Files

150 lines
6.3 KiB
Python

"""Custom / Ollama (local) provider profile.
Covers any endpoint registered as provider="custom", including local
Ollama instances and OpenAI-compatible reasoning endpoints (GLM-5.2 on
Volcengine ARK, vLLM, llama.cpp). Key quirks:
- ollama_num_ctx → extra_body.options.num_ctx (local context window)
- reasoning_config disabled → top-level reasoning_effort="none"
(Ollama /v1/chat/completions ignores think=False — ollama#14820)
+ extra_body.think = False only on Ollama URLs (/api/chat and proxies)
- reasoning_config enabled + effort → top-level reasoning_effort
(the native OpenAI-compatible format GLM/ARK expect; unset omits it
so the endpoint's server default applies)
"""
from typing import Any
from urllib.parse import urlparse
from providers import register_provider
from providers.base import ProviderProfile
def _looks_like_ollama_endpoint(base_url: str | None) -> bool:
"""True when ``base_url`` is an Ollama host, not a generic OpenAI-compat relay.
``think`` is an Ollama-native extra_body field. Strict hosts (Mistral
``extra=forbid``, Groq, …) reject it with HTTP 422. Match only explicit
Ollama signatures — default port 11434, or ``ollama`` as a hostname
label — not arbitrary localhost (llama.cpp / vLLM / LM Studio).
"""
raw = (base_url or "").strip()
if not raw:
return False
parsed = urlparse(raw if "://" in raw else f"//{raw}")
# urlparse raises ValueError for non-integer / out-of-range ports
# ("http://host:99999/v1" parses fine in the OpenAI client, so the URL
# is reachable here). Treat a malformed port as "not Ollama" instead of
# killing the whole kwargs build — same try/except shape the 11434
# check in hermes_cli/models.py uses, not the same detection logic.
try:
if parsed.port == 11434:
return True
except ValueError:
return False
host = (parsed.hostname or "").lower().rstrip(".")
if not host:
return False
if host == "ollama.com" or host.endswith(".ollama.com"):
return True
return "ollama" in host.split(".")
class CustomProfile(ProviderProfile):
"""Custom/Ollama local provider — think=false and num_ctx support."""
def build_api_kwargs_extras(
self,
*,
reasoning_config: dict | None = None,
ollama_num_ctx: int | None = None,
**ctx: Any,
) -> tuple[dict[str, Any], dict[str, Any]]:
extra_body: dict[str, Any] = {}
top_level: dict[str, Any] = {}
# Ollama context window
if ollama_num_ctx:
options = extra_body.get("options", {})
options["num_ctx"] = ollama_num_ctx
extra_body["options"] = options
# Reasoning / thinking control for custom OpenAI-compatible endpoints
# (GLM-5.2 on Volcengine ARK, vLLM, Ollama, llama.cpp, …).
#
# - disabled → top-level reasoning_effort="none"; extra_body.think
# = False only on Ollama URLs (Ollama's thinking-off flag)
# - enabled + effort set → TOP-LEVEL reasoning_effort string, the
# format GLM-5.2/ARK and other OpenAI-compatible reasoning APIs
# expect (GLM documents "high" and "max"; "max" is its default).
# - enabled + no effort → omit both, so the endpoint applies its own
# server-side default (do NOT force a level the user didn't pick).
#
# We deliberately do NOT emit ``think=True`` on enable: it is an
# Ollama-only flag and thinking is already server-default-on for these
# backends, so forcing it risks a 400 on GLM/vLLM endpoints that don't
# recognize it. Mirrors the DeepSeek/Zai profile precedent. The same
# constraint applies to ``think=False`` on disable — Mistral/Groq
# reject unknown fields (HTTP 422 extra_forbidden) rather than ignoring
# them, so that flag stays Ollama-URL-gated.
if reasoning_config and isinstance(reasoning_config, dict):
_effort = (reasoning_config.get("effort") or "").strip().lower()
_enabled = reasoning_config.get("enabled", True)
if _effort == "none" or _enabled is False:
# Ollama's /v1/chat/completions silently ignores
# extra_body.think (only /api/chat honours it — ollama#14820)
# but respects the top-level reasoning_effort field (#25758).
# Always emit reasoning_effort="none"; only add think=False
# when the URL is actually Ollama.
top_level["reasoning_effort"] = "none"
if _looks_like_ollama_endpoint(ctx.get("base_url")):
extra_body["think"] = False
elif _effort:
# Clamp the internal ladder onto the widest OpenAI-compatible
# wire vocabulary (shared policy in agent.reasoning_effort) —
# GLM/ARK, vLLM and SGLang all top out at "max"; forwarding
# "ultra" verbatim is a guaranteed 400 (#89503).
from agent.reasoning_effort import (
OPENAI_COMPAT_WIRE_EFFORTS,
clamp_effort,
)
top_level["reasoning_effort"] = clamp_effort(
_effort, OPENAI_COMPAT_WIRE_EFFORTS
)
return extra_body, top_level
def fetch_models(
self,
*,
api_key: str | None = None,
base_url: str | None = None,
timeout: float = 8.0,
) -> list[str] | None:
"""Custom/Ollama: base_url is user-configured; fetch if set."""
if not (base_url or self.base_url):
return None
return super().fetch_models(api_key=api_key, base_url=base_url, timeout=timeout)
custom = CustomProfile(
name="custom",
aliases=(
"ollama",
"local",
"vllm",
"llamacpp",
"llama.cpp",
"llama-cpp",
),
env_vars=(), # No fixed key — custom endpoint
base_url="", # User-configured
# Without this, no max_tokens is sent and Ollama falls back to its internal
# num_predict=128, truncating responses after a few tokens (#39281). This is
# only a floor used when the user hasn't set model.max_tokens — they can
# override per-model — so we set it generously rather than lowballing it.
default_max_tokens=65536,
)
register_provider(custom)