llama_cpp 0.2.2 → 0.3.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: e5e221d4831be790a990b121e6ac780d10b4cbfb85b2a9b4284d9c216f6e5604
-  data.tar.gz: fba76ac1a70bfd7b02b8d123c57e4c8096a29ac7f658bb090cda91c6a54752d2
+  metadata.gz: 7a1f299e21bfe5b12d517a4254657cbc5bf9af6d0571285e2a5aff67b9175646
+  data.tar.gz: 62dd6e0d4f0b052a912d87b52cd0cff5bb873ab12378413a3ee0af5671331ef6
 SHA512:
-  metadata.gz: 994029383219077e134d170177954251c20ede6d1c83843ecd22c42eeae83584079d124b41702f55add7f3f237e9bdb14382fbd37dde2d0e74f8cffcfed1715b
-  data.tar.gz: ca4e94b6ddf4e4e9ddabbb2b8309cf4b2b06a881df09fdf4ad96e27c4f1f620ca0024ac46f69d9b474849c074a5c9ba9b0440777a0b52a12413bc356457a02f3
+  metadata.gz: b12dc73914e5c7ecdd951fd57b70e01aae1926a2adc88030b5f5310f99c789e129cf552811363ec99525b37b9ca167a708cb756057b94f5cf4dd2a0100b06b6e
+  data.tar.gz: d1d79696b08f89894de02a02fac91f0783c432efa641b21ee59f6987946b045681a60113392db6c85fe97bd0e1fc9860235faa358fb805bb0de21eb85926edd5

data/CHANGELOG.md

@@ -1,3 +1,37 @@
+## [[0.3.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.3.0...v0.3.1)] - 2023-07-02
+
+- Bump bundled llama.cpp from master-9d23589 to master-b8c8dda.
+  - Use unsigned values for random seed.
+- Add `eval_embd` method to `Context` class.
+
+## [[0.3.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.2.2...v0.3.0)] - 2023-06-30
+
+- Add no_k_quants and qkk_64 config options:
+  ```
+  $ gem install llama_cpp -- --with-no_k_quants
+  ```
+  ```
+  $ gem install llama_cpp -- --with-qkk_64
+  ```
+
+**Breaking Changes**
+- Remove `Client` class to concentrate on developing bindings.
+- Bump bundled llama.cpp from master-7487137 to master-9d23589.
+  - llama_init_from_file and llama_apply_lora_from_file are deprecated.
+- Add `Model` class for wrapping llama_model.
+- Move the `apply_lora_from_file method`, `free`, `load`, and `empty?` methods to `Model` class from `Context` class.
+- Change arguments of initialize method of Context. Its initialize method requires Model object instead of the model's file path.
+  ```ruby
+  requre 'llama_cpp'
+
+  params = LLaMACpp::ContextParams.new
+
+  model = LLaMACpp::Model.new(model_path: '/path/to/quantized-model.bin', params: params)
+  context = LLaMACpp::Context.new(model: model)
+
+  LLaMACpp.generate(context, 'Hello, world.')
+  ```
+
 ## [[0.2.2](https://github.com/yoshoku/llama_cpp.rb/compare/v0.2.1...v0.2.2)] - 2023-06-24
 
 - Bump bundled llama.cpp from master-a09f919 to master-7487137.

data/README.md

@@ -20,21 +20,54 @@ If bundler is not being used to manage dependencies, install the gem by executin
 
 ## Usage
 
-Prepare the quantized model by refering to [the usage section on the llama.cpp README](https://github.com/ggerganov/llama.cpp#usage) or
-download the qunatized model, for example [ggml-vicuna-7b-4bit](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5541351), from Hugging Face.
+Prepare the quantized model by refering to [the usage section on the llama.cpp README](https://github.com/ggerganov/llama.cpp#usage).
+For example, preparing the quatization model based on [open_llama_7b](https://huggingface.co/openlm-research/open_llama_7b) is as follows:
+
+```sh
+$ cd ~/
+$ brew install git-lfs
+$ git lfs install
+$ git clone https://github.com/ggerganov/llama.cpp.git
+$ cd llama.cpp
+$ python3 -m pip install -r requirements.txt
+$ cd models
+$ git clone https://huggingface.co/openlm-research/open_llama_7b
+$ cd ../
+$ python3 convert.py models/open_llama_7b
+$ make
+$ ./quantize ./models/open_llama_7b/ggml-model-f16.bin ./models/open_llama_7b/ggml-model-q4_0.bin q4_0
+```
+
+An example of Ruby code that generates sentences with the quantization model is as follows:
 
 ```ruby
 require 'llama_cpp'
 
 params = LLaMACpp::ContextParams.new
-params.seed = 12
+params.seed = 42
 
-context = LLaMACpp::Context.new(model_path: '/path/to/quantized-model.bin', params: params)
+model = LLaMACpp::Model.new(model_path: '/home/user/llama.cpp/models/open_llama_7b/ggml-model-q4_0.bin', params: params)
+context = LLaMACpp::Context.new(model: model)
 
-puts LLaMACpp.generate(context, 'Please tell me the largest city in Japan.', n_threads: 4)
-# => "There are two major cities in Japan, Tokyo and Osaka, which have about 30 million populations."
+puts LLaMACpp.generate(context, 'Hello, World.', n_threads: 4)
 ```
 
+## Examples
+There is a sample program in the [examples](https://github.com/yoshoku/llama_cpp.rb/tree/main/examples) directory that allow interactvie communication like ChatGPT.
+
+```sh
+$ git clone https://github.com/yoshoku/llama_cpp.rb.git
+$ cd examples
+$ bundle install
+$ ruby chat.rb --model /home/user/llama.cpp/models/open_llama_7b/ggml-model-q4_0.bin --seed 2023
+...
+User: Who is the originator of the Ruby programming language?
+Bob: The originator of the Ruby programming language is Mr. Yukihiro Matsumoto.
+User:
+```
+
+![llama_cpp_chat_example](https://github.com/yoshoku/llama_cpp.rb/assets/5562409/374ae3d8-63a6-498f-ae6e-5552b464bdda)
+
 ## Contributing
 
 Bug reports and pull requests are welcome on GitHub at https://github.com/yoshoku/llama_cpp.rb.

data/examples/chat.rb

@@ -35,7 +35,8 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
     params = LLaMACpp::ContextParams.new
     params.seed = options[:seed]
     params.n_gpu_layers = options[:n_gpu_layers]
-    context = LLaMACpp::Context.new(model_path: options[:model], params: params)
+    model = LLaMACpp::Model.new(model_path: options[:model], params: params)
+    context = LLaMACpp::Context.new(model: model)
 
     antiprompt = options[:reverse_prompt] || 'User:'
     start_prompt = read_prompt(options[:file]) || default_prompt(antiprompt)

data/examples/embedding.rb

@@ -16,12 +16,13 @@ class Embedding < Thor # rubocop:disable Style/Documentation
   option :model, type: :string, aliases: '-m', desc: 'path to model file', required: true
   option :prompt, type: :string, aliases: '-p', desc: 'prompt to generate embedding', required: true
   option :n_gpu_layers, type: :numeric, desc: 'number of layers on GPU', default: 0
-  def main # rubocop:disable Metrics/AbcSize
+  def main # rubocop:disable Metrics/AbcSize, Metrics/MethodLength
     params = LLaMACpp::ContextParams.new
     params.seed = options[:seed]
     params.n_gpu_layers = options[:n_gpu_layers]
     params.embedding = true
-    context = LLaMACpp::Context.new(model_path: options[:model], params: params)
+    model = LLaMACpp::Model.new(model_path: options[:model], params: params)
+    context = LLaMACpp::Context.new(model: model)
 
     embd_input = context.tokenize(text: options[:prompt], add_bos: true)
 

data/ext/llama_cpp/extconf.rb

@@ -17,6 +17,17 @@ if RUBY_PLATFORM.match?(/darwin|linux|bsd/) && try_compile('#include <stdio.h>',
   $CXXFLAGS << ' -pthread'
 end
 
+unless with_config('no_k_quants')
+  $CFLAGS << ' -DGGML_USE_K_QUANTS'
+  $CXXFLAGS << ' -DGGML_USE_K_QUANTS'
+  $srcs << 'k_quants.c'
+end
+
+if with_config('qkk_64')
+  $CFLAGS << ' -DGGML_QKK_64'
+  $CXXFLAGS << ' -DGGML_QKK_64'
+end
+
 if with_config('openblas')
   abort 'libopenblas is not found.' unless have_library('openblas')
   abort 'cblas.h is not found.' unless have_header('cblas.h')
@@ -42,6 +53,7 @@ if with_config('metal')
   $CXXFLAGS << ' -DGGML_USE_METAL'
   $LDFLAGS << ' -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders'
   $objs = %w[ggml.o llama.o llama_cpp.o ggml-metal.o]
+  $objs << 'k_quants.o' unless with_config('no_k_quants')
 end
 
 if with_config('cublas')
@@ -49,6 +61,7 @@ if with_config('cublas')
   $CXXFLAGS << ' -DGGML_USE_CUBLAS -I/usr/local/cuda/include'
   $LDFLAGS << ' -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64'
   $objs = %w[ggml-cuda.o ggml.o llama.o llama_cpp.o]
+  $objs << 'k_quants.o' unless with_config('no_k_quants')
 end
 
 if with_config('clblast')