Class: LLaMACpp::Client

Inherits:
Object
  • Object
show all
Defined in:
lib/llama_cpp/client.rb

Overview

Client provides a high-level interface to the LLM model.

Instance Method Summary collapse

Constructor Details

#initialize(model_path:, lora_adapter_path: nil, lora_base_path: nil, n_ctx: 512, n_parts: -1,, memory_f16: false, use_mmap: true, use_mlock: false, embedding: false, n_threads: 1, seed: 0) ⇒ Client

Creates a new client.

rubocop:disable Metrics/MethodLength, Metrics/ParameterLists

Parameters:

  • model_path (String) —

    The path to the model file.

  • lora_adapter_path (String) (defaults to: nil) —

    The path to the LoRA adapter file.

  • lora_base_path (String) (defaults to: nil) —

    The path to the LoRA base model file.

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

    The context size.

  • n_parts (Integer) (defaults to: -1,) —

    The amount of model parts (-1 = determine from model dimensions).

  • memory_f16 (Boolean) (defaults to: false) —

    The flag wheter to use f16 instead of f32 for memory kv.

  • use_mmap (Boolean) (defaults to: true) —

    The flag whether to use mmap.

  • use_mlock (Boolean) (defaults to: false) —

    The flag hether to use mlock.

  • embedding (Boolean) (defaults to: false) —

    The flag whether to calculate embedding.

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

    The number of threads to use.

  • seed (Integer) (defaults to: 0) —

    The seed for the random number generator.



21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
 
# File 'lib/llama_cpp/client.rb', line 21

def initialize(model_path:, lora_adapter_path: nil, lora_base_path: nil,
               n_ctx: 512, n_parts: -1, memory_f16: false, use_mmap: true, use_mlock: false,
               embedding: false,
               n_threads: 1, seed: 0)
  @params = {
    model_path: model_path,
    lora_adapter_path: lora_adapter_path,
    lora_base_path: lora_base_path,
    n_ctx: n_ctx,
    n_parts: n_parts,
    memory_f16: memory_f16,
    use_mmap: use_mmap,
    use_mlock: use_mlock,
    embedding: embedding,
    n_threads: n_threads,
    seed: seed
  }
  @context_params = ContextParams.new
  @context_params.n_ctx = n_ctx
  @context_params.n_parts = n_parts
  @context_params.f16_kv = memory_f16
  @context_params.use_mmap = use_mmap
  @context_params.use_mlock = use_mlock
  @context_params.embedding = embedding
  @context_params.seed = seed
  @context = Context.new(model_path: model_path, params: @context_params)
  return unless lora_adapter_path.is_a?(String)

  if lora_base_path.is_a?(String)
    @context.apply_lora_from_file(lora_path: lora_adapter_path, base_model_path: lora_base_path, n_threads: n_threads)
  else
    @context.apply_lora_from_file(lora_path: lora_adapter_path, n_threads: n_threads)
  end
end

Instance Method Details

#completions(prompt, max_tokens: 64, n_keep: 10, repeat_last_n: 64, n_batch: 512, frequency: 0.0, presence: 0.0, top_k: 40, top_p: 0.95, tfs_z: 1.0, typical_p: 1.0, temperature: 0.8, repeat_penalty: 1.1) ⇒ String

Generates completions for a given prompt.

rubocop:disable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/ParameterLists, Metrics/PerceivedComplexity

Parameters:

  • prompt (String) —

    The prompt to generate completions for.

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

    The maximum number of tokens to generate.

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

    The number of tokens to keep from the initial prompt.

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

    The number of tokens to use for repeat penalty.

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

    The batch size.

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

    The frequency penalty value.

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

    The presence penalty value.

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

    The top-k value.

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

    The top-p value.

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

    The tail free sampling parameter.

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

    The typical probability value.

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

    The temperature value.

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

    The repeat penalty value.

Returns:

  • (String)

Raises:

  • (ArgumentError) 


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
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
 
# File 'lib/llama_cpp/client.rb', line 74

def completions(prompt, max_tokens: 64, n_keep: 10, repeat_last_n: 64, n_batch: 512,
                frequency: 0.0, presence: 0.0,
                top_k: 40, top_p: 0.95, tfs_z: 1.0, typical_p: 1.0, temperature: 0.8, repeat_penalty: 1.1)
  embd_input = tokenize_prompt(prompt)

  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 = max_tokens
  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: @params[: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)

      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_str(token) }

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

  output.join.delete_prefix(" #{prompt}").strip
end

#embeddings(text) ⇒ Array<Float>

Obtains the embedding for a given text.

Parameters:

  • text (String) —

    The text to obtain the embedding for.

Returns:

  • (Array<Float>) 


158
159
160
161
162
163
164
165
166
 
# File 'lib/llama_cpp/client.rb', line 158

def embeddings(text)
  raise 'The embedding option is set to false' unless @params[:embedding]

  embd_input = tokenize_prompt(text)
  raise 'The result of tokenizing the input text is empty' unless embd_input.size.positive?

  @context.eval(tokens: embd_input, n_past: 0, n_threads: @params[:n_threads])
  @context.embeddings
end