sync update from LAN repo (source dhk-1 @ 3b1ec5b4)
This commit is contained in:
1 parent
da6521b6cb
commit
264e2200d6
995 files changed
+148105
-13399
No files matched your search
@@ -6,6 +6,7 @@ from deerflow.config import get_app_config
|
||||
from deerflow.config.app_config import AppConfig
|
||||
from deerflow.config.model_config import WriteFileTransport
|
||||
from deerflow.config.system_settings import get_write_file_compatibility_settings
|
||||
from deerflow.models.chat_stream_state import needs_qwen_flash_next_integrity
|
||||
from deerflow.reflection import resolve_class
|
||||
from deerflow.tracing import build_tracing_callbacks
|
||||
|
||||
@@ -86,7 +87,6 @@ def _apply_openai_thinking_control(
|
||||
model_settings_from_config: dict,
|
||||
*,
|
||||
thinking_enabled: bool,
|
||||
reasoning_effort: object,
|
||||
) -> None:
|
||||
"""Inject vLLM/Qwen thinking flags when the UI toggle is on or off."""
|
||||
if "extra_body" not in getattr(model_class, "model_fields", {}):
|
||||
@@ -99,9 +99,6 @@ def _apply_openai_thinking_control(
|
||||
"enable_thinking": True,
|
||||
"chat_template_kwargs": {"enable_thinking": True},
|
||||
}
|
||||
mapped_effort = _map_qwen_reasoning_effort(reasoning_effort)
|
||||
if mapped_effort:
|
||||
enable_body["chat_template_kwargs"]["reasoning_effort"] = mapped_effort
|
||||
model_settings_from_config["extra_body"] = _deep_merge_dicts(
|
||||
model_settings_from_config.get("extra_body"),
|
||||
enable_body,
|
||||
@@ -139,7 +136,16 @@ def _enable_stream_usage_by_default(model_use_path: str, model_settings_from_con
|
||||
model_settings_from_config["stream_usage"] = True
|
||||
|
||||
|
||||
def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *, app_config: AppConfig | None = None, force_disable_thinking: bool = False, **kwargs) -> BaseChatModel:
|
||||
def create_chat_model(
|
||||
name: str | None = None,
|
||||
thinking_enabled: bool = False,
|
||||
*,
|
||||
app_config: AppConfig | None = None,
|
||||
force_disable_thinking: bool = False,
|
||||
write_file_compatibility: bool | None = None,
|
||||
prompt_suffix_enabled: bool = True,
|
||||
**kwargs,
|
||||
) -> BaseChatModel:
|
||||
"""Create a chat model instance from the config.
|
||||
|
||||
Args:
|
||||
@@ -172,8 +178,8 @@ def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *
|
||||
from deerflow.models.patched_openai import PatchedChatOpenAI
|
||||
|
||||
model_class = PatchedChatOpenAI
|
||||
model_settings_from_config = model_config.model_dump(
|
||||
exclude_none=True,
|
||||
model_settings_from_config = model_config.model_dump(
|
||||
exclude_none=True,
|
||||
exclude={
|
||||
"use",
|
||||
"name",
|
||||
@@ -186,15 +192,37 @@ def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *
|
||||
"thinking",
|
||||
"supports_vision",
|
||||
"write_file_transport",
|
||||
"tool_call_transport",
|
||||
},
|
||||
)
|
||||
configured_write_file_transport = model_config.write_file_transport
|
||||
runtime_write_file_compatibility = get_write_file_compatibility_settings()
|
||||
# ``None`` follows the global switch, so a roundtable run with its own
|
||||
# switch off still uses compatibility when the global switch is on.
|
||||
# ``True`` is the roundtable opt-in used when the global switch is off.
|
||||
compatibility_enabled = (
|
||||
get_write_file_compatibility_settings().enabled
|
||||
if write_file_compatibility is None
|
||||
else write_file_compatibility
|
||||
)
|
||||
effective_write_file_transport = (
|
||||
WriteFileTransport.BUFFERED_GATEWAY
|
||||
if runtime_write_file_compatibility.enabled
|
||||
if compatibility_enabled
|
||||
else configured_write_file_transport
|
||||
)
|
||||
if model_config.tool_call_transport.value != "native":
|
||||
from deerflow.models.patched_openai import PatchedChatOpenAI
|
||||
|
||||
if not issubclass(model_class, PatchedChatOpenAI):
|
||||
raise ValueError(
|
||||
"tool_call_transport=deepseek_dsml is supported only by "
|
||||
"langchain_openai:ChatOpenAI or PatchedChatOpenAI models"
|
||||
)
|
||||
if model_config.use_responses_api:
|
||||
raise ValueError(
|
||||
"tool_call_transport=deepseek_dsml requires the OpenAI-compatible "
|
||||
"chat-completions API, not use_responses_api"
|
||||
)
|
||||
model_settings_from_config["tool_call_transport"] = model_config.tool_call_transport.value
|
||||
if effective_write_file_transport in {
|
||||
WriteFileTransport.BUFFERED_GATEWAY,
|
||||
WriteFileTransport.ARTIFACT_RENDERER,
|
||||
@@ -235,18 +263,18 @@ def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *
|
||||
if model_config.thinking is not None:
|
||||
merged_thinking = {**(effective_wte.get("thinking") or {}), **model_config.thinking}
|
||||
effective_wte = {**effective_wte, "thinking": merged_thinking}
|
||||
if thinking_enabled and has_thinking_settings:
|
||||
if not model_config.supports_thinking:
|
||||
raise ValueError(f"Model {name} does not support thinking. Set `supports_thinking` to true in the `config.yaml` to enable thinking.") from None
|
||||
if effective_wte:
|
||||
model_settings_from_config.update(effective_wte)
|
||||
elif thinking_enabled and model_config.supports_thinking and model_config.use == "langchain_openai:ChatOpenAI":
|
||||
# vLLM/Qwen OpenAI-compatible endpoints commonly require this explicit
|
||||
# request flag before they emit delta.reasoning.
|
||||
model_settings_from_config["extra_body"] = _deep_merge_dicts(
|
||||
model_settings_from_config.get("extra_body"),
|
||||
{"enable_thinking": True},
|
||||
)
|
||||
if thinking_enabled and has_thinking_settings:
|
||||
if not model_config.supports_thinking:
|
||||
raise ValueError(f"Model {name} does not support thinking. Set `supports_thinking` to true in the `config.yaml` to enable thinking.") from None
|
||||
if effective_wte:
|
||||
model_settings_from_config.update(effective_wte)
|
||||
elif thinking_enabled and model_config.supports_thinking and model_config.use == "langchain_openai:ChatOpenAI":
|
||||
# vLLM/Qwen OpenAI-compatible endpoints commonly require this explicit
|
||||
# request flag before they emit delta.reasoning.
|
||||
model_settings_from_config["extra_body"] = _deep_merge_dicts(
|
||||
model_settings_from_config.get("extra_body"),
|
||||
{"enable_thinking": True},
|
||||
)
|
||||
if not thinking_enabled:
|
||||
if model_config.when_thinking_disabled is not None:
|
||||
# User-provided disable settings take full precedence
|
||||
@@ -267,17 +295,24 @@ def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *
|
||||
elif has_thinking_settings and effective_wte.get("thinking", {}).get("type"):
|
||||
# Native langchain_anthropic: thinking is a direct constructor parameter
|
||||
model_settings_from_config["thinking"] = {"type": "disabled"}
|
||||
# The Gateway validates the database-backed capability before constructing
|
||||
# the graph. Honor that server-validated runtime value without consulting a
|
||||
# legacy config.yaml flag. Qwen/vLLM accepts low/medium/xhigh; normalize the
|
||||
# old UI aliases defensively for internal callers that do not traverse the
|
||||
# public run boundary.
|
||||
requested_reasoning_effort = kwargs.get("reasoning_effort")
|
||||
if not model_config.supports_reasoning_effort:
|
||||
kwargs.pop("reasoning_effort", None)
|
||||
if requested_reasoning_effort is not None:
|
||||
model_settings_from_config.pop("reasoning_effort", None)
|
||||
if _is_vllm_provider(model_config.use) or _is_qwen38_model(model_config):
|
||||
mapped_effort = _map_qwen_reasoning_effort(requested_reasoning_effort)
|
||||
if mapped_effort is not None:
|
||||
kwargs["reasoning_effort"] = mapped_effort
|
||||
|
||||
_apply_openai_thinking_control(
|
||||
model_config,
|
||||
model_class,
|
||||
model_settings_from_config,
|
||||
thinking_enabled=thinking_enabled,
|
||||
reasoning_effort=requested_reasoning_effort,
|
||||
)
|
||||
|
||||
# Hard override: force-disable thinking regardless of the per-model config shape.
|
||||
@@ -291,12 +326,7 @@ def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *
|
||||
# Only models that explicitly opt into thinking should receive vendor-specific
|
||||
# disable payloads. Generic OpenAI-compatible gateways (including LiteLLM)
|
||||
# may reject these otherwise-unknown fields instead of forwarding them.
|
||||
if (
|
||||
force_disable_thinking
|
||||
and not thinking_enabled
|
||||
and model_config.supports_thinking
|
||||
and "extra_body" in getattr(model_class, "model_fields", {})
|
||||
):
|
||||
if force_disable_thinking and not thinking_enabled and model_config.supports_thinking and "extra_body" in getattr(model_class, "model_fields", {}):
|
||||
model_settings_from_config["extra_body"] = _deep_merge_dicts(
|
||||
model_settings_from_config.get("extra_body"),
|
||||
{"enable_thinking": False, "chat_template_kwargs": {"enable_thinking": False}, "thinking": {"type": "disabled"}},
|
||||
@@ -335,7 +365,23 @@ def create_chat_model(name: str | None = None, thinking_enabled: bool = False, *
|
||||
if "stream_usage" in getattr(model_class, "model_fields", {}):
|
||||
model_settings_from_config["stream_usage"] = True
|
||||
|
||||
model_instance = model_class(**kwargs, **model_settings_from_config)
|
||||
# The deployment may expose the affected model under a generic provider id.
|
||||
# Enable its local stream checks by either id or configured name, while
|
||||
# honoring an explicit override and keeping other models on their old path.
|
||||
if "strict_stream_integrity" in getattr(model_class, "model_fields", {}):
|
||||
if "strict_stream_integrity" in kwargs:
|
||||
model_settings_from_config["strict_stream_integrity"] = kwargs.pop("strict_stream_integrity")
|
||||
elif needs_qwen_flash_next_integrity(model_config.name, model_config.model):
|
||||
model_settings_from_config.setdefault("strict_stream_integrity", True)
|
||||
|
||||
from deerflow.models.prompt_suffix import with_model_prompt_suffix
|
||||
|
||||
model_instance = with_model_prompt_suffix(model_class)(
|
||||
**kwargs,
|
||||
**model_settings_from_config,
|
||||
deerflow_prompt_model_name=name,
|
||||
deerflow_prompt_suffix_enabled=prompt_suffix_enabled,
|
||||
)
|
||||
|
||||
# Always emit a compact model lifecycle trace. In particular, the first
|
||||
# streamed callback proves that the upstream endpoint accepted the request
|
||||
|
||||
Reference in new issue
Block a user