.gitattributes

@@ -1,35 +0,0 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text

1/model.py

@@ -0,0 +1,255 @@
+"""
+[Transformer-LLM 백엔드 가이드]
+
+본 파일은 NVIDIA Triton Server에서 Hugging Face `AutoModelForCausalLM` 기반 모델을 손쉽게 배포하기 위해 제공되는 커스텀 Python 백엔드 템플릿입니다.
+
+1. 모델 호환성
+   - Hugging Face의 `AutoModelForCausalLM` 클래스와 호환되는 모든 Causal Language Model을 지원합니다.
+   - [확인] 배포할 모델 `config.json`의 `architectures` 항목이 `...ForCausalLM` 형식인지 확인.
+   - [확인] 모델이 Hugging Face 공식 문서의 AutoModelForCausalLM이 지원하는 모델인지 확인.
+           (https://huggingface.co/docs/transformers/en/model_doc/auto#transformers.AutoModelForCausalLM.from_pretrained)
+
+2. 토크나이저 호환성
+   - `AutoTokenizer`와 호환되는 토크나이저를 지원하며, 모델과 동일한 경로에서 자동으로 로드됩니다.
+
+3. 커스터마이징 안내
+   - 본 템플릿은 범용적인 사용을 위해 작성되었습니다.
+   - 특정 모델의 동작 방식이나 예외 처리가 필요한 경우, 이 파일(`model.py`)과 설정 파일(`config.pbtxt`)을 직접 수정하여 사용하시기 바랍니다.
+"""
+
+import json
+import torch
+import numpy as np
+import triton_python_backend_utils as pb_utils
+import uuid
+from typing import List, Dict, Any, Union, Tuple
+
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    GenerationConfig,
+    BitsAndBytesConfig,
+)
+from peft import PeftModel, PeftConfig
+
+class TritonPythonModel:
+    def initialize(self, args: Dict[str, str]):
+        """
+        모델 초기화: 설정 로드, 로거 설정, 모델 및 토크나이저 로드
+        """
+        self.logger = pb_utils.Logger
+        self.model_config = json.loads(args["model_config"])
+        self.model_name = args["model_name"]
+
+        # 설정 파라미터 로드
+        self.base_model_path = self._get_config_param("base_model_path")
+        self.is_adapter_model = self._get_config_param("is_adapter_model", "false").lower() == "true"
+        self.adapter_model_path = self._get_config_param("adapter_model_path")
+        self.quantization = self._get_config_param("quantization")
+        self.device = "cuda" if torch.cuda.is_available() else "cpu"
+
+        # 설정 로그 출력
+        self.logger.log_info(f"================ {self.model_name} Setup ================")
+        self.logger.log_info(f"Base Model: {self.base_model_path}")
+        self.logger.log_info(f"Adapter Mode: {self.is_adapter_model} ({self.adapter_model_path})")
+        self.logger.log_info(f"Quantization: {self.quantization}")
+        self.logger.log_info(f"Device: {self.device}")
+
+        self._load_model_and_tokenizer()
+        self.logger.log_info(f"Model initialized successfully.")
+
+    def _load_model_and_tokenizer(self):
+        """모델과 토크나이저를 로드하고 설정합니다."""
+        # 1. Quantization 설정
+        bnb_config = self._get_bnb_config()
+
+        # 2. Base Model 로드
+        load_path = self.base_model_path
+        if self.is_adapter_model:
+            peft_config = PeftConfig.from_pretrained(self.adapter_model_path)
+            load_path = peft_config.base_model_name_or_path
+
+        try:
+            self.model = AutoModelForCausalLM.from_pretrained(
+                load_path,
+                torch_dtype="auto",
+                quantization_config=bnb_config,
+                device_map="auto",
+                local_files_only=True,
+                trust_remote_code=True
+            )
+        except Exception as e:
+            self.logger.log_error(f"Failed to load base model: {e}")
+            raise e
+
+        # 3. Adapter 병합 (필요 시)
+        if self.is_adapter_model:
+            self.model = PeftModel.from_pretrained(self.model, self.adapter_model_path)
+
+        self.model.eval()
+
+        # 4. Tokenizer 로드
+        self.tokenizer = AutoTokenizer.from_pretrained(load_path, trust_remote_code=True)
+
+        if self.tokenizer.pad_token is None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+            self.logger.log_info("Pad token was None. Set to EOS token.")
+
+        self.supports_chat_template = (
+            hasattr(self.tokenizer, "chat_template") and
+            self.tokenizer.chat_template is not None
+        )
+
+        self.logger.log_info(f"Supports Chat Template: {self.supports_chat_template}")
+        if self.supports_chat_template:
+            self.logger.log_info(f"Chat Template Content:\n{self.tokenizer.chat_template}")
+
+    def _get_bnb_config(self) -> Union[BitsAndBytesConfig, None]:
+        if self.quantization == "int4":
+            return BitsAndBytesConfig(
+                load_in_4bit=True,
+                bnb_4bit_use_double_quant=True,
+                bnb_4bit_quant_type="nf4",
+                bnb_4bit_compute_dtype=torch.float16
+            )
+        elif self.quantization == "int8":
+            return BitsAndBytesConfig(
+                load_in_8bit=True,
+                llm_int8_threshold=6.0,
+                llm_int8_has_fp16_weight=True
+            )
+        return None
+
+    def execute(self, requests):
+        """Triton Inference Request 처리 메인 루프"""
+        responses = []
+
+        for request in requests:
+            # [ID 생성 로직] - 로그 추적용으로 유지 (Response에는 포함 X)
+            request_id = request.request_id()
+            if not request_id:
+                request_id = str(uuid.uuid4())
+
+            try:
+                # 1. 입력 데이터 파싱
+                input_data, is_chat = self._parse_input(request)
+
+                # [LOGGING] Request ID 포함하여 로그 출력
+                log_input_str = json.dumps(input_data, ensure_ascii=False) if isinstance(input_data, (list, dict)) else str(input_data)
+                self.logger.log_info(f"\n[RID: {request_id}] >>> [{'CHAT' if is_chat else 'TEXT'}][Input]: {log_input_str}")
+
+                # 2. Generation Config 생성
+                gen_config = self._create_generation_config(request)
+
+                # 3. 토크나이징
+                inputs = self._tokenize(input_data, is_chat)
+
+                # 4. 모델 추론 (Generate)
+                output_text = self._generate(inputs, gen_config)
+
+                # [LOGGING] Request ID 포함하여 결과 출력
+                self.logger.log_info(f"\n[RID: {request_id}] <<< [Output]: {output_text}")
+
+                # 5. 응답 생성
+                responses.append(self._create_response(output_text, request_id))
+
+            except Exception as e:
+                self.logger.log_error(f"[RID: {request_id}] Error during execution: {e}")
+                err_tensor = pb_utils.Tensor("text_output", np.array([str(e).encode('utf-8')], dtype=np.bytes_))
+                responses.append(pb_utils.InferenceResponse(output_tensors=[err_tensor]))
+
+        return responses
+
+    def _parse_input(self, request) -> Tuple[Union[str, List[Dict]], bool]:
+        input_text = self._get_input_scalar(request, "text_input")
+        try:
+            conversation = json.loads(input_text)
+            if isinstance(conversation, list):
+                return conversation, True
+        except (json.JSONDecodeError, TypeError):
+            pass
+        return input_text, False
+
+    def _tokenize(self, input_data, is_chat: bool):
+        if self.supports_chat_template and is_chat:
+            return self.tokenizer.apply_chat_template(
+                input_data,
+                tokenize=True,
+                add_generation_prompt=True,
+                return_tensors="pt",
+                return_dict=True
+            ).to(self.device)
+        else:
+            if is_chat:
+                 input_data = str(input_data)
+            return self.tokenizer(input_data, return_tensors="pt").to(self.device)
+
+    def _generate(self, inputs, gen_config: GenerationConfig) -> str:
+        input_ids = inputs["input_ids"]
+        input_len = input_ids.shape[-1]
+
+        with torch.no_grad():
+            outputs = self.model.generate(
+                **inputs,
+                generation_config=gen_config,
+                pad_token_id=self.tokenizer.pad_token_id,
+                eos_token_id=self.tokenizer.eos_token_id
+            )
+
+        generated_tokens = outputs[0][input_len:]
+        decoded_output = self.tokenizer.decode(generated_tokens, skip_special_tokens=True)
+        return decoded_output.strip()
+
+    def _create_generation_config(self, request) -> GenerationConfig:
+        def get_param(name, default=None, cast_type=None):
+            val = self._get_input_scalar(request, name, default)
+            if val is not None and cast_type:
+                return cast_type(val)
+            return val
+
+        return GenerationConfig(
+            max_length=get_param("max_length", 1024, int),
+            max_new_tokens=get_param("max_new_tokens", 256, int),
+            temperature=get_param("temperature", 1.0, float),
+            do_sample=get_param("do_sample", False, bool),
+            top_k=get_param("top_k", 50, int),
+            top_p=get_param("top_p", 1.0, float),
+            repetition_penalty=get_param("repetition_penalty", 1.0, float),
+        )
+
+    def _create_response(self, output_text: str, request_id: str):
+        """생성된 텍스트를 Triton Response 객체로 변환"""
+        output_tensor = pb_utils.Tensor(
+            "text_output",
+            np.array([output_text.encode('utf-8')], dtype=np.bytes_)
+        )
+        return pb_utils.InferenceResponse(output_tensors=[output_tensor])
+
+    def _get_config_param(self, key: str, default: str = None) -> str:
+        params = self.model_config.get('parameters', {})
+        if key in params:
+            return params[key].get('string_value', default)
+        return default
+
+    def _get_input_scalar(self, request, name: str, default=None):
+        tensor = pb_utils.get_input_tensor_by_name(request, name)
+        if tensor is None:
+            return default
+        return self._np_decoder(tensor.as_numpy()[0])
+
+    def _np_decoder(self, obj):
+        if isinstance(obj, bytes):
+           return obj.decode('utf-8')
+        if np.issubdtype(obj, np.integer):
+           return int(obj)
+        if np.issubdtype(obj, np.floating):
+           return round(float(obj), 3)
+        if isinstance(obj, np.bool_):
+           return bool(obj)
+
+    def finalize(self):
+        self.logger.log_info(f"Finalizing model {self.model_name}")
+        self.model = None
+        self.tokenizer = None
+        torch.cuda.empty_cache()

README.md

@@ -1,66 +0,0 @@
----
-license: apache-2.0
-datasets:
-- cerebras/SlimPajama-627B
-- bigcode/starcoderdata
-- HuggingFaceH4/ultrachat_200k
-- HuggingFaceH4/ultrafeedback_binarized
-language:
-- en
-widget:
-  - example_title: Fibonacci (Python)
-    messages:
-    - role: system
-      content: You are a chatbot who can help code!
-    - role: user
-      content: Write me a function to calculate the first 10 digits of the fibonacci sequence in Python and print it out to the CLI.
----
-<div align="center">
-
-# TinyLlama-1.1B
-</div>
-
-https://github.com/jzhang38/TinyLlama
-
-The TinyLlama project aims to **pretrain** a **1.1B Llama model on 3 trillion tokens**. With some proper optimization, we can achieve this within a span of "just" 90 days using 16 A100-40G GPUs 🚀🚀. The training has started on 2023-09-01. 
-
-
-We adopted exactly the same architecture and tokenizer as Llama 2. This means TinyLlama can be plugged and played in many open-source projects built upon Llama. Besides, TinyLlama is compact with only 1.1B parameters. This compactness allows it to cater to a multitude of applications demanding a restricted computation and memory footprint.
-
-#### This Model
-This is the chat model finetuned on top of [TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T](https://huggingface.co/TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T). **We follow [HF's Zephyr](https://huggingface.co/HuggingFaceH4/zephyr-7b-alpha)'s training recipe.** The model was " initially fine-tuned on a variant of the [`UltraChat`](https://huggingface.co/datasets/stingning/ultrachat) dataset, which contains a diverse range of synthetic dialogues generated by ChatGPT. 
-We then further aligned the model with [🤗 TRL's](https://github.com/huggingface/trl) `DPOTrainer` on the [openbmb/UltraFeedback](https://huggingface.co/datasets/openbmb/UltraFeedback) dataset, which contain 64k prompts and model completions that are ranked by GPT-4." 
-
-
-#### How to use
-You will need the transformers>=4.34
-Do check the [TinyLlama](https://github.com/jzhang38/TinyLlama) github page for more information.
-
-```python
-# Install transformers from source - only needed for versions <= v4.34
-# pip install git+https://github.com/huggingface/transformers.git
-# pip install accelerate
-
-import torch
-from transformers import pipeline
-
-pipe = pipeline("text-generation", model="TinyLlama/TinyLlama-1.1B-Chat-v1.0", torch_dtype=torch.bfloat16, device_map="auto")
-
-# We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating
-messages = [
-    {
-        "role": "system",
-        "content": "You are a friendly chatbot who always responds in the style of a pirate",
-    },
-    {"role": "user", "content": "How many helicopters can a human eat in one sitting?"},
-]
-prompt = pipe.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
-outputs = pipe(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
-print(outputs[0]["generated_text"])
-# <|system|>
-# You are a friendly chatbot who always responds in the style of a pirate.</s>
-# <|user|>
-# How many helicopters can a human eat in one sitting?</s>
-# <|assistant|>
-# ...
-```

config.json

@@ -1,26 +0,0 @@
-{
-  "architectures": [
-    "LlamaForCausalLM"
-  ],
-  "attention_bias": false,
-  "bos_token_id": 1,
-  "eos_token_id": 2,
-  "hidden_act": "silu",
-  "hidden_size": 2048,
-  "initializer_range": 0.02,
-  "intermediate_size": 5632,
-  "max_position_embeddings": 2048,
-  "model_type": "llama",
-  "num_attention_heads": 32,
-  "num_hidden_layers": 22,
-  "num_key_value_heads": 4,
-  "pretraining_tp": 1,
-  "rms_norm_eps": 1e-05,
-  "rope_scaling": null,
-  "rope_theta": 10000.0,
-  "tie_word_embeddings": false,
-  "torch_dtype": "bfloat16",
-  "transformers_version": "4.35.0",
-  "use_cache": true,
-  "vocab_size": 32000
-}

config.pbtxt

@@ -0,0 +1,131 @@
+# Triton Backend for TransformerLLM.
+backend: "python"
+max_batch_size: 0
+
+# Triton should expect as input a single string
+# input of variable length named 'text_input'
+input [
+
+  {
+    name: "text_input"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+
+
+  },
+  {
+    name: "max_length"
+    data_type: TYPE_INT32
+    dims: [ 1 ]
+
+    optional: true
+
+
+  },
+  {
+    name: "max_new_tokens"
+    data_type: TYPE_INT32
+    dims: [ 1 ]
+
+    optional: true
+
+
+  },
+  {
+    name: "do_sample"
+    data_type: TYPE_BOOL
+    dims: [ 1 ]
+
+    optional: true
+
+
+  },
+  {
+    name: "top_k"
+    data_type: TYPE_INT32
+    dims: [ 1 ]
+
+    optional: true
+
+
+  },
+  {
+    name: "top_p"
+    data_type: TYPE_FP32
+    dims: [ 1 ]
+
+    optional: true
+
+
+  },
+  {
+    name: "temperature"
+    data_type: TYPE_FP32
+    dims: [ 1 ]
+
+    optional: true
+
+
+  },
+  {
+    name: "repetition_penalty"
+    data_type: TYPE_FP32
+    dims: [ 1 ]
+
+    optional: true
+
+
+  },
+  {
+    name: "stream"
+    data_type: TYPE_BOOL
+    dims: [ 1 ]
+
+    optional: true
+
+
+  }
+
+]
+
+
+# Triton should expect to respond with a single string
+# output of variable length named 'text_output'
+output [
+
+  {
+    name: "text_output"
+    data_type: TYPE_STRING
+    dims: [ 1 ]
+
+  }
+
+]
+
+parameters: [
+  {
+    key: "base_model_path",
+    value: {string_value: "/cheetah/input/model/groupuser/TinyLlama-1.1B-Chat-v1.0"}
+  },
+  {
+    key: "is_adapter_model",
+    value: {string_value: "false"}
+  },
+  {
+    key: "adapter_model_path",
+    value: {string_value: ""}
+  },
+  
+  {
+    key: "quantization",
+    value: {string_value: "none"}
+  }
+]
+
+instance_group [
+  {
+    kind: KIND_AUTO
+    count: 1
+  }
+]
+

eval_results.json

@@ -1,16 +0,0 @@
-{
-    "epoch": 3.0,
-    "eval_logits/chosen": -2.707406759262085,
-    "eval_logits/rejected": -2.656524419784546,
-    "eval_logps/chosen": -370.1297607421875,
-    "eval_logps/rejected": -296.0738525390625,
-    "eval_loss": 0.513750433921814,
-    "eval_rewards/accuracies": 0.738095223903656,
-    "eval_rewards/chosen": -0.02744222804903984,
-    "eval_rewards/margins": 1.0087225437164307,
-    "eval_rewards/rejected": -1.03616464138031,
-    "eval_runtime": 93.5908,
-    "eval_samples": 2000,
-    "eval_samples_per_second": 21.37,
-    "eval_steps_per_second": 0.673
-}

generation_config.json

@@ -1,7 +0,0 @@
-{
-  "bos_token_id": 1,
-  "eos_token_id": 2,
-  "max_length": 2048,
-  "pad_token_id": 0,
-  "transformers_version": "4.35.0"
-}

special_tokens_map.json

@@ -1,30 +0,0 @@
-{
-  "bos_token": {
-    "content": "<s>",
-    "lstrip": false,
-    "normalized": false,
-    "rstrip": false,
-    "single_word": false
-  },
-  "eos_token": {
-    "content": "</s>",
-    "lstrip": false,
-    "normalized": false,
-    "rstrip": false,
-    "single_word": false
-  },
-  "pad_token": {
-    "content": "</s>",
-    "lstrip": false,
-    "normalized": false,
-    "rstrip": false,
-    "single_word": false
-  },
-  "unk_token": {
-    "content": "<unk>",
-    "lstrip": false,
-    "normalized": false,
-    "rstrip": false,
-    "single_word": false
-  }
-}

tokenizer_config.json

@@ -1,40 +0,0 @@
-{
-  "added_tokens_decoder": {
-    "0": {
-      "content": "<unk>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "1": {
-      "content": "<s>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    },
-    "2": {
-      "content": "</s>",
-      "lstrip": false,
-      "normalized": false,
-      "rstrip": false,
-      "single_word": false,
-      "special": true
-    }
-  },
-  "bos_token": "<s>",
-  "chat_template": "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\n'  + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}",
-  "clean_up_tokenization_spaces": false,
-  "eos_token": "</s>",
-  "legacy": false,
-  "model_max_length": 2048,
-  "pad_token": "</s>",
-  "padding_side": "right",
-  "sp_model_kwargs": {},
-  "tokenizer_class": "LlamaTokenizer",
-  "unk_token": "<unk>",
-  "use_default_system_prompt": false
-}