github/OpenHands
1
0
Fork 0
mirror of https://github.com/OpenHands/OpenHands.git synced 2026-03-22 13:47:19 +08:00
Code Issues Packages Projects Releases Wiki Activity

Centralize model feature checks (#10414)

Browse Source 
Co-authored-by: OpenHands-GPT-5 <openhands@all-hands.dev>
This commit is contained in:
Engel Nyst
2025-08-19 22:30:07 +02:00
committed by GitHub
parent aa6b454772
commit bb0e24d23b
8 changed files with 596 additions and 112 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
openhands/llm/async_llm.py
View File

@@ -9,8 +9,8 @@ from openhands.core.logger import openhands_logger as logger
from openhands.llm.llm import (
LLM,
LLM_RETRY_EXCEPTIONS,
REASONING_EFFORT_SUPPORTED_MODELS,
)
from openhands.llm.model_features import get_features
from openhands.utils.shutdown_listener import should_continue
@@ -63,7 +63,7 @@ class AsyncLLM(LLM):
messages = kwargs['messages']
# Set reasoning effort for models that support it
if self.config.model.lower() in REASONING_EFFORT_SUPPORTED_MODELS:
if get_features(self.config.model).supports_reasoning_effort:
kwargs['reasoning_effort'] = self.config.reasoning_effort
# ensure we work with a list of messages

108
openhands/llm/llm.py
View File

@@ -9,6 +9,7 @@ import httpx
from openhands.core.config import LLMConfig
from openhands.llm.metrics import Metrics
from openhands.llm.model_features import get_features
with warnings.catch_warnings():
warnings.simplefilter('ignore')
@@ -49,79 +50,6 @@ LLM_RETRY_EXCEPTIONS: tuple[type[Exception], ...] = (
LLMNoResponseError,
)
# cache prompt supporting models
# remove this when we gemini and deepseek are supported
CACHE_PROMPT_SUPPORTED_MODELS = [
'claude-3-7-sonnet-20250219',
'claude-sonnet-3-7-latest',
'claude-3.7-sonnet',
'claude-3-5-sonnet-20241022',
'claude-3-5-sonnet-20240620',
'claude-3-5-haiku-20241022',
'claude-3-haiku-20240307',
'claude-3-opus-20240229',
'claude-sonnet-4-20250514',
'claude-sonnet-4',
'claude-opus-4-20250514',
'claude-opus-4-1-20250805',
]
# function calling supporting models
FUNCTION_CALLING_SUPPORTED_MODELS = [
'claude-3-7-sonnet-20250219',
'claude-sonnet-3-7-latest',
'claude-3-5-sonnet',
'claude-3-5-sonnet-20240620',
'claude-3-5-sonnet-20241022',
'claude-3.5-haiku',
'claude-3-5-haiku-20241022',
'claude-sonnet-4-20250514',
'claude-sonnet-4',
'claude-opus-4-20250514',
'claude-opus-4-1-20250805',
'gpt-4o-mini',
'gpt-4o',
'o1-2024-12-17',
'o3-mini-2025-01-31',
'o3-mini',
'o3',
'o3-2025-04-16',
'o4-mini',
'o4-mini-2025-04-16',
'gemini-2.5-pro',
'gpt-4.1',
'kimi-k2-0711-preview',
'kimi-k2-instruct',
'Qwen3-Coder-480B-A35B-Instruct',
'qwen3-coder', # this will match both qwen3-coder-480b (openhands provider) and qwen3-coder (for openrouter)
'gpt-5',
'gpt-5-2025-08-07',
]
REASONING_EFFORT_SUPPORTED_MODELS = [
'o1-2024-12-17',
'o1',
'o3',
'o3-2025-04-16',
'o3-mini-2025-01-31',
'o3-mini',
'o4-mini',
'o4-mini-2025-04-16',
'gemini-2.5-flash',
'gemini-2.5-pro',
'gpt-5',
'gpt-5-2025-08-07',
'claude-opus-4-1-20250805', # we need to remove top_p for opus 4.1
]
MODELS_WITHOUT_STOP_WORDS = [
'o1-mini',
'o1-preview',
'o1',
'o1-2024-12-17',
'xai/grok-4-0709',
]
class LLM(RetryMixin, DebugMixin):
"""The LLM class represents a Language Model instance.
@@ -154,6 +82,7 @@ class LLM(RetryMixin, DebugMixin):
)
self.model_info: ModelInfo | None = None
self._function_calling_active: bool = False
self.retry_listener = retry_listener
if self.config.log_completions:
if self.config.log_completions_folder is None:
@@ -202,10 +131,8 @@ class LLM(RetryMixin, DebugMixin):
f'Rewrote openhands/{model_name} to {self.config.model} with base URL {self.config.base_url}'
)
if (
self.config.model.lower() in REASONING_EFFORT_SUPPORTED_MODELS
or self.config.model.split('/')[-1] in REASONING_EFFORT_SUPPORTED_MODELS
):
features = get_features(self.config.model)
if features.supports_reasoning_effort:
# For Gemini models, only map 'low' to optimized thinking budget
# Let other reasoning_effort values pass through to API as-is
if 'gemini-2.5-pro' in self.config.model:
@@ -312,7 +239,7 @@ class LLM(RetryMixin, DebugMixin):
# add stop words if the model supports it and stop words are not disabled
if (
self.config.model not in MODELS_WITHOUT_STOP_WORDS
get_features(self.config.model).supports_stop_words
and not self.config.disable_stop_word
):
kwargs['stop'] = STOP_WORDS
@@ -556,17 +483,10 @@ class LLM(RetryMixin, DebugMixin):
):
self.config.max_output_tokens = self.model_info['max_tokens']
# Initialize function calling capability
# Check if model name is in our supported list
model_name_supported = (
self.config.model in FUNCTION_CALLING_SUPPORTED_MODELS
or self.config.model.split('/')[-1] in FUNCTION_CALLING_SUPPORTED_MODELS
or any(m in self.config.model for m in FUNCTION_CALLING_SUPPORTED_MODELS)
)
# Handle native_tool_calling user-defined configuration
# Initialize function calling using centralized model features
features = get_features(self.config.model)
if self.config.native_tool_calling is None:
self._function_calling_active = model_name_supported
self._function_calling_active = features.supports_function_calling
else:
self._function_calling_active = self.config.native_tool_calling
@@ -601,14 +521,10 @@ class LLM(RetryMixin, DebugMixin):
Returns:
boolean: True if prompt caching is supported and enabled for the given model.
"""
return (
self.config.caching_prompt is True
and (
self.config.model in CACHE_PROMPT_SUPPORTED_MODELS
or self.config.model.split('/')[-1] in CACHE_PROMPT_SUPPORTED_MODELS
)
# We don't need to look-up model_info, because only Anthropic models needs the explicit caching breakpoint
)
if not self.config.caching_prompt:
return False
# We don't need to look-up model_info, because only Anthropic models need explicit caching breakpoints
return get_features(self.config.model).supports_prompt_cache
def is_function_calling_active(self) -> bool:
"""Returns whether function calling is supported and enabled for this LLM instance.

139
openhands/llm/model_features.py Normal file
View File

@@ -0,0 +1,139 @@
from __future__ import annotations
from dataclasses import dataclass
from fnmatch import fnmatch
def normalize_model_name(model: str) -> str:
"""Normalize a model string to a canonical, comparable name.
Strategy:
- Trim whitespace
- Lowercase
- If there is a '/', keep only the basename after the last '/'
(handles prefixes like openrouter/, litellm_proxy/, anthropic/, etc.)
and treat ':' inside that basename as an Ollama-style variant tag to be removed
- There is no provider:model form; providers, when present, use 'provider/model'
- Drop a trailing "-gguf" suffix if present
"""
raw = (model or '').strip().lower()
if '/' in raw:
name = raw.split('/')[-1]
if ':' in name:
# Drop Ollama-style variant tag in basename
name = name.split(':', 1)[0]
else:
# No '/', keep the whole raw name (we do not support provider:model)
name = raw
if name.endswith('-gguf'):
name = name[: -len('-gguf')]
return name
def model_matches(model: str, patterns: list[str]) -> bool:
"""Return True if the model matches any of the glob patterns.
If a pattern contains a '/', it is treated as provider-qualified and matched
against the full, lowercased model string (including provider prefix).
Otherwise, it is matched against the normalized basename.
"""
raw = (model or '').strip().lower()
name = normalize_model_name(model)
for pat in patterns:
pat_l = pat.lower()
if '/' in pat_l:
if fnmatch(raw, pat_l):
return True
else:
if fnmatch(name, pat_l):
return True
return False
@dataclass(frozen=True)
class ModelFeatures:
supports_function_calling: bool
supports_reasoning_effort: bool
supports_prompt_cache: bool
supports_stop_words: bool
# Pattern tables capturing current behavior. Keep patterns lowercase.
FUNCTION_CALLING_PATTERNS: list[str] = [
# Anthropic families
'claude-3-7-sonnet*',
'claude-3.7-sonnet*',
'claude-sonnet-3-7-latest',
'claude-3-5-sonnet*',
'claude-3.5-haiku*',
'claude-3-5-haiku*',
'claude-sonnet-4*',
'claude-opus-4*',
# OpenAI families
'gpt-4o*',
'gpt-4.1',
'gpt-5*',
# o-series (keep exact o1 support per existing list)
'o1-2024-12-17',
'o3*',
'o4-mini*',
# Google Gemini
'gemini-2.5-pro*',
# Others
'kimi-k2-0711-preview',
'kimi-k2-instruct',
'qwen3-coder*',
'qwen3-coder-480b-a35b-instruct',
]
REASONING_EFFORT_PATTERNS: list[str] = [
# Mirror main behavior exactly (no unintended expansion), plus DeepSeek support
'o1-2024-12-17',
'o1',
'o3',
'o3-2025-04-16',
'o3-mini-2025-01-31',
'o3-mini',
'o4-mini',
'o4-mini-2025-04-16',
'gemini-2.5-flash',
'gemini-2.5-pro',
'gpt-5',
'gpt-5-2025-08-07',
'claude-opus-4-1-20250805',
# DeepSeek reasoning family
'deepseek-r1-0528*',
]
PROMPT_CACHE_PATTERNS: list[str] = [
'claude-3-7-sonnet*',
'claude-3.7-sonnet*',
'claude-sonnet-3-7-latest',
'claude-3-5-sonnet*',
'claude-3-5-haiku*',
'claude-3.5-haiku*',
'claude-3-haiku-20240307',
'claude-3-opus-20240229',
'claude-sonnet-4*',
'claude-opus-4*',
]
SUPPORTS_STOP_WORDS_FALSE_PATTERNS: list[str] = [
# o1 family doesn't support stop words
'o1*',
# grok-4 specific model name (basename)
'grok-4-0709',
# DeepSeek R1 family
'deepseek-r1-0528*',
]
def get_features(model: str) -> ModelFeatures:
return ModelFeatures(
supports_function_calling=model_matches(model, FUNCTION_CALLING_PATTERNS),
supports_reasoning_effort=model_matches(model, REASONING_EFFORT_PATTERNS),
supports_prompt_cache=model_matches(model, PROMPT_CACHE_PATTERNS),
supports_stop_words=not model_matches(
model, SUPPORTS_STOP_WORDS_FALSE_PATTERNS
),
)

4
openhands/llm/streaming_llm.py
View File

@@ -5,7 +5,7 @@ from typing import Any, Callable
from openhands.core.exceptions import UserCancelledError
from openhands.core.logger import openhands_logger as logger
from openhands.llm.async_llm import LLM_RETRY_EXCEPTIONS, AsyncLLM
from openhands.llm.llm import REASONING_EFFORT_SUPPORTED_MODELS
from openhands.llm.model_features import get_features
class StreamingLLM(AsyncLLM):
@@ -65,7 +65,7 @@ class StreamingLLM(AsyncLLM):
)
# Set reasoning effort for models that support it
if self.config.model.lower() in REASONING_EFFORT_SUPPORTED_MODELS:
if get_features(self.config.model).supports_reasoning_effort:
kwargs['reasoning_effort'] = self.config.reasoning_effort
self.log_prompt(messages)