mirror of
https://github.com/BerriAI/litellm.git
synced 2026-11-01 07:03:02 +00:00
d4e69e48a4
DeepSeek V4 (`deepseek-v4-pro` / `deepseek-v4-flash`, released 2026-04) is a
reasoner-only model family that rejects:
- `tool_choice="required"`
- `tool_choice="any"`
- `tool_choice={"type":"function","function":{"name":"..."}}`
with HTTP 400 from the DeepSeek API:
deepseek-reasoner does not support this tool_choice
Only `"auto"` and `"none"` are accepted in thinking mode, and V4 enables
thinking mode by default — so this rejection happens even when the caller
did NOT explicitly send `thinking={"type":"enabled"}`.
This breaks any SDK that hardcodes `tool_choice="required"` for forced
tool selection. A real-world example: OpenAI Agents SDK 0.17+ uses
`ModelSettings(tool_choice="required")` on its delegation /
sub-agent paths to force the parent agent to actually dispatch a tool.
With DeepSeek V4 as the routed model, every delegation call 400s.
## Fix
In `DeepSeekChatConfig.map_openai_params`, when the model is recognised
as a V4 reasoner (either by model-name substring match — needed for the
first turn before `thinking` propagates — OR by `thinking={"type":
"enabled"}` already being set), rewrite unsupported `tool_choice` values
to `"auto"`. `"auto"` and `"none"` pass through unchanged. Non-V4 models
(e.g. `deepseek-chat` / V3) are unaffected.
## Tests
Added `TestDeepSeekV4ToolChoiceConstraint` with 7 cases:
- v4-pro + tool_choice="required" → "auto"
- v4-flash + tool_choice="any" → "auto"
- v4-pro + dict tool_choice → "auto"
- v4-pro + "auto" → unchanged
- v4-pro + "none" → unchanged
- non-V4 deepseek-chat + "required" → unchanged (no regression)
- explicit thinking={"type":"enabled"} on any model + "required" → "auto"
All 18 tests in this file pass (11 existing + 7 new):
pytest tests/litellm/llms/deepseek/chat/test_deepseek_chat_transformation.py
18 passed in 2.38s
Real-world verified: re-ran a multi-turn delegation e2e against
`deepseek/deepseek-v4-pro` (Lead `Financial Advisor` → SubAgent
`market_analyst` with `tool_choice="required"` forced by the SDK) — 3
prior 400s now zero, full 7-section market-review output completes.
## Refs
- DeepSeek docs: https://api-docs.deepseek.com/guides/thinking_mode
- Related upstream context: V4 series is reasoner-only by design, mirroring
`deepseek-reasoner` (R1) tool_choice semantics.
|
||
|---|---|---|
| .. | ||
| agent_tests | ||
| audio_tests | ||
| basic_proxy_startup_tests | ||
| batches_tests | ||
| benchmarks | ||
| code_coverage_tests | ||
| documentation_tests | ||
| enterprise | ||
| guardrails_tests | ||
| image_gen_tests | ||
| litellm | ||
| litellm-proxy-extras | ||
| litellm_core_utils | ||
| litellm_utils_tests | ||
| llm_responses_api_testing | ||
| llm_translation | ||
| load_tests | ||
| local_testing | ||
| logging_callback_tests | ||
| mcp_tests | ||
| multi_instance_e2e_tests | ||
| ocr_tests | ||
| old_proxy_tests/tests | ||
| openai_endpoints_tests | ||
| otel_tests | ||
| pass_through_tests | ||
| pass_through_unit_tests | ||
| proxy_admin_ui_tests | ||
| proxy_e2e_anthropic_messages_tests | ||
| proxy_security_tests | ||
| proxy_unit_tests | ||
| router_unit_tests | ||
| scim_tests | ||
| search_tests | ||
| spend_tracking_tests | ||
| store_model_in_db_tests | ||
| test_litellm | ||
| unified_google_tests | ||
| vector_store_tests | ||
| windows_tests | ||
| __init__.py | ||
| _flush_vcr_cache.py | ||
| _vcr_conftest_common.py | ||
| _vcr_redis_persister.py | ||
| eval_swe_bench.py | ||
| gettysburg.wav | ||
| large_text.py | ||
| openai_batch_completions.jsonl | ||
| README.MD | ||
| test_budget_management.py | ||
| test_callbacks_on_proxy.py | ||
| test_config.py | ||
| test_debug_warning.py | ||
| test_default_encoding_non_root.py | ||
| test_end_users.py | ||
| test_entrypoint.py | ||
| test_fallbacks.py | ||
| test_gpt5_azure_temperature_support.py | ||
| test_health.py | ||
| test_keys.py | ||
| test_litellm_proxy_responses_config.py | ||
| test_logging.conf | ||
| test_models.py | ||
| test_new_vector_store_endpoints.py | ||
| test_openai_endpoints.py | ||
| test_organizations.py | ||
| test_otel_thread_leak.py | ||
| test_passthrough_endpoints.py | ||
| test_presidio_latency.py | ||
| test_proxy_server_non_root.py | ||
| test_ratelimit.py | ||
| test_resource_cleanup.py | ||
| test_service_logger_otel.py | ||
| test_spend_logs.py | ||
| test_team.py | ||
| test_team_logging.py | ||
| test_team_members.py | ||
| test_users.py | ||
In total litellm runs 1000+ tests
[02/20/2025] Update:
To make it easier to contribute and map what behavior is tested,
we've started mapping the litellm directory in tests/test_litellm
This folder can only run mock tests.