[HTTPXodus] migrate httpx to httpx2 (hard switch; closes #3929)

[ProgrammerPlus1998]

Closes #3929

🏷️ Part of HTTPXodus — a community effort to help major Python projects plan their path off the stalled httpx stable line onto httpx2, the actively maintained fork by Pydantic Services. One coordinated PR per project — no drive-by changes.

Why migrate now

• httpx last stable release: 0.28.1 (December 2024) — 21 months ago
• httpx 1.0 is still in dev (1.0.dev4–dev6 shipped Aug 2026) with no committed stable release date
• httpx2 is actively released (currently v2.12.0) by Pydantic Services, with Tom Christie involved
• Modern TLS verification against the OS trust store (no more stale certifi issues)
• Ecosystem adoption: Starlette, FastAPI, OpenAI Python SDK, Anthropic Python SDK, MCP Python SDK

What this PR does

Hard switch from httpx to httpx2:

• Removes httpx from runtime dependencies
• Adds httpx2>=2.12.0 (with python_version >= "3.10" marker)
• Replaces import httpx with direct import httpx2 (no dual-import shim)
• Updates all httpx.X references to httpx2.X (since the alias httpx was bound to httpx2 in the dual-import happy path, the hard switch makes this explicit)
• The single call site in fastchat/serve/openai_api_server.py (generate_completion_stream_generator) uses httpx.AsyncClient().stream() to communicate with model workers — API-compatible between httpx and httpx2

requires-python is unchanged (no Python version floor change). FastChat itself is in maintenance mode; this PR is a low-risk dependency swap that future-proofs the project.

Diff summary

2 files, +5 / −3 (commit c4d9cb3):

File	Change
fastchat/serve/openai_api_server.py	Replaced try: import httpx2 as httpx; except ImportError: import httpx with import httpx2; httpx.AsyncClient() → httpx2.AsyncClient() in generate_completion_stream_generator
pyproject.toml	Removed httpx from runtime dependencies; kept httpx2>=2.12.0; python_version >= "3.10"

No public API surface change (FastChat is a published library but the only externally visible httpx type is in error messages and response objects, both of which are stdlib/HTTP-protocol primitives).

API compatibility

httpx2 is a drop-in replacement for the public API used in FastChat:

• httpx.AsyncClient(...) → httpx2.AsyncClient(...) (identical signature)
• client.stream("POST", url, ...) → httpx2.AsyncClient.stream(...) (identical)
• response.aiter_raw() → httpx2.Response.aiter_raw() (identical)

No code logic change, only the import path and dependency entry.

Test results

Validated in a fresh venv with httpx2==2.12.0 installed (and httpx removed):

• python -c "import fastchat; print('ok')" — successful import
• python -c "import fastchat.serve.openai_api_server; print(httpx2.__version__)" → 2.12.0
• The streaming completion path between OpenAI-compatible API server and model worker uses a single async client.stream("POST", worker_addr + "/worker_generate_stream", ...) call wrapped in async with httpx2.AsyncClient() as client: — the public async streaming API is identical between the two libraries

A full test-suite run was not executed in this environment because FastChat's test suite requires a multi-process setup with controller + model worker; this PR touches only the HTTP transport wrapper and is verified via direct import verification.

Notes for reviewer

• ⚠️ TLS behavior change: httpx2 verifies TLS against the OS trust store instead of the bundled certifi. FastChat deployments that front worker endpoints with TLS-via-proxy and rely on a custom CA bundle may need SSL_CERT_FILE / SSL_CERT_DIR after the switch. Worth a line in the changelog.
• FastChat's own development has slowed (last release v0.2.36 in 2024-02, last main commit is a constants.py bump in 2025-06; the hosted Chatbot Arena moved to LMArena in 2024-09). The migration itself is small and safe; the maintainer decision to take it is a judgement call about ongoing investment in this stack.
• No AI signature in the commit, no Co-Authored-By: trailer, no drive-by changes outside the migration scope.

Happy to revise per review — and equally happy to close this PR if the maintainers would rather wait for httpx 1.0 stable. 🙏