eval with merged model

competition/12_InterLM_Merged_Model_M3.ipynb

@@ -0,0 +1 @@
+{"cells":[{"cell_type":"code","execution_count":1,"metadata":{"executionInfo":{"elapsed":476,"status":"ok","timestamp":1720679526275,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"uWKRSV6eZsCn"},"outputs":[],"source":["%load_ext autoreload\n","%autoreload 2"]},{"cell_type":"code","execution_count":2,"metadata":{"application/vnd.databricks.v1+cell":{"cellMetadata":{"byteLimit":2048000,"rowLimit":10000},"inputWidgets":{},"nuid":"eb33b19f-1206-41ee-84e2-e6258a12eef7","showTitle":false,"title":""},"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":2534,"status":"ok","timestamp":1720679529344,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"xwFh14uiZBrI","outputId":"d767799c-34c2-46a5-f052-378146a55321"},"outputs":[],"source":["from pathlib import Path\n","\n","try:\n","    from google.colab import drive\n","\n","    drive.mount(\"/content/drive\")\n","    workding_dir = \"/content/drive/MyDrive/logical-reasoning/\"\n","except ModuleNotFoundError:\n","    workding_dir = str(Path.cwd().parent)"]},{"cell_type":"code","execution_count":3,"metadata":{"application/vnd.databricks.v1+cell":{"cellMetadata":{"byteLimit":2048000,"rowLimit":10000},"inputWidgets":{},"nuid":"6d394937-6c99-4a7c-9d32-7600a280032f","showTitle":false,"title":""},"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":5,"status":"ok","timestamp":1720679529345,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"G5pNu3zgZBrL","outputId":"160a554f-fb08-4aa0-bc00-0422fb7c1fac"},"outputs":[{"name":"stdout","output_type":"stream","text":["workding dir: /Users/inflaton/code/engd/projects/logical-reasoning\n"]}],"source":["import os\n","import sys\n","from pathlib import Path\n","\n","os.chdir(workding_dir)\n","sys.path.append(workding_dir)\n","print(\"workding dir:\", workding_dir)"]},{"cell_type":"code","execution_count":4,"metadata":{"application/vnd.databricks.v1+cell":{"cellMetadata":{"byteLimit":2048000,"rowLimit":10000},"inputWidgets":{},"nuid":"9f67ec60-2f24-411c-84eb-0dd664b44775","showTitle":false,"title":""},"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":3,"status":"ok","timestamp":1720679529345,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"hPCC-6m7ZBrM","outputId":"c7aa2c96-5e99-440a-c148-201d79465ff9"},"outputs":[{"name":"stdout","output_type":"stream","text":["loading env vars from: /Users/inflaton/code/engd/projects/logical-reasoning/.env\n"]},{"data":{"text/plain":["True"]},"execution_count":4,"metadata":{},"output_type":"execute_result"}],"source":["from dotenv import find_dotenv, load_dotenv\n","\n","found_dotenv = find_dotenv(\".env\")\n","\n","if len(found_dotenv) == 0:\n","    found_dotenv = find_dotenv(\".env.example\")\n","print(f\"loading env vars from: {found_dotenv}\")\n","load_dotenv(found_dotenv, override=True)"]},{"cell_type":"code","execution_count":5,"metadata":{"application/vnd.databricks.v1+cell":{"cellMetadata":{"byteLimit":2048000,"rowLimit":10000},"inputWidgets":{},"nuid":"f1597656-8042-4878-9d3b-9ebfb8dd86dc","showTitle":false,"title":""},"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":3,"status":"ok","timestamp":1720679529345,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"1M3IraVtZBrM","outputId":"29ab35f6-2970-4ade-d85d-3174acf8cda0"},"outputs":[{"name":"stdout","output_type":"stream","text":["llama-factory/merged_models/internlm2_5-7b-chat-1m_sft_bf16_p2_full None False datasets/mgtv results/mgtv-results_merged_model.csv\n"]}],"source":["import os\n","\n","model_name = os.getenv(\"MODEL_NAME\")\n","adapter_name_or_path = os.getenv(\"ADAPTER_NAME_OR_PATH\")\n","load_in_4bit = os.getenv(\"LOAD_IN_4BIT\") == \"true\"\n","data_path = os.getenv(\"LOGICAL_REASONING_DATA_PATH\")\n","results_path = os.getenv(\"LOGICAL_REASONING_RESULTS_PATH\")\n","use_english_datasets = os.getenv(\"USE_ENGLISH_DATASETS\") == \"true\"\n","\n","\n","print(model_name, adapter_name_or_path, load_in_4bit, data_path, results_path)"]},{"cell_type":"code","execution_count":6,"metadata":{"application/vnd.databricks.v1+cell":{"cellMetadata":{"byteLimit":2048000,"rowLimit":10000},"inputWidgets":{},"nuid":"b2a43943-9324-4839-9a47-cfa72de2244b","showTitle":false,"title":""},"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":564,"status":"ok","timestamp":1720679529907,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"UgMvt6dIZBrM","outputId":"ce37581c-fd26-46c2-ad87-d933d99f68f7"},"outputs":[{"name":"stdout","output_type":"stream","text":["Python 3.11.9\n","\u001b[33mWARNING: Package(s) not found: flash-attn\u001b[0m\u001b[33m\n","\u001b[0mCPU times: user 3.5 ms, sys: 6.71 ms, total: 10.2 ms\n","Wall time: 634 ms\n"]}],"source":["%%time\n","!python --version\n","!pip show flash-attn"]},{"cell_type":"code","execution_count":7,"metadata":{"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":1685,"status":"ok","timestamp":1720679531591,"user":{"displayName":"HUANG DONGHAO _","userId":"00977795705617022768"},"user_tz":-480},"id":"ZuS_FsLyZBrN","outputId":"2cba0105-c505-4395-afbd-2f2fee6581d0"},"outputs":[{"name":"stdout","output_type":"stream","text":["loading /Users/inflaton/code/engd/projects/logical-reasoning/llm_toolkit/logical_reasoning_utils.py\n","MPS is available\n"]}],"source":["from llm_toolkit.llm_utils import *\n","from llm_toolkit.logical_reasoning_utils import *\n","\n","device = check_gpu()"]},{"cell_type":"code","execution_count":8,"metadata":{},"outputs":[{"name":"stderr","output_type":"stream","text":["Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.\n"]},{"name":"stdout","output_type":"stream","text":["loading model: llama-factory/merged_models/internlm2_5-7b-chat-1m_sft_bf16_p2_full\n"]},{"data":{"application/vnd.jupyter.widget-view+json":{"model_id":"6f6cf944680d466d817f8f57fd56cd5b","version_major":2,"version_minor":0},"text/plain":["Loading checkpoint shards:   0%|          | 0/8 [00:00<?, ?it/s]"]},"metadata":{},"output_type":"display_data"},{"name":"stdout","output_type":"stream","text":["CPU times: user 782 ms, sys: 3.73 s, total: 4.51 s\n","Wall time: 30.3 s\n"]}],"source":["%%time\n","\n","model, tokenizer = load_model(model_name)"]},{"cell_type":"code","execution_count":9,"metadata":{},"outputs":[{"name":"stdout","output_type":"stream","text":["loading train/test data files\n","DatasetDict({\n","    train: Dataset({\n","        features: ['text', 'label', 'answer', 'title', 'puzzle', 'truth', 'train_text', 'prompt'],\n","        num_rows: 25000\n","    })\n","    test: Dataset({\n","        features: ['text', 'label', 'answer', 'title', 'puzzle', 'truth', 'train_text', 'prompt'],\n","        num_rows: 3000\n","    })\n","})\n"]}],"source":["datasets = load_logical_reasoning_dataset(\n","    data_path,\n","    tokenizer=tokenizer,\n","    chinese_prompt=not use_english_datasets,\n","    using_p1=False,\n",")"]},{"cell_type":"code","execution_count":10,"metadata":{},"outputs":[{"name":"stdout","output_type":"stream","text":["<s><|im_start|>system\n","You are an expert in logical reasoning.<|im_end|>\n","<|im_start|>user\n","你是一个情景猜谜游戏的主持人。游戏规则如下：\n","\n","1. 参与者会得到一个谜面，谜面会描述一个简单又难以理解的事件。\n","2. 主持人知道谜底，谜底是谜面的答案。\n","3. 参与者可以询问任何封闭式问题来找寻事件的真相。\n","4. 对于每个问题，主持人将根据实际情况回答以下五个选项之一：是、不是、不重要、回答正确、问法错误。各回答的判断标准如下：\n","   - 若谜面和谜底能找到问题的答案，回答：是或者不是\n","   - 若谜面和谜底不能直接或者间接推断出问题的答案，回答：不重要\n","   - 若参与者提问不是一个封闭式问题或者问题难以理解，回答：问法错误\n","   - 若参与者提问基本还原了谜底真相，回答：回答正确\n","5. 回答中不能添加任何其它信息，也不能省略选项中的任何一个字。例如，不可以把“不是”省略成“不”。\n","\n","请严格按照这些规则回答参与者提出的问题。\n","\n","**谜面:** 在一个安静的夜晚，小村庄的湖边突然传来了阵阵哭泣声。第二天早晨，村长甄锐发现湖边的石头上放着一顶破旧的帽子，但没有人知道这顶帽子是从哪里来的，哭泣声又是为何。请还原故事真相。\n","\n","**谜底:** 原来，这顶破旧的帽子属于一个小男孩，他小时候与爷爷在湖边生活。爷爷教他钓鱼、游泳，还告诉他湖中的海龟是他们的朋友。后来，小男孩随父母去了城市生活，但每年夏天都会回到村子探望爷爷。然而，去年夏天，爷爷因病去世，小男孩伤心欲绝。今年夏天，他回到村子，来到湖边，想起和爷爷的美好回忆，忍不住哭泣。他将爷爷的帽子放在湖边的石头上，希望能让爷爷的在天之灵得到安慰。那晚的哭泣声正是小男孩在祭莫他亲爱的爷爷。\n","\n","**参与者提出的问题:** 哭泣和村庄有关系吗\n","<|im_end|>\n","<|im_start|>assistant\n","\n","----------------------------------------\n","<s><s><|im_start|>system\n","You are an expert in logical reasoning.<|im_end|>\n","<|im_start|>user\n","你是一个情景猜谜游戏的主持人。游戏规则如下：\n","\n","1. 参与者会得到一个谜面，谜面会描述一个简单又难以理解的事件。\n","2. 主持人知道谜底，谜底是谜面的答案。\n","3. 参与者可以询问任何封闭式问题来找寻事件的真相。\n","4. 对于每个问题，主持人将根据实际情况回答以下五个选项之一：是、不是、不重要、回答正确、问法错误。各回答的判断标准如下：\n","   - 若谜面和谜底能找到问题的答案，回答：是或者不是\n","   - 若谜���和谜底不能直接或者间接推断出问题的答案，回答：不重要\n","   - 若参与者提问不是一个封闭式问题或者问题难以理解，回答：问法错误\n","   - 若参与者提问基本还原了谜底真相，回答：回答正确\n","5. 回答中不能添加任何其它信息，也不能省略选项中的任何一个字。例如，不可以把“不是”省略成“不”。\n","\n","请严格按照这些规则回答参与者提出的问题。\n","\n","**谜面:** 在一个安静的夜晚，小村庄的湖边突然传来了阵阵哭泣声。第二天早晨，村长甄锐发现湖边的石头上放着一顶破旧的帽子，但没有人知道这顶帽子是从哪里来的，哭泣声又是为何。请还原故事真相。\n","\n","**谜底:** 原来，这顶破旧的帽子属于一个小男孩，他小时候与爷爷在湖边生活。爷爷教他钓鱼、游泳，还告诉他湖中的海龟是他们的朋友。后来，小男孩随父母去了城市生活，但每年夏天都会回到村子探望爷爷。然而，去年夏天，爷爷因病去世，小男孩伤心欲绝。今年夏天，他回到村子，来到湖边，想起和爷爷的美好回忆，忍不住哭泣。他将爷爷的帽子放在湖边的石头上，希望能让爷爷的在天之灵得到安慰。那晚的哭泣声正是小男孩在祭莫他亲爱的爷爷。\n","\n","**参与者提出的问题:** 哭泣和村庄有关系吗\n","<|im_end|>\n","<|im_start|>assistant\n","是</s>\n","CPU times: user 624 ms, sys: 865 ms, total: 1.49 s\n","Wall time: 8.11 s\n"]}],"source":["%%time\n","\n","prompt1 = datasets[\"test\"][\"prompt\"][1000]\n","print(prompt1)\n","print(\"--\" * 20)\n","test_model(model, tokenizer, prompt1, device=device)"]},{"cell_type":"code","execution_count":11,"metadata":{},"outputs":[{"name":"stderr","output_type":"stream","text":[" 12%|█▎        | 1/8 [00:02<00:18,  2.59s/it]"]},{"name":"stdout","output_type":"stream","text":["--------\n","step 1: 不是</s>\n","--------\n","step 2: 不是\n","--------\n","step 3: 不是\n","--------\n","step 4: 不是\n","--------\n","step 5: 不是\n"]},{"name":"stderr","output_type":"stream","text":["100%|██████████| 8/8 [00:17<00:00,  2.17s/it]"]},{"name":"stdout","output_type":"stream","text":["CPU times: user 2.59 s, sys: 483 ms, total: 3.07 s\n","Wall time: 17.4 s\n"]},{"name":"stderr","output_type":"stream","text":["\n"]},{"data":{"text/plain":["['不是', '是', '是', '是', '不是', '是', '是', '不是']"]},"execution_count":11,"metadata":{},"output_type":"execute_result"}],"source":["%%time\n","\n","eval_model(model, tokenizer, datasets[\"test\"][:10], device=device)"]},{"cell_type":"code","execution_count":14,"metadata":{"id":"L370pvGTZBrN"},"outputs":[],"source":["def evaluate_model(model, tokenizer, model_name, dataset):\n","    print(\"Evaluating model: \" + model_name)\n","    predictions = eval_model(model, tokenizer, dataset, device=device)\n","\n","    model_name += \"_M3\"\n","\n","    save_results(\n","        model_name,\n","        results_path,\n","        dataset,\n","        predictions,\n","        debug=True,\n","    )\n","\n","    metrics = calc_metrics(dataset[\"label\"], predictions, debug=True)\n","    print(metrics)"]},{"cell_type":"code","execution_count":16,"metadata":{"id":"WUFjhxmiZBrN"},"outputs":[{"name":"stdout","output_type":"stream","text":["Evaluating model: llama-factory/merged_models/internlm2_5-7b-chat-1m_sft_bf16_p2_full\n"]},{"name":"stderr","output_type":"stream","text":["  0%|          | 1/3000 [00:02<1:55:37,  2.31s/it]"]},{"name":"stdout","output_type":"stream","text":["--------\n","step 1: 不是</s>\n","--------\n","step 2: 不是\n","--------\n","step 3: 不是\n","--------\n","step 4: 不是\n","--------\n","step 5: 不是\n"]},{"name":"stderr","output_type":"stream","text":["  5%|▍         | 148/3000 [09:35<2:17:04,  2.88s/it]"]}],"source":["%%time\n","\n","evaluate_model(model, tokenizer, model_name, datasets[\"test\"])"]}],"metadata":{"accelerator":"GPU","application/vnd.databricks.v1+notebook":{"dashboards":[],"environmentMetadata":null,"language":"python","notebookMetadata":{"mostRecentlyExecutedCommandWithImplicitDF":{"commandId":-1,"dataframes":["_sqldf"]},"pythonIndentUnit":4},"notebookName":"10_eval-lf-medium-py3.11","widgets":{}},"colab":{"gpuType":"L4","provenance":[]},"kernelspec":{"display_name":"Python 3","name":"python3"},"language_info":{"codemirror_mode":{"name":"ipython","version":3},"file_extension":".py","mimetype":"text/x-python","name":"python","nbconvert_exporter":"python","pygments_lexer":"ipython3","version":"3.11.9"}},"nbformat":4,"nbformat_minor":0}

llm_toolkit/llm_utils.py

@@ -1,14 +1,13 @@
 import os
 import re
-import sys
 import torch
-from llamafactory.chat import ChatModel
 from transformers import (
     AutoModelForCausalLM,
     AutoTokenizer,
     BitsAndBytesConfig,
     TextStreamer,
 )
+from tqdm import tqdm
 
 
 def load_model(
@@ -21,6 +20,8 @@ def load_model(
     print(f"loading model: {model_name}")
 
     if adapter_name_or_path:
+        from llamafactory.chat import ChatModel
+
         template = "llama3" if "llama-3" in model_name.lower() else "chatml"
 
         args = dict(
@@ -32,7 +33,9 @@ def load_model(
         )
         chat_model = ChatModel(args)
         if os.getenv("RESIZE_TOKEN_EMBEDDINGS") == "true":
-            chat_model.engine.model.resize_token_embeddings(len(chat_model.engine.tokenizer), pad_to_multiple_of=32)
+            chat_model.engine.model.resize_token_embeddings(
+                len(chat_model.engine.tokenizer), pad_to_multiple_of=32
+            )
         return chat_model.engine.model, chat_model.engine.tokenizer
 
     tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
@@ -63,11 +66,28 @@ def load_model(
     return model, tokenizer
 
 
-def test_model(model, tokenizer, prompt):
+def check_gpu():
+    # torch.cuda.is_available() checks and returns a Boolean True if a GPU is available, else it'll return False
+    is_cuda = torch.cuda.is_available()
+
+    # If we have a GPU available, we'll set our device to GPU. We'll use this device variable later in our code.
+    if is_cuda:
+        device = torch.device("cuda")
+        print("GPU is available")
+    elif torch.backends.mps.is_available():
+        device = torch.device("mps")
+        print("MPS is available")
+    else:
+        device = torch.device("cpu")
+        print("GPU/MPS not available, CPU used")
+    return device
+
+
+def test_model(model, tokenizer, prompt, device="cuda"):
     inputs = tokenizer(
         [prompt],
         return_tensors="pt",
-    ).to("cuda")
+    ).to(device)
 
     text_streamer = TextStreamer(tokenizer)
 
@@ -114,14 +134,14 @@ def extract_answer(text, debug=False):
     return text
 
 
-def eval_model(model, tokenizer, eval_dataset):
+def eval_model(model, tokenizer, eval_dataset, device="cuda"):
     total = len(eval_dataset)
     predictions = []
     for i in tqdm(range(total)):
         inputs = tokenizer(
             eval_dataset["prompt"][i : i + 1],
             return_tensors="pt",
-        ).to("cuda")
+        ).to(device)
 
         outputs = model.generate(**inputs, max_new_tokens=4096, use_cache=False)
         decoded_output = tokenizer.batch_decode(outputs)

llm_toolkit/logical_reasoning_utils.py

@@ -4,7 +4,6 @@ import pandas as pd
 import seaborn as sns
 import matplotlib.pyplot as plt
 from datasets import load_dataset
-from tqdm import tqdm
 
 print(f"loading {__file__}")
 
@@ -203,87 +202,6 @@ Please strictly follow these rules when answering the participant's questions.
     return datasets
 
 
-def eval_model(model, tokenizer, eval_dataset):
-    total = len(eval_dataset)
-    predictions = []
-    for i in tqdm(range(total)):
-        inputs = tokenizer(
-            eval_dataset["prompt"][i : i + 1],
-            return_tensors="pt",
-        ).to("cuda")
-
-        outputs = model.generate(**inputs, max_new_tokens=4096, use_cache=False)
-        decoded_output = tokenizer.batch_decode(outputs)
-        debug = i == 0
-        decoded_output = [
-            extract_answer(output, debug=debug) for output in decoded_output
-        ]
-        predictions.extend(decoded_output)
-
-    return predictions
-
-
-def save_model(
-    model,
-    tokenizer,
-    include_gguf=True,
-    include_merged=True,
-    publish=True,
-):
-    try:
-        token = os.getenv("HF_TOKEN") or None
-        model_name = os.getenv("MODEL_NAME")
-
-        save_method = "lora"
-        quantization_method = "q5_k_m"
-
-        model_names = get_model_names(
-            model_name, save_method=save_method, quantization_method=quantization_method
-        )
-
-        model.save_pretrained(model_names["local"])
-        tokenizer.save_pretrained(model_names["local"])
-
-        if publish:
-            model.push_to_hub(
-                model_names["hub"],
-                token=token,
-            )
-            tokenizer.push_to_hub(
-                model_names["hub"],
-                token=token,
-            )
-
-        if include_merged:
-            model.save_pretrained_merged(
-                model_names["local"] + "-merged", tokenizer, save_method=save_method
-            )
-            if publish:
-                model.push_to_hub_merged(
-                    model_names["hub"] + "-merged",
-                    tokenizer,
-                    save_method="lora",
-                    token="",
-                )
-
-        if include_gguf:
-            model.save_pretrained_gguf(
-                model_names["local-gguf"],
-                tokenizer,
-                quantization_method=quantization_method,
-            )
-
-            if publish:
-                model.push_to_hub_gguf(
-                    model_names["hub-gguf"],
-                    tokenizer,
-                    quantization_method=quantization_method,
-                    token=token,
-                )
-    except Exception as e:
-        print(e)
-
-
 def get_metrics(df):
     metrics_df = pd.DataFrame(df.columns.T)[2:]
     metrics_df.rename(columns={0: "model"}, inplace=True)

requirements.txt

@@ -12,5 +12,7 @@ ipywidgets
 packaging
 langchain_openai==0.1.13
 wandb==0.17.4
-# triton
-# xformers
+transformers==4.41.2
+sentencepiece==0.2.0
+einops==0.8.0
+accelerate==0.32.1