apply penalties
      last_n_repeat = [last_n_tokens.size, repeat_last_n, n_ctx].min
      context.sample_repetition_penalty(candidates, last_n_tokens[-last_n_repeat..], penalty: repeat_penalty)
      context.sample_frequency_and_presence_penalties(
        candidates, last_n_tokens[-last_n_repeat..], frequency: frequency, presence: presence
      )

      # temperature sampling
      context.sample_top_k(candidates, k: top_k)
      context.sample_tail_free(candidates, z: tfs_z)
      context.sample_typical(candidates, prob: typical_p)
      context.sample_top_p(candidates, prob: top_p)
      context.sample_temperature(candidates, temperature: temperature)
      id = context.sample_token(candidates)

      last_n_tokens.shift
      last_n_tokens.push(id)

      embd.push(id)
      n_remain -= 1
    else
      while embd_input.size > n_consumed
        embd.push(embd_input[n_consumed])
        last_n_tokens.shift
        last_n_tokens.push(embd_input[n_consumed])
        n_consumed += 1
        break if embd.size >= n_batch
      end
    end

    embd.each { |token| output << context.token_to_piece(token) }

    break if !embd.empty? && embd[-1] == context.token_eos
  end

  output.join.delete_prefix(spaced_prompt).strip
end

.max_devices ⇒ Object 



2523
2524
2525
 
# File 'ext/llama_cpp/llama_cpp.cpp', line 2523

static VALUE rb_llama_max_devices(VALUE self) {
  return INT2NUM(llama_max_devices());
}

.mlock_supported? ⇒ Boolean

Returns:

  • (Boolean) 


2519
2520
2521
 
# File 'ext/llama_cpp/llama_cpp.cpp', line 2519

static VALUE rb_llama_mlock_supported(VALUE self) {
  return llama_mlock_supported() ? Qtrue : Qfalse;
}

.mmap_supported? ⇒ Boolean

Returns:

  • (Boolean) 


2515
2516
2517
 
# File 'ext/llama_cpp/llama_cpp.cpp', line 2515

static VALUE rb_llama_mmap_supported(VALUE self) {
  return llama_mmap_supported() ? Qtrue : Qfalse;
}

.model_quantize(*args) ⇒ Object 



2474
2475
2476
2477
2478
2479
2480
2481
2482
2483
2484
2485
2486
2487
2488
2489
2490
2491
2492
2493
2494
2495
2496
2497
2498
2499
2500
2501
2502
2503
2504
 
# File 'ext/llama_cpp/llama_cpp.cpp', line 2474

static VALUE rb_llama_model_quantize(int argc, VALUE* argv, VALUE self) {
  VALUE kw_args = Qnil;
  ID kw_table[3] = { rb_intern("input_path"), rb_intern("output_path"), rb_intern("params") };
  VALUE kw_values[3] = { Qundef, Qundef, Qundef };
  rb_scan_args(argc, argv, ":", &kw_args);
  rb_get_kwargs(kw_args, kw_table, 3, 0, kw_values);

  if (!RB_TYPE_P(kw_values[0], T_STRING)) {
    rb_raise(rb_eArgError, "input_path must be a string");
    return Qnil;
  }
  if (!RB_TYPE_P(kw_values[1], T_STRING)) {
    rb_raise(rb_eArgError, "output_path must be a string");
    return Qnil;
  }
  if (!rb_obj_is_kind_of(kw_values[2], rb_cLLaMAModelQuantizeParams)) {
    rb_raise(rb_eArgError, "params must be a ModelQuantizeParams");
    return Qnil;
  }

  const char* input_path = StringValueCStr(kw_values[0]);
  const char* output_path = StringValueCStr(kw_values[1]);
  LLaMAModelQuantizeParamsWrapper* wrapper = RbLLaMAModelQuantizeParams::get_llama_model_quantize_params(kw_values[2]);

  if (llama_model_quantize(input_path, output_path, &(wrapper->params)) != 0) {
    rb_raise(rb_eRuntimeError, "Failed to quantize model");
    return Qnil;
  }

  return Qnil;
}

.print_system_info ⇒ Object 



2506
2507
2508
2509
 
# File 'ext/llama_cpp/llama_cpp.cpp', line 2506

static VALUE rb_llama_print_system_info(VALUE self) {
  const char* result = llama_print_system_info();
  return rb_utf8_str_new_cstr(result);
}

.time_us ⇒ Object 



2511
2512
2513
 
# File 'ext/llama_cpp/llama_cpp.cpp', line 2511

static VALUE rb_llama_time_us(VALUE self) {
  return LONG2NUM(llama_time_us());
}