Merge pull request #25 from CodeWithKyrian/add-llama-support

Add support for Llama models

Author: CodeWithKyrian

composer.json

@@ -21,8 +21,7 @@
     "codewithkyrian/onnxruntime-downloader-plugin": "^1.1",
     "symfony/console": "^6.4|^7.0",
     "imagine/imagine": "^1.3",
-    "rokka/imagine-vips": "^0.31.0",
-    "spatie/fork": "^1.2"
+    "rokka/imagine-vips": "^0.31.0"
   },
   "require-dev": {
     "pestphp/pest": "^2.31",

examples/pipelines/text-generation.php

@@ -12,13 +12,14 @@
 ini_set('memory_limit', -1);
 //
 //$generator = pipeline('text-generation', 'Xenova/gpt2');
-$generator = pipeline('text-generation', 'Xenova/Qwen1.5-0.5B-Chat');
-//
+//$generator = pipeline('text-generation', 'Xenova/Qwen1.5-0.5B-Chat');
+$generator = pipeline('text-generation', 'Xenova/TinyLlama-1.1B-Chat-v1.0');
+
 $streamer = StdOutStreamer::make();
 
 $messages = [
     ['role' => 'system', 'content' => 'You are a helpful assistant.'],
-    ['role' => 'user', 'content' => 'What is diffusion?'],
+    ['role' => 'user', 'content' => 'What is diffusion in chemistry?'],
 ];
 
 $input = $generator->tokenizer->applyChatTemplate($messages, addGenerationPrompt: true, tokenize: false);

src/Decoders/ByteFallback.php

@@ -18,8 +18,8 @@ public function __construct(array $config)
 
     protected function decodeChain(array $tokens): array
     {
-        $new_tokens = [];
-        $previous_byte_tokens = [];
+        $newTokens = [];
+        $previousByteTokens = [];
 
         foreach ($tokens as $token) {
             $bytes = null;
@@ -30,22 +30,22 @@ protected function decodeChain(array $tokens): array
                 }
             }
             if ($bytes !== null) {
-                $previous_byte_tokens[] = $bytes;
+                $previousByteTokens[] = $bytes;
             } else {
-                if (count($previous_byte_tokens) > 0) {
-                    $string = $this->bytesToString($previous_byte_tokens);
-                    $new_tokens[] = $string;
-                    $previous_byte_tokens = [];
+                if (count($previousByteTokens) > 0) {
+                    $string = $this->bytesToString($previousByteTokens);
+                    $newTokens[] = $string;
+                    $previousByteTokens = [];
                 }
-                $new_tokens[] = $token;
+                $newTokens[] = $token;
             }
         }
-        if (count($previous_byte_tokens) > 0) {
-            $string = $this->bytesToString($previous_byte_tokens);
-            $new_tokens[] = $string;
+        if (count($previousByteTokens) > 0) {
+            $string = $this->bytesToString($previousByteTokens);
+            $newTokens[] = $string;
         }
 
-        return $new_tokens;
+        return $newTokens;
     }
 
     /**
@@ -56,9 +56,7 @@ protected function decodeChain(array $tokens): array
      */
     protected function bytesToString(array $bytes): string
     {
-        $chars = array_map(function ($byte) {
-            return chr($byte);
-        }, $bytes);
-        return implode('', $chars);
+        $binaryString = pack('C*', ...$bytes);
+        return mb_convert_encoding($binaryString, 'ISO-8859-1');
     }
 }

src/Decoders/FuseDecoder.php

@@ -18,6 +18,6 @@ public function __construct(array $config)
 
     protected function decodeChain(array $tokens): array
     {
-       return [implode('', $tokens)];
+        return [implode('', $tokens)];
     }
 }

src/Decoders/StripDecoder.php

@@ -28,7 +28,7 @@ protected function decodeChain(array $tokens): array
         return array_map(function ($token) {
             $startCut = 0;
             for ($i = 0; $i < $this->start; ++$i) {
-                if ($token[$i] === $this->content) {
+                if ($token[$i] ?? null === $this->content) {
                     $startCut = $i + 1;
                     continue;
                 } else {
@@ -39,7 +39,7 @@ protected function decodeChain(array $tokens): array
             $stopCut = strlen($token);
             for ($i = 0; $i < $this->stop; ++$i) {
                 $index = strlen($token) - $i - 1;
-                if ($token[$index] === $this->content) {
+                if ($token[$index] ?? null === $this->content) {
                     $stopCut = $index;
                     continue;
                 } else {

src/Models/Auto/AutoModel.php

@@ -38,6 +38,7 @@ class AutoModel extends PretrainedMixin
         "gptj" => \Codewithkyrian\Transformers\Models\Pretrained\GPTJModel::class,
         "gpt_bigcode" => \Codewithkyrian\Transformers\Models\Pretrained\GPTBigCodeModel::class,
         "codegen" => \Codewithkyrian\Transformers\Models\Pretrained\CodeGenModel::class,
+        "llama" => \Codewithkyrian\Transformers\Models\Pretrained\LlamaModel::class,
         "qwen2" => \Codewithkyrian\Transformers\Models\Pretrained\Qwen2Model::class,
     ];
 

src/Models/Auto/AutoModelForCausalLM.php

@@ -12,6 +12,7 @@ class AutoModelForCausalLM extends PretrainedMixin
         'gptj' => \Codewithkyrian\Transformers\Models\Pretrained\GPTJForCausalLM::class,
         'gpt_bigcode' => \Codewithkyrian\Transformers\Models\Pretrained\GPTBigCodeForCausalLM::class,
         'codegen' => \Codewithkyrian\Transformers\Models\Pretrained\CodeGenForCausalLM::class,
+        'llama' => \Codewithkyrian\Transformers\Models\Pretrained\LlamaForCausalLM::class,
         'trocr' => \Codewithkyrian\Transformers\Models\Pretrained\TrOCRForCausalLM::class,
         'qwen2' => \Codewithkyrian\Transformers\Models\Pretrained\Qwen2ForCausalLM::class
     ];

src/Models/ModelArchitecture.php

@@ -119,6 +119,8 @@ protected function decoderRunBeam(PretrainedModel $model, array &$beam): array
             'past_key_values' => $beam['prev_model_outputs']['past_key_values'] ?? null,
         ];
 
+
+        // 2. Run
         $output = $model->forward($modelInputs);
 
         // 3. Update

src/Models/Pretrained/LlamaForCausalLM.php

@@ -0,0 +1,11 @@
+<?php
+
+declare(strict_types=1);
+
+
+namespace Codewithkyrian\Transformers\Models\Pretrained;
+
+class LlamaForCausalLM extends LlamaPretrainedModel
+{
+
+}

src/Models/Pretrained/LlamaModel.php

@@ -0,0 +1,14 @@
+<?php
+
+declare(strict_types=1);
+
+
+namespace Codewithkyrian\Transformers\Models\Pretrained;
+
+/**
+ * The bare LLaMA Model outputting raw hidden-states without any specific head on top.
+ */
+class LlamaModel extends LlamaPretrainedModel
+{
+
+}