gpt_neox_client 0.1.0 → 0.2.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: d77a992f9cbba2e7a91141c859194cc0a200b9c5dd4e58aeedc51482ec75b8c0
-  data.tar.gz: 302c37b125a0875463859b62fb7918b47c38cd521ad444fe758a36ba06e27ade
+  metadata.gz: e76735b1c4c6a4e228620bd4cd3ab20d02d0b20505eb85acbcab263301ad4e49
+  data.tar.gz: 05d285d7b1daa24408c1087f0c748a456a8398d45c59b3b311e1d0a4413df00a
 SHA512:
-  metadata.gz: a3ab1eb43db87f08e24fb16181d83feeb4b1421fd601165a233cc48eccfd0403de2e1d98042e89e259c843e5c3021e3931d57f4742ee20ae8c7e55cf45f6c0d4
-  data.tar.gz: 88f961e5a901ea5896486b4612ab3ecc8ebcbad12726fd76700bb1248e31c11716ae4a7248592f7657fb056e36af19f466f95860862a4d165fc026df3d4cb04f
+  metadata.gz: dda9974e3d4d1023ec0e8783922c6cb779b41d0083aa26bdfb73e69778de353eee9b26d5185ea0f160bec07df89ec9f34267dd2c42e02f0d95bc224fb4b4a43a
+  data.tar.gz: 0a2c389774a0e49b8b6f4ee2dac8bea96f5d1608ef4b4ac9cef95f29480fc23f4cbc19c14be97781e9076eee3a9fd247883c979390ff1fd2385159c42a12189e

data/CHANGELOG.md

@@ -1,5 +1,10 @@
 ## [Unreleased]
 
-## [0.1.0] - 2023-09-xx
+## [0.2.0] - 2023-09-02
+
+- Add Accelerate framework and Metal build option for macOS.
+- Add OpenBLAS build option for platforms other than macOS.
+
+## [0.1.0] - 2023-09-01
 
 - Initial release

data/README.md

@@ -38,15 +38,15 @@ japanese-large-lm-3.6b-instruction-sft/ggml-model-f16.bin
 ```
 
 ```ruby
-require "gpt_neox_client"
+require 'gpt_neox_client'
 
-client = GPTNeoXClient.new('japanese-large-lm-3.6b-instruction-sft/ggml-model-f16.bin', seed: 123456789, n_threads: 4)
+client = GPTNeoXClient.new(path: 'japanese-large-lm-3.6b-instruction-sft/ggml-model-f16.bin', seed: 123456789, n_threads: 4)
 puts client.completions(
   'ユーザー:四国の県名を全て列挙してください。<0x0A>システム:',
   top_p: 0.9,
   top_k: 1,
   temperature: 0.7
-).gsub("<0x0A>", "\n").gsub("</s>", " ")
+).gsub('<0x0A>', "\n").gsub('</s>', '')
 #
 # ユーザー:四国の県名を全て列挙してください。
 # システム:徳島県、香川県、愛媛県、高知県

data/ext/gpt_neox_client/extconf.rb

@@ -22,4 +22,30 @@ $INCFLAGS << ' -I$(srcdir)/src/ggml'
 $VPATH << '$(srcdir)/src'
 $VPATH << '$(srcdir)/src/ggml'
 
+if RUBY_PLATFORM.match?(/darwin|linux|bsd/) && try_compile('#include <stdio.h>', '-pthread')
+  $CFLAGS << ' -pthread'
+  $CXXFLAGS << ' -pthread'
+end
+
+if RUBY_PLATFORM.match?(/darwin/)
+  if have_framework('Accelerate')
+    $CFLAGS << ' -DGGML_USE_ACCELERATE'
+  else
+    warning 'Accelerate framework is not found.'
+  end
+end
+
+$CFLAGS << ' -DGGML_USE_OPENBLAS' if !RUBY_PLATFORM.match?(/darwin/) && (have_library('openblas') && have_header('cblas.h'))
+
 create_makefile('gpt_neox_client/gpt_neox_client')
+
+if RUBY_PLATFORM.match?(/darwin/)
+  File.open('Makefile', 'a') do |f|
+    f.puts "\nggml-metal.o: ggml-metal.m ggml-metal.h"
+    f.puts "\t$(CC) $(CFLAGS) -c $< -o $@"
+  end
+
+  metal_path = File.expand_path("#{__dir__}/src/ggml/ggml-metal.metal")
+  dest_path = File.expand_path("#{__dir__}/../../lib/gpt_neox_client/")
+  FileUtils.cp(metal_path, dest_path)
+end

data/ext/gpt_neox_client/gpt_neox_client.cpp

@@ -208,6 +208,7 @@ static VALUE gpt_neox_client_completions(int argc, VALUE* argv, VALUE self) {
   std::mt19937 rng(seed);
   std::vector<gpt_vocab::id> embd;
   std::vector<int32_t> last_n_tokens(model->hparams.n_ctx, 0);
+  gpt_vocab::id token_eos = vocab->token_to_id["</s>"];
 
   while (n_sampled < n_predict) {
     if (embd.size() > 0) {
@@ -240,7 +241,7 @@ static VALUE gpt_neox_client_completions(int argc, VALUE* argv, VALUE self) {
     }
 
     for (auto id : embd) completions += vocab->id_to_token[id];
-    if (embd.back() == 0) break;
+    if (!embd.empty() && embd.back() == token_eos) break;
   }
 
   RB_GC_GUARD(prompt_);

data/lib/gpt_neox_client/version.rb

@@ -3,5 +3,5 @@
 # GPTNeoXClient is a Ruby client for GPT-NeoX.
 class GPTNeoXClient
   # The version of GPTNeoXClient you are using.
-  VERSION = '0.1.0'
+  VERSION = '0.2.0'
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: gpt_neox_client
 version: !ruby/object:Gem::Version
-  version: 0.1.0
+  version: 0.2.0
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2023-09-01 00:00:00.000000000 Z
+date: 2023-09-02 00:00:00.000000000 Z
 dependencies: []
 description: gpt_neox_client is a simple client for GPT-NeoX.
 email: