diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 9b5a7b42d0..f6edcf7efd 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -10756,14 +10756,22 @@ "input_cost_per_token": 2.8e-07, "input_cost_per_token_cache_hit": 2.8e-08, "litellm_provider": "deepseek", - "max_input_tokens": 128000, + "max_input_tokens": 131072, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", "output_cost_per_token": 4.2e-07, + "source": "https://api-docs.deepseek.com/quick_start/pricing", + "supported_endpoints": [ + "/v1/chat/completions" + ], "supports_assistant_prefill": true, "supports_function_calling": true, + "supports_native_streaming": true, + "supports_parallel_function_calling": true, "supports_prompt_caching": true, + "supports_response_schema": true, + "supports_system_messages": true, "supports_tool_choice": true }, "deepseek/deepseek-coder": { @@ -10800,16 +10808,24 @@ "input_cost_per_token": 2.8e-07, "input_cost_per_token_cache_hit": 2.8e-08, "litellm_provider": "deepseek", - "max_input_tokens": 128000, - "max_output_tokens": 8192, - "max_tokens": 8192, + "max_input_tokens": 131072, + "max_output_tokens": 65536, + "max_tokens": 65536, "mode": "chat", "output_cost_per_token": 4.2e-07, + "source": "https://api-docs.deepseek.com/quick_start/pricing", + "supported_endpoints": [ + "/v1/chat/completions" + ], "supports_assistant_prefill": true, - "supports_function_calling": true, + "supports_function_calling": false, + "supports_native_streaming": true, + "supports_parallel_function_calling": false, "supports_prompt_caching": true, "supports_reasoning": true, - "supports_tool_choice": true + "supports_response_schema": true, + "supports_system_messages": true, + "supports_tool_choice": false }, "deepseek/deepseek-v3": { "cache_creation_input_token_cost": 0.0, diff --git a/litellm/proxy/policy_engine/policy_resolve_endpoints.py b/litellm/proxy/policy_engine/policy_resolve_endpoints.py index a81d440567..eb4d3fc584 100644 --- a/litellm/proxy/policy_engine/policy_resolve_endpoints.py +++ b/litellm/proxy/policy_engine/policy_resolve_endpoints.py @@ -12,7 +12,6 @@ from fastapi import APIRouter, Depends, HTTPException, Query from litellm._logging import verbose_proxy_logger from litellm.constants import MAX_POLICY_ESTIMATE_IMPACT_ROWS from litellm.proxy._types import UserAPIKeyAuth -from litellm.proxy.auth.route_checks import RouteChecks from litellm.proxy.auth.user_api_key_auth import user_api_key_auth from litellm.proxy.policy_engine.attachment_registry import get_attachment_registry from litellm.proxy.policy_engine.policy_registry import get_policy_registry diff --git a/litellm/utils.py b/litellm/utils.py index ed0d6ee930..291969cc5d 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -2506,6 +2506,16 @@ def _supports_factory(model: str, custom_llm_provider: Optional[str], key: str) if model_info.get(key, False) is True: return True elif model_info.get(key) is None: # don't check if 'False' explicitly set + # Fallback: when the provider-prefixed entry (e.g. + # "deepseek/deepseek-chat") exists but is missing a capability + # field, check the bare model-name entry (e.g. "deepseek-chat") + # which may carry the complete metadata. See #20885. + bare_model_key = _get_model_cost_key(model) + if bare_model_key is not None: + bare_entry = litellm.model_cost.get(bare_model_key) or {} + if bare_entry.get(key, False) is True: + return True + supported_by_provider = _supports_provider_info_factory( model, custom_llm_provider, key ) diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 9b5a7b42d0..f6edcf7efd 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -10756,14 +10756,22 @@ "input_cost_per_token": 2.8e-07, "input_cost_per_token_cache_hit": 2.8e-08, "litellm_provider": "deepseek", - "max_input_tokens": 128000, + "max_input_tokens": 131072, "max_output_tokens": 8192, "max_tokens": 8192, "mode": "chat", "output_cost_per_token": 4.2e-07, + "source": "https://api-docs.deepseek.com/quick_start/pricing", + "supported_endpoints": [ + "/v1/chat/completions" + ], "supports_assistant_prefill": true, "supports_function_calling": true, + "supports_native_streaming": true, + "supports_parallel_function_calling": true, "supports_prompt_caching": true, + "supports_response_schema": true, + "supports_system_messages": true, "supports_tool_choice": true }, "deepseek/deepseek-coder": { @@ -10800,16 +10808,24 @@ "input_cost_per_token": 2.8e-07, "input_cost_per_token_cache_hit": 2.8e-08, "litellm_provider": "deepseek", - "max_input_tokens": 128000, - "max_output_tokens": 8192, - "max_tokens": 8192, + "max_input_tokens": 131072, + "max_output_tokens": 65536, + "max_tokens": 65536, "mode": "chat", "output_cost_per_token": 4.2e-07, + "source": "https://api-docs.deepseek.com/quick_start/pricing", + "supported_endpoints": [ + "/v1/chat/completions" + ], "supports_assistant_prefill": true, - "supports_function_calling": true, + "supports_function_calling": false, + "supports_native_streaming": true, + "supports_parallel_function_calling": false, "supports_prompt_caching": true, "supports_reasoning": true, - "supports_tool_choice": true + "supports_response_schema": true, + "supports_system_messages": true, + "supports_tool_choice": false }, "deepseek/deepseek-v3": { "cache_creation_input_token_cost": 0.0, diff --git a/tests/test_litellm/test_deepseek_model_metadata.py b/tests/test_litellm/test_deepseek_model_metadata.py new file mode 100644 index 0000000000..4900af5d97 --- /dev/null +++ b/tests/test_litellm/test_deepseek_model_metadata.py @@ -0,0 +1,180 @@ +""" +Regression tests for #20885 – ``supports_response_schema`` (and related +capability flags) must be consistent between the bare model-name entry +(e.g. ``deepseek-chat``) and the provider-prefixed entry +(e.g. ``deepseek/deepseek-chat``) in the model-cost map. + +The bug caused ``supports_response_schema("deepseek/deepseek-chat")`` to +return ``False`` even though the canonical ``deepseek-chat`` entry has the +field set to ``True``. +""" + +import json +import os +import sys + +sys.path.insert( + 0, os.path.abspath("../..") +) # Adds the parent directory to the system path + +import litellm +from litellm.utils import ( + _supports_factory, + supports_response_schema, +) + + +# --------------------------------------------------------------------------- +# Data-level tests – verify the JSON files are in sync +# --------------------------------------------------------------------------- + + +def _load_backup_json() -> dict: + """Load the backup JSON directly from disk.""" + backup_path = os.path.join( + os.path.dirname(litellm.__file__), + "model_prices_and_context_window_backup.json", + ) + with open(backup_path, encoding="utf-8") as f: + return json.load(f) + + +class TestDeepSeekModelCostEntries: + """Verify that provider-prefixed DeepSeek entries contain the same + capability flags as their bare-name counterparts in the JSON files.""" + + def test_deepseek_chat_supports_response_schema_in_backup(self): + data = _load_backup_json() + entry = data.get("deepseek/deepseek-chat", {}) + assert entry.get("supports_response_schema") is True + + def test_deepseek_reasoner_supports_response_schema_in_backup(self): + data = _load_backup_json() + entry = data.get("deepseek/deepseek-reasoner", {}) + assert entry.get("supports_response_schema") is True + + def test_deepseek_chat_supports_system_messages_in_backup(self): + data = _load_backup_json() + entry = data.get("deepseek/deepseek-chat", {}) + assert entry.get("supports_system_messages") is True + + def test_deepseek_reasoner_supports_system_messages_in_backup(self): + data = _load_backup_json() + entry = data.get("deepseek/deepseek-reasoner", {}) + assert entry.get("supports_system_messages") is True + + def test_deepseek_chat_max_input_tokens_matches_bare_in_backup(self): + data = _load_backup_json() + bare = data.get("deepseek-chat", {}) + prefixed = data.get("deepseek/deepseek-chat", {}) + assert prefixed.get("max_input_tokens") == bare.get("max_input_tokens") + + def test_deepseek_reasoner_max_output_tokens_matches_bare_in_backup(self): + data = _load_backup_json() + bare = data.get("deepseek-reasoner", {}) + prefixed = data.get("deepseek/deepseek-reasoner", {}) + assert prefixed.get("max_output_tokens") == bare.get("max_output_tokens") + + def test_main_json_deepseek_chat_supports_response_schema(self): + main_path = os.path.join( + os.path.dirname(os.path.dirname(litellm.__file__)), + "model_prices_and_context_window.json", + ) + with open(main_path, encoding="utf-8") as f: + data = json.load(f) + entry = data.get("deepseek/deepseek-chat", {}) + assert entry.get("supports_response_schema") is True + + def test_main_json_deepseek_reasoner_supports_response_schema(self): + main_path = os.path.join( + os.path.dirname(os.path.dirname(litellm.__file__)), + "model_prices_and_context_window.json", + ) + with open(main_path, encoding="utf-8") as f: + data = json.load(f) + entry = data.get("deepseek/deepseek-reasoner", {}) + assert entry.get("supports_response_schema") is True + + +# --------------------------------------------------------------------------- +# API-level tests – verify supports_response_schema returns True +# --------------------------------------------------------------------------- + + +class TestSupportsResponseSchemaDeepSeek: + """All calling conventions for DeepSeek should return True for + ``supports_response_schema``.""" + + def test_provider_slash_model(self): + assert supports_response_schema(model="deepseek/deepseek-chat") is True + + def test_explicit_provider(self): + assert ( + supports_response_schema( + model="deepseek-chat", custom_llm_provider="deepseek" + ) + is True + ) + + def test_reasoner_provider_slash_model(self): + assert supports_response_schema(model="deepseek/deepseek-reasoner") is True + + def test_reasoner_explicit_provider(self): + assert ( + supports_response_schema( + model="deepseek-reasoner", custom_llm_provider="deepseek" + ) + is True + ) + + +# --------------------------------------------------------------------------- +# Fallback-logic test – bare model entry used when prefixed is incomplete +# --------------------------------------------------------------------------- + + +class TestBareModelFallback: + """When a provider-prefixed entry is missing a capability flag, the + ``_supports_factory`` fallback should consult the bare model-name + entry in ``litellm.model_cost``.""" + + def test_fallback_uses_bare_entry(self): + """Temporarily remove ``supports_response_schema`` from the prefixed + entry and verify the fallback still returns True.""" + key = "deepseek/deepseek-chat" + original = litellm.model_cost.get(key, {}).get("supports_response_schema") + try: + # Simulate the pre-fix state: field missing from prefixed entry + if key in litellm.model_cost: + litellm.model_cost[key].pop("supports_response_schema", None) + result = _supports_factory( + model="deepseek-chat", + custom_llm_provider="deepseek", + key="supports_response_schema", + ) + assert result is True + finally: + # Restore + if key in litellm.model_cost and original is not None: + litellm.model_cost[key]["supports_response_schema"] = original + + def test_no_fallback_when_explicitly_false(self): + """If the prefixed entry explicitly sets a capability to ``False``, + the fallback must NOT override it.""" + key = "deepseek/deepseek-reasoner" + # After the data fix, deepseek/deepseek-reasoner has + # supports_function_calling=false (matching the bare entry). + # Explicitly set it to False to test the guard. + original = litellm.model_cost.get(key, {}).get("supports_function_calling") + try: + if key in litellm.model_cost: + litellm.model_cost[key]["supports_function_calling"] = False + result = _supports_factory( + model="deepseek-reasoner", + custom_llm_provider="deepseek", + key="supports_function_calling", + ) + assert result is False + finally: + if key in litellm.model_cost and original is not None: + litellm.model_cost[key]["supports_function_calling"] = original