Teknium
9ec5750aca
The chat_completions chokepoint fix (ultra->max for every model, cherry-picked from #89509) has siblings with the same bug shape: - codex.py: ultra->max was gated on gpt-5.6 only; now baseline for all Responses-API models (backend-specific branches still override). - Kimi top-level reasoning_effort: K3 accepts low/high/max only — 'medium' and upper-ladder levels were dropped to the medium default (400s on K3, ladder inversion on K2). Full ladder mapped per family, mirroring the kimi-coding plugin's K3 map. - TokenHub: 'minimal' fell through to the 'high' default (asked least, got most); full ladder now mapped onto low/medium/high. - auxiliary_client Responses path: ultra->max alongside the existing minimal->low clamp. - custom provider plugin: ultra capped at max instead of forwarded verbatim to GLM/vLLM/SGLang backends that reject it. - copilot plugin: ad-hoc downgrade rules replaced with the shared clamp_reasoning_effort_to_supported ladder walk so ultra/max resolve to the strongest supported level instead of medium (#74295). Sabotage-verified: new sibling-site tests fail 6/10 without the fixes.
110 lines
4.6 KiB
Python
110 lines
4.6 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 for /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 providers import register_provider
|
|
from providers.base import ProviderProfile
|
|
|
|
|
|
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 → extra_body.think = False (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.
|
|
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, so both
|
|
# are needed to actually stop a thinking-capable model from
|
|
# reasoning (#25758). Endpoints that recognize neither simply
|
|
# ignore them.
|
|
top_level["reasoning_effort"] = "none"
|
|
extra_body["think"] = False
|
|
elif _effort:
|
|
# "ultra" is Hermes-internal ladder vocabulary — no known
|
|
# OpenAI-compatible backend accepts it verbatim (GLM/ARK,
|
|
# vLLM and SGLang all top out at "max"); cap it at the wire
|
|
# ceiling instead of forwarding a guaranteed 400 (#89503).
|
|
top_level["reasoning_effort"] = (
|
|
"max" if _effort == "ultra" else _effort
|
|
)
|
|
|
|
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)
|