feat: add tabby-eval package to run e2e evaluation on modal (#893)

* updating predict.py, adding tabby_python_client package

* update predict.py

* move to tabby-eval directory

* delete old files

* update predict.py

* delete other folder files

python/tabby-eval/modal/predict.py

@@ -0,0 +1,212 @@
+import asyncio
+import json
+import modal
+import os
+import pandas as pd
+
+from collections import namedtuple
+from datetime import datetime
+from modal import Image, Mount, Secret, Stub, asgi_app, gpu, method
+from pathlib import Path
+from typing import Union, List, Optional, Any, Tuple
+
+
+GPU_CONFIG = gpu.A10G()
+
+MODEL_ID = os.environ.get("MODEL_ID")
+LAUNCH_FLAGS = ["serve", "--model", MODEL_ID, "--port", "8000", "--device", "cuda"]
+
+
+def download_model():
+    import subprocess
+    import os
+    MODEL_ID = os.environ.get("MODEL_ID")
+    print(f'MODEL_ID={MODEL_ID}')
+    subprocess.run(
+        [
+            "/opt/tabby/bin/tabby",
+            "download",
+            "--model",
+            MODEL_ID,
+        ]
+    )
+
+
+image = (
+    Image.from_registry(
+        "tabbyml/tabby:0.5.5",
+        add_python="3.11",
+    )
+    .env({"MODEL_ID": os.environ.get("MODEL_ID")})
+    .dockerfile_commands("ENTRYPOINT []")
+    .copy_local_dir(local_path='./modal/tabby_python_client/tabby_python_client', remote_path='/root/tabby_python_client')
+    .pip_install(
+        "httpx",
+        "pandas"
+    )
+    .run_function(download_model)
+)
+
+stub = Stub("tabby-" + MODEL_ID.split("/")[-1], image=image)
+
+
+@stub.cls(
+    gpu=GPU_CONFIG,
+    concurrency_limit=10,
+    allow_concurrent_inputs=2,
+    container_idle_timeout=60 * 10,
+    timeout=600,
+)
+class Model:
+
+    def __enter__(self):
+        import socket
+        import subprocess, os
+        import time
+
+        from tabby_python_client import Client
+
+
+        my_env = os.environ.copy()
+        my_env["TABBY_DISABLE_USAGE_COLLECTION"] = "1"
+        MODEL_ID = os.environ.get("MODEL_ID")
+        print(f'MODEL_ID={MODEL_ID}')
+
+        LAUNCH_FLAGS = ["serve", "--model", MODEL_ID, "--port", "8000", "--device", "cuda"]
+        self.launcher = subprocess.Popen(["/opt/tabby/bin/tabby"] + LAUNCH_FLAGS, env=my_env)
+        self.client = Client("http://127.0.0.1:8000", timeout=240)
+
+        # Poll until webserver at 127.0.0.1:8000 accepts connections before running inputs.
+        def webserver_ready():
+            try:
+                socket.create_connection(("127.0.0.1", 8000), timeout=1).close()
+                return True
+            except (socket.timeout, ConnectionRefusedError):
+                # Check if launcher webserving process has exited.
+                # If so, a connection can never be made.
+                retcode = self.launcher.poll()
+                if retcode is not None:
+                    raise RuntimeError(
+                        f"launcher exited unexpectedly with code {retcode}"
+                    )
+                return False
+
+        while not webserver_ready():
+            time.sleep(1.0)
+
+        print("Tabby server ready!")
+
+    def __exit__(self, _exc_type, _exc_value, _traceback):
+        self.launcher.terminate()
+
+    @method()
+    async def health(self):
+        from tabby_python_client.api.v1 import health
+
+        resp = await health.asyncio(client=self.client)
+        return resp.to_dict()
+
+    @method()
+    async def complete(self, language: str, index: int, prompt: str) -> Tuple[int, Optional[str], Optional[str]]:
+        from tabby_python_client.api.v1 import completion
+        from tabby_python_client.models import (
+            CompletionRequest,
+            DebugOptions,
+            CompletionResponse,
+            Segments,
+        )
+        from tabby_python_client.types import Response
+
+
+        request = CompletionRequest(
+            language=language, debug_options=DebugOptions(raw_prompt=prompt)
+        )
+
+        try:
+            resp: Response = await completion.asyncio_detailed(
+                client=self.client, json_body=request
+            )
+
+            if resp.parsed != None:
+                return index, resp.parsed.choices[0].text, None
+            else:
+                return index, None, f"<{resp.status_code}>"
+        except errors.UnexpectedStatus as e:
+            return index, None, f"error: code={e.status_code} content={e.content} error={e}"
+        except Exception as e:
+            return index, None, f"error type: {type(e)}"
+
+def write_log(log: str):
+    now = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+    with open('./modal/log.txt', 'a') as f:
+        f.write(f"{now} : {log}")
+        f.write("\n")
+
+def chunker(seq, size) -> List:
+    return (seq[pos:pos + size] for pos in range(0, len(seq), size))
+
+
+def read_dataframe_from_file(language: str, file: str) -> pd.DataFrame:
+    whole_path_file = "./data/" + MODEL_ID.split("/")[-1] + "/" + language + "/" + file
+    objs = []
+    with open(whole_path_file) as fin:
+        for line in fin:
+            obj = json.loads(line)
+            if 'crossfile_context' in obj.keys():
+                obj['raw_prompt'] = obj['crossfile_context']['text'] + obj['prompt']
+            else:
+                obj['raw_prompt'] = obj['prompt']
+            objs.append(obj)
+
+    df = pd.DataFrame(objs)
+    return df
+
+@stub.local_entrypoint()
+async def main(language: str, files: str):
+    #Multiple files seperated by ','
+
+    model = Model()
+
+    health_resp = model.health.remote()
+    print(f'model info:\n{health_resp}')
+    assert(health_resp['model'] == MODEL_ID)
+
+    files = files.split(',')
+
+    for file in files:
+
+        df = read_dataframe_from_file(language, file.strip())
+
+        write_log(f'model: {MODEL_ID}; language: {language}; file: {file}: length = {len(df)}')
+
+
+        if 'prediction' in df.columns:
+            df_no_prediction = df[df['prediction'].isna()]
+        else:
+            df_no_prediction = df
+
+        skipped = len(df) - len(df_no_prediction)
+        success = 0
+        error = 0
+
+        for group in chunker(df_no_prediction, 30):
+            outputs = await asyncio.gather(*[model.complete.remote.aio(language, index, row['raw_prompt']) for index, row in group.iterrows()])
+
+            for index, prediction, error_msg in outputs:
+                if prediction is not None:
+                    df.loc[index, 'prediction'] = prediction
+                    success += 1
+                else:
+                    df.loc[index, 'error'] = error_msg
+                    error += 1
+
+        write_log(f"Skipped {skipped} rows, {success} rows with predictions, {error} rows with errors")
+
+        whole_path_file = "./data/" + MODEL_ID.split("/")[-1] + "/" + language + "/" + file
+        with open(whole_path_file, 'w') as fout:
+            for index, row in df.iterrows():
+                row_dict = row.to_dict()
+                json.dump(row_dict, fout)
+                fout.write('\n')
+
+        write_log(f"model: {MODEL_ID}; language: {language}; file: {file}: end!\n")

python/tabby-eval/modal/tabby_python_client/.gitignore

@@ -0,0 +1,23 @@
+__pycache__/
+build/
+dist/
+*.egg-info/
+.pytest_cache/
+
+# pyenv
+.python-version
+
+# Environments
+.env
+.venv
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# JetBrains
+.idea/
+
+/coverage.xml
+/.coverage

python/tabby-eval/modal/tabby_python_client/README.md

@@ -0,0 +1,89 @@
+# tabby-python-client
+A client library for accessing Tabby Server
+
+## Usage
+First, create a client:
+
+```python
+from tabby_python_client import Client
+
+client = Client(base_url="https://api.example.com")
+```
+
+If the endpoints you're going to hit require authentication, use `AuthenticatedClient` instead:
+
+```python
+from tabby_python_client import AuthenticatedClient
+
+client = AuthenticatedClient(base_url="https://api.example.com", token="SuperSecretToken")
+```
+
+Now call your endpoint and use your models:
+
+```python
+from tabby_python_client.models import MyDataModel
+from tabby_python_client.api.my_tag import get_my_data_model
+from tabby_python_client.types import Response
+
+my_data: MyDataModel = get_my_data_model.sync(client=client)
+# or if you need more info (e.g. status_code)
+response: Response[MyDataModel] = get_my_data_model.sync_detailed(client=client)
+```
+
+Or do the same thing with an async version:
+
+```python
+from tabby_python_client.models import MyDataModel
+from tabby_python_client.api.my_tag import get_my_data_model
+from tabby_python_client.types import Response
+
+my_data: MyDataModel = await get_my_data_model.asyncio(client=client)
+response: Response[MyDataModel] = await get_my_data_model.asyncio_detailed(client=client)
+```
+
+By default, when you're calling an HTTPS API it will attempt to verify that SSL is working correctly. Using certificate verification is highly recommended most of the time, but sometimes you may need to authenticate to a server (especially an internal server) using a custom certificate bundle.
+
+```python
+client = AuthenticatedClient(
+    base_url="https://internal_api.example.com", 
+    token="SuperSecretToken",
+    verify_ssl="/path/to/certificate_bundle.pem",
+)
+```
+
+You can also disable certificate validation altogether, but beware that **this is a security risk**.
+
+```python
+client = AuthenticatedClient(
+    base_url="https://internal_api.example.com", 
+    token="SuperSecretToken", 
+    verify_ssl=False
+)
+```
+
+There are more settings on the generated `Client` class which let you control more runtime behavior, check out the docstring on that class for more info.
+
+Things to know:
+1. Every path/method combo becomes a Python module with four functions:
+    1. `sync`: Blocking request that returns parsed data (if successful) or `None`
+    1. `sync_detailed`: Blocking request that always returns a `Request`, optionally with `parsed` set if the request was successful.
+    1. `asyncio`: Like `sync` but async instead of blocking
+    1. `asyncio_detailed`: Like `sync_detailed` but async instead of blocking
+
+1. All path/query params, and bodies become method arguments.
+1. If your endpoint had any tags on it, the first tag will be used as a module name for the function (my_tag above)
+1. Any endpoint which did not have a tag will be in `tabby_python_client.api.default`
+
+## Building / publishing this Client
+This project uses [Poetry](https://python-poetry.org/) to manage dependencies  and packaging.  Here are the basics:
+1. Update the metadata in pyproject.toml (e.g. authors, version)
+1. If you're using a private repository, configure it with Poetry
+    1. `poetry config repositories.<your-repository-name> <url-to-your-repository>`
+    1. `poetry config http-basic.<your-repository-name> <username> <password>`
+1. Publish the client with `poetry publish --build -r <your-repository-name>` or, if for public PyPI, just `poetry publish --build`
+
+If you want to install this client into another project without publishing it (e.g. for development) then:
+1. If that project **is using Poetry**, you can simply do `poetry add <path-to-this-client>` from that project
+1. If that project is not using Poetry:
+    1. Build a wheel with `poetry build -f wheel`
+    1. Install that wheel from the other project `pip install <path-to-wheel>`

python/tabby-eval/modal/tabby_python_client/pyproject.toml

@@ -0,0 +1,16 @@
+[tool.black]
+line-length = 120
+target_version = ['py38', 'py39', 'py310', 'py311']
+exclude = '''
+(
+  /(
+    | \.git
+    | \.venv
+    | \.mypy_cache
+  )/
+)
+'''
+
+[tool.isort]
+line_length = 120
+profile = "black"

python/tabby-eval/modal/tabby_python_client/setup.py

@@ -0,0 +1,18 @@
+import pathlib
+
+from setuptools import find_packages, setup
+
+here = pathlib.Path(__file__).parent.resolve()
+long_description = (here / "README.md").read_text(encoding="utf-8")
+
+setup(
+    name="tabby-python-client",
+    version="0.4.0-dev",
+    description="A client library for accessing Tabby Server",
+    long_description=long_description,
+    long_description_content_type="text/markdown",
+    packages=find_packages(),
+    python_requires=">=3.8, <4",
+    install_requires=["httpx >= 0.15.0, < 0.25.0", "attrs >= 21.3.0", "python-dateutil >= 2.8.0, < 3"],
+    package_data={"tabby_python_client": ["py.typed"]},
+)

python/tabby-eval/modal/tabby_python_client/tabby_python_client/__init__.py

@@ -0,0 +1,7 @@
+""" A client library for accessing Tabby Server """
+from .client import AuthenticatedClient, Client
+
+__all__ = (
+    "AuthenticatedClient",
+    "Client",
+)

python/tabby-eval/modal/tabby_python_client/tabby_python_client/api/__init__.py

@@ -0,0 +1 @@
+""" Contains methods for accessing the API """