diff --git a/litellm/litellm_core_utils/thinking_param_translation.py b/litellm/litellm_core_utils/thinking_param_translation.py new file mode 100644 index 00000000000..0ebdb5b3667 --- /dev/null +++ b/litellm/litellm_core_utils/thinking_param_translation.py @@ -0,0 +1,238 @@ +from __future__ import annotations + +from collections.abc import Mapping, Sequence +from dataclasses import dataclass +from types import MappingProxyType +from typing import Final + +from pydantic import TypeAdapter, ValidationError + +_STR_KEYED_MAPPING: Final = TypeAdapter(Mapping[str, object]) +_OBJECT_TUPLE: Final = TypeAdapter(tuple[object, ...]) +_EMPTY: Final[Mapping[str, object]] = MappingProxyType({}) + +_SEND_VIA_EXTRA_BODY: Final = "extra_body" +_SEND_VIA_PROVIDER_MAPPED: Final = "provider_mapped" +_SEND_VIA_VALUES: Final = frozenset((_SEND_VIA_EXTRA_BODY, _SEND_VIA_PROVIDER_MAPPED)) +_THINKING_ENABLED_STRINGS: Final = frozenset(("enabled", "true", "1", "auto")) +_THINKING_TYPE_ENABLED: Final = frozenset(("enabled", "auto", "true")) + +_EFFORT_FALLBACKS: Final[Mapping[str, tuple[str, ...]]] = MappingProxyType( + { + "xhigh": ("max", "high"), + "max": ("xhigh", "high"), + "medium": ("high", "low"), + "minimal": ("low", "none"), + "none": ("low",), + } +) + + +@dataclass(frozen=True, slots=True) +class ThinkingParamsState: + thinking: object | None + reasoning_effort: object | None + extra_body: Mapping[str, object] + + +def str_keyed_mapping_or_none(value: object) -> Mapping[str, object] | None: + if not isinstance(value, Mapping): + return None + try: + return _STR_KEYED_MAPPING.validate_python(value) + except ValidationError: + return None + + +def _as_str_tuple(value: object) -> tuple[str, ...]: + if not isinstance(value, (list, tuple)): + return () + return tuple(item for item in _OBJECT_TUPLE.validate_python(value) if isinstance(item, str)) + + +def _thinking_enabled(thinking: object) -> bool: + if isinstance(thinking, bool): + return thinking + if isinstance(thinking, str): + return thinking.lower() in _THINKING_ENABLED_STRINGS + mapping: Final = str_keyed_mapping_or_none(thinking) + if mapping is None: + return False + typ: Final = mapping.get("type") + if isinstance(typ, str): + return typ.lower() in _THINKING_TYPE_ENABLED + enabled: Final = mapping.get("enabled") + return enabled if isinstance(enabled, bool) else False + + +def _thinking_type_candidate(thinking: object) -> str | None: + match thinking: + case bool(): + return "enabled" if thinking else "disabled" + case str(): + return thinking + case _: + mapping: Final = str_keyed_mapping_or_none(thinking) + raw: Final = mapping.get("type") if mapping is not None else None + return raw if isinstance(raw, str) else None + + +def _thinking_type_value(thinking: object, allowed: Sequence[str]) -> str | None: + candidate: Final = _thinking_type_candidate(thinking) + if candidate is None: + return None + if not allowed or candidate in allowed: + return candidate + if candidate == "auto" and "enabled" in allowed: + return "enabled" + return None + + +def _thinking_payload(thinking: object, typ: str) -> Mapping[str, object]: + mapping: Final = str_keyed_mapping_or_none(thinking) + if mapping is None: + return MappingProxyType({"type": typ}) + return MappingProxyType({**mapping, "type": typ}) + + +def _clamp_effort(value: object, allowed: Sequence[str]) -> str | None: + if not isinstance(value, str): + return None + if not allowed: + return None + if value in allowed: + return value + for fallback in _EFFORT_FALLBACKS.get(value, ()): + if fallback in allowed: + return fallback + return None + + +def _merged_mapping_value(left: Mapping[str, object], right: Mapping[str, object], key: str) -> object: + if key not in right: + return left[key] + if key not in left: + return right[key] + left_map: Final = str_keyed_mapping_or_none(left[key]) + right_map: Final = str_keyed_mapping_or_none(right[key]) + if left_map is not None and right_map is not None: + return _deep_merge_pair(left_map, right_map) + return right[key] + + +def _deep_merge_pair(left: Mapping[str, object], right: Mapping[str, object]) -> Mapping[str, object]: + keys: Final = (*left, *(key for key in right if key not in left)) + return MappingProxyType({key: _merged_mapping_value(left, right, key) for key in keys}) + + +def _thawed(value: object) -> object: + mapping: Final = str_keyed_mapping_or_none(value) + return value if mapping is None else thaw_mapping(mapping) + + +def thaw_mapping(mapping: Mapping[str, object]) -> dict[str, object]: # mutable-ok: JSON request body + return {key: _thawed(item) for key, item in mapping.items()} # mutable-ok: JSON request body + + +def _map_thinking_to_extra_body( + *, + thinking_param: str | None, + thinking: object, + thinking_values: Sequence[str], +) -> Mapping[str, object]: + match thinking_param: + case "thinking.type": + typ: Final = _thinking_type_value(thinking, thinking_values) + if typ is None: + return _EMPTY + return MappingProxyType({"thinking": _thinking_payload(thinking, typ)}) + case "thinking": + thinking_mapping: Final = str_keyed_mapping_or_none(thinking) + if thinking_mapping is not None: + return MappingProxyType({"thinking": MappingProxyType(thinking_mapping)}) + typ_only: Final = _thinking_type_value(thinking, thinking_values or ("enabled", "disabled")) + if typ_only is None: + return _EMPTY + return MappingProxyType({"thinking": MappingProxyType({"type": typ_only})}) + case "enable_thinking": + return MappingProxyType({"enable_thinking": _thinking_enabled(thinking)}) + case "chat_template_kwargs": + return MappingProxyType( + {"chat_template_kwargs": MappingProxyType({"enable_thinking": _thinking_enabled(thinking)})} + ) + case _: + return _EMPTY + + +def _map_effort_to_extra_body( + *, + effort: object, + effort_values: Sequence[str], + send_via: object, +) -> Mapping[str, object]: + clamped: Final = _clamp_effort(effort, effort_values) + if clamped is not None: + return MappingProxyType({"reasoning_effort": clamped}) + if not effort_values and send_via == _SEND_VIA_EXTRA_BODY and isinstance(effort, str): + return MappingProxyType({"reasoning_effort": effort}) + return _EMPTY + + +def translate_thinking_params( + *, + model_info: Mapping[str, object] | None, + state: ThinkingParamsState, +) -> ThinkingParamsState: + if model_info is None: + return state + + send_via: Final = model_info.get("thinking_send_via") + if send_via not in _SEND_VIA_VALUES: + return state + + supports_reasoning: Final = model_info.get("supports_reasoning") is True + thinking_param_raw: Final = model_info.get("thinking_param") + thinking_param: Final = thinking_param_raw if isinstance(thinking_param_raw, str) else None + thinking_values: Final = _as_str_tuple(model_info.get("thinking_values")) + effort_values: Final = _as_str_tuple(model_info.get("reasoning_effort_values")) + + if not supports_reasoning and send_via != _SEND_VIA_PROVIDER_MAPPED: + return state + + thinking: Final = state.thinking + effort: Final = state.reasoning_effort + if thinking is None and effort is None: + return state + + keep_thinking: Final = send_via == _SEND_VIA_PROVIDER_MAPPED + thinking_patch: Final = ( + _map_thinking_to_extra_body( + thinking_param=thinking_param, + thinking=thinking, + thinking_values=thinking_values, + ) + if thinking is not None and send_via == _SEND_VIA_EXTRA_BODY + else _EMPTY + ) + effort_patch: Final = ( + _map_effort_to_extra_body( + effort=effort, + effort_values=effort_values, + send_via=send_via, + ) + if effort is not None + else _EMPTY + ) + patch: Final = MappingProxyType({**thinking_patch, **effort_patch}) + if not patch: + return state + + thinking_mapped: Final = any(key in patch for key in ("thinking", "enable_thinking", "chat_template_kwargs")) + next_thinking: Final = thinking if (keep_thinking or not thinking_mapped) else None + next_effort: Final = None if "reasoning_effort" in patch else effort + merged_extra: Final = _deep_merge_pair(patch, state.extra_body) + return ThinkingParamsState( + thinking=next_thinking, + reasoning_effort=next_effort, + extra_body=merged_extra, + ) diff --git a/litellm/main.py b/litellm/main.py index 6c85adf3ae8..a125c792846 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -5621,6 +5621,7 @@ def completion( "prompt_cache_key": prompt_cache_key, "allowed_openai_params": allowed_openai_params, "base_model": base_model, + "model_info": model_info if isinstance(model_info, dict) else None, } optional_params = get_optional_params(**optional_param_args, **non_default_params) processed_non_default_params: Final = pre_process_non_default_params( diff --git a/litellm/utils.py b/litellm/utils.py index 13a46840431..636e0c9a49f 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -294,6 +294,12 @@ from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, cast, runtime_c from typing_extensions import assert_never from litellm import utils as litellm_utils +from litellm.litellm_core_utils.thinking_param_translation import ( + ThinkingParamsState, + str_keyed_mapping_or_none, + thaw_mapping, + translate_thinking_params, +) # These are lazy loaded via __getattr__ from litellm.llms.base_llm.base_utils import ( @@ -4312,6 +4318,39 @@ def remove_sensitive_keys_from_dict(d: dict) -> dict: return d +def _translate_thinking_in_params( + *, + model_info: Mapping[str, object] | None, + passed_params: dict[str, object], # mutable-ok: get_optional_params hands over its legacy mutable params + non_default_params: dict[str, object], # mutable-ok: get_optional_params hands over its legacy mutable params +) -> tuple[dict[str, object], dict[str, object]]: # mutable-ok: get_optional_params keeps mutating both downstream + state: Final = ThinkingParamsState( + thinking=non_default_params.get("thinking"), + reasoning_effort=non_default_params.get("reasoning_effort"), + extra_body=str_keyed_mapping_or_none(passed_params.get("extra_body")) or MappingProxyType({}), + ) + translated: Final = translate_thinking_params(model_info=model_info, state=state) + if translated is state: + return passed_params, non_default_params + thinking_values: Final = MappingProxyType( + {"thinking": translated.thinking, "reasoning_effort": translated.reasoning_effort} + ) + untouched_non_default: Final = MappingProxyType( + {key: value for key, value in non_default_params.items() if key not in thinking_values} + ) + surviving_thinking_values: Final = MappingProxyType( + {key: value for key, value in thinking_values.items() if value is not None} + ) + return ( + { # mutable-ok: get_optional_params keeps mutating passed_params downstream + **passed_params, + **thinking_values, + "extra_body": thaw_mapping(translated.extra_body), + }, + {**untouched_non_default, **surviving_thinking_values}, # mutable-ok: _check_valid_arg pops unsupported keys + ) + + def pre_process_optional_params(passed_params: dict, non_default_params: dict, custom_llm_provider: str) -> dict: """For .completion(), preprocess optional params""" optional_params: dict = {} @@ -4434,15 +4473,17 @@ def get_optional_params( store: bool | None = None, prompt_cache_key: str | None = None, base_model: str | None = None, + model_info: Mapping[str, object] | None = None, **kwargs, ): drop_params = normalize_drop_params(drop_params) # rebind-ok: config and DB deployments pass "true" as a string - passed_params: Final = locals().copy() - special_params: Final = passed_params.pop("kwargs") + untranslated_passed_params: Final = locals().copy() + special_params: Final = untranslated_passed_params.pop("kwargs") # Remove base_model from passed_params so it doesn't interfere with # non_default_params / _check_valid_arg — it's a routing hint, not an # OpenAI param. - passed_params.pop("base_model", None) + untranslated_passed_params.pop("base_model", None) + untranslated_passed_params.pop("model_info", None) provider_config: BaseConfig | None = None if custom_llm_provider is not None and custom_llm_provider in [provider.value for provider in LlmProviders]: provider_config = ProviderConfigManager.get_provider_chat_config( @@ -4450,14 +4491,19 @@ def get_optional_params( provider=LlmProviders(custom_llm_provider), base_model=base_model, ) - non_default_params: Final = pre_process_non_default_params( - passed_params=passed_params, + untranslated_non_default_params: Final = pre_process_non_default_params( + passed_params=untranslated_passed_params, special_params=special_params, custom_llm_provider=custom_llm_provider, additional_drop_params=additional_drop_params, model=model, provider_config=provider_config, ) + passed_params, non_default_params = _translate_thinking_in_params( + model_info=model_info, + passed_params=untranslated_passed_params, + non_default_params=untranslated_non_default_params, + ) optional_params = pre_process_optional_params( passed_params=passed_params, non_default_params=non_default_params, diff --git a/tests/unit/litellm_core_utils/test_thinking_param_translation.py b/tests/unit/litellm_core_utils/test_thinking_param_translation.py new file mode 100644 index 00000000000..956022ba7f9 --- /dev/null +++ b/tests/unit/litellm_core_utils/test_thinking_param_translation.py @@ -0,0 +1,361 @@ +import copy +import json +from collections.abc import Mapping +from types import MappingProxyType +from typing import Final + +import httpx +import openai +import pytest + +import litellm +from litellm.litellm_core_utils.thinking_param_translation import ( + ThinkingParamsState, + translate_thinking_params, +) +from litellm.utils import get_optional_params + +_EXTRA_BODY_MODEL_INFO: Final = MappingProxyType( + { + "supports_reasoning": True, + "thinking_param": "thinking.type", + "thinking_values": ["enabled", "disabled"], + "reasoning_effort_values": ["low", "high", "max"], + "thinking_send_via": "extra_body", + } +) + +_CHAT_COMPLETION_RESPONSE: Final = MappingProxyType( + { + "id": "chatcmpl-thinking", + "object": "chat.completion", + "created": 0, + "model": "deepseek-v4-flash", + "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}], + "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2}, + } +) + + +def _model_info(**overrides: object) -> Mapping[str, object]: + return MappingProxyType({**_EXTRA_BODY_MODEL_INFO, **overrides}) + + +def _state( + *, + thinking: object = None, + reasoning_effort: object = None, + extra_body: Mapping[str, object] = MappingProxyType({}), +) -> ThinkingParamsState: + return ThinkingParamsState(thinking=thinking, reasoning_effort=reasoning_effort, extra_body=extra_body) + + +def _as_plain(value: object) -> object: + if isinstance(value, Mapping): + return {key: _as_plain(item) for key, item in value.items()} + return value + + +def _recording_client(bodies: list[object]) -> openai.OpenAI: + def respond(request: httpx.Request) -> httpx.Response: + bodies.append(json.loads(request.content)) + return httpx.Response(200, json=dict(_CHAT_COMPLETION_RESPONSE)) + + return openai.OpenAI(api_key="test-key", http_client=httpx.Client(transport=httpx.MockTransport(respond))) + + +@pytest.mark.parametrize( + ("model_info", "thinking", "reasoning_effort", "expected_extra_body"), + [ + pytest.param( + _model_info(), + {"type": "enabled", "budget_tokens": 1024, "clear_thinking": False}, + "high", + { + "thinking": {"type": "enabled", "budget_tokens": 1024, "clear_thinking": False}, + "reasoning_effort": "high", + }, + id="thinking_type_keeps_caller_thinking_keys", + ), + pytest.param( + _model_info(), + {"type": "auto"}, + None, + {"thinking": {"type": "enabled"}}, + id="thinking_type_auto_falls_back_to_enabled", + ), + pytest.param( + _model_info(), + False, + None, + {"thinking": {"type": "disabled"}}, + id="thinking_type_from_bool", + ), + pytest.param( + _model_info(), + "enabled", + None, + {"thinking": {"type": "enabled"}}, + id="thinking_type_from_string", + ), + pytest.param( + _model_info(thinking_param="thinking", thinking_values=[]), + {"type": "enabled", "budget_tokens": 2048}, + None, + {"thinking": {"type": "enabled", "budget_tokens": 2048}}, + id="thinking_dict_passthrough", + ), + pytest.param( + _model_info(thinking_param="thinking", thinking_values=[]), + True, + None, + {"thinking": {"type": "enabled"}}, + id="thinking_from_bool", + ), + pytest.param( + _model_info(thinking_param="enable_thinking"), + {"type": "enabled"}, + None, + {"enable_thinking": True}, + id="enable_thinking_from_type", + ), + pytest.param( + _model_info(thinking_param="enable_thinking"), + {"enabled": True}, + None, + {"enable_thinking": True}, + id="enable_thinking_from_enabled_flag", + ), + pytest.param( + _model_info(thinking_param="enable_thinking"), + "true", + None, + {"enable_thinking": True}, + id="enable_thinking_from_string", + ), + pytest.param( + _model_info(thinking_param="enable_thinking"), + False, + None, + {"enable_thinking": False}, + id="enable_thinking_from_bool", + ), + pytest.param( + _model_info(thinking_param="enable_thinking"), + 1, + None, + {"enable_thinking": False}, + id="enable_thinking_unrecognized_value_is_disabled", + ), + pytest.param( + _model_info(thinking_param="chat_template_kwargs", reasoning_effort_values=["low", "medium", "high"]), + {"type": "disabled"}, + "medium", + {"chat_template_kwargs": {"enable_thinking": False}, "reasoning_effort": "medium"}, + id="chat_template_kwargs", + ), + pytest.param( + _model_info(reasoning_effort_values=["low", "high", "max"]), + None, + "xhigh", + {"reasoning_effort": "max"}, + id="effort_clamped_to_alias", + ), + pytest.param( + _model_info(reasoning_effort_values=["low"]), + None, + "minimal", + {"reasoning_effort": "low"}, + id="effort_clamped_down", + ), + pytest.param( + _model_info(reasoning_effort_values=[]), + None, + "high", + {"reasoning_effort": "high"}, + id="effort_passthrough_without_allowed_values", + ), + pytest.param( + _model_info(reasoning_effort_values=None), + None, + "high", + {"reasoning_effort": "high"}, + id="effort_passthrough_when_allowed_values_missing", + ), + ], +) +def test_translate_moves_params_into_extra_body( + model_info: Mapping[str, object], + thinking: object, + reasoning_effort: object, + expected_extra_body: dict[str, object], +): + result = translate_thinking_params( + model_info=model_info, state=_state(thinking=thinking, reasoning_effort=reasoning_effort) + ) + + assert (result.thinking, result.reasoning_effort) == (None, None) + assert _as_plain(result.extra_body) == expected_extra_body + + +@pytest.mark.parametrize( + ("model_info", "thinking", "reasoning_effort"), + [ + pytest.param(None, {"type": "enabled"}, "high", id="no_model_info"), + pytest.param(_model_info(thinking_send_via="n/a"), {"type": "enabled"}, "high", id="send_via_not_applicable"), + pytest.param(_model_info(supports_reasoning=False), {"type": "enabled"}, "high", id="reasoning_unsupported"), + pytest.param(_model_info(), None, None, id="nothing_requested"), + pytest.param(_model_info(thinking_param="unknown"), {"type": "enabled"}, None, id="unknown_thinking_param"), + pytest.param(_model_info(thinking_param=None), {"type": "enabled"}, None, id="thinking_param_missing"), + pytest.param(_model_info(), {"type": "adaptive"}, None, id="thinking_type_not_allowed"), + pytest.param(_model_info(), {"budget_tokens": 1024}, None, id="thinking_type_unreadable"), + pytest.param(_model_info(), 1, None, id="thinking_type_unsupported_value"), + pytest.param(_model_info(), {1: "enabled"}, None, id="thinking_mapping_with_non_string_keys"), + pytest.param( + _model_info(thinking_param="thinking", thinking_values=[]), "adaptive", None, id="thinking_value_unmapped" + ), + pytest.param(_model_info(reasoning_effort_values=["low"]), None, "ultra", id="effort_without_fallback"), + pytest.param(_model_info(), None, 5, id="effort_not_a_string"), + ], +) +def test_translate_returns_state_unchanged_when_nothing_applies( + model_info: Mapping[str, object] | None, + thinking: object, + reasoning_effort: object, +): + state = _state(thinking=thinking, reasoning_effort=reasoning_effort) + + assert translate_thinking_params(model_info=model_info, state=state) is state + + +def test_translate_provider_mapped_keeps_thinking_and_moves_effort(): + thinking = {"type": "enabled"} + + result = translate_thinking_params( + model_info=_model_info(thinking_send_via="provider_mapped", supports_reasoning=False), + state=_state(thinking=thinking, reasoning_effort="high"), + ) + + assert result.thinking is thinking + assert result.reasoning_effort is None + assert _as_plain(result.extra_body) == {"reasoning_effort": "high"} + + +def test_translate_keeps_caller_extra_body_values_over_translated_ones(): + result = translate_thinking_params( + model_info=_model_info(thinking_param="chat_template_kwargs", thinking_values=[]), + state=_state( + thinking={"type": "enabled"}, + reasoning_effort="high", + extra_body={"chat_template_kwargs": {"enable_thinking": False, "reasoning_budget": 512}, "top_k": 20}, + ), + ) + + assert (result.thinking, result.reasoning_effort) == (None, None) + assert _as_plain(result.extra_body) == { + "chat_template_kwargs": {"enable_thinking": False, "reasoning_budget": 512}, + "reasoning_effort": "high", + "top_k": 20, + } + + +def test_get_optional_params_moves_thinking_into_extra_body(): + optional_params = get_optional_params( + model="deepseek-v4-flash", + custom_llm_provider="openai", + drop_params=True, + thinking={"type": "enabled"}, + reasoning_effort="high", + model_info=_model_info(), + ) + + assert "thinking" not in optional_params + assert "reasoning_effort" not in optional_params + assert optional_params["extra_body"] == {"thinking": {"type": "enabled"}, "reasoning_effort": "high"} + + +def test_get_optional_params_without_model_info_drops_thinking(): + optional_params = get_optional_params( + model="gpt-4o", + custom_llm_provider="openai", + drop_params=True, + thinking={"type": "enabled"}, + reasoning_effort="high", + ) + + assert "thinking" not in optional_params + assert "thinking" not in (optional_params.get("extra_body") or {}) + + +def test_get_optional_params_does_not_reintroduce_dropped_thinking(): + optional_params = get_optional_params( + model="deepseek-v4-flash", + custom_llm_provider="openai", + drop_params=True, + thinking={"type": "enabled"}, + additional_drop_params=["thinking"], + model_info=_model_info(thinking_param="chat_template_kwargs", thinking_values=[]), + ) + + assert "thinking" not in optional_params + assert optional_params.get("extra_body") in (None, {}) + + +def test_get_optional_params_leaves_caller_extra_body_untouched_and_serializable(): + caller_extra_body = {"chat_template_kwargs": {"reasoning_budget": 512}} + + optional_params = get_optional_params( + model="deepseek-v4-flash", + custom_llm_provider="openai", + thinking={"type": "enabled"}, + extra_body=caller_extra_body, + model_info=_model_info(thinking_param="chat_template_kwargs", thinking_values=[]), + ) + + expected_extra_body = {"chat_template_kwargs": {"reasoning_budget": 512, "enable_thinking": True}} + assert optional_params["extra_body"] == expected_extra_body + assert json.loads(json.dumps(optional_params["extra_body"])) == expected_extra_body + assert copy.deepcopy(optional_params)["extra_body"] == expected_extra_body + assert caller_extra_body == {"chat_template_kwargs": {"reasoning_budget": 512}} + + +def test_completion_sends_translated_thinking_on_the_wire(): + bodies: list[object] = [] + + litellm.completion( + model="openai/deepseek-v4-flash", + messages=[{"role": "user", "content": "hi"}], + thinking={"type": "enabled", "budget_tokens": 1024}, + reasoning_effort="high", + model_info=dict(_model_info()), + client=_recording_client(bodies), + num_retries=0, + ) + + assert bodies == [ + { + "model": "deepseek-v4-flash", + "messages": [{"role": "user", "content": "hi"}], + "thinking": {"type": "enabled", "budget_tokens": 1024}, + "reasoning_effort": "high", + } + ] + + +def test_batch_completion_translates_every_request_like_completion(): + bodies: list[object] = [] + + litellm.batch_completion( + model="openai/deepseek-v4-flash", + messages=[[{"role": "user", "content": "one"}], [{"role": "user", "content": "two"}]], + thinking={"type": "enabled"}, + model_info=dict(_model_info(thinking_param="enable_thinking")), + client=_recording_client(bodies), + num_retries=0, + max_workers=1, + ) + + assert sorted(bodies, key=lambda body: json.dumps(body, sort_keys=True)) == [ + {"model": "deepseek-v4-flash", "messages": [{"role": "user", "content": "one"}], "enable_thinking": True}, + {"model": "deepseek-v4-flash", "messages": [{"role": "user", "content": "two"}], "enable_thinking": True}, + ]