Streaming response wrapping for custom transport layers

[psymbio]

Not sure if this should be raised in urllib3 or this library, apologies in advance.

I wanted to understand how to wrap a streaming response back into httpx if we use urllib3 - particularly for the openai library. So, for the simple non-streaming response the code is as follows. But since urllib3 supports streaming, I would like to wrap the response back in httpx and return it.

Without streaming:

class URLLib3Transport(httpx.BaseTransport):
    def __init__(self):
        self.pool = urllib3.PoolManager()

    def handle_request(self, request: httpx.Request):
        payload = json.loads(request.content.decode("utf-8").replace("'",'"'))
        urllib3_response = self.pool.request(request.method, str(request.url), headers=request.headers, json=payload)
        content = json.loads(urllib3_response.data.decode('utf-8'))
        stream = httpx.ByteStream(json.dumps(content).encode("utf-8"))
        headers = [(b"content-type", b"application/json")]
        return httpx.Response(200, headers=headers, stream=stream)

With streaming (incomplete code):

class URLLib3Transport(httpx.BaseTransport):
    def __init__(self):
        pass

    def handle_request(self, request: httpx.Request):
        payload = json.loads(request.content.decode("utf-8").replace("'",'"'))
        urllib3_response = urllib3.request(request.method, str(request.url), headers=request.headers, json=payload, preload_content=False)
        for chunk in urllib3_response.stream():
            chunk_dict = json.loads(chunk.decode('utf-8'))
            # print(chunk_dict)
            print(chunk_dict['choices'][0]['message']['content'])

This results in the following output:

(Verse 1)
In a world full of chaos, where troubles never cease,
I'll sing a little melody, to bring you some peace.

(Verse 1)
In a world full of wonder, I'll sing you a song,
Where melodies dance and the words flow along.
With every
Sure! Here's a short song for you:

(Verse 1)
In the morning light, a brand new day,
I'll sing a song

Which, is not a truly a streamed response which would look like: (Verse 1)Oh, there once was a time when dreams were just out of sight, But now I realize they can come true with all coming word by word using this code - instead of printing Verse(1) three times like the above code.

Also, I wanted to understand how well this integrates with openai code (@RobertCraigie, @rattrayalex), i.e., the stream parameter in openai_client.chat.completions.with_raw_response.create() to able to switch in the custom transport layer based on this parameter's value. This would add value to the openai library capability when we move from a server to a client-based solution which can be supported on the web browser.

Link for the entire code - which fails for streaming.

[lovelydinosaur]

Something like this...

class URLLib3Stream(httpx.SyncByteStream):
    def __init__(self, stream):
        self._stream = stream

    def __iter__(self):
        ...

    def close(self):
        ...


class URLLib3Transport(httpx.BaseTransport):
    def __init__(self):
        self._pool = urllib3.ConnectionPool()

    def handle_request(self, request: httpx.Request):
        urllib3_response = self._pool.request(
            method=...,
            url=...,
            headers=...,
            stream=...,
        )
        return Response(
            status=...,
            headers=...,
            stream=URLLib3Stream(urllib3_response.stream)
        )

    def close(self):
        self._pool.close()

[psymbio]

So, my current code in the JupyterLite environment looks like:

import micropip
await micropip.install('https://raw.githubusercontent.com/psymbio/pyodide_wheels/main/multidict/multidict-4.7.6-py3-none-any.whl', keep_going=True)
await micropip.install('https://raw.githubusercontent.com/psymbio/pyodide_wheels/main/frozenlist/frozenlist-1.4.0-py3-none-any.whl', keep_going=True)
await micropip.install('https://raw.githubusercontent.com/psymbio/pyodide_wheels/main/aiohttp/aiohttp-3.9.1-py3-none-any.whl', keep_going=True)
await micropip.install('https://raw.githubusercontent.com/psymbio/pyodide_wheels/main/openai/openai-1.3.7-py3-none-any.whl', keep_going=True)
await micropip.install('https://raw.githubusercontent.com/psymbio/pyodide_wheels/main/urllib3/urllib3-2.1.0-py3-none-any.whl', keep_going=True)

await micropip.install("ssl")
import ssl
await micropip.install("httpx", keep_going=True)
import httpx
await micropip.install('https://raw.githubusercontent.com/psymbio/pyodide_wheels/main/urllib3/urllib3-2.1.0-py3-none-any.whl', keep_going=True)
import urllib3
urllib3.disable_warnings(urllib3.exceptions.InsecureRequestWarning)
import json

class URLLib3Stream(httpx.SyncByteStream):
    def __init__(self, stream):
        self._stream = stream

    def __iter__(self):
        with httpx._transports.default.map_httpcore_exceptions():
            for part in self._stream:
                yield part

    def close(self):
        if hasattr(self._stream, "close"):
            self._stream.close()


class URLLib3Transport(httpx.BaseTransport):
    def __init__(self):
        # self._pool = urllib3.PoolManager()
        self._pool = urllib3.connectionpool.ConnectionPool(host="localhost")

    def handle_request(self, request: httpx.Request):
        payload = json.loads(request.content.decode("utf-8").replace("'",'"'))
        urllib3_response = self._pool.request(
            method=request.method,
            url=str(request.url),
            headers=request.headers,
            stream=True,
        )
        headers = [(b"content-type", b"application/json")]
        return Response(
            status=200,
            headers=headers,
            stream=URLLib3Stream(urllib3_response.stream)
        )

client = httpx.Client(transport=URLLib3Transport())

from openai import OpenAI
from openai import AzureOpenAI
import openai
import os

os.environ['AZURE_OPENAI_API_KEY'] = "xxx"
openai_client = AzureOpenAI(
    api_version="2023-07-01-preview",
    azure_endpoint="https://xxx.openai.azure.com/",
    http_client=client
)

response = openai_client.chat.completions.with_raw_response.create(
    messages=[{
        "role": "user",
        "content": "sing me a song",
    }],
    model="gpt-35-turbo",
    max_tokens=30,
    temperature=0.7,
    # stream=True
)

completion = response.parse()
print(completion)

And get the following error: APIConnectionError: Connection error., complete error message.

And I've also tried another variation (because AttributeError: 'ConnectionPool' object has no attribute 'request'), but this didn't work out as well:

class URLLib3Stream(httpx.SyncByteStream):
    def __init__(self, stream):
        self._stream = stream

    def __iter__(self):
        with httpx._transports.default.map_httpcore_exceptions():
            for part in self._stream:
                yield part

    def close(self):
        if hasattr(self._stream, "close"):
            self._stream.close()

class URLLib3Transport(httpx.BaseTransport):
    def __init__(self):
        self._pool = urllib3.connectionpool.ConnectionPool(host="localhost")
        self._http = urllib3.PoolManager(self._pool)

    def handle_request(self, request: httpx.Request):
        payload = json.loads(request.content.decode("utf-8").replace("'",'"'))
        urllib3_request = self._http.request(
            method=request.method,
            url=str(request.url),
            headers=request.headers,
            body=payload,
            fields=None,
            stream=True,
        )
        headers = [(b"content-type", b"application/json")]
        return Response(
            status=200,
            headers=headers,
            stream=URLLib3Stream(urllib3_request.stream)
        )

client = httpx.Client(transport=URLLib3Transport())

This results in the error message: APIConnectionError: Connection error. and the complete error message.

[psymbio]

Hi @tomchristie,

I rectified some more of the above code but have noticed some issues in the way that streaming works:

Particularly this part of the code:

class URLLib3Stream(httpx.SyncByteStream):
    def __init__(self, stream):
        self._stream = stream

    def __iter__(self):
        print(self._stream.__self__.data)
        with httpx._transports.default.map_httpcore_exceptions():
            for part in self._stream:
                yield part

    def close(self):
        if hasattr(self._stream, "close"):
            self._stream.close()

printing out the data in the before yielding the output give me:

b'{"id":"chatcmpl-8ucZ1UsLPxKy6SMYpE1JkJ4SsaSwT","object":"chat.completion","created":1708505299,"model":"gpt-35-turbo","prompt_filter_results":[{"prompt_index":0,"content_filter_results":{"hate":{"filtered":false,"severity":"safe"},"self_harm":{"filtered":false,"severity":"safe"},"sexual":{"filtered":false,"severity":"safe"},"violence":{"filtered":false,"severity":"safe"}}}],"choices":[{"finish_reason":"stop","index":0,"message":{"role":"assistant","content":"I can\'t sing, but I can certainly help you find a song to listen to! What kind of music are you in the mood for?"},"content_filter_results":{"hate":{"filtered":false,"severity":"safe"},"self_harm":{"filtered":false,"severity":"safe"},"sexual":{"filtered":false,"severity":"safe"},"violence":{"filtered":false,"severity":"safe"}}}],"usage":{"prompt_tokens":11,"completion_tokens":29,"total_tokens":40},"system_fingerprint":"fp_68a7d165bf"}\n'
b'{"id":"chatcmpl-8ucZ2tRwvhcyTgINUlrsZScHuWIXu","object":"chat.completion","created":1708505300,"model":"gpt-35-turbo","prompt_filter_results":[{"prompt_index":0,"content_filter_results":{"hate":{"filtered":false,"severity":"safe"},"self_harm":{"filtered":false,"severity":"safe"},"sexual":{"filtered":false,"severity":"safe"},"violence":{"filtered":false,"severity":"safe"}}}],"choices":[{"finish_reason":"length","index":0,"message":{"role":"assistant","content":"I\'m just a digital assistant, so I can\'t sing, but I can certainly help you find the lyrics to a song or recommend some good tunes"},"content_filter_results":{"hate":{"filtered":false,"severity":"safe"},"self_harm":{"filtered":false,"severity":"safe"},"sexual":{"filtered":false,"severity":"safe"},"violence":{"filtered":false,"severity":"safe"}}}],"usage":{"prompt_tokens":11,"completion_tokens":30,"total_tokens":41},"system_fingerprint":"fp_68a7d165bf"}\n'
b'{"id":"chatcmpl-8ucZ5to0cm2ZVFGtIDK9RcvuN3iRL","object":"chat.completion","created":1708505303,"model":"gpt-35-turbo","prompt_filter_results":[{"prompt_index":0,"content_filter_results":{"hate":{"filtered":false,"severity":"safe"},"self_harm":{"filtered":false,"severity":"safe"},"sexual":{"filtered":false,"severity":"safe"},"violence":{"filtered":false,"severity":"safe"}}}],"choices":[{"finish_reason":"length","index":0,"message":{"role":"assistant","content":"I\'m just a virtual assistant, so I can\'t sing, but I can help you find the lyrics to a song or recommend some music for you"},"content_filter_results":{"hate":{"filtered":false,"severity":"safe"},"self_harm":{"filtered":false,"severity":"safe"},"sexual":{"filtered":false,"severity":"safe"},"violence":{"filtered":false,"severity":"safe"}}}],"usage":{"prompt_tokens":11,"completion_tokens":30,"total_tokens":41},"system_fingerprint":"fp_68a7d165bf"}\n'

Which is three different responses for a single request.

Complete code can be found here.