feat: generate prompts from datasets

README.md

@@ -50,7 +50,7 @@ tracestorm --model "Qwen/Qwen2.5-1.5B-Instruct" --pattern azure_code
 #### Example Command for Loading Prompts from Datasets
 
 ```bash
-tracestorm --model "Qwen/Qwen2.5-1.5B-Instruct" --duration 30 --datasets-config-file ./examples/datasets_config_hf.json
+tracestorm --model "Qwen/Qwen2.5-1.5B-Instruct" --duration 30 --datasets-config ./examples/datasets_config_hf.json
 ```
 
 
@@ -60,7 +60,7 @@ tracestorm --model "Qwen/Qwen2.5-1.5B-Instruct" --duration 30 --datasets-config-
 - Refer to `./examples/datasets_config_local.json` for an example configuration.
 - If you want to test loading from local files, please run `./examples/test_data_loader.py` first to download and save two datasets.
 
-2. Remote datasets from Hugging Face 
+2. Remote datasets from Hugging Face
 - Refer to `./examples/datasets_config_hf.json` for an example configuration.
 
 **Sorting Strategy**: Defines how prompts from multiple datasets are ordered
@@ -85,6 +85,6 @@ Please check `./examples/datasets_config_default.json` for required fields in `d
 - `--base-url`: Optional. OpenAI Base URL (default is `http://localhost:8000/v1`).
 - `--api-key`: Optional. OpenAI API Key (default is `none`).
 - `--seed`: Optional. Random seed for trace pattern reproducibility (default is `none`).
-- `--datasets-config-file`: Optional. Configuration file for loading prompt messages from provided datasets. Uses `DEFAULT_MESSAGES` is not specified.
+- `--datasets-config`: Optional. Configuration file for loading prompt messages from provided datasets. Uses `DEFAULT_MESSAGES` is not specified.
 
 Make sure to adjust the parameters according to your testing needs!

examples/datasets_config_default.json

@@ -0,0 +1,14 @@
+{
+    "sort_strategy": "random",
+    "dataset_1": {
+        "file_name": "",
+        "prompt_field": "",
+        "select_ratio": 1,
+        "split": "train"
+    },
+    "dataset_2": {
+        "file_name": "",
+        "prompt_field": "",
+        "select_ratio": 1
+    }
+}

examples/datasets_config_hf.json

@@ -0,0 +1,15 @@
+{
+    "sort_strategy": "original",
+    "dataset_1": {
+        "file_name": "hf://datasets/fka/awesome-chatgpt-prompts/prompts.csv",
+        "prompt_field": "prompt",
+        "select_ratio": 2,
+        "split": "train"
+    },
+    "dataset_2": {
+        "file_name": "MAsad789565/Coding_GPT4_Data",
+        "prompt_field": "user",
+        "select_ratio": 8,
+        "split": "train"
+    }
+}

examples/datasets_config_local.json

@@ -0,0 +1,13 @@
+{
+    "sort_strategy": "random",
+    "dataset_1": {
+        "file_name": "Conversational_dataset.jsonl",
+        "prompt_field": "messages",
+        "select_ratio": 6
+    },
+    "dataset_2": {
+        "file_name": "~/.cache/tracestorm/GPT4_coding_sample.csv",
+        "prompt_field": "user",
+        "select_ratio": 4
+    }
+}

examples/save_test_datasets.py

@@ -0,0 +1,28 @@
+import os
+
+import pandas as pd
+
+from tracestorm.constants import DEFAULT_DATASET_FOLDER
+
+
+def prepare_test_datasets():
+    df1 = pd.read_json(
+        "hf://datasets/MAsad789565/Coding_GPT4_Data/Data/GPT_4_Coding.json"
+    )
+    df2 = pd.read_json(
+        "hf://datasets/olathepavilion/Conversational-datasets-json/Validation.jsonl",
+        lines=True,
+    )
+
+    # save the pre-processed dataset to the default folder for test
+    os.makedirs(DEFAULT_DATASET_FOLDER, exist_ok=True)
+    path1 = os.path.join(DEFAULT_DATASET_FOLDER, "GPT4_coding_sample.csv")
+    path2 = os.path.join(DEFAULT_DATASET_FOLDER, "Conversational_dataset.jsonl")
+
+    # test with different file formats
+    df1.to_csv(path1, index=False)
+    df2.to_json(path2, orient="records", lines=True)
+
+
+if __name__ == "__main__":
+    prepare_test_datasets()

pyproject.toml

@@ -12,9 +12,11 @@ authors = [
 ]
 
 dependencies = [
+    "datasets>=3.3.2",
     "openai>=1.58.0",
     "numpy>=1.26.4",
     "pandas>=2.2.3",
+    "requests>=2.31.0",
     "seaborn>=0.13.2",
     "matplotlib>=3.9",
     "click>=8.1.8"
@@ -43,4 +45,7 @@ ignore = ["B007"]  # Loop control variable not used within loop body
 
 [tool.isort]
 use_parentheses = true
-skip_gitignore = true
+skip_gitignore = true
+
+[tool.setuptools]
+packages = { find = { exclude = ["examples"] } }

requirements.txt

@@ -1,5 +1,8 @@
+click>=8.1.8
+datasets>=3.3.2
 matplotlib>=3.9
 numpy>=1.26.4
+openai>=1.58.0
 pandas>=2.2.3
 requests>=2.31.0
 seaborn>=0.13.2

tests/test_cli.py

@@ -81,7 +81,7 @@ def test_cli_invalid_pattern(self):
         )
 
         self.assertNotEqual(result.exit_code, 0)
-        self.assertIn("Invalid pattern", result.output)
+        self.assertIn("Invalid value for '--pattern'", result.output)
 
 
 if __name__ == "__main__":

tests/test_data_loader.py

@@ -0,0 +1,63 @@
+import os
+import unittest
+
+import pandas as pd
+
+from tracestorm.constants import DEFAULT_DATASET_FOLDER
+from tracestorm.data_loader import Dataset, load_datasets
+
+
+class TestDataLoader(unittest.TestCase):
+    def test_remote_files(self):
+        """
+        Test loading datasets from hugging face.
+        There are 2 datasets, testing for:
+        1. loading with datasets.load_dataset
+        2. loading csv format with pandas
+        """
+        datasets, sort = load_datasets("examples/datasets_config_hf.json")
+        assert isinstance(datasets, list)
+        assert isinstance(datasets[0], Dataset) and isinstance(
+            datasets[1], Dataset
+        )
+        assert sort == "original"
+        assert len(datasets) == 2
+        assert datasets[0].select_ratio == 2 and datasets[1].select_ratio == 8
+        assert datasets[0].length > 0 and datasets[1].length > 0
+
+    def test_local_files(self):
+        """Test loading from local files"""
+
+        os.makedirs(DEFAULT_DATASET_FOLDER, exist_ok=True)
+        # testing datasets
+        df1 = pd.read_json(
+            "hf://datasets/MAsad789565/Coding_GPT4_Data/Data/GPT_4_Coding.json"
+        )
+        df2 = pd.read_json(
+            "hf://datasets/olathepavilion/Conversational-datasets-json/Validation.jsonl",
+            lines=True,
+        )
+
+        # test with different file formats
+        path1 = os.path.join(DEFAULT_DATASET_FOLDER, "GPT4_coding_sample.csv")
+        path2 = os.path.join(
+            DEFAULT_DATASET_FOLDER, "Conversational_dataset.jsonl"
+        )
+
+        # save the pre-processed dataset to the default folder for test
+        df1.to_csv(path1, index=False)
+        df2.to_json(path2, orient="records", lines=True)
+
+        datasets, sort = load_datasets("examples/datasets_config_local.json")
+        assert isinstance(datasets, list)
+        assert isinstance(datasets[0], Dataset) and isinstance(
+            datasets[1], Dataset
+        )
+        assert sort == "random"
+        assert len(datasets) == 2
+        assert datasets[0].select_ratio == 6 and datasets[1].select_ratio == 4
+        assert datasets[0].length > 0 and datasets[1].length > 0
+
+
+if __name__ == "__main__":
+    unittest.main()

tracestorm/cli.py

@@ -1,9 +1,10 @@
 import os
-from typing import Tuple
+from typing import Optional, Tuple
 
 import click
 
 from tracestorm.core import run_load_test
+from tracestorm.data_loader import load_datasets
 from tracestorm.logger import init_logger
 from tracestorm.trace_generator import (
     AzureTraceGenerator,
@@ -14,13 +15,13 @@
 logger = init_logger(__name__)
 
 # Valid patterns
-SYNTHETIC_PATTERNS = {"uniform"}
+SYNTHETIC_PATTERNS = {"uniform", "poisson", "random"}
 AZURE_PATTERNS = {"azure_code", "azure_conv"}
 VALID_PATTERNS = SYNTHETIC_PATTERNS | AZURE_PATTERNS
 
 
 def create_trace_generator(
-    pattern: str, rps: int, duration: int
+    pattern: str, rps: int, duration: int, seed: Optional[int] = None
 ) -> Tuple[TraceGenerator, str]:
     """
     Create appropriate trace generator based on pattern and validate parameters.
@@ -29,6 +30,7 @@ def create_trace_generator(
         pattern: Pattern for trace generation
         rps: Requests per second (only for synthetic patterns)
         duration: Duration in seconds (only for synthetic patterns)
+        seed: Random seed for reproducibility of trace patterns
 
     Returns:
         Tuple of (TraceGenerator instance, Warning message or empty string)
@@ -50,7 +52,9 @@ def create_trace_generator(
             raise ValueError(
                 "Duration must be non-negative for synthetic patterns"
             )
-        return SyntheticTraceGenerator(rps, pattern, duration), warning_msg
+        return SyntheticTraceGenerator(
+            rps, pattern, duration, seed
+        ), warning_msg
 
     # Azure patterns
     if rps != 1:
@@ -75,6 +79,7 @@ def create_trace_generator(
 @click.option(
     "--pattern",
     default="uniform",
+    type=click.Choice(sorted(VALID_PATTERNS), case_sensitive=False),
     help=f"Pattern for generating trace. Valid patterns: {sorted(VALID_PATTERNS)}",
 )
 @click.option(
@@ -83,6 +88,12 @@ def create_trace_generator(
     default=10,
     help="Duration in seconds (only used with synthetic patterns)",
 )
+@click.option(
+    "--seed",
+    type=int,
+    default=None,
+    help="Random seed for reproducibility of trace patterns",
+)
 @click.option(
     "--subprocesses", type=int, default=1, help="Number of subprocesses"
 )
@@ -98,21 +109,43 @@ def create_trace_generator(
     default=lambda: os.environ.get("OPENAI_API_KEY", "none"),
     help="OpenAI API Key",
 )
-def main(model, rps, pattern, duration, subprocesses, base_url, api_key):
+@click.option(
+    "--datasets-config", default=None, help="Config file for datasets"
+)
+def main(
+    model,
+    rps,
+    pattern,
+    duration,
+    seed,
+    subprocesses,
+    base_url,
+    api_key,
+    datasets_config,
+):
     """Run trace-based load testing for OpenAI API endpoints."""
     try:
         trace_generator, warning_msg = create_trace_generator(
-            pattern, rps, duration
+            pattern, rps, duration, seed
         )
         if warning_msg:
             logger.warning(warning_msg)
 
+        if datasets_config is None:
+            datasets = []
+            sort_strategy = None
+        else:
+            datasets, sort_strategy = load_datasets(datasets_config)
+
         _, result_analyzer = run_load_test(
             trace_generator=trace_generator,
             model=model,
             subprocesses=subprocesses,
             base_url=base_url,
             api_key=api_key,
+            datasets=datasets,
+            sort_strategy=sort_strategy,
+            seed=seed,
         )
 
         print(result_analyzer)

tracestorm/constants.py

@@ -1,3 +1,5 @@
+import os
+
 AZURE_REPO_URL = "Azure/AzurePublicDataset"
 
 AZURE_DATASET_PATHS = {
@@ -11,3 +13,5 @@
 DEFAULT_SUBPROCESSES = 1
 
 DEFAULT_MESSAGES = "Tell me a story"
+
+DEFAULT_DATASET_FOLDER = os.path.expanduser("~/.cache/tracestorm")

tracestorm/core.py

@@ -1,5 +1,5 @@
 import multiprocessing
-from typing import List, Tuple
+from typing import List, Optional, Tuple
 
 from tracestorm.logger import init_logger
 from tracestorm.request_generator import generate_request
@@ -17,6 +17,9 @@ def run_load_test(
     subprocesses: int,
     base_url: str,
     api_key: str,
+    datasets: List,
+    sort_strategy: Optional[str] = None,
+    seed: Optional[int] = None,
 ) -> Tuple[List[Tuple], ResultAnalyzer]:
     """
     Run load test with given configuration.
@@ -27,6 +30,9 @@ def run_load_test(
         subprocesses: Number of subprocesses to use
         base_url: Base URL for API calls
         api_key: API key for authentication
+        datasets: List of datasets to generate prompts
+        sort_strategy: Sorting strategy for prompts in datasets.
+        seed: Random seed for sorting.
 
     Returns:
         Tuple of (List of results, ResultAnalyzer instance)
@@ -38,7 +44,13 @@ def run_load_test(
         logger.warning("No requests to process. Trace is empty.")
         return [], ResultAnalyzer()
 
-    requests = generate_request(model, total_requests)
+    requests = generate_request(
+        model_name=model,
+        nums=total_requests,
+        datasets=datasets,
+        sort_strategy=sort_strategy,
+        seed=seed,
+    )
     ipc_queue = multiprocessing.Queue()
     processes = []