From 9675c13cb02b2e183deb04b0ae912ed85b20f1e2 Mon Sep 17 00:00:00 2001 From: m6atlas <349022726+m6atlas@users.noreply.github.com> Date: Wed, 28 Oct 2026 12:38:36 -0700 Subject: [PATCH 1/3] fix(openai)!: send the whole reasoning_effort dict when chat is bridged to Responses (#48807) * fix(openai): send the whole reasoning_effort dict when chat is bridged to Responses * fix(openai): drop only a refused effort and honor drop params on the bridged reasoning dict * fix(openai): drop generate_summary and forward a non-catalog model's reasoning dict whole on the bridge * fix(responses): drop a mode or context the model does not take under drop_params Read each OpenAI model's reasoning_modes and reasoning_contexts from the model map when drop_params filters the bridged reasoning object, send no reasoning object when validation dropped the effort, and ignore an additional_drop_params value that is not a list. * fix(responses): drop only a refused effort and forward an unlisted mode or context on the bridge --------- Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com> --- .../transformation.py | 19 +- litellm/main.py | 13 +- ...odel_prices_and_context_window_backup.json | 318 +++++++++++++++++ litellm/responses/utils.py | 101 +++++- litellm/types/utils.py | 2 + litellm/utils.py | 2 + model_prices_and_context_window.json | 318 +++++++++++++++++ tests/unit/test_main.py | 324 ++++++++++++++++++ tests/unit/test_utils.py | 2 + 9 files changed, 1083 insertions(+), 16 deletions(-) diff --git a/litellm/completion_extras/litellm_responses_transformation/transformation.py b/litellm/completion_extras/litellm_responses_transformation/transformation.py index 98322aecf6a..47a3a903d82 100644 --- a/litellm/completion_extras/litellm_responses_transformation/transformation.py +++ b/litellm/completion_extras/litellm_responses_transformation/transformation.py @@ -3,7 +3,6 @@ Handler for transforming /chat/completions api requests to litellm.responses req """ import json -import os from collections.abc import AsyncIterator, Callable, Iterable, Iterator, Mapping, Sequence from itertools import accumulate, chain from types import MappingProxyType @@ -43,8 +42,8 @@ from litellm.responses.sse_output_recovery import ( ) from litellm.responses.utils import ( ResponsesAPIRequestUtils, - auto_reasoning_summary, normalize_responses_api_stream_options, + opt_in_auto_reasoning_summary, ) from litellm.types.llms.openai import ( ChatCompletionAnnotation, @@ -1745,19 +1744,11 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge): **cast(Reasoning, reasoning_effort) # cast-ok: dict is forwarded verbatim to the provider ) - # Check if auto-summary is enabled via flag or environment variable - # Priority: litellm.reasoning_auto_summary flag > LITELLM_REASONING_AUTO_SUMMARY env var - auto_summary_enabled: Final = ( - litellm.reasoning_auto_summary or os.getenv("LITELLM_REASONING_AUTO_SUMMARY", "false").lower() == "true" - ) - + summary: Final = opt_in_auto_reasoning_summary(model, custom_llm_provider) return ( - Reasoning( - effort=reasoning_effort, - summary="detailed" if model is None else auto_reasoning_summary(model, custom_llm_provider), - ) - if auto_summary_enabled - else Reasoning(effort=reasoning_effort) + Reasoning(effort=reasoning_effort) + if summary is None + else Reasoning(effort=reasoning_effort, summary=summary) ) _map_reasoning_effort = map_reasoning_effort diff --git a/litellm/main.py b/litellm/main.py index 75bd4431120..41c4302499e 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -6124,7 +6124,18 @@ def completion( optional_params, rs_val = strip_reasoning_summary_aliases_from_optional_params(optional_params) - if isinstance(resolved_reasoning_effort, dict) and "summary" in resolved_reasoning_effort: + if custom_llm_provider == "openai" and isinstance(resolved_reasoning_effort, dict): + from litellm.responses.utils import openai_bridge_reasoning + + optional_params["reasoning_effort"] = openai_bridge_reasoning( + requested=resolved_reasoning_effort, + effort_kept="reasoning_effort" in optional_params, + summary_alias=rs_val, + model=model, + drop_params=drop_unsupported_reasoning, + additional_drop_params=routing_kwargs.get("additional_drop_params"), + ) + elif isinstance(resolved_reasoning_effort, dict) and "summary" in resolved_reasoning_effort: optional_params["reasoning_effort"] = resolved_reasoning_effort elif rs_val is not None: eff: Final = optional_params.get("reasoning_effort", resolved_reasoning_effort) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 7199a075612..825de5ac3ee 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -35598,6 +35598,13 @@ "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35675,6 +35682,13 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35851,6 +35865,13 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35902,6 +35923,13 @@ "mode": "chat", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36017,6 +36045,13 @@ "mode": "chat", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36066,6 +36101,13 @@ "mode": "responses", "output_cost_per_token": 0.000168, "output_cost_per_token_batches": 8.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36107,6 +36149,13 @@ "mode": "responses", "output_cost_per_token": 0.000168, "output_cost_per_token_batches": 8.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36194,6 +36243,15 @@ "output_cost_per_token_flex": 2.5e-05, "output_cost_per_token_priority": 0.0001, "output_cost_per_token_ultrafast": 0.0003, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36278,6 +36336,15 @@ "output_cost_per_token_above_272k_tokens_batches": 7.5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36370,6 +36437,15 @@ "output_cost_per_token_above_272k_tokens_batches": 3.75e-07, "output_cost_per_token_flex": 2.5e-07, "output_cost_per_token_priority": 1e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36453,6 +36529,15 @@ "output_cost_per_token_above_272k_tokens_batches": 1.5e-05, "output_cost_per_token_flex": 1e-05, "output_cost_per_token_priority": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36534,6 +36619,15 @@ "output_cost_per_token_above_272k_tokens_batches": 1.5e-05, "output_cost_per_token_flex": 1e-05, "output_cost_per_token_priority": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36617,6 +36711,15 @@ "output_cost_per_token_above_272k_tokens_batches": 9e-06, "output_cost_per_token_flex": 6e-06, "output_cost_per_token_priority": 2.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36699,6 +36802,15 @@ "output_cost_per_token_above_272k_tokens_batches": 9e-07, "output_cost_per_token_flex": 6e-07, "output_cost_per_token_priority": 2.4e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36936,6 +37048,14 @@ "reasoning_effort_levels": [ "medium" ], + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", @@ -36983,6 +37103,14 @@ "output_cost_per_token_batches": 1.5e-05, "output_cost_per_token_above_272k_tokens_batches": 2.25e-05, "output_cost_per_token_priority": 7.5e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37045,6 +37173,14 @@ "output_cost_per_token_batches": 1.5e-05, "output_cost_per_token_above_272k_tokens_batches": 2.25e-05, "output_cost_per_token_priority": 7.5e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37099,6 +37235,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37150,6 +37294,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37209,6 +37361,14 @@ "output_cost_per_token_batches": 7.5e-06, "output_cost_per_token_above_272k_tokens_batches": 1.125e-05, "output_cost_per_token_priority": 3e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "supported_endpoints": [ @@ -37266,6 +37426,14 @@ "output_cost_per_token_batches": 7.5e-06, "output_cost_per_token_above_272k_tokens_batches": 1.125e-05, "output_cost_per_token_priority": 3e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "supported_endpoints": [ @@ -37315,6 +37483,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37367,6 +37543,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37421,6 +37605,14 @@ "output_cost_per_token_flex": 2.25e-06, "output_cost_per_token_batches": 2.25e-06, "output_cost_per_token_priority": 9e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37475,6 +37667,14 @@ "output_cost_per_token_flex": 2.25e-06, "output_cost_per_token_batches": 2.25e-06, "output_cost_per_token_priority": 9e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37526,6 +37726,14 @@ "output_cost_per_token": 1.25e-06, "output_cost_per_token_flex": 6.25e-07, "output_cost_per_token_batches": 6.25e-07, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37578,6 +37786,14 @@ "output_cost_per_token": 1.25e-06, "output_cost_per_token_flex": 6.25e-07, "output_cost_per_token_batches": 6.25e-07, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37624,6 +37840,13 @@ "mode": "responses", "output_cost_per_token": 0.00012, "output_cost_per_token_batches": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37666,6 +37889,13 @@ "mode": "responses", "output_cost_per_token": 0.00012, "output_cost_per_token_batches": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37716,6 +37946,13 @@ "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37826,6 +38063,13 @@ "mode": "responses", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37908,6 +38152,13 @@ "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37959,6 +38210,13 @@ "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -38007,6 +38265,13 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -38055,6 +38320,13 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -43821,6 +44093,13 @@ "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -43929,6 +44208,13 @@ "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -44008,6 +44294,13 @@ "mode": "responses", "output_cost_per_token": 8e-05, "output_cost_per_token_batches": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -44045,6 +44338,13 @@ "mode": "responses", "output_cost_per_token": 8e-05, "output_cost_per_token_batches": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -73288,6 +73588,15 @@ "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, "output_cost_per_token_ultrafast": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -76787,6 +77096,15 @@ "output_cost_per_token_batches": 2.5e-06, "output_cost_per_token_flex": 2.5e-06, "output_cost_per_token_priority": 1e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": {"search_context_size_high":0.01,"search_context_size_low":0.01,"search_context_size_medium":0.01}, diff --git a/litellm/responses/utils.py b/litellm/responses/utils.py index 02d07932533..cf993c69814 100644 --- a/litellm/responses/utils.py +++ b/litellm/responses/utils.py @@ -1,8 +1,10 @@ import base64 +import os import re from collections.abc import Callable, Collection, Iterable, Mapping, Sequence from functools import reduce -from typing import Any, Final, Optional, TypeVar, Union, cast, get_type_hints, overload +from types import MappingProxyType +from typing import Any, Final, Optional, TypeVar, Union, cast, get_args, get_type_hints, overload from pydantic import BaseModel, TypeAdapter from typing_extensions import TypedDict as ExtraItemsTypedDict @@ -14,6 +16,7 @@ from litellm.litellm_core_utils.core_helpers import as_str_mapping, attribute_or from litellm.litellm_core_utils.dot_notation_indexing import delete_nested_value, is_nested_path from litellm.litellm_core_utils.get_provider_specific_headers import ProviderSpecificHeaderUtils from litellm.llms.base_llm.responses.transformation import BaseResponsesAPIConfig, ReasoningSummaryLevel +from litellm.llms.openai.chat.gpt_transformation import OpenAIGPTConfig from litellm.types.llms.openai import ( AllMessageValues, InputTokensDetails, @@ -29,6 +32,7 @@ from litellm.types.responses.main import DecodedResponseId from litellm.types.utils import ( CacheCreationTokenDetails, CompletionTokensDetailsWrapper, + ModelInfo, PromptTokensDetailsWrapper, ProviderSpecificHeader, SpecialEnums, @@ -1382,3 +1386,98 @@ def auto_reasoning_summary(model: str, custom_llm_provider: object) -> Reasoning provider, model.removeprefix(f"{provider}/") ) return "detailed" if responses_config is None else responses_config.auto_reasoning_summary() + + +def opt_in_auto_reasoning_summary(model: str | None, custom_llm_provider: object) -> ReasoningSummaryLevel | None: + if not (litellm.reasoning_auto_summary or os.getenv("LITELLM_REASONING_AUTO_SUMMARY", "false").lower() == "true"): + return None + return "detailed" if model is None else auto_reasoning_summary(model, custom_llm_provider) + + +def reasoning_with_summary_default( + reasoning: Mapping[str, object], summary_alias: object, model: str, custom_llm_provider: object +) -> Mapping[str, object]: + if "summary" in reasoning: + return {**reasoning} + summary: Final = ( + summary_alias if summary_alias is not None else opt_in_auto_reasoning_summary(model, custom_llm_provider) + ) + return {**reasoning} if summary is None else {**reasoning, "summary": summary} + + +_SUMMARY_VALUES: Final = frozenset((*get_args(ReasoningSummaryLevel), None)) + + +def _openai_model_info(model: str) -> ModelInfo | None: + from litellm.utils import ModelInfoLookupMiss, lookup_model_info_for_registration + + model_info: Final = lookup_model_info_for_registration(model.removeprefix("responses/"), "openai") + return None if isinstance(model_info, ModelInfoLookupMiss) else model_info + + +def _declared_values(values: Sequence[str] | None) -> frozenset[str | None] | None: + return None if values is None else frozenset(values) + + +def _openai_reasoning_field_values(model: str) -> Mapping[str, frozenset[str | None] | None]: + model_info: Final = _openai_model_info(model) + if model_info is None: + return MappingProxyType({"summary": _SUMMARY_VALUES, "mode": None, "context": None}) + if not model_info.get("supports_reasoning"): + return MappingProxyType({}) + return MappingProxyType( + { + "summary": _SUMMARY_VALUES, + "mode": _declared_values(model_info.get("reasoning_modes")), + "context": _declared_values(model_info.get("reasoning_contexts")), + } + ) + + +def _openai_accepts_reasoning_field( + key: str, value: object, values_by_field: Mapping[str, frozenset[str | None] | None] +) -> bool: + if key == "effort": + return True + if key not in values_by_field: + return False + allowed: Final = values_by_field[key] + return allowed is None or (isinstance(value, str | None) and value in allowed) + + +def _openai_accepted_reasoning(reasoning: Mapping[str, object], model: str) -> Mapping[str, object]: + values_by_field: Final = _openai_reasoning_field_values(model) + return { + key: value for key, value in reasoning.items() if _openai_accepts_reasoning_field(key, value, values_by_field) + } + + +def _additional_drop_paths(additional_drop_params: object) -> tuple[str, ...]: + if additional_drop_params.__class__ is not list: + return () + return tuple(path for path in iter_items(additional_drop_params) if isinstance(path, str)) + + +def openai_bridge_reasoning( + *, + requested: Mapping[str, object], + effort_kept: bool, + summary_alias: object, + model: str, + drop_params: bool, + additional_drop_params: object, +) -> Mapping[str, object] | None: + drop_paths: Final = _additional_drop_paths(additional_drop_params) + if "reasoning_effort" in drop_paths: + return None + kept: Final = requested if effort_kept else {key: value for key, value in requested.items() if key != "effort"} + with_summary: Final = reasoning_with_summary_default(kept, summary_alias, model, "openai") + accepted: Final = ( + _openai_accepted_reasoning(with_summary, model) + if drop_params and OpenAIGPTConfig.is_openai_catalog_model(model) + else {**with_summary} + ) + remaining: Final = as_str_mapping( + _apply_nested_drop_params({"reasoning_effort": accepted}, list(drop_paths)).get("reasoning_effort") + ) + return remaining or None diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 055c18f2e36..4ce35bf86c1 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -225,6 +225,8 @@ class ProviderSpecificModelInfo(TypedDict, total=False): supports_xhigh_reasoning_effort: bool | None supports_max_reasoning_effort: bool | None reasoning_effort_levels: ReadOnly[Sequence[str] | None] + reasoning_modes: ReadOnly[Sequence[str] | None] + reasoning_contexts: ReadOnly[Sequence[str] | None] default_reasoning_effort: ReadOnly[Literal["none", "minimal", "low", "medium", "high", "xhigh"] | None] supports_output_config: bool | None supports_image_size: bool | None diff --git a/litellm/utils.py b/litellm/utils.py index 61b0e17e197..a231644addf 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -7000,6 +7000,8 @@ def get_model_info_helper( supports_xhigh_reasoning_effort=_model_info.get("supports_xhigh_reasoning_effort", None), supports_max_reasoning_effort=_model_info.get("supports_max_reasoning_effort", None), reasoning_effort_levels=_model_info.get("reasoning_effort_levels", None), + reasoning_modes=_model_info.get("reasoning_modes", None), + reasoning_contexts=_model_info.get("reasoning_contexts", None), default_reasoning_effort=_model_info.get("default_reasoning_effort", None), bedrock_output_config_effort_ceiling=_model_info.get("bedrock_output_config_effort_ceiling", None), bedrock_converse_supports_strict_tools=_model_info.get("bedrock_converse_supports_strict_tools", None), diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 7199a075612..825de5ac3ee 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -35598,6 +35598,13 @@ "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35675,6 +35682,13 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35851,6 +35865,13 @@ "mode": "chat", "output_cost_per_token": 1e-05, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -35902,6 +35923,13 @@ "mode": "chat", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36017,6 +36045,13 @@ "mode": "chat", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36066,6 +36101,13 @@ "mode": "responses", "output_cost_per_token": 0.000168, "output_cost_per_token_batches": 8.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36107,6 +36149,13 @@ "mode": "responses", "output_cost_per_token": 0.000168, "output_cost_per_token_batches": 8.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -36194,6 +36243,15 @@ "output_cost_per_token_flex": 2.5e-05, "output_cost_per_token_priority": 0.0001, "output_cost_per_token_ultrafast": 0.0003, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36278,6 +36336,15 @@ "output_cost_per_token_above_272k_tokens_batches": 7.5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36370,6 +36437,15 @@ "output_cost_per_token_above_272k_tokens_batches": 3.75e-07, "output_cost_per_token_flex": 2.5e-07, "output_cost_per_token_priority": 1e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36453,6 +36529,15 @@ "output_cost_per_token_above_272k_tokens_batches": 1.5e-05, "output_cost_per_token_flex": 1e-05, "output_cost_per_token_priority": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36534,6 +36619,15 @@ "output_cost_per_token_above_272k_tokens_batches": 1.5e-05, "output_cost_per_token_flex": 1e-05, "output_cost_per_token_priority": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36617,6 +36711,15 @@ "output_cost_per_token_above_272k_tokens_batches": 9e-06, "output_cost_per_token_flex": 6e-06, "output_cost_per_token_priority": 2.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36699,6 +36802,15 @@ "output_cost_per_token_above_272k_tokens_batches": 9e-07, "output_cost_per_token_flex": 6e-07, "output_cost_per_token_priority": 2.4e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -36936,6 +37048,14 @@ "reasoning_effort_levels": [ "medium" ], + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "source": "https://developers.openai.com/api/docs/pricing", "supported_endpoints": [ "/v1/chat/completions", @@ -36983,6 +37103,14 @@ "output_cost_per_token_batches": 1.5e-05, "output_cost_per_token_above_272k_tokens_batches": 2.25e-05, "output_cost_per_token_priority": 7.5e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37045,6 +37173,14 @@ "output_cost_per_token_batches": 1.5e-05, "output_cost_per_token_above_272k_tokens_batches": 2.25e-05, "output_cost_per_token_priority": 7.5e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37099,6 +37235,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37150,6 +37294,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37209,6 +37361,14 @@ "output_cost_per_token_batches": 7.5e-06, "output_cost_per_token_above_272k_tokens_batches": 1.125e-05, "output_cost_per_token_priority": 3e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "supported_endpoints": [ @@ -37266,6 +37426,14 @@ "output_cost_per_token_batches": 7.5e-06, "output_cost_per_token_above_272k_tokens_batches": 1.125e-05, "output_cost_per_token_priority": 3e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "supported_endpoints": [ @@ -37315,6 +37483,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37367,6 +37543,14 @@ "output_cost_per_token_flex": 9e-05, "output_cost_per_token_batches": 9e-05, "output_cost_per_token_above_272k_tokens_batches": 0.000135, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37421,6 +37605,14 @@ "output_cost_per_token_flex": 2.25e-06, "output_cost_per_token_batches": 2.25e-06, "output_cost_per_token_priority": 9e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37475,6 +37667,14 @@ "output_cost_per_token_flex": 2.25e-06, "output_cost_per_token_batches": 2.25e-06, "output_cost_per_token_priority": 9e-06, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37526,6 +37726,14 @@ "output_cost_per_token": 1.25e-06, "output_cost_per_token_flex": 6.25e-07, "output_cost_per_token_batches": 6.25e-07, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37578,6 +37786,14 @@ "output_cost_per_token": 1.25e-06, "output_cost_per_token_flex": 6.25e-07, "output_cost_per_token_batches": 6.25e-07, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -37624,6 +37840,13 @@ "mode": "responses", "output_cost_per_token": 0.00012, "output_cost_per_token_batches": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37666,6 +37889,13 @@ "mode": "responses", "output_cost_per_token": 0.00012, "output_cost_per_token_batches": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37716,6 +37946,13 @@ "output_cost_per_token_batches": 5e-06, "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37826,6 +38063,13 @@ "mode": "responses", "output_cost_per_token": 1.4e-05, "output_cost_per_token_priority": 2.8e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37908,6 +38152,13 @@ "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -37959,6 +38210,13 @@ "output_cost_per_token_batches": 1e-06, "output_cost_per_token_flex": 1e-06, "output_cost_per_token_priority": 3.6e-06, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -38007,6 +38265,13 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -38055,6 +38320,13 @@ "output_cost_per_token": 4e-07, "output_cost_per_token_batches": 2e-07, "output_cost_per_token_flex": 2e-07, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -43821,6 +44093,13 @@ "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -43929,6 +44208,13 @@ "output_cost_per_token_batches": 4e-06, "output_cost_per_token_flex": 4e-06, "output_cost_per_token_priority": 1.4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "standard" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -44008,6 +44294,13 @@ "mode": "responses", "output_cost_per_token": 8e-05, "output_cost_per_token_batches": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -44045,6 +44338,13 @@ "mode": "responses", "output_cost_per_token": 8e-05, "output_cost_per_token_batches": 4e-05, + "reasoning_contexts": [ + "auto", + "current_turn" + ], + "reasoning_modes": [ + "pro" + ], "search_context_cost_per_query": { "search_context_size_high": 0.01, "search_context_size_low": 0.01, @@ -73288,6 +73588,15 @@ "output_cost_per_token_flex": 5e-06, "output_cost_per_token_priority": 2e-05, "output_cost_per_token_ultrafast": 6e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": { @@ -76787,6 +77096,15 @@ "output_cost_per_token_batches": 2.5e-06, "output_cost_per_token_flex": 2.5e-06, "output_cost_per_token_priority": 1e-05, + "reasoning_contexts": [ + "auto", + "current_turn", + "all_turns" + ], + "reasoning_modes": [ + "standard", + "pro" + ], "regional_processing_uplift_multiplier_eu": 1.1, "regional_processing_uplift_multiplier_us": 1.1, "search_context_cost_per_query": {"search_context_size_high":0.01,"search_context_size_low":0.01,"search_context_size_medium":0.01}, diff --git a/tests/unit/test_main.py b/tests/unit/test_main.py index ba39fc3f5ac..f38bbcf4404 100644 --- a/tests/unit/test_main.py +++ b/tests/unit/test_main.py @@ -2127,6 +2127,330 @@ def test_responses_bridge_preserves_reasoning_effort_with_drop_params( assert request_body["reasoning"] == {"effort": "high"} +@pytest.fixture +def openai_responses_route( + restore_model_registry, respx_mock: respx.MockRouter, monkeypatch: pytest.MonkeyPatch +) -> respx.Route: + monkeypatch.setattr(litellm, "disable_aiohttp_transport", True) + monkeypatch.delenv("LITELLM_REASONING_AUTO_SUMMARY", raising=False) + litellm.register_model( + { + "gpt-6.1-terra": { + "litellm_provider": "openai", + "mode": "chat", + "supports_reasoning": True, + "supports_minimal_reasoning_effort": False, + "reasoning_modes": ["standard", "pro"], + "reasoning_contexts": ["auto", "current_turn", "all_turns"], + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + } + }, + persist_across_reloads=False, + ) + return respx_mock.post("https://api.openai.com/v1/responses").respond( + json={ + "id": "resp_test", + "object": "response", + "created_at": 1734366691, + "status": "completed", + "model": "gpt-6.1-terra", + "output": [ + { + "type": "message", + "id": "msg_1", + "status": "completed", + "role": "assistant", + "content": [{"type": "output_text", "text": "Done.", "annotations": []}], + } + ], + "usage": {"input_tokens": 1, "output_tokens": 1, "total_tokens": 2}, + } + ) + + +@pytest.mark.parametrize( + ("reasoning_effort", "summary_alias", "expected_reasoning"), + [ + ( + {"effort": "low", "mode": "pro", "context": "current_turn"}, + {}, + {"effort": "low", "mode": "pro", "context": "current_turn"}, + ), + ( + {"effort": "low", "mode": "pro", "context": "current_turn"}, + {"reasoning_summary": "auto"}, + {"effort": "low", "mode": "pro", "context": "current_turn", "summary": "auto"}, + ), + ( + {"effort": "low", "mode": "pro", "summary": "concise"}, + {"reasoningSummary": "auto"}, + {"effort": "low", "mode": "pro", "summary": "concise"}, + ), + ( + {"effort": "low", "context": "current_turn", "summary": None}, + {"reasoning_summary": "auto"}, + {"effort": "low", "context": "current_turn", "summary": None}, + ), + ({"mode": "pro", "context": "all_turns"}, {}, {"mode": "pro", "context": "all_turns"}), + ("low", {}, {"effort": "low"}), + ], +) +def test_openai_responses_bridge_sends_the_whole_reasoning_effort_dict( + reasoning_effort: str | dict[str, str | None], + summary_alias: dict[str, str], + expected_reasoning: dict[str, str | None], + openai_responses_route: respx.Route, +): + caller_reasoning_effort: Final = copy.deepcopy(reasoning_effort) + + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort=reasoning_effort, + api_key="fake-key", + **summary_alias, + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == expected_reasoning + assert reasoning_effort == caller_reasoning_effort + + +@pytest.mark.parametrize( + "reasoning_effort, summary_alias, expected_reasoning", + [ + ( + {"effort": "minimal", "mode": "pro", "context": "current_turn"}, + {}, + {"mode": "pro", "context": "current_turn"}, + ), + ({"effort": "minimal", "mode": "pro", "summary": "auto"}, {}, {"mode": "pro", "summary": "auto"}), + ({"effort": "minimal", "mode": "pro"}, {"reasoning_summary": "auto"}, {"mode": "pro", "summary": "auto"}), + ], +) +def test_openai_responses_bridge_drops_only_a_refused_effort_under_drop_params( + reasoning_effort: dict[str, str], + summary_alias: dict[str, str], + expected_reasoning: dict[str, str], + openai_responses_route: respx.Route, +): + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort=reasoning_effort, + drop_params=True, + api_key="fake-key", + **summary_alias, + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == expected_reasoning + + +def test_openai_responses_bridge_drops_a_reasoning_effort_dict_for_a_model_without_reasoning( + openai_responses_route: respx.Route, +): + litellm.completion( + model="openai/responses/gpt-4.1", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "current_turn"}, + drop_params=True, + api_key="fake-key", + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert "reasoning" not in request_body + + +@pytest.mark.parametrize( + "drop_params, expected_reasoning", + [ + (True, {"effort": "low", "context": "current_turn"}), + ( + False, + { + "effort": "low", + "exclude": False, + "mode": "turbo", + "context": "current_turn", + "generate_summary": "concise", + }, + ), + ], +) +def test_openai_responses_bridge_drops_reasoning_fields_openai_refuses_only_under_drop_params( + drop_params: bool, + expected_reasoning: dict[str, object], + openai_responses_route: respx.Route, +): + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={ + "effort": "low", + "exclude": False, + "mode": "turbo", + "context": "current_turn", + "generate_summary": "concise", + }, + drop_params=drop_params, + api_key="fake-key", + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == expected_reasoning + + +@pytest.mark.parametrize( + "declared, drop_params, expected_reasoning", + [ + ( + {"reasoning_modes": ["standard"], "reasoning_contexts": ["auto", "current_turn"]}, + True, + {"effort": "low"}, + ), + ( + {"reasoning_modes": ["pro"], "reasoning_contexts": ["all_turns"]}, + True, + {"effort": "low", "mode": "pro", "context": "all_turns"}, + ), + ({}, True, {"effort": "low", "mode": "pro", "context": "all_turns"}), + ({"reasoning_modes": ["standard"]}, True, {"effort": "low", "context": "all_turns"}), + ( + {"reasoning_modes": ["standard"], "reasoning_contexts": ["auto", "current_turn"]}, + False, + {"effort": "low", "mode": "pro", "context": "all_turns"}, + ), + ], +) +def test_openai_responses_bridge_drops_a_mode_or_context_the_model_does_not_list_only_under_drop_params( + declared: dict[str, list[str]], + drop_params: bool, + expected_reasoning: dict[str, str], + openai_responses_route: respx.Route, +): + litellm.register_model( + { + "gpt-6.1-reasoning-test": { + "litellm_provider": "openai", + "mode": "chat", + "supports_reasoning": True, + "input_cost_per_token": 0.0, + "output_cost_per_token": 0.0, + **declared, + } + }, + persist_across_reloads=False, + ) + + litellm.completion( + model="openai/responses/gpt-6.1-reasoning-test", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "all_turns"}, + drop_params=drop_params, + api_key="fake-key", + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == expected_reasoning + + +def test_openai_responses_bridge_forwards_mode_and_context_under_drop_params_for_a_catalog_model_with_no_cost_entry( + openai_responses_route: respx.Route, +): + litellm.open_ai_chat_completion_models.add("unpriced-reasoning-model") + + litellm.completion( + model="openai/responses/unpriced-reasoning-model", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "all_turns", "generate_summary": "auto"}, + drop_params=True, + api_key="fake-key", + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == {"effort": "low", "mode": "pro", "context": "all_turns"} + + +def test_openai_responses_bridge_forwards_a_reasoning_effort_dict_whole_for_a_model_outside_the_openai_catalog( + openai_responses_route: respx.Route, +): + reasoning_effort: Final = {"effort": "high", "exclude": True, "budget_tokens": 2000} + + litellm.completion( + model="openai/responses/my-proxy-alias", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort=reasoning_effort, + drop_params=True, + api_key="fake-key", + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body["reasoning"] == reasoning_effort + + +@pytest.mark.parametrize( + "additional_drop_params, summary_alias, expected_reasoning", + [ + (["reasoning_effort.mode"], {}, {"effort": "low", "context": "current_turn"}), + ( + ["reasoning_effort.summary"], + {"reasoning_summary": "auto"}, + {"effort": "low", "mode": "pro", "context": "current_turn"}, + ), + (["reasoning_effort"], {"reasoning_summary": "auto"}, None), + ("reasoning_effort", {}, {"effort": "low", "mode": "pro", "context": "current_turn"}), + ], +) +def test_openai_responses_bridge_applies_additional_drop_params_to_a_reasoning_effort_dict( + additional_drop_params: list[str] | str, + summary_alias: dict[str, str], + expected_reasoning: dict[str, str] | None, + openai_responses_route: respx.Route, +): + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "current_turn"}, + additional_drop_params=additional_drop_params, + api_key="fake-key", + **summary_alias, + ) + + request_body: Final = json.loads(openai_responses_route.calls[0].request.content) + assert request_body.get("reasoning") == expected_reasoning + + +def test_openai_responses_bridge_gives_a_reasoning_effort_dict_the_auto_summary_a_string_gets( + openai_responses_route: respx.Route, monkeypatch: pytest.MonkeyPatch +): + monkeypatch.setenv("LITELLM_REASONING_AUTO_SUMMARY", "true") + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort="low", + api_key="fake-key", + ) + litellm.completion( + model="openai/responses/gpt-6.1-terra", + messages=[{"role": "user", "content": "hello"}], + reasoning_effort={"effort": "low", "mode": "pro", "context": "current_turn"}, + api_key="fake-key", + ) + + string_reasoning, dict_reasoning = ( + json.loads(call.request.content)["reasoning"] for call in openai_responses_route.calls + ) + assert string_reasoning["summary"] is not None + assert dict_reasoning == { + "effort": "low", + "mode": "pro", + "context": "current_turn", + "summary": string_reasoning["summary"], + } + + _BRIDGED_RESPONSE_ANSWER: Final = json.dumps( { "id": "resp_slow", diff --git a/tests/unit/test_utils.py b/tests/unit/test_utils.py index 38287b25169..146e2df6413 100644 --- a/tests/unit/test_utils.py +++ b/tests/unit/test_utils.py @@ -1195,6 +1195,8 @@ _COST_MAP_ROW_SCHEMA: Final = { "type": "array", "items": {"type": "string", "enum": ["none", "minimal", "low", "medium", "high", "xhigh", "max"]}, }, + "reasoning_modes": {"type": "array", "items": {"type": "string"}}, + "reasoning_contexts": {"type": "array", "items": {"type": "string"}}, "default_reasoning_effort": { "type": "string", "enum": ["none", "minimal", "low", "medium", "high", "xhigh"], From ca6c83022193c563b42c347d7e54ed8f66edc009 Mon Sep 17 00:00:00 2001 From: tin-berri Date: Wed, 28 Oct 2026 15:39:41 -0400 Subject: [PATCH 2/3] fix(proxy): account for media in combined TPM reservations (#48865) --- .../litellm_core_utils/audio_utils/utils.py | 43 +++ .../prompt_templates/common_utils.py | 155 +++++++++ litellm/litellm_core_utils/token_counter.py | 232 +++++++++++++- .../hooks/parallel_request_limiter_v3.py | 91 +++++- litellm/router.py | 38 ++- ...ore_utils_prompt_templates_common_utils.py | 116 +++++++ .../litellm_core_utils/test_audio_utils.py | 64 ++++ .../litellm_core_utils/test_token_counter.py | 131 ++++++++ tests/unit/proxy/hooks/test_tpm_concurrent.py | 299 +++++++++++++++++- tests/unit/test_router/test_router.py | 31 ++ 10 files changed, 1169 insertions(+), 31 deletions(-) diff --git a/litellm/litellm_core_utils/audio_utils/utils.py b/litellm/litellm_core_utils/audio_utils/utils.py index dab3e48f91a..640559ac7e2 100644 --- a/litellm/litellm_core_utils/audio_utils/utils.py +++ b/litellm/litellm_core_utils/audio_utils/utils.py @@ -2,8 +2,12 @@ Utils used for litellm.transcription() and litellm.atranscription() """ +import base64 +import binascii import hashlib +import math import os +from collections.abc import Sequence from dataclasses import dataclass from typing import Final @@ -331,6 +335,45 @@ def calculate_request_duration(file: FileTypes) -> float | None: return None +DEFAULT_AUDIO_TOKEN_ESTIMATE: Final = 300 +AUDIO_TOKENS_PER_SECOND: Final = 10 +AUDIO_BYTES_PER_TOKEN: Final = 1600 +_MAX_AUDIO_ESTIMATE_BYTES: Final = 16 * 1024 * 1024 + + +def _audio_token_rate(model: str) -> int | None: + if model.startswith( + ("gemini-", "gemini/gemini-", "vertex_ai/gemini-", "vertex_ai_beta/gemini-", "openrouter/google/gemini-") + ): + family: Final = model.rsplit("/", 1)[-1] + return 25 if family.startswith(("gemini-3-", "gemini-3.")) else 32 + if model.startswith(("gpt-", "openai/gpt-", "azure/gpt-", "openrouter/openai/gpt-")): + return AUDIO_TOKENS_PER_SECOND + return None + + +def estimate_audio_tokens(data: str | None, models: Sequence[str] = (), fallback_model: str | None = None) -> int: + """Estimate audio tokens from container duration within a bounded decoding budget.""" + if not data: + return DEFAULT_AUDIO_TOKEN_ESTIMATE + fallback_estimate: Final = max(len(data) * 3 // 4 // AUDIO_BYTES_PER_TOKEN, DEFAULT_AUDIO_TOKEN_ESTIMATE) + padding: Final = 2 if data.endswith("==") else 1 if data.endswith("=") else 0 + if len(data) * 3 // 4 - padding > _MAX_AUDIO_ESTIMATE_BYTES: + return fallback_estimate + try: + duration: Final = calculate_request_duration(base64.b64decode(data, validate=True)) + except (ValueError, binascii.Error): + return fallback_estimate + if duration is None or not math.isfinite(duration) or duration <= 0: + return fallback_estimate + fallback_rate: Final = _audio_token_rate(fallback_model) if fallback_model is not None else None + rate: Final = max( + (_audio_token_rate(model) or fallback_rate or AUDIO_TOKENS_PER_SECOND for model in models), + default=fallback_rate or AUDIO_TOKENS_PER_SECOND, + ) + return max(1, math.ceil(duration * rate)) + + DEFAULT_SPEECH_MEDIA_TYPE: Final = "audio/mpeg" diff --git a/litellm/litellm_core_utils/prompt_templates/common_utils.py b/litellm/litellm_core_utils/prompt_templates/common_utils.py index 3c06c522dd6..cf6bdc69556 100644 --- a/litellm/litellm_core_utils/prompt_templates/common_utils.py +++ b/litellm/litellm_core_utils/prompt_templates/common_utils.py @@ -9,6 +9,7 @@ import json import mimetypes import re from collections.abc import Callable, Iterable, Iterator, Mapping, Sequence +from dataclasses import dataclass from graphlib import CycleError, TopologicalSorter from itertools import chain, groupby, islice from os import PathLike @@ -236,6 +237,160 @@ _IMAGE_CONTENT_PART_TYPES: Final = frozenset({"image_url", "input_image", "image _IMAGE_SCAN_MAX_DEPTH: Final = 4 +@dataclass(frozen=True, slots=True) +class _TPMMediaProjection: + content: object + media: tuple[Mapping[str, object], ...] = () + + +def _tpm_image_part(url: str, detail: object = "auto") -> Mapping[str, object]: + return {"type": "image_url", "image_url": {"url": url, "detail": detail}} + + +def _tpm_blank_fields(fields: Mapping[str, object], names: tuple[str, ...]) -> Mapping[str, object]: + return {key: "" if key in names and isinstance(value, str) else value for key, value in fields.items()} + + +def _tpm_file_media(part: Mapping[str, object]) -> _TPMMediaProjection: + from litellm.litellm_core_utils.prompt_templates.factory import parse_mime_type + + nested: Final = part.get("type") == "file" + fields: Final = as_str_mapping(part.get("file")) if nested else part + if fields is None: + return _TPMMediaProjection(part) + raw_data: Final = fields.get("file_data") + data: Final = raw_data if isinstance(raw_data, str) else "" + declared: Final = next( + ( + value + for key in ("format", "mime_type", "content_type") + if isinstance(value := fields.get(key), str) and value + ), + "", + ) + mime: Final = (declared or parse_mime_type(data[:256]) or "").split(";", 1)[0].strip().lower() + if not mime.startswith(("image/", "audio/")): + return _TPMMediaProjection(part) + raw_url: Final = fields.get("file_url") or fields.get("file_id") + url: Final = data or (raw_url if isinstance(raw_url, str) else "") + header_end: Final = data.find(",", 0, 256) + encoded: Final = ( + data[header_end + 1 :] + if header_end >= 0 and data[:header_end].endswith(";base64") + else "" + if data.startswith(("data:", "https://", "http://", "gs://")) + else data + ) + media: Final = ( + _tpm_image_part( + url + if not data or url.startswith(("data:", "https://", "http://", "gs://")) + else f"data:{mime};base64,{url}", + fields.get("detail", "auto"), + ) + if mime.startswith("image/") + else {"type": "input_audio", "input_audio": {"data": encoded, "format": mime.removeprefix("audio/")}} + ) + projected: Final = _tpm_blank_fields(fields, ("file_data", "file_url")) + return _TPMMediaProjection({**part, "file": projected} if nested else projected, (media,)) + + +def _tpm_media_leaf(value: object) -> _TPMMediaProjection: + part: Final = as_str_mapping(value) + if part is None: + return _TPMMediaProjection(value) + kind: Final = part.get("type") + if kind in ("file", "input_file"): + return _tpm_file_media(part) + if kind == "input_audio": + audio: Final = as_str_mapping(part.get("input_audio")) + return _TPMMediaProjection( + {**part, "input_audio": _tpm_blank_fields(audio, ("data",))} if audio is not None else value, + ({"type": "input_audio", "input_audio": part.get("input_audio")},), + ) + if kind in ("image_url", "input_image", "image", "computer_screenshot"): + image: Final = as_str_mapping(part.get("image_url")) + source: Final = as_str_mapping(part.get("source")) + canonical: Final = _tpm_image_part( + _image_part_url({**part, "type": "input_image"} if kind == "computer_screenshot" else part) or "", + (image or part).get("detail", part.get("detail", "auto")), + ) + projected: Final = ( + {**part, "image_url": _tpm_blank_fields(image, ("url",))} + if image is not None and kind in ("image_url", "input_image", "computer_screenshot") + else {**part, "source": _tpm_blank_fields(source, ("data", "url"))} + if source is not None and kind == "image" and source.get("type") in ("base64", "url") + else _tpm_blank_fields(part, ("image_url", "url")) + ) + return _TPMMediaProjection(projected, (canonical,)) + return _TPMMediaProjection(value) + + +def _tpm_media_list(value: object, split: Callable[[object], _TPMMediaProjection]) -> _TPMMediaProjection: + if not isinstance(value, (list, tuple)): + return _TPMMediaProjection(value) + parts: Final = cast(Sequence[object], value) # cast-ok: finite JSON content arrays + projected: Final = tuple(map(split, parts)) + media: Final = tuple(chain.from_iterable(part.media for part in projected)) + content: Final = tuple(part.content for part in projected) + return _TPMMediaProjection(parts if not media else content if isinstance(value, tuple) else list(content), media) + + +def _tpm_replace_field(part: Mapping[str, object], key: str, projection: _TPMMediaProjection) -> _TPMMediaProjection: + return _TPMMediaProjection({**part, key: projection.content} if projection.media else part, projection.media) + + +def _tpm_document_media(value: object) -> _TPMMediaProjection: + part: Final = as_str_mapping(value) + source: Final = as_str_mapping(part.get("source")) if part is not None else None + if part is None or part.get("type") != "document" or source is None or source.get("type") != "content": + return _tpm_media_leaf(value) + content: Final = _tpm_media_list(source.get("content"), _tpm_media_leaf) + return _tpm_replace_field(part, "source", _tpm_replace_field(source, "content", content)) + + +def _tpm_tool_result_media(value: object) -> _TPMMediaProjection: + part: Final = as_str_mapping(value) + if part is None or part.get("type") != "tool_result": + return _tpm_document_media(value) + return _tpm_replace_field(part, "content", _tpm_media_list(part.get("content"), _tpm_document_media)) + + +def _tpm_code_output_media(value: object) -> _TPMMediaProjection: + part: Final = as_str_mapping(value) + url: Final = part.get("url") if part is not None else None + if part is None or part.get("type") != "image" or not isinstance(url, str) or not url: + return _TPMMediaProjection(value) + return _TPMMediaProjection(_tpm_blank_fields(part, ("url",)), (_tpm_image_part(url),)) + + +def _tpm_input_media(value: object) -> _TPMMediaProjection: + part: Final = as_str_mapping(value) + if part is None: + return _TPMMediaProjection(value) + kind: Final = part.get("type") + if kind in ("function_call_output", "custom_tool_call_output"): + return _tpm_replace_field(part, "output", _tpm_media_list(part.get("output"), _tpm_tool_result_media)) + if kind == "computer_call_output": + return _tpm_replace_field(part, "output", _tpm_media_leaf(part.get("output"))) + if kind == "code_interpreter_call": + return _tpm_replace_field(part, "outputs", _tpm_media_list(part.get("outputs"), _tpm_code_output_media)) + result: Final = part.get("result") + if kind == "image_generation_call" and isinstance(result, str) and result: + return _TPMMediaProjection( + _tpm_blank_fields(part, ("result",)), + (_tpm_image_part("data:application/octet-stream;base64," + result),), + ) + if kind == "message" or (kind is None and isinstance(part.get("role"), str)): + return _tpm_replace_field(part, "content", _tpm_media_list(part.get("content"), _tpm_tool_result_media)) + return _tpm_tool_result_media(value) + + +def split_media_for_tpm(value: object) -> tuple[object, tuple[Mapping[str, object], ...]]: + projection: Final = _tpm_media_list(value, _tpm_input_media) + return projection.content, projection.media + + def _content_parts_contain_image(parts: Sequence[object]) -> bool: """Depth-bounded frontier walk over nested content lists, iterative because the repo bans recursion; an Anthropic tool_result nests its image parts exactly one level down.""" diff --git a/litellm/litellm_core_utils/token_counter.py b/litellm/litellm_core_utils/token_counter.py index 6d50e4b065e..f42e4eae08f 100644 --- a/litellm/litellm_core_utils/token_counter.py +++ b/litellm/litellm_core_utils/token_counter.py @@ -3,9 +3,12 @@ import base64 import io import math +import re import struct +from bisect import bisect_right from collections.abc import Awaitable, Callable, Iterable, Mapping, Sequence from itertools import accumulate +from types import MappingProxyType from typing import TYPE_CHECKING, Final, Literal, cast import anyio @@ -35,6 +38,7 @@ from litellm.constants import ( TOKEN_COUNTER_MAX_EXACT_CHARS, ) from litellm.litellm_core_utils.asyncify import asyncify +from litellm.litellm_core_utils.core_helpers import as_str_mapping from litellm.litellm_core_utils.prompt_templates.chat_template_tools import ( ChatTemplateToolSection, chat_template_tool_section, @@ -188,22 +192,23 @@ def resize_image_high_res( # Test the function with the given example def calculate_tiles_needed( - resized_width, - resized_height, - tile_width=MAX_TILE_WIDTH, - tile_height=MAX_TILE_HEIGHT, -): + resized_width: int, + resized_height: int, + tile_width: int = MAX_TILE_WIDTH, + tile_height: int = MAX_TILE_HEIGHT, +) -> int: tiles_across: Final = (resized_width + tile_width - 1) // tile_width tiles_down: Final = (resized_height + tile_height - 1) // tile_height total_tiles: Final = tiles_across * tiles_down return total_tiles -def high_detail_image_token_upper_bound(base_tokens: int = 85) -> int: +def high_detail_image_token_upper_bound(base_tokens: int = 85, tile_tokens: int | None = None) -> int: largest_tile_count: Final = calculate_tiles_needed( MAX_LONG_SIDE_FOR_IMAGE_HIGH_RES, MAX_SHORT_SIDE_FOR_IMAGE_HIGH_RES ) - return base_tokens + (base_tokens * 2) * largest_tile_count + per_tile: Final = base_tokens * 2 if tile_tokens is None else tile_tokens + return base_tokens + per_tile * largest_tile_count def _unpack_ints(fmt: str, buffer: bytes) -> tuple[int, ...]: @@ -381,6 +386,219 @@ def calculate_img_tokens( return total_tokens +_IMAGE_HEADER_BASE64_CHARS: Final = 65_536 +_MAX_IMAGE_PATCHES: Final = 30_000 +_ANTHROPIC_IMAGE_LIMITS: Final = MappingProxyType( + {"claude-standard": (ANTHROPIC_IMAGE_MAX_LONG_EDGE_PX, 1568), "claude-high": (2576, 4784)} +) +_TILE_IMAGE_TOKEN_COSTS: Final = MappingProxyType( + { + "gpt-4o": (85, 170), + "gpt-4.1": (85, 170), + "gpt-4o-mini": (2833, 5667), + "gpt-5": (70, 140), + "gpt-5.1": (70, 140), + "o1": (75, 150), + "o1-pro": (75, 150), + "o3": (75, 150), + } +) +_PATCH_IMAGE_MULTIPLIERS: Final = MappingProxyType( + { + "gpt-6-astra": 1.2, + "gpt-5.6-sol": 1.2, + "gpt-5.6-terra": 1.2, + "gpt-5.6-luna": 1.2, + "gpt-5.5": 1.2, + "gpt-5.4": 1.2, + "gpt-5.4-mini": 1.2, + "gpt-5.4-nano": 1.2, + "gpt-5.2": 1.2, + "gpt-4.1-mini": 1.62, + } +) + + +def _inline_image_dimensions(data: str) -> tuple[int, int] | None: + separator: Final = data.find(",", 0, 256) + if not data.startswith("data:") or separator < 0 or not data[:separator].endswith(";base64"): + return None + encoded_header: Final = data[separator + 1 : separator + 1 + _IMAGE_HEADER_BASE64_CHARS] + try: + header: Final = base64.b64decode(encoded_header, validate=True) + image_type: Final = get_image_type(header) + if image_type not in ("png", "gif", "jpeg", "webp"): + return None + if image_type == "webp" and header[12:16] not in (b"VP8X", b"VP8 ", b"VP8L"): + return None + dimensions: Final = _image_dimensions_from_bytes(header) + except (ValueError, TypeError, struct.error): + return None + width, height = dimensions + return (width, height) if width > 0 and height > 0 else None + + +def _patch_image_limits(model: str, detail: str) -> tuple[int, int | None] | None: + if model in ("gpt-5.2", "gpt-4.1-mini"): + return None if detail == "original" else (2048, 6144) + original_default: Final = model in ("gpt-6-astra", "gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna", "gpt-5.5") + effective_detail: Final = ("original" if original_default else "high") if detail == "auto" else detail + if effective_detail == "low": + return (2048, 6144) if model in ("gpt-5.4", "gpt-5.4-mini", "gpt-5.4-nano") else (512, None) + if effective_detail == "high": + return (65_535 if model == "gpt-6-astra" else 2048, 2500) + return ( + (65_535, None) if model in ("gpt-6-astra", "gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna") else (6000, 10_000) + ) + + +def _image_patch_count(width: int, height: int, max_dimension: int, budget: int | None) -> int: + dimension_scale: Final = min(1.0, max_dimension / max(width, height)) + fitted_width: Final = max(1, round(width * dimension_scale)) + fitted_height: Final = max(1, round(height * dimension_scale)) + patches: Final = math.ceil(fitted_width / 32) * math.ceil(fitted_height / 32) + if budget is None or patches <= budget: + return patches + shrink: Final = math.sqrt(32**2 * budget / (fitted_width * fitted_height)) + width_patches: Final = fitted_width * shrink / 32 + height_patches: Final = fitted_height * shrink / 32 + aligned_shrink: Final = shrink * min( + math.floor(width_patches) / width_patches, math.floor(height_patches) / height_patches + ) + resized_width: Final = max(1, math.floor(fitted_width * aligned_shrink)) + resized_height: Final = max(1, math.floor(fitted_height * aligned_shrink)) + return math.ceil(resized_width / 32) * math.ceil(resized_height / 32) + + +def _image_policy_model_name(model: str) -> str | None: + provider, separator, provider_model = model.partition("/") + if not separator: + return model + if provider == "openai": + return provider_model + if provider == "openrouter": + return provider_model.removeprefix("openai/") if provider_model.startswith("openai/") else None + if provider not in ("azure", "azure_ai"): + return None + catalog_entry: Final = as_str_mapping( + cast( # cast-ok: read-only view of the legacy string-keyed registry; as_str_mapping checks the entry + Mapping[str, object], litellm.model_cost + ).get(model) + ) + if catalog_entry is not None and catalog_entry.get("litellm_provider") in ("azure", "azure_ai"): + return provider_model.rsplit("/", 1)[-1] + return provider_model.removeprefix("responses/").removeprefix("o_series/").removeprefix("openai/") + + +def _anthropic_image_family(model: str) -> str | None: + from litellm.utils import get_potential_model_names + + provider, separator, _ = model.partition("/") + if separator and provider not in ( + "anthropic", + "bedrock", + "bedrock_converse", + "bedrock_mantle", + "vertex_ai", + "vertex_ai_beta", + "azure_ai", + "openrouter", + ): + return None + candidates: Final = get_potential_model_names(model, provider if separator else "anthropic") + catalog: Final = cast(Mapping[str, object], litellm.model_cost) # cast-ok: read-only view of the legacy registry + for candidate in dict.fromkeys(candidates.values()): + if not isinstance(candidate, str): + continue + if (entry := as_str_mapping(catalog.get(candidate))) is None or entry.get("supports_vision") is not True: + continue + if ( + version := re.search( + r"(?:^|[./])claude-(?:(?:haiku|sonnet|opus)-)?(\d+)(?:[-.](\d{1,2})(?!\d))?", candidate + ) + ) is None: + continue + major, minor = int(version[1]), int(version[2] or 0) + return "claude-high" if (major, minor) >= (4, 7) else "claude-standard" + return None + + +def _anthropic_image_tokens(dimensions: tuple[int, int] | None, limits: tuple[int, int]) -> int: + max_edge, token_budget = limits + if dimensions is None: + return token_budget + long_edge, short_edge = max(dimensions), min(dimensions) + aspect_ratio: Final = long_edge / short_edge + + def tokens_at_edge(edge: int) -> int: + return math.ceil(edge / 28) * math.ceil(max(round(edge / aspect_ratio), 1) / 28) + + fitted_edge: Final = bisect_right(range(1, min(long_edge, max_edge) + 1), token_budget, key=tokens_at_edge) + return tokens_at_edge(max(fitted_edge, 1)) + + +def _image_model_family(model: str) -> str | None: + provider_model: Final = _image_policy_model_name(model) + if provider_model is None: + return _anthropic_image_family(model) + family: Final = re.sub(r"-\d{4}-\d{2}-\d{2}$", "", provider_model) + canonical: Final = litellm.azure_llms.get(f"azure/{family}", family) if model.startswith("azure/") else family + return ( + canonical + if canonical in _TILE_IMAGE_TOKEN_COSTS or canonical in _PATCH_IMAGE_MULTIPLIERS + else _anthropic_image_family(model) + ) + + +def _image_policy_tokens(family: str | None, dimensions: tuple[int, int] | None, detail: str) -> int: + if family is None: + return DEFAULT_IMAGE_TOKEN_COUNT + anthropic_limits: Final = _ANTHROPIC_IMAGE_LIMITS.get(family) + if anthropic_limits is not None: + return _anthropic_image_tokens(dimensions, anthropic_limits) + tile_costs: Final = _TILE_IMAGE_TOKEN_COSTS.get(family) + multiplier: Final = _PATCH_IMAGE_MULTIPLIERS.get(family) + if tile_costs is not None: + base_tokens, tile_tokens = tile_costs + if detail == "original": + return DEFAULT_IMAGE_TOKEN_COUNT + if detail == "low": + return base_tokens + if dimensions is None: + return high_detail_image_token_upper_bound(base_tokens, tile_tokens) + width, height = dimensions + scale: Final = min(1.0, 2048 / max(width, height), 768 / min(width, height)) + tiles: Final = calculate_tiles_needed(max(1, math.floor(width * scale)), max(1, math.floor(height * scale))) + return base_tokens + tiles * tile_tokens + limits: Final = _patch_image_limits(family, detail) + if limits is None or multiplier is None: + return DEFAULT_IMAGE_TOKEN_COUNT + max_dimension, patch_budget = limits + patches: Final = ( + min(math.ceil(max_dimension / 32) ** 2, patch_budget if patch_budget is not None else _MAX_IMAGE_PATCHES) + if dimensions is None + else _image_patch_count(*dimensions, max_dimension, patch_budget) + ) + return math.ceil(patches * multiplier) + + +def estimate_image_tokens( + model: str | Sequence[str], data: str, detail: str = "auto", *, fallback_model: str | None = None +) -> int: + """Estimate image policies from at most 48 KiB of inline headers, without fetching media.""" + if detail not in ("low", "high", "auto", "original"): + return DEFAULT_IMAGE_TOKEN_COUNT + models: Final = (model,) if isinstance(model, str) else model + fallback_family: Final = _image_model_family(fallback_model) if fallback_model is not None else None + families: Final = tuple( + dict.fromkeys(_image_model_family(candidate) or fallback_family for candidate in models or ("",)) + ) + if all(family is None for family in families): + return DEFAULT_IMAGE_TOKEN_COUNT + dimensions: Final = _inline_image_dimensions(data) + return max(_image_policy_tokens(family, dimensions, detail) for family in families) + + TokenCounterFunction = Callable[[str], int] """ Type for a function that counts tokens in a string. diff --git a/litellm/proxy/hooks/parallel_request_limiter_v3.py b/litellm/proxy/hooks/parallel_request_limiter_v3.py index 567b8452ae2..e44d9304ac1 100644 --- a/litellm/proxy/hooks/parallel_request_limiter_v3.py +++ b/litellm/proxy/hooks/parallel_request_limiter_v3.py @@ -45,11 +45,14 @@ from litellm.caching.redis_cache import log_redis_failure from litellm.constants import DYNAMIC_RATE_LIMIT_ERROR_THRESHOLD_PER_MINUTE, INTERNAL_CALL_ORIGIN_METADATA_KEY from litellm.cost_calculator import get_usage_object from litellm.integrations.custom_logger import CustomLogger +from litellm.litellm_core_utils.audio_utils.utils import estimate_audio_tokens +from litellm.litellm_core_utils.core_helpers import as_str_mapping from litellm.litellm_core_utils.prompt_templates.common_utils import ( get_str_from_messages, + split_media_for_tpm, ) from litellm.litellm_core_utils.ptu_pricing import is_ptu_cost_attribution_enabled -from litellm.litellm_core_utils.token_counter import offload_token_count +from litellm.litellm_core_utils.token_counter import estimate_image_tokens, offload_token_count from litellm.llms.azure.ptu_capacity import normalized_tokens from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.auth.auth_utils import ( @@ -98,6 +101,7 @@ if TYPE_CHECKING: from opentelemetry.trace import Span as _Span from litellm.proxy.utils import InternalUsageCache as _InternalUsageCache + from litellm.router import Router from litellm.types.agents import AgentResponse from litellm.types.caching import RedisPipelineIncrementOperation @@ -135,6 +139,18 @@ class RateLimitedModel: return requested_limit if requested_limit is not None else limits.get(self.group) +def _resolve_media_models(router: "Router | None", model: str, team_id: str | None) -> tuple[str, ...]: + if router is None: + return (model,) + return tuple(dict.fromkeys(router.resolved_litellm_models(model, team_id, use_base_model=True))) or (model,) + + +def _resolve_media_models_via_proxy_router(model: str, team_id: str | None) -> tuple[str, ...]: + from litellm.proxy.proxy_server import llm_router + + return _resolve_media_models(llm_router, model, team_id) + + def _resolve_model_group_alias_via_proxy_router(model: str) -> str | None: from litellm.proxy.proxy_server import llm_router @@ -862,6 +878,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): time_provider: Callable[[], datetime] | None = None, tag_rate_limit_resolver: TagRateLimitResolver = resolve_tag_rate_limits_from_db, model_group_resolver: Callable[[str], str | None] = _resolve_model_group_alias_via_proxy_router, + media_model_resolver: Callable[[str, str | None], tuple[str, ...]] = _resolve_media_models_via_proxy_router, ptu_team_ceiling_resolver: Callable[ [str, str], PTUTeamCeiling | None ] = _resolve_ptu_team_ceiling_via_proxy_router, @@ -872,6 +889,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): self._time_provider = time_provider or datetime.now self._tag_rate_limit_resolver = tag_rate_limit_resolver self._model_group_resolver = model_group_resolver + self._media_model_resolver = media_model_resolver self._ptu_team_ceiling_resolver = ptu_team_ceiling_resolver self._fail_closed_resolver = fail_closed_resolver self._force_hash_tag_grouping_resolver = force_hash_tag_grouping_resolver @@ -1133,11 +1151,12 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): def _estimate_tokens_for_request( self, - data: dict, + data: object, model: str | None = None, min_configured_tpm_limit: int | None = None, call_type: str | None = None, configured_output_tokens: int | None = None, + team_id: str | None = None, ) -> int: """ Estimate total tokens this request will consume so we can reserve them @@ -1164,6 +1183,8 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): min_configured_tpm_limit=min_configured_tpm_limit, call_type=call_type, configured_output_tokens=configured_output_tokens, + team_id=team_id, + media_model_resolver=self._media_model_resolver, ) total_estimated: Final = estimated_input_tokens + max_tokens_estimate @@ -1184,6 +1205,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): call_type: str | None, configured_output_tokens: int | None, raw_estimate: int, + team_id: str | None = None, ) -> int: """The team PTU ceiling counts Azure normalized tokens, so its reservation weighs the output budget the way the ceiling does instead of the raw sum the other scopes reserve.""" @@ -1194,6 +1216,8 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): min_configured_tpm_limit=min_configured_tpm_limit, call_type=call_type, configured_output_tokens=configured_output_tokens, + team_id=team_id, + media_model_resolver=self._media_model_resolver, ) normalized: Final = normalized_tokens( ceiling, prompt_tokens=estimated_input_tokens, completion_tokens=max_tokens_estimate @@ -1206,6 +1230,9 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): min_configured_tpm_limit: int | None = None, call_type: str | None = None, configured_output_tokens: int | None = None, + include_media: bool = True, + team_id: str | None = None, + media_model_resolver: Callable[[str, str | None], tuple[str, ...]] | None = None, ) -> tuple[int, int]: """ Estimate input tokens and output (max_tokens) budget separately, so @@ -1229,10 +1256,12 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): floor entirely, so the reservation reflects what this tenant's model actually emits rather than one constant shared by every tenant. """ - if not isinstance(data, dict): + request_data: Final = as_str_mapping(data) + if not isinstance(data, dict) or request_data is None: return 0, 0 + request_model: Final = request_data.get("model") translated_data: Final = self._translate_google_genai_native_request(data, call_type) - estimable_data: Final = translated_data if translated_data is not None else data + estimable_data: Final = translated_data if translated_data is not None else request_data selected_fields: Final[tuple[object | None, object | None, object | None]] = ( (None, None, estimable_data.get("input")) if call_type in RESPONSES_API_CALL_TYPES or call_type in EMBEDDING_API_CALL_TYPES @@ -1249,10 +1278,27 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): ) ) messages, prompt, raw_input = selected_fields - input_text: Final = ( + original_input: Final = ( billable_input_since_newest_compaction(raw_input) if call_type in RESPONSES_API_CALL_TYPES else raw_input ) + media_source: Final = messages if messages is not None else original_input + projected, media = ( + split_media_for_tpm(media_source) + if include_media and (isinstance(messages, list) or call_type in RESPONSES_API_CALL_TYPES) + else (original_input, ()) + ) + input_text: Final = projected if call_type in RESPONSES_API_CALL_TYPES else original_input + fallback_model: Final = request_model if isinstance(request_model, str) else None + resolved_models: Final = ( + media_model_resolver(fallback_model or "", team_id) + if media and media_model_resolver is not None + else (fallback_model or "",) + ) + media_tokens: Final = sum( + self._estimate_media_block_tokens(block, resolved_models, fallback_model) for block in media + ) + total_chars: Final = ( len(get_str_from_messages(messages)) if isinstance(messages, list) and messages @@ -1267,7 +1313,9 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): else 0 ) - estimated_input_tokens: Final = max(1, total_chars // DEFAULT_CHARS_PER_TOKEN) if total_chars > 0 else 0 + estimated_input_tokens: Final = ( + max(1, total_chars // DEFAULT_CHARS_PER_TOKEN) if total_chars > 0 else 0 + ) + media_tokens explicit_max_tokens: Final = self._get_explicit_output_cap(data, call_type) generates_no_output: Final = ( @@ -1283,7 +1331,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): max_tokens_estimate: Final = ( 0 if generates_no_output - or (explicit_max_tokens is None and total_chars == 0 and configured_output_tokens is None) + or (explicit_max_tokens is None and estimated_input_tokens == 0 and configured_output_tokens is None) else explicit_max_tokens if explicit_max_tokens is not None else configured_output_tokens @@ -1293,6 +1341,24 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): return estimated_input_tokens, max_tokens_estimate * self.get_output_candidate_count(data, call_type) + @staticmethod + def _estimate_media_block_tokens( + block: Mapping[str, object], models: tuple[str, ...], fallback_model: str | None + ) -> int: + if block.get("type") == "input_audio": + audio: Final = _callback_kwargs_mapping(block.get("input_audio")) + data: Final = audio.get("data") + return estimate_audio_tokens(data if isinstance(data, str) else None, models, fallback_model=fallback_model) + image: Final = _callback_kwargs_mapping(block.get("image_url")) + url: Final = image.get("url") + detail: Final = image.get("detail") + return estimate_image_tokens( + models, + url if isinstance(url, str) else "", + detail if isinstance(detail, str) else "auto", + fallback_model=fallback_model, + ) + def _is_redis_cluster(self) -> bool: """ Check if the dual cache is using Redis cluster. @@ -3958,7 +4024,9 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): except Exception: # noqa: BLE001 # tokenizer failures degrade to the cheap estimate if call_type in RERANK_API_CALL_TYPES and isinstance(selected_text, str): return max(0, len(selected_text) // DEFAULT_CHARS_PER_TOKEN) - estimated_input_tokens, _ = self._estimate_input_and_output_tokens(data=data, call_type=call_type) + estimated_input_tokens, _ = self._estimate_input_and_output_tokens( + data=data, call_type=call_type, include_media=False + ) return estimated_input_tokens + audio_token_estimate async def _reserve_project_io_tokens_or_raise( @@ -4003,6 +4071,7 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): data=data, min_configured_tpm_limit=min_configured_otpm_limit, call_type=call_type, + include_media=False, ) raw_estimated_input_tokens: Final = await offload_token_count(self._estimate_precise_input_tokens)( data=data, model=requested_model, call_type=call_type @@ -4360,12 +4429,13 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): # Post-call reconciliation refunds the over-reservation # delta when actual usage comes in below the floor. estimated_tokens: Final = max( - self._estimate_tokens_for_request( + await offload_token_count(self._estimate_tokens_for_request)( data=data, model=requested_model, min_configured_tpm_limit=min_configured_tpm_limit, call_type=call_type, configured_output_tokens=configured_output_tokens, + team_id=user_api_key_dict.team_id, ), 1, ) @@ -4383,13 +4453,14 @@ class _PROXY_MaxParallelRequestsHandler_v3(CustomLogger): ) capped_request: Final = _REQUEST_RATE_LIMIT_DATA.validate_python(data) - ptu_estimated_tokens: Final = self._estimate_ptu_tokens_for_request( + ptu_estimated_tokens: Final = await offload_token_count(self._estimate_ptu_tokens_for_request)( ceiling=stash.ptu_ceiling, data=capped_request, min_configured_tpm_limit=min_configured_tpm_limit, call_type=call_type, configured_output_tokens=configured_output_tokens, raw_estimate=estimated_tokens, + team_id=user_api_key_dict.team_id, ) if stash.ptu_ceiling is not None and ptu_estimated_tokens > stash.ptu_ceiling.tpm_limit: if not reuses_request_slot: diff --git a/litellm/router.py b/litellm/router.py index e8d83e1203e..d712ed4ac52 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -10364,28 +10364,38 @@ class Router: model_info[field] = backend_value @staticmethod - def _cost_map_backend_model(deployment: Deployment) -> str: - model_info_base_model: Final = deployment.model_info.base_model + def _cost_map_backend_model(deployment: Deployment | DeploymentTypedDict) -> str: + model_info_base_model: Final = ( + deployment.model_info.base_model + if isinstance(deployment, Deployment) + else _configured_model_info(deployment).get("base_model") + ) if isinstance(model_info_base_model, str) and model_info_base_model: return model_info_base_model - params_base_model: Final = deployment.litellm_params.get("base_model") + params: Final = _LITELLM_PARAMS_ADAPTER.validate_python( + deployment.litellm_params.model_dump(exclude_none=True) + if isinstance(deployment, Deployment) + else deployment["litellm_params"] + ) + params_base_model: Final = params.get("base_model") if isinstance(params_base_model, str) and params_base_model: return params_base_model - params: Final = _LITELLM_PARAMS_ADAPTER.validate_python(deployment.litellm_params.model_dump(exclude_none=True)) + model: Final = ( + deployment.litellm_params.model + if isinstance(deployment, Deployment) + else deployment["litellm_params"].get("model", "") + ) invoked_model: Final = params.get("model_id") if ( isinstance(invoked_model, str) and invoked_model - and ( - deployment.litellm_params.custom_llm_provider == "bedrock" - or deployment.litellm_params.model.startswith("bedrock/") - ) + and (params.get("custom_llm_provider") == "bedrock" or model.startswith("bedrock/")) ): profile_model: Final = invoked_model if invoked_model.startswith("bedrock/") else f"bedrock/{invoked_model}" profile_info: Final = lookup_model_info_for_registration(model=profile_model, custom_llm_provider="bedrock") if not isinstance(profile_info, ModelInfoLookupMiss): return profile_model - return deployment.litellm_params.model + return model @staticmethod def _inherit_builtin_service_tier_pricing( @@ -13945,7 +13955,9 @@ class Router: ] return named or self._get_wildcard_deployments(model_name=model_group) - def resolved_litellm_models(self, model_name: str, team_id: str | None = None) -> tuple[str, ...]: + def resolved_litellm_models( + self, model_name: str, team_id: str | None = None, *, use_base_model: bool = False + ) -> tuple[str, ...]: """The provider model strings `model_name` can actually be served by on this proxy. `get_model_list` composes every channel the request path itself uses (exact name, @@ -13958,6 +13970,12 @@ class Router: deployment behind it - so the fallback for an empty result is the caller's policy, never this function's. """ + if use_base_model: + return tuple( + backend_model + for deployment in self.get_model_list(model_name=model_name, team_id=team_id) or () + if (backend_model := self._cost_map_backend_model(deployment)) + ) return tuple( litellm_model for deployment in self.get_model_list(model_name=model_name, team_id=team_id) or () diff --git a/tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py b/tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py index d50ab908b43..b9675f6fd5c 100644 --- a/tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py +++ b/tests/unit/litellm_core_utils/prompt_templates/test_litellm_core_utils_prompt_templates_common_utils.py @@ -37,6 +37,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import ( schema_has_unresolvable_ref, schema_property_count, split_concatenated_json_objects, + split_media_for_tpm, strip_encrypted_reasoning_from_messages, strip_litellm_response_only_fields, system_messages_first, @@ -2512,3 +2513,118 @@ def test_shared_image_url_agrees_with_responses_normalization( expected: Final = {"type": "image_url", "image_url": {"url": "data:image/png;base64,IMAGE"}} assert as_openai_image_part(part) == expected assert as_openai_image_part(normalized) == expected + + +@pytest.mark.parametrize( + "part,projected,media", + [ + ( + {"type": "image_url", "image_url": {"url": "data:image/png;base64,IMAGE", "detail": "high"}}, + {"type": "image_url", "image_url": {"url": "", "detail": "high"}}, + {"type": "image_url", "image_url": {"url": "data:image/png;base64,IMAGE", "detail": "high"}}, + ), + ( + {"type": "input_image", "image_url": "https://example.com/image.png", "detail": "low"}, + {"type": "input_image", "image_url": "", "detail": "low"}, + {"type": "image_url", "image_url": {"url": "https://example.com/image.png", "detail": "low"}}, + ), + ( + {"type": "computer_screenshot", "image_url": "data:image/png;base64,IMAGE"}, + {"type": "computer_screenshot", "image_url": ""}, + {"type": "image_url", "image_url": {"url": "data:image/png;base64,IMAGE", "detail": "auto"}}, + ), + ( + {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": "IMAGE"}}, + {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": ""}}, + {"type": "image_url", "image_url": {"url": "data:image/png;base64,IMAGE", "detail": "auto"}}, + ), + ( + {"type": "input_audio", "input_audio": {"data": "AUDIO", "format": "wav"}}, + {"type": "input_audio", "input_audio": {"data": "", "format": "wav"}}, + {"type": "input_audio", "input_audio": {"data": "AUDIO", "format": "wav"}}, + ), + ( + {"type": "file", "file": {"file_data": "IMAGE", "format": "image/png", "filename": "image.png"}}, + {"type": "file", "file": {"file_data": "", "format": "image/png", "filename": "image.png"}}, + {"type": "image_url", "image_url": {"url": "data:image/png;base64,IMAGE", "detail": "auto"}}, + ), + ( + {"type": "input_file", "file_data": "data:audio/wav;base64,AUDIO", "filename": "audio.wav"}, + {"type": "input_file", "file_data": "", "filename": "audio.wav"}, + {"type": "input_audio", "input_audio": {"data": "AUDIO", "format": "wav"}}, + ), + ], +) +def test_tpm_projection_blanks_only_declared_media_transport( + part: Mapping[str, object], projected: Mapping[str, object], media: Mapping[str, object], +) -> None: + text: Final = {"type": "text", "text": "data:image/png;base64,IMAGE"} + request: Final = [{"role": "user", "content": [text, {**part, "cache_control": {"type": "ephemeral"}}]}] + original: Final = copy.deepcopy(request) + assert split_media_for_tpm(request) == ( + [{"role": "user", "content": [text, {**projected, "cache_control": {"type": "ephemeral"}}]}], + (media,), + ) + assert request == original + + +def test_tpm_projection_retains_nested_and_native_history_around_media() -> None: + image: Final = {"type": "image_url", "image_url": {"url": "data:image/png;base64,IMAGE", "detail": "high"}} + empty_image: Final = {"type": "image_url", "image_url": {"url": "", "detail": "high"}} + document: Final = {"type": "document", "title": "keep", "source": {"type": "content", "content": [image]}} + empty_document: Final = {**document, "source": {"type": "content", "content": [empty_image]}} + tool: Final = {"type": "tool_result", "tool_use_id": "call", "content": [document, {"type": "text", "text": "keep"}]} + empty_tool: Final = {**tool, "content": [empty_document, {"type": "text", "text": "keep"}]} + request: Final = [ + {"role": "user", "content": [tool]}, + {"type": "function_call_output", "call_id": "call", "output": [image]}, + {"type": "custom_tool_call_output", "call_id": "custom", "output": [image]}, + {"type": "computer_call_output", "call_id": "computer", "output": { + "type": "computer_screenshot", "image_url": "data:image/png;base64,IMAGE", + }}, + {"type": "image_generation_call", "id": "ig", "result": "IMAGE", "status": "completed"}, + {"type": "code_interpreter_call", "id": "ci", "code": "keep()", "outputs": [ + {"type": "logs", "logs": "keep"}, {"type": "image", "url": "https://example.com/image.png"}, + ]}, + ] + original: Final = copy.deepcopy(request) + projected, media = split_media_for_tpm(request) + assert projected == [ + {"role": "user", "content": [empty_tool]}, + {"type": "function_call_output", "call_id": "call", "output": [empty_image]}, + {"type": "custom_tool_call_output", "call_id": "custom", "output": [empty_image]}, + {"type": "computer_call_output", "call_id": "computer", "output": { + "type": "computer_screenshot", "image_url": "", + }}, + {"type": "image_generation_call", "id": "ig", "result": "", "status": "completed"}, + {"type": "code_interpreter_call", "id": "ci", "code": "keep()", "outputs": [ + {"type": "logs", "logs": "keep"}, {"type": "image", "url": ""}, + ]}, + ] + assert media == ( + image, image, image, + {"type": "image_url", "image_url": {"url": "data:image/png;base64,IMAGE", "detail": "auto"}}, + {"type": "image_url", "image_url": {"url": "data:application/octet-stream;base64,IMAGE", "detail": "auto"}}, + {"type": "image_url", "image_url": {"url": "https://example.com/image.png", "detail": "auto"}}, + ) + assert request == original + + +def test_tpm_projection_preserves_non_media_identity_and_opaque_values() -> None: + image: Final = {"type": "input_image", "image_url": "data:image/png;base64,OPAQUE"} + request: Final = [ + {"role": "user", "content": json.dumps(image)}, + {"type": "custom_tool_call_output", "call_id": "call", "output": json.dumps(image)}, + {"type": "function_call", "call_id": "call", "name": "keep", "arguments": {"content": [image]}}, + {"type": "file_search_call", "results": [{"text": "keep", "content": [image]}]}, + {"type": "image_generation_call", "id": "ig", "status": "in_progress", "result": None}, + {"role": "user", "content": [ + {"type": "document", "source": {"type": "base64", "media_type": "application/pdf", "data": "PDF"}}, + {"type": "document", "source": {"type": "text", "data": json.dumps(image)}}, + {"type": "file", "file": {"format": "application/pdf", "file_data": "PDF"}}, + {"type": "text", "text": "keep", "metadata": {"content": [image]}}, + ]}, + ] + projected, media = split_media_for_tpm(request) + assert projected is request + assert media == () diff --git a/tests/unit/litellm_core_utils/test_audio_utils.py b/tests/unit/litellm_core_utils/test_audio_utils.py index 155f6680416..122445e8e30 100644 --- a/tests/unit/litellm_core_utils/test_audio_utils.py +++ b/tests/unit/litellm_core_utils/test_audio_utils.py @@ -2,16 +2,23 @@ Test the audio utils functionality in litellm_core_utils/audio_utils/utils.py """ +import base64 import io import os import tempfile +from typing import Final from unittest.mock import mock_open, patch import pytest +import soundfile as sf from litellm.litellm_core_utils.audio_utils.utils import ( + AUDIO_BYTES_PER_TOKEN, + AUDIO_TOKENS_PER_SECOND, + DEFAULT_AUDIO_TOKEN_ESTIMATE, ProcessedAudioFile, calculate_request_duration, + estimate_audio_tokens, get_audio_file_content_hash, get_audio_file_for_health_check, get_audio_file_name, @@ -19,6 +26,63 @@ from litellm.litellm_core_utils.audio_utils.utils import ( ) +def _encoded_audio(seconds: int, sample_rate: int, channels: int, container: str) -> str: + buffer: Final = io.BytesIO() + with sf.SoundFile(buffer, "w", samplerate=sample_rate, channels=channels, format=container) as audio: + audio.buffer_write(b"\0" * (seconds * sample_rate * channels * 2), dtype="int16") + return base64.b64encode(buffer.getvalue()).decode() + + +@pytest.mark.parametrize( + ("sample_rate", "channels", "container"), [(8000, 1, "WAV"), (48000, 2, "WAV"), (24000, 1, "FLAC")] +) +def test_audio_estimate_uses_duration_independent_of_sample_rate_channels_and_compression( + sample_rate: int, channels: int, container: str +) -> None: + assert estimate_audio_tokens(_encoded_audio(1, sample_rate, channels, container)) == AUDIO_TOKENS_PER_SECOND + assert estimate_audio_tokens(_encoded_audio(31, sample_rate, channels, container)) == 31 * AUDIO_TOKENS_PER_SECOND + + +# Audio rates verified 2026-10-28: https://ai.google.dev/gemini-api/docs/generate-content/media-resolution and +# https://developers.openai.com/api/docs/guides/voice-latency-cost +@pytest.mark.parametrize( + ("models", "fallback", "expected"), + [ + (("openai/gpt-audio-1.5",), "gemini-3.8-flash", 10), + (("gpt-audio-1.5", "gemini/gemini-3.8-flash"), None, 25), + (("vertex_ai/gemini-3.8-flash",), None, 25), + (("openrouter/google/gemini-3.8-flash",), None, 25), + (("azure/audio-deployment",), "gpt-audio-1.5", 10), + (("opaque-deployment",), "gemini-3.8-flash", 25), + ((), "gemini-3.8-flash", 25), + (("gemini/gemini-2.5-flash", "gemini/gemini-3-pro-preview"), None, 32), + (("other/gemini-3.8-flash",), None, 10), + ], +) +def test_audio_estimate_prefers_each_backend_rate_over_the_requested_model( + models: tuple[str, ...], fallback: str | None, expected: int +) -> None: + assert estimate_audio_tokens(_encoded_audio(1, 8000, 1, "WAV"), models, fallback) == expected + + +@pytest.mark.parametrize("data", [None, "", "invalid-base64!", base64.b64encode(b"unreadable audio").decode()]) +def test_audio_estimate_retains_nonzero_fallback_for_unreadable_audio(data: str | None) -> None: + assert estimate_audio_tokens(data) == DEFAULT_AUDIO_TOKEN_ESTIMATE + + +def test_audio_estimate_retains_size_based_fallback_for_large_unreadable_audio() -> None: + data: Final = base64.b64encode(b"\0" * (AUDIO_BYTES_PER_TOKEN * 400)).decode() + assert estimate_audio_tokens(data) == 400 + + +def test_audio_estimate_bounds_base64_allocation_before_decoding() -> None: + data: Final = base64.b64encode(b"\0" * (16 * 1024 * 1024 + 1)).decode() + with patch("base64.b64decode", side_effect=AssertionError("oversized audio must not be decoded")) as decode: + estimate: Final = estimate_audio_tokens(data) + assert estimate == len(data) * 3 // 4 // AUDIO_BYTES_PER_TOKEN + decode.assert_not_called() + + class TestProcessAudioFile: """Test the process_audio_file function with various input types""" diff --git a/tests/unit/litellm_core_utils/test_token_counter.py b/tests/unit/litellm_core_utils/test_token_counter.py index 92f971f64e9..b02e0732817 100644 --- a/tests/unit/litellm_core_utils/test_token_counter.py +++ b/tests/unit/litellm_core_utils/test_token_counter.py @@ -37,6 +37,7 @@ from litellm.litellm_core_utils.token_counter import ( _get_extrapolating_count_function, _get_tiktoken_count_function, calculate_img_tokens, + estimate_image_tokens, high_detail_image_token_upper_bound, offload_token_count, ) @@ -2152,6 +2153,136 @@ def test_high_detail_image_token_upper_bound_is_reached_by_the_largest_high_res_ assert calculate_img_tokens(_png_data_url(1, 1), mode="high") < high_detail_image_token_upper_bound() +# Image sizing verified 2026-10-28: https://developers.openai.com/api/docs/guides/images-vision +# and https://platform.claude.com/docs/en/build-with-claude/vision-coordinates +@pytest.mark.parametrize( + ("model", "width", "height", "detail", "expected"), + [ + ("gpt-4o-mini", 1024, 1024, "low", 2833), + ("gpt-4o-mini", 1024, 1024, "high", 25501), + ("gpt-5.1", 512, 512, "auto", 210), + ("gpt-6-astra", 2048, 2048, "high", 3000), + ("gpt-6-astra", 4096, 512, "high", 2458), + ("gpt-6-astra", 2048, 2048, "auto", 4916), + ("gpt-5.6-sol", 2048, 2048, "low", 308), + ("gpt-5.5", 6400, 6400, "original", 12000), + ("gpt-5.4-mini", 2048, 2048, "low", 4916), + ("gpt-5.4-nano", 2048, 2048, "auto", 3000), + ("gpt-5.2", 4096, 512, "auto", 615), + ("gpt-4.1-mini-2025-04-14", 1024, 1024, "low", 1659), + ("claude-haiku-5-5", 200, 200, "auto", 64), + ("anthropic/claude-haiku-5-5", 1024, 1024, "auto", 1369), + ("claude-haiku-5-5", 1920, 1080, "low", 2691), + ("claude-haiku-4-5", 1920, 1080, "auto", 1560), + ("claude-haiku-4-5", 1075, 1520, "auto", 1551), + ("claude-haiku-5-5", 3840, 2160, "auto", 4784), + ("claude-haiku-5-5", 8000, 100, "auto", 184), + ], +) +def test_image_estimate_uses_model_resolution_and_detail( + model: str, width: int, height: int, detail: str, expected: int +) -> None: + assert estimate_image_tokens(model, _png_data_url(width, height), detail) == expected + + +@pytest.mark.parametrize( + ("models", "fallback", "expected_model"), + [ + (("azure/deployment-prod",), "gpt-4o", "gpt-4o"), + (("openai/gpt-4o",), "gpt-4o-mini", "gpt-4o"), + (("azure/deployment-prod", "openai/gpt-4o"), "gpt-4o-mini", "gpt-4o-mini"), + (("openai/gpt-4o", "openrouter/openai/gpt-4o-mini"), None, "gpt-4o-mini"), + (("azure_ai/openai/gpt-6-astra",), None, "gpt-6-astra"), + (("azure/gpt-41",), None, "gpt-4.1"), + (("azure/eu/gpt-4o-mini",), None, "gpt-4o-mini"), + (("unknown", "gpt-4o-mini"), None, "gpt-4o-mini"), + ((), "gpt-6-astra", "gpt-6-astra"), + (("other/gpt-4o-mini",), None, "unregistered-model"), + (("azure/opaque/gpt-4o-mini",), None, "unregistered-model"), + (("bedrock/anthropic.claude-haiku-5-5",), None, "claude-haiku-5-5"), + (("bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0",), None, "claude-haiku-4-5"), + (("vertex_ai/claude-haiku-5-5@default",), None, "claude-haiku-5-5"), + (("azure_ai/claude-haiku-5-5",), None, "claude-haiku-5-5"), + (("anthropic/claude-haiku-5-5", "openai/gpt-4o"), "gpt-4o-mini", "claude-haiku-5-5"), + (("other/claude-haiku-5-5",), None, "unregistered-model"), + ], +) +def test_image_estimate_prefers_each_backend_policy_over_the_requested_model( + models: tuple[str, ...], fallback: str | None, expected_model: str +) -> None: + data: Final = _png_data_url(1024, 1024) + assert estimate_image_tokens(models, data, "high", fallback_model=fallback) == estimate_image_tokens( + expected_model, data, "high" + ) + + +@pytest.mark.parametrize( + ("model", "detail"), + [ + ("unknown", "high"), + ("claude-unregistered", "auto"), + ("gpt-4o", "original"), + ("gpt-5.2", "original"), + ("gpt-6-astra", "invalid"), + ], +) +def test_image_estimate_preserves_default_when_policy_is_unavailable(model: str, detail: str) -> None: + assert ( + estimate_image_tokens(model, _png_data_url(1024, 1024), detail) == litellm.constants.DEFAULT_IMAGE_TOKEN_COUNT + ) + + +@pytest.mark.parametrize( + "data", + [ + "https://image.invalid/photo.png", + "data:image/png;base64,invalid!", + "data:application/octet-stream;base64,bm90LWFuLWltYWdl", + "data:image/jpeg;base64,/9j/4AAQ", + "data:image/webp;base64,UklGRgAAAABXRUJQ", + _png_data_url(0, 1024), + ], +) +# Accepted patch limits verified 2026-10-28: https://developers.openai.com/api/docs/guides/images-vision +# and https://platform.claude.com/docs/en/build-with-claude/vision +@pytest.mark.parametrize( + ("model", "detail", "bound"), + [ + ("gpt-4o", "high", high_detail_image_token_upper_bound()), + ("gpt-6-astra", "high", 3000), + ("gpt-6-astra", "auto", 36000), + ("claude-haiku-5-5", "auto", 4784), + ("claude-haiku-4-5", "auto", 1568), + ], +) +def test_image_estimate_uses_bounded_fallback_without_fetching_media( + data: str, model: str, detail: str, bound: int +) -> None: + with patch("httpx.Client.send") as send: + estimate: Final = estimate_image_tokens(model, data, detail) + send.assert_not_called() + assert estimate == bound + assert estimate > estimate_image_tokens(model, _png_data_url(1024, 1024), detail) + + +@pytest.mark.parametrize("container", ["PNG", "JPEG", "GIF", "WEBP"]) +def test_image_estimate_depends_on_dimensions_and_not_compressed_payload_size(container: str) -> None: + output: Final = io.BytesIO() + Image.new("RGB", (1024, 1024)).save(output, format=container) + data: Final = "data:application/octet-stream;base64," + base64.b64encode(output.getvalue()).decode() + assert estimate_image_tokens("gpt-6-astra", data) == estimate_image_tokens("gpt-6-astra", _png_data_url(1024, 1024)) + + +def test_image_estimate_bounds_header_decoding_for_large_inline_payloads() -> None: + header: Final = base64.b64decode(_png_data_url(1024, 1024).split(",", 1)[1]) + data: Final = "data:image/png;base64," + base64.b64encode(header + b"\0" * 4_000_000).decode() + with patch("base64.b64decode", wraps=base64.b64decode) as decode_header: + estimate: Final = estimate_image_tokens("gpt-6-astra", data) + assert estimate == estimate_image_tokens("gpt-6-astra", _png_data_url(1024, 1024)) + assert decode_header.call_count == 1 + assert len(decode_header.call_args.args[0]) <= 65_536 + + HUB_SETUP_SCRIPT: Final = """ import json import sys diff --git a/tests/unit/proxy/hooks/test_tpm_concurrent.py b/tests/unit/proxy/hooks/test_tpm_concurrent.py index 3d1e494e36c..8245c261a12 100644 --- a/tests/unit/proxy/hooks/test_tpm_concurrent.py +++ b/tests/unit/proxy/hooks/test_tpm_concurrent.py @@ -15,28 +15,38 @@ Redis. """ import asyncio +import base64 +import copy +import io import json from datetime import datetime, timedelta from typing import Any, Dict, Final import pytest +import soundfile as sf +from fastapi import HTTPException +from PIL import Image +from litellm import token_counter from litellm.caching.caching import DualCache +from litellm.litellm_core_utils.token_counter import UnsupportedContentTypeError, estimate_image_tokens from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.hooks.parallel_request_limiter_v3 import ( + _AUDIO_BYTES_PER_TOKEN, + DEFAULT_AUDIO_TOKEN_ESTIMATE, PROJECT_ITPM_DESCRIPTOR_KEY, PROJECT_OTPM_DESCRIPTOR_KEY, RateLimitedModel, - _AUDIO_BYTES_PER_TOKEN, - PROXY_MaxParallelRequestsHandler_v3 as RateLimitHandler, -) -from litellm.proxy.hooks.parallel_request_limiter_v3 import ( _call_id_from_callback_kwargs, _request_stash, get_or_create_request_stash, get_request_stash, ) +from litellm.proxy.hooks.parallel_request_limiter_v3 import ( + PROXY_MaxParallelRequestsHandler_v3 as RateLimitHandler, +) from litellm.proxy.utils import InternalUsageCache, hash_token +from litellm.router_utils.ptu_shares import PTUTeamCeiling from litellm.types.llms.openai import ( InputTokensDetails, ResponseAPIUsage, @@ -3775,3 +3785,284 @@ def test_reservation_counts_a_compaction_item_by_its_summary(rate_limiter): ) == handler._estimate_input_and_output_tokens( data={"model": "gpt-4o", "input": _COMPACTED_INPUT[2:], "max_output_tokens": 1}, call_type="aresponses" ) + + +_QUOTA_TEXT: Final = "a" * 400 +_QUOTA_JSON_TEXT: Final = '{"type":"input_image","image_url":"data:image/png;base64,AAAA"}' + + +@pytest.mark.parametrize( + ("call_type", "request_data", "expected"), + [ + ("acompletion", {"messages": [{"role": "user", "content": _QUOTA_TEXT}], "max_tokens": 16}, (100, 16)), + ("acompletion", {"messages": [{"role": "user", "content": _QUOTA_JSON_TEXT}], "max_tokens": 16}, (15, 16)), + ( + "anthropic_messages", + { + "messages": [ + { + "role": "user", + "content": [ + { + "type": "tool_result", + "tool_use_id": "t", + "content": [{"type": "text", "text": _QUOTA_TEXT}], + } + ], + } + ], + "max_tokens": 16, + }, + (0, 16), + ), + ( + "acompletion", + { + "messages": [ + { + "role": "assistant", + "content": None, + "tool_calls": [ + {"id": "t", "type": "function", "function": {"name": "f", "arguments": _QUOTA_TEXT}} + ], + } + ], + "max_tokens": 16, + }, + (0, 16), + ), + ("aresponses", {"input": _QUOTA_TEXT, "max_output_tokens": 16}, (100, 16)), + ("aresponses", {"input": _QUOTA_TEXT, "instructions": "b" * 400, "max_output_tokens": 16}, (100, 16)), + ( + "aresponses", + {"input": [{"role": "user", "content": _QUOTA_TEXT}], "instructions": "b" * 400, "max_output_tokens": 16}, + (107, 16), + ), + ( + "aresponses", + { + "input": [ + { + "type": "file_search_call", + "id": "f", + "status": "completed", + "queries": ["q"], + "results": [ + { + "file_id": "file_1", + "filename": "result.txt", + "attributes": {}, + "score": 1, + "text": _QUOTA_TEXT, + } + ], + } + ], + "max_output_tokens": 16, + }, + (146, 16), + ), + ( + "aresponses", + { + "input": [{"type": "custom_tool_call_output", "call_id": "t", "output": _QUOTA_JSON_TEXT}], + "max_output_tokens": 16, + }, + (32, 16), + ), + ("aresponses", {"input": "", "instructions": "b" * 400}, (0, 0)), + ("aembedding", {"input": [_QUOTA_TEXT]}, (100, 0)), + ("atext_completion", {"prompt": [_QUOTA_TEXT], "max_tokens": 16}, (100, 16)), + ], +) +def test_media_correction_preserves_exact_non_media_reservations( + rate_limiter: tuple[RateLimitHandler, DualCache], + call_type: str, + request_data: dict[str, object], + expected: tuple[int, int], +) -> None: + handler, _ = rate_limiter + assert handler._estimate_input_and_output_tokens(request_data, call_type=call_type) == expected + + +def _quota_media_payload(kind: str, compressed: bool) -> str: + output: Final = io.BytesIO() + if kind == "image": + Image.new("RGB", (1024, 1024)).save(output, format="PNG", compress_level=9 if compressed else 0) + else: + sample_rate: Final = 8000 if compressed else 48000 + with sf.SoundFile(output, "w", samplerate=sample_rate, channels=1, format="WAV") as audio: + audio.buffer_write(b"\0" * (sample_rate * 2), dtype="int16") + return base64.b64encode(output.getvalue()).decode() + + +def _quota_media_request(kind: str, payload: str, call_type: str) -> dict[str, object]: + block: Final = ( + {"type": "input_audio", "input_audio": {"data": payload, "format": "wav"}} + if kind == "audio" + else {"type": "input_image", "image_url": "data:image/png;base64," + payload} + if call_type == "aresponses" + else {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": payload}} + if call_type == "anthropic_messages" + else {"type": "image_url", "image_url": {"url": "data:image/png;base64," + payload}} + ) + content: Final = ( + [{"type": "tool_result", "tool_use_id": "t", "content": [block]}] + if call_type == "anthropic_messages" + else [block] + ) + field: Final = "input" if call_type == "aresponses" else "messages" + return { + "model": "gpt-6-astra", + field: [{"role": "user", "content": content}], + "max_tokens": 16, + "max_output_tokens": 16, + } + + +@pytest.mark.parametrize( + ("kind", "call_type"), + [ + ("image", "acompletion"), + ("image", "aresponses"), + ("image", "anthropic_messages"), + ("audio", "acompletion"), + ("audio", "aresponses"), + ], +) +def test_media_reservation_is_independent_of_encoding_and_preserves_request( + rate_limiter: tuple[RateLimitHandler, DualCache], kind: str, call_type: str +) -> None: + handler, _ = rate_limiter + compressed: Final = _quota_media_request(kind, _quota_media_payload(kind, True), call_type) + uncompressed: Final = _quota_media_request(kind, _quota_media_payload(kind, False), call_type) + original: Final = copy.deepcopy(uncompressed) + estimate: Final = handler._estimate_input_and_output_tokens(uncompressed, call_type=call_type) + assert estimate == handler._estimate_input_and_output_tokens(compressed, call_type=call_type) + assert estimate[0] > 0 and estimate[1] == 16 + assert uncompressed == original + + +@pytest.mark.asyncio +@pytest.mark.parametrize("call_type", ["acompletion", "aresponses"]) +@pytest.mark.parametrize("limit", [1000, 20000]) +async def test_media_pre_call_rejects_tight_quota_and_reserves_sufficient_quota( + rate_limiter: tuple[RateLimitHandler, DualCache], call_type: str, limit: int +) -> None: + handler, cache = rate_limiter + request: Final = { + key: value + for key, value in _quota_media_request("image", _quota_media_payload("image", True), call_type).items() + if key not in ("max_tokens", "max_output_tokens") + } + auth: Final = UserAPIKeyAuth(api_key=hash_token("sk-media-quota"), tpm_limit=limit) + if limit == 1000: + with pytest.raises(HTTPException) as error: + await handler.async_pre_call_hook(user_api_key_dict=auth, cache=cache, data=request, call_type=call_type) + assert error.value.status_code == 429 + return + await handler.async_pre_call_hook(user_api_key_dict=auth, cache=cache, data=request, call_type=call_type) + stash: Final = get_request_stash() + assert stash is not None + expected: Final = handler._estimate_input_and_output_tokens( + request, call_type=call_type, min_configured_tpm_limit=limit + ) + assert stash.reserved_tokens == sum(expected) + assert expected[0] > 1000 and expected[1] >= expected[0] + + +@pytest.mark.asyncio +async def test_ptu_pre_call_weights_media_input_and_output() -> None: + cache: Final = DualCache() + ceiling: Final = PTUTeamCeiling( + model_group="gpt-6-astra", tpm_limit=10000, output_to_input_ratio=4, cached_input_ratio=0 + ) + + def resolve_ceiling(team: str, model: str) -> PTUTeamCeiling | None: + return ceiling if (team, model) == ("test-team", "gpt-6-astra") else None + + handler: Final = RateLimitHandler( + internal_usage_cache=InternalUsageCache(cache), ptu_team_ceiling_resolver=resolve_ceiling + ) + request: Final = _quota_media_request("image", _quota_media_payload("image", True), "acompletion") + await handler.async_pre_call_hook( + user_api_key_dict=UserAPIKeyAuth(api_key=hash_token("sk-media-ptu"), team_id="test-team"), + cache=cache, + data=request, + call_type="acompletion", + ) + stash: Final = get_request_stash() + assert stash is not None + assert ( + stash.ptu_reserved_tokens + == estimate_image_tokens("gpt-6-astra", "data:image/png;base64," + _quota_media_payload("image", True)) + 4 * 16 + ) + + +@pytest.mark.asyncio +async def test_project_fallback_keeps_legacy_audio_estimate_and_output_policy( + rate_limiter: tuple[RateLimitHandler, DualCache], +) -> None: + handler, cache = rate_limiter + unsupported: Final = {"type": "unsupported_content"} + with pytest.raises(UnsupportedContentTypeError): + token_counter(model="gpt-6-astra", messages=[{"role": "user", "content": [unsupported]}]) + request: Final = { + "model": "gpt-6-astra", + "messages": [ + { + "role": "user", + "content": [ + unsupported, + { + "type": "input_audio", + "input_audio": {"data": _quota_media_payload("audio", True), "format": "wav"}, + }, + ], + } + ], + } + auth: Final = UserAPIKeyAuth( + api_key=hash_token("sk-media-project"), + project_id="media-project", + project_metadata={"model_itpm_limit": {"gpt-6-astra": 1000}, "model_otpm_limit": {"gpt-6-astra": 1000}}, + ) + await handler.async_pre_call_hook(user_api_key_dict=auth, cache=cache, data=request, call_type="acompletion") + stash: Final = get_request_stash() + assert stash is not None + assert stash.itpm_reserved_tokens == DEFAULT_AUDIO_TOKEN_ESTIMATE + assert stash.otpm_reserved_tokens == 1 + + +@pytest.mark.asyncio +async def test_single_anthropic_image_exceeds_tight_quota_with_explicit_output_cap( + rate_limiter: tuple[RateLimitHandler, DualCache], +) -> None: + handler, cache = rate_limiter + request: Final = { + "model": "anthropic/claude-haiku-5-5", + "messages": [ + { + "role": "user", + "content": [ + { + "type": "image", + "source": { + "type": "base64", + "media_type": "image/png", + "data": _quota_media_payload("image", True), + }, + } + ], + } + ], + "max_tokens": 64, + } + with pytest.raises(HTTPException) as error: + await handler.async_pre_call_hook( + user_api_key_dict=UserAPIKeyAuth(api_key=hash_token("sk-claude-media-quota"), tpm_limit=500), + cache=cache, + data=request, + call_type="anthropic_messages", + ) + assert error.value.status_code == 429 diff --git a/tests/unit/test_router/test_router.py b/tests/unit/test_router/test_router.py index 63ee1c43edb..566024b5d94 100644 --- a/tests/unit/test_router/test_router.py +++ b/tests/unit/test_router/test_router.py @@ -20630,6 +20630,37 @@ def test_resolved_litellm_models_answers_through_every_channel_a_request_uses( assert set(_resolution_router().resolved_litellm_models(model_name)) == set(expected) +def test_resolved_litellm_models_base_model_precedence_is_opt_in() -> None: + router: Final = Router( + model_list=[ + { + "model_name": "pooled", + "litellm_params": {"model": "openai/deployment-a", "api_key": "sk-test", "base_model": "gpt-5.5"}, + "model_info": {"base_model": "gpt-6-astra"}, + }, + { + "model_name": "pooled", + "litellm_params": {"model": "openai/deployment-b", "api_key": "sk-test", "base_model": "gpt-5.6-sol"}, + }, + { + "model_name": "team-deployment", + "litellm_params": {"model": "openai/deployment-c", "api_key": "sk-test"}, + "model_info": { + "base_model": "gpt-6-astra", + "team_id": "test-team", + "team_public_model_name": "team-model", + }, + }, + ], + model_group_alias={"nickname": "pooled"}, + ) + assert router.resolved_litellm_models("nickname") == ("openai/deployment-a", "openai/deployment-b") + assert router.resolved_litellm_models("nickname", use_base_model=True) == ("gpt-6-astra", "gpt-5.6-sol") + assert router.resolved_litellm_models("team-model", "test-team") == ("openai/deployment-c",) + assert router.resolved_litellm_models("team-model", "test-team", use_base_model=True) == ("gpt-6-astra",) + assert router.resolved_litellm_models("never-configured", use_base_model=True) == () + + class TestTierParamsTheTargetAccepts: """A tier's litellm_params are applied to every request that tier routes, so one the target cannot take raised UnsupportedParamsError before the request left the proxy, turning the whole From 7c7304c22b4a855242a101303a33215ca7f1c466 Mon Sep 17 00:00:00 2001 From: Amaan Date: Thu, 29 Oct 2026 01:10:32 +0530 Subject: [PATCH 3/3] feat(providers): add ZeroGPU as an OpenAI-compatible provider (#43874) * feat(providers): add ZeroGPU as an OpenAI-compatible provider Registers zerogpu through the JSON provider registry so zerogpu/ routes to https://api.zerogpu.ai/v1/chat/completions with ZEROGPU_API_KEY and an optional ZEROGPU_API_BASE. Adds LlmProviders.ZEROGPU, the constants.py provider and endpoint lists, cost map entries for the 10 ZeroGPU text generation models, the endpoint matrix, the Add Model form fields, the README row and the dashboard provider entry * feat(zerogpu): add capability flags verified against the live API Live calls to https://api.zerogpu.ai/v1/chat/completions confirmed tools, tool_choice and json_schema response_format on 8 of the 10 catalog text models, cached token reporting on deepseek-v4.1-flash, gpt-4.1-mini, gpt-5.4-nano and gpt-5.6-luna, and reasoning_effort on gpt-5.6-luna. LFM2.5-1.2B-Instruct answers stream requests with plain JSON instead of SSE and LFM2.5-1.2B-Thinking returns 500 on every request, so both stay out of the cost map until the API serves them correctly * feat(zerogpu): price LFM2.5-1.2B-Instruct Adds zerogpu/LFM2.5-1.2B-Instruct to the cost map with the catalog prices and context window. It gets no capability flags because the live API ignores tools, response_format and max_completion_tokens for this model, so LiteLLM drops tool params for it instead of sending them. LFM2.5-1.2B-Thinking stays out of the cost map * fix(zerogpu): reject transcription requests before calling the upstream API ZeroGPU only serves chat completions, so drop it from the OpenAI transcription transport set. A zerogpu transcription request now fails locally instead of hitting an /audio/transcriptions endpoint that does not exist * test(zerogpu): cover the responses and messages bridges Both endpoints reach ZeroGPU through chat completions, so assert the request each one sends upstream and the reply it returns * refactor(zerogpu): move zerogpu out of the provider list tails New upstream providers append to the end of openai_compatible_endpoints, openai_compatible_providers and providers.json, so an entry at the tail conflicts with every one of them. Placing zerogpu next to long-stable entries keeps those merges clean. No behavior change * fix(zerogpu): price cached input for qwen3-30b-a3b-fp8 The ZeroGPU model catalog (https://docs.zerogpu.ai/docs/model-catalog, checked 2026-10-09) lists cached input for qwen3-30b-a3b-fp8 at $0.05 per 1M tokens. Without cache_read_input_token_cost, cached tokens were billed at the full $0.10 input rate * revert(zerogpu): leave the audio transcription provider set unchanged Per review, keep OPENAI_AUDIO_TRANSCRIPTION_PROVIDERS as upstream defines it. This PR only covers chat completions, and ZeroGPU now serves /audio/transcriptions anyway, so the local refusal and its test are removed * fix(zerogpu): add max_tokens so trim_messages can size ZeroGPU history trim_messages reads max_tokens as the fallback for max_input_tokens and evaluates it eagerly, so without it the lookup raised, got swallowed, and the full history came back untrimmed. max_tokens matches each model's maxTokens in ZeroGPU's models API (https://api-dashboard.zerogpu.ai/api/models, checked 2026-10-10) * test(zerogpu): record mocked requests on a recorder object The mock handler appended to a list the caller passed in, which AGENTS.md forbids. A recorder now owns the captured requests as a tuple * fix(zerogpu): cap max_tokens at each model's output limit Each cap is the max_output_tokens ZeroGPU stores for the model in its production models table (read 2026-10-20), the same column its /v1/models serves as max_length. That gives 65536 for deepseek-v4.1-flash, 32768 for gpt-4.1-mini and gpt-oss-120b, 128000 for gpt-5.4-nano and gpt-5.6-luna, 8192 for LFM2.5-1.2B-Instruct and llama-3.1-8b-instruct-fast, and 16384 for qwen3-30b-a3b-fp8. glm-5.3-flash is stored as 32786, which reads as a typo for 32768, so that row uses 32768 trim_messages sizes history from max_input_tokens, so lowering max_tokens leaves it unaffected * fix(zerogpu): flag prompt caching on glm-5.3-flash Repeating the same ~4.3K-token prompt against ZeroGPU on 2026-10-20 returned cached_tokens of 4320 on some glm-5.3-flash calls. gpt-oss-120b and llama-3.1-8b-instruct-fast returned 0 on all nine calls, and qwen3-30b-a3b-fp8 never returned cached_tokens, so they stay unflagged. Their cached prices stay as ZeroGPU publishes them, so any cache hit is still billed at the right rate * fix(zerogpu): set gpt-5.4-nano context window to 272000 ZeroGPU accepts a 271001-token gpt-5.4-nano prompt and rejects 273000 tokens as over the context window (checked 2026-10-21), so max_input_tokens is 272000 rather than the 400000 its catalog lists * fix(zerogpu): list zerogpu models for zerogpu/* wildcards zerogpu had no models_by_provider entry, so get_provider_models("zerogpu") returned None and a zerogpu/* wildcard listed no models. zerogpu_models is now filled from the cost map rows like openinfer_models and registered in models_by_provider * fix(zerogpu): flag gpt-5.4-nano as a reasoning model ZeroGPU honors reasoning_effort on gpt-5.4-nano (checked 2026-10-21). Without it the model spends 0 reasoning tokens, and with low or high it spends 32 or more. LiteLLM only treated it as a reasoning model by falling back to OpenAI's bare gpt-5.4-nano row, so get_model_info and /model/info reported supports_reasoning as null. The row now carries the flag itself * fix(zerogpu): flag deepseek, glm, gpt-oss and qwen3 as reasoning models ZeroGPU's /v1/models marks these four as reasoning models and accepts reasoning_effort on all of them (checked 2026-10-21), but LiteLLM rejected the param with UnsupportedParamsError because their rows had no supports_reasoning flag. glm reports reasoning_tokens under reasoning_effort, and gpt-oss and qwen3 spend hidden reasoning tokens on every reply. The gpt-oss test case now expects reasoning_effort to reach ZeroGPU * fix(zerogpu): set gpt-oss-120b context window to 128000 ZeroGPU routes gpt-oss-120b across several upstream backends, and at least one of them returns 413 past a 128000-token window, so max_input_tokens uses that as the safe floor. The backends reached on 2026-10-21 accepted up to about 131K prompt tokens, so this only gives up a little headroom on the larger ones * feat(zerogpu): call ZeroGPU's Responses API natively ZeroGPU serves /v1/responses for all nine catalog models, so LiteLLM no longer bridges Responses requests through chat completions. Native calls return ZeroGPU's reasoning output items for glm, gpt-oss and qwen3, which the bridge dropped. Checked live on 2026-10-22 through litellm.responses for every model, non-streaming and streaming, plus a gpt-oss tool call. LFM2.5-1.2B-Instruct streams correctly now that ZeroGPU's text_generation lane emits Responses SSE * fix(zerogpu): stop inferring the zerogpu provider from api_base alone Listing https://api.zerogpu.ai/v1 in openai_compatible_endpoints rerouted bare model names on that api_base from openai to zerogpu. A bare gpt-5.6-luna call with verbosity used to return 200 and started raising UnsupportedParamsError. Drop the entry and its api_base-alone test, so zerogpu is only selected through the zerogpu/ prefix * fix(zerogpu): price gpt-6-luna, llama-3.2-3b, LFM2.5-1.2B-Thinking and qwen3.6 A zerogpu/* wildcard served these four chat models with no cost-map row, so their calls logged $0 spend. Add rows at ZeroGPU's catalog prices from https://api-dashboard.zerogpu.ai/api/models, with context windows and output caps from ZeroGPU's models table. LFM2.5-1.2B-Thinking has no cached price, like LFM2.5-1.2B-Instruct. gpt-6-luna is flagged supports_reasoning because LiteLLM already routes gpt-6 names as reasoning models. Checked live on 2026-10-28, all four now report nonzero response_cost * fix(zerogpu): drop the retired max_tokens field from the zerogpu rows Upstream #48471 retired max_tokens from the cost map in favor of max_input_tokens and max_output_tokens, and test_no_row_carries_the_retired_max_tokens_field failed on all 13 zerogpu rows. Every row already carries the same value in max_output_tokens, so nothing is lost. The zerogpu limits test now checks that each row's output limit fits its context window * chore(zerogpu): regenerate provider endpoint support for zerogpu Upstream #48619 generates provider_endpoints_support.json from the cost map and checks it in CI, and the hand-written zerogpu entry no longer matched. Regenerated with scripts/generate_provider_endpoints_support.py. Only the zerogpu entry changed, and it now matches what the generator emits for other JSON-registry providers such as openinfer * fix(zerogpu): flag tools, json_schema and prompt caching on gpt-6-luna ZeroGPU serves function calling, tool_choice, json_schema response_format and prompt caching on gpt-6-luna, so set supports_function_calling, supports_tool_choice, supports_response_schema and supports_prompt_caching on its row in both cost maps. Checked live on 2026-10-28, a weather tool call came back with the right arguments, json_schema output validated, and repeated prompts reported 3260 of 3263 tokens cached * fix(zerogpu): set qwen3.6-35b-a3b context to 262144 and flag its tools and reasoning ZeroGPU accepts up to 262,144 input tokens on qwen3.6-35b-a3b and serves tools, tool_choice, json_schema response_format and reasoning_effort on it. Raise max_input_tokens from 16384 and set supports_function_calling, supports_tool_choice, supports_response_schema and supports_reasoning in both cost maps. Before this, LiteLLM rejected tools and reasoning_effort for the model. Checked live on 2026-10-28, a 209,905-token prompt succeeded, an oversized one was refused with a 262144-token limit, and tool calls, json_schema output and reasoning_effort all worked * fix(zerogpu): flag vision on gpt-6-luna and qwen3.6-35b-a3b ZeroGPU reads image inputs on gpt-6-luna and qwen3.6-35b-a3b, so set supports_vision on both rows in both cost maps. Checked live on 2026-10-28, both models named the animal in an image_url photo, and the image showed up in their prompt token counts --- README.md | 1 + .../src/exception_mapping_utils/original.rs | 1 + litellm/__init__.py | 5 + litellm/constants.py | 1 + litellm/llms/openai_like/providers.json | 6 + ...odel_prices_and_context_window_backup.json | 174 +++++++++++++ .../provider_endpoints_support_backup.json | 26 ++ .../provider_create_fields.json | 28 ++ litellm/types/utils.py | 1 + model_prices_and_context_window.json | 174 +++++++++++++ provider_endpoints_support.json | 26 ++ .../llms/openai_like/test_zerogpu_provider.py | 239 ++++++++++++++++++ .../public_endpoints/test_public_endpoints.py | 20 ++ .../test_model_prices_and_context_window.py | 17 ++ .../components/provider_info_helpers.test.tsx | 5 + .../src/components/provider_info_helpers.tsx | 3 + 16 files changed, 727 insertions(+) create mode 100644 tests/unit/llms/openai_like/test_zerogpu_provider.py diff --git a/README.md b/README.md index 66b6acb7a71..a7311882bd0 100644 --- a/README.md +++ b/README.md @@ -427,6 +427,7 @@ Set `LITELLM_PROXY_API_BASE` and `LITELLM_PROXY_API_KEY` and every model call th | [Watsonx Text (`watsonx_text`)](https://docs.litellm.ai/docs/providers/watsonx) | ✅ | ✅ | ✅ | | | | | | | | | [xAI (`xai`)](https://docs.litellm.ai/docs/providers/xai) | ✅ | ✅ | ✅ | | | | | | | | | [Xinference (`xinference`)](https://docs.litellm.ai/docs/providers/xinference) | | | | ✅ | | | | | | | +| [ZeroGPU (`zerogpu`)](https://docs.litellm.ai/docs/providers/zerogpu) | ✅ | ✅ | ✅ | | | | | | | | [**Read the Docs**](https://docs.litellm.ai/docs/) diff --git a/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs b/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs index 6517b093bbd..615d530af19 100644 --- a/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs +++ b/litellm-rust/crates/core-utils/src/exception_mapping_utils/original.rs @@ -113,6 +113,7 @@ const OPENAI_COMPATIBLE_PROVIDERS: &[&str] = &[ "meta", "cognition", "scx-ai", + "zerogpu", ]; impl ExceptionFamily { diff --git a/litellm/__init__.py b/litellm/__init__.py index fc8f6b57254..8dfe5adca1c 100644 --- a/litellm/__init__.py +++ b/litellm/__init__.py @@ -681,6 +681,7 @@ moonshot_models: Set = set() publicai_models: Set = set() darkbloom_models: Set = set() openinfer_models: Final[Set[str]] = set() +zerogpu_models: Final[Set[str]] = set() v0_models: Set = set() morph_models: Set = set() opencode_go_models: Final[set[str]] = set() # mutable-ok: populated when the shared model catalog loads @@ -969,6 +970,8 @@ def _populate_provider_model_sets(model_cost_map: Mapping[str, Mapping[str, obje darkbloom_models.add(key) elif value.get("litellm_provider") == "openinfer": openinfer_models.add(key) + elif value.get("litellm_provider") == "zerogpu": + zerogpu_models.add(key) elif value.get("litellm_provider") == "v0": v0_models.add(key) elif value.get("litellm_provider") == "morph": @@ -1142,6 +1145,7 @@ model_list = list( | publicai_models | darkbloom_models | openinfer_models + | zerogpu_models | v0_models | morph_models | opencode_go_models @@ -1261,6 +1265,7 @@ def _build_models_by_provider() -> dict: "publicai": publicai_models, "darkbloom": darkbloom_models, "openinfer": openinfer_models, + "zerogpu": zerogpu_models, "v0": v0_models, "morph": morph_models, "opencode_go": opencode_go_models, diff --git a/litellm/constants.py b/litellm/constants.py index 2a20e812e68..2b083b03766 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -1096,6 +1096,7 @@ openai_compatible_providers: Final[list[str]] = [ "llamafile", "lm_studio", "galadriel", + "zerogpu", "github_copilot", # GitHub Copilot Chat API "chatgpt", # ChatGPT subscription API "novita", diff --git a/litellm/llms/openai_like/providers.json b/litellm/llms/openai_like/providers.json index 0034ed2b9d4..2b71707239b 100644 --- a/litellm/llms/openai_like/providers.json +++ b/litellm/llms/openai_like/providers.json @@ -58,6 +58,12 @@ "max_completion_tokens": "max_tokens" } }, + "zerogpu": { + "base_url": "https://api.zerogpu.ai/v1", + "api_key_env": "ZEROGPU_API_KEY", + "api_base_env": "ZEROGPU_API_BASE", + "supported_endpoints": ["/v1/chat/completions", "/v1/responses"] + }, "chutes": { "base_url": "https://llm.chutes.ai/v1/", "api_key_env": "CHUTES_API_KEY", diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 825de5ac3ee..4e022e1a4b2 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -66163,6 +66163,180 @@ "supports_reasoning": true, "supports_tool_choice": true }, + "zerogpu/deepseek-v4.1-flash": { + "cache_read_input_token_cost": 3e-09, + "input_cost_per_token": 3e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 1048576, + "max_output_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/glm-5.3-flash": { + "cache_read_input_token_cost": 1.5e-08, + "input_cost_per_token": 1e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 1048576, + "max_output_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 3.5e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/gpt-4.1-mini": { + "cache_read_input_token_cost": 1e-07, + "input_cost_per_token": 4e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 1047576, + "max_output_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 1.6e-06, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/gpt-5.4-nano": { + "cache_read_input_token_cost": 2e-08, + "input_cost_per_token": 2e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 1.25e-06, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/gpt-5.6-luna": { + "cache_read_input_token_cost": 2e-07, + "input_cost_per_token": 2e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/gpt-6-luna": { + "cache_read_input_token_cost": 1e-08, + "input_cost_per_token": 1e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 5e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "zerogpu/gpt-oss-120b": { + "cache_read_input_token_cost": 3e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 128000, + "max_output_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 6e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/LFM2.5-1.2B-Instruct": { + "input_cost_per_token": 2e-08, + "litellm_provider": "zerogpu", + "max_input_tokens": 32768, + "max_output_tokens": 8192, + "mode": "chat", + "output_cost_per_token": 5e-08, + "source": "https://docs.zerogpu.ai/docs/model-catalog" + }, + "zerogpu/LFM2.5-1.2B-Thinking": { + "input_cost_per_token": 2e-08, + "litellm_provider": "zerogpu", + "max_input_tokens": 32768, + "max_output_tokens": 8192, + "mode": "chat", + "output_cost_per_token": 5e-08, + "source": "https://docs.zerogpu.ai/docs/model-catalog" + }, + "zerogpu/llama-3.1-8b-instruct-fast": { + "cache_read_input_token_cost": 2.5e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 131072, + "max_output_tokens": 8192, + "mode": "chat", + "output_cost_per_token": 2.8e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/llama-3.2-3b-instruct-fast": { + "cache_read_input_token_cost": 2.5e-08, + "input_cost_per_token": 5e-08, + "litellm_provider": "zerogpu", + "max_input_tokens": 131072, + "max_output_tokens": 8192, + "mode": "chat", + "output_cost_per_token": 3e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog" + }, + "zerogpu/qwen3-30b-a3b-fp8": { + "cache_read_input_token_cost": 5e-08, + "input_cost_per_token": 1e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 32768, + "max_output_tokens": 16384, + "mode": "chat", + "output_cost_per_token": 4.5e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/qwen3.6-35b-a3b": { + "cache_read_input_token_cost": 2e-07, + "input_cost_per_token": 2e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 262144, + "max_output_tokens": 16384, + "mode": "chat", + "output_cost_per_token": 1.5e-06, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, "together_ai/Qwen/Qwen3.8-Flash": { "input_cost_per_token": 1.5e-07, "litellm_provider": "together_ai", diff --git a/litellm/provider_endpoints_support_backup.json b/litellm/provider_endpoints_support_backup.json index 1ab03fb3fdc..90395421109 100644 --- a/litellm/provider_endpoints_support_backup.json +++ b/litellm/provider_endpoints_support_backup.json @@ -4437,6 +4437,32 @@ "generateContent": true } }, + "zerogpu": { + "display_name": "ZeroGPU (`zerogpu`)", + "url": "https://docs.litellm.ai/docs/providers/zerogpu", + "endpoints": { + "chat_completions": true, + "responses": true, + "text_completion": true, + "embeddings": false, + "image_generations": false, + "image_edits": false, + "audio_transcriptions": false, + "audio_speech": false, + "rerank": false, + "ocr": false, + "search": false, + "moderations": false, + "realtime": false, + "video_generations": false, + "batches": false, + "messages": true, + "interactions": true, + "systemone": false, + "a2a": true, + "generateContent": true + } + }, "ragflow": { "display_name": "RAGFlow (`ragflow`)", "url": "https://docs.litellm.ai/docs/providers/ragflow", diff --git a/litellm/proxy/public_endpoints/provider_create_fields.json b/litellm/proxy/public_endpoints/provider_create_fields.json index 0dd96c856f8..f2d88702470 100644 --- a/litellm/proxy/public_endpoints/provider_create_fields.json +++ b/litellm/proxy/public_endpoints/provider_create_fields.json @@ -4145,6 +4145,34 @@ ], "default_model_placeholder": "gpt-3.5-turbo" }, + { + "provider": "ZeroGPU", + "provider_display_name": "ZeroGPU", + "litellm_provider": "zerogpu", + "credential_fields": [ + { + "key": "api_base", + "label": "API Base", + "placeholder": "https://api.zerogpu.ai/v1", + "tooltip": null, + "required": false, + "field_type": "text", + "options": null, + "default_value": null + }, + { + "key": "api_key", + "label": "API Key", + "placeholder": null, + "tooltip": null, + "required": true, + "field_type": "password", + "options": null, + "default_value": null + } + ], + "default_model_placeholder": "zerogpu/gpt-oss-120b" + }, { "provider": "CURSOR", "provider_display_name": "Cursor", diff --git a/litellm/types/utils.py b/litellm/types/utils.py index 4ce35bf86c1..b21bf0e14d4 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -4588,6 +4588,7 @@ class LlmProviders(str, Enum): PRISM = "prism" OPENINFER = "openinfer" INFERENCEAPIS = "inferenceapis" + ZEROGPU = "zerogpu" DARKBLOOM = "darkbloom" META = "meta" SAIL = "sail" diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 825de5ac3ee..4e022e1a4b2 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -66163,6 +66163,180 @@ "supports_reasoning": true, "supports_tool_choice": true }, + "zerogpu/deepseek-v4.1-flash": { + "cache_read_input_token_cost": 3e-09, + "input_cost_per_token": 3e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 1048576, + "max_output_tokens": 65536, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/glm-5.3-flash": { + "cache_read_input_token_cost": 1.5e-08, + "input_cost_per_token": 1e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 1048576, + "max_output_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 3.5e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/gpt-4.1-mini": { + "cache_read_input_token_cost": 1e-07, + "input_cost_per_token": 4e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 1047576, + "max_output_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 1.6e-06, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/gpt-5.4-nano": { + "cache_read_input_token_cost": 2e-08, + "input_cost_per_token": 2e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 1.25e-06, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/gpt-5.6-luna": { + "cache_read_input_token_cost": 2e-07, + "input_cost_per_token": 2e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 1.2e-06, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/gpt-6-luna": { + "cache_read_input_token_cost": 1e-08, + "input_cost_per_token": 1e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 272000, + "max_output_tokens": 128000, + "mode": "chat", + "output_cost_per_token": 5e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_prompt_caching": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, + "zerogpu/gpt-oss-120b": { + "cache_read_input_token_cost": 3e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 128000, + "max_output_tokens": 32768, + "mode": "chat", + "output_cost_per_token": 6e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/LFM2.5-1.2B-Instruct": { + "input_cost_per_token": 2e-08, + "litellm_provider": "zerogpu", + "max_input_tokens": 32768, + "max_output_tokens": 8192, + "mode": "chat", + "output_cost_per_token": 5e-08, + "source": "https://docs.zerogpu.ai/docs/model-catalog" + }, + "zerogpu/LFM2.5-1.2B-Thinking": { + "input_cost_per_token": 2e-08, + "litellm_provider": "zerogpu", + "max_input_tokens": 32768, + "max_output_tokens": 8192, + "mode": "chat", + "output_cost_per_token": 5e-08, + "source": "https://docs.zerogpu.ai/docs/model-catalog" + }, + "zerogpu/llama-3.1-8b-instruct-fast": { + "cache_read_input_token_cost": 2.5e-08, + "input_cost_per_token": 1.5e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 131072, + "max_output_tokens": 8192, + "mode": "chat", + "output_cost_per_token": 2.8e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/llama-3.2-3b-instruct-fast": { + "cache_read_input_token_cost": 2.5e-08, + "input_cost_per_token": 5e-08, + "litellm_provider": "zerogpu", + "max_input_tokens": 131072, + "max_output_tokens": 8192, + "mode": "chat", + "output_cost_per_token": 3e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog" + }, + "zerogpu/qwen3-30b-a3b-fp8": { + "cache_read_input_token_cost": 5e-08, + "input_cost_per_token": 1e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 32768, + "max_output_tokens": 16384, + "mode": "chat", + "output_cost_per_token": 4.5e-07, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true + }, + "zerogpu/qwen3.6-35b-a3b": { + "cache_read_input_token_cost": 2e-07, + "input_cost_per_token": 2e-07, + "litellm_provider": "zerogpu", + "max_input_tokens": 262144, + "max_output_tokens": 16384, + "mode": "chat", + "output_cost_per_token": 1.5e-06, + "source": "https://docs.zerogpu.ai/docs/model-catalog", + "supports_function_calling": true, + "supports_reasoning": true, + "supports_response_schema": true, + "supports_tool_choice": true, + "supports_vision": true + }, "together_ai/Qwen/Qwen3.8-Flash": { "input_cost_per_token": 1.5e-07, "litellm_provider": "together_ai", diff --git a/provider_endpoints_support.json b/provider_endpoints_support.json index 1ab03fb3fdc..90395421109 100644 --- a/provider_endpoints_support.json +++ b/provider_endpoints_support.json @@ -4437,6 +4437,32 @@ "generateContent": true } }, + "zerogpu": { + "display_name": "ZeroGPU (`zerogpu`)", + "url": "https://docs.litellm.ai/docs/providers/zerogpu", + "endpoints": { + "chat_completions": true, + "responses": true, + "text_completion": true, + "embeddings": false, + "image_generations": false, + "image_edits": false, + "audio_transcriptions": false, + "audio_speech": false, + "rerank": false, + "ocr": false, + "search": false, + "moderations": false, + "realtime": false, + "video_generations": false, + "batches": false, + "messages": true, + "interactions": true, + "systemone": false, + "a2a": true, + "generateContent": true + } + }, "ragflow": { "display_name": "RAGFlow (`ragflow`)", "url": "https://docs.litellm.ai/docs/providers/ragflow", diff --git a/tests/unit/llms/openai_like/test_zerogpu_provider.py b/tests/unit/llms/openai_like/test_zerogpu_provider.py new file mode 100644 index 00000000000..1021237c335 --- /dev/null +++ b/tests/unit/llms/openai_like/test_zerogpu_provider.py @@ -0,0 +1,239 @@ +import json +from collections.abc import Mapping +from typing import Final + +import httpx +import pytest +import respx +from openai import OpenAI + +import litellm +from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider +from litellm.proxy.auth.model_checks import get_provider_models + +ZEROGPU_API_BASE: Final = "https://api.zerogpu.ai/v1" + +ZEROGPU_MODELS: Final = ( + "deepseek-v4.1-flash", + "glm-5.3-flash", + "gpt-4.1-mini", + "gpt-5.4-nano", + "gpt-5.6-luna", + "gpt-6-luna", + "gpt-oss-120b", + "LFM2.5-1.2B-Instruct", + "LFM2.5-1.2B-Thinking", + "llama-3.1-8b-instruct-fast", + "llama-3.2-3b-instruct-fast", + "qwen3-30b-a3b-fp8", + "qwen3.6-35b-a3b", +) + + +WEATHER_TOOL: Final = { + "type": "function", + "function": { + "name": "get_weather", + "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}, + }, +} + + +class _RecordingZeroGPU: + def __init__(self, usage: Mapping[str, object]) -> None: + self.usage: Final = usage + self.requests: tuple[httpx.Request, ...] = () + + def respond(self, request: httpx.Request) -> httpx.Response: + self.requests = (*self.requests, request) + return httpx.Response( + 200, + json={ + "id": "chatcmpl-zerogpu", + "object": "chat.completion", + "created": 1234567890, + "model": json.loads(request.content)["model"], + "choices": [{"index": 0, "message": {"role": "assistant", "content": "hi"}, "finish_reason": "stop"}], + "usage": dict(self.usage), + }, + ) + + def client(self) -> OpenAI: + return OpenAI( + api_key="zerogpu-test-key", + base_url=ZEROGPU_API_BASE, + http_client=httpx.Client(transport=httpx.MockTransport(self.respond)), + ) + + +def test_prefixed_model_reaches_zerogpu_chat_completions_without_the_prefix(): + zerogpu: Final = _RecordingZeroGPU({"prompt_tokens": 3, "completion_tokens": 1, "total_tokens": 4}) + + with zerogpu.client() as client: + litellm.completion( + model="zerogpu/gpt-oss-120b", + messages=[{"role": "user", "content": "hello"}], + client=client, + ) + + assert len(zerogpu.requests) == 1 + assert str(zerogpu.requests[0].url) == f"{ZEROGPU_API_BASE}/chat/completions" + assert json.loads(zerogpu.requests[0].content) == { + "model": "gpt-oss-120b", + "messages": [{"role": "user", "content": "hello"}], + } + + +def test_responses_request_reaches_zerogpu_responses_natively(): + with respx.mock() as upstream: + route: Final = upstream.post(f"{ZEROGPU_API_BASE}/responses").respond( + 200, + json={ + "id": "resp_zerogpu", + "object": "response", + "created_at": 1_792_000_000, + "model": "gpt-oss-120b", + "status": "completed", + "output": [ + { + "id": "msg_zerogpu", + "type": "message", + "role": "assistant", + "status": "completed", + "content": [{"type": "output_text", "text": "hi", "annotations": []}], + } + ], + "usage": {"input_tokens": 3, "output_tokens": 1, "total_tokens": 4}, + }, + ) + response: Final = litellm.responses(model="zerogpu/gpt-oss-120b", input="hello", api_key="zerogpu-test-key") + + body: Final = json.loads(route.calls.last.request.content) + assert body["model"] == "gpt-oss-120b" + assert body["input"] == "hello" + assert response.output[0].content[0].text == "hi" + + +def test_anthropic_messages_request_is_bridged_to_zerogpu_chat_completions(): + zerogpu: Final = _RecordingZeroGPU({"prompt_tokens": 3, "completion_tokens": 1, "total_tokens": 4}) + + with zerogpu.client() as client: + response: Final = litellm.anthropic.messages.create( + model="zerogpu/gpt-oss-120b", + messages=[{"role": "user", "content": "hello"}], + max_tokens=10, + client=client, + ) + + assert [str(request.url) for request in zerogpu.requests] == [f"{ZEROGPU_API_BASE}/chat/completions"] + assert json.loads(zerogpu.requests[0].content) == { + "model": "gpt-oss-120b", + "messages": [{"role": "user", "content": "hello"}], + "max_tokens": 10, + } + assert response["content"] == [{"type": "text", "text": "hi"}] + + +@pytest.mark.parametrize("model", ZEROGPU_MODELS) +def test_zerogpu_completion_is_charged_at_the_cost_map_rates(model: str): + usage: Final = { + "prompt_tokens": 1000, + "completion_tokens": 200, + "total_tokens": 1200, + "prompt_tokens_details": {"cached_tokens": 400}, + } + zerogpu: Final = _RecordingZeroGPU(usage) + + with zerogpu.client() as client: + response: Final = litellm.completion( + model=f"zerogpu/{model}", + messages=[{"role": "user", "content": "hello"}], + client=client, + ) + + rates: Final = litellm.model_cost[f"zerogpu/{model}"] + cached_rate: Final = rates.get("cache_read_input_token_cost", rates["input_cost_per_token"]) + expected_cost: Final = ( + 600 * rates["input_cost_per_token"] + 400 * cached_rate + 200 * rates["output_cost_per_token"] + ) + assert response._hidden_params["response_cost"] == pytest.approx(expected_cost) + assert expected_cost > 0 + + +@pytest.mark.parametrize("model", ZEROGPU_MODELS) +def test_trim_messages_fits_zerogpu_history_into_the_model_context_window(model: str): + budget: Final = 20 + trim_ratio: Final = budget / litellm.model_cost[f"zerogpu/{model}"]["max_input_tokens"] + messages: Final = [{"role": "user", "content": "lorem ipsum " * 300}] + + trimmed: Final = litellm.utils.trim_messages(messages, model=f"zerogpu/{model}", trim_ratio=trim_ratio) + + trimmed_tokens: Final = litellm.token_counter(model=f"zerogpu/{model}", messages=trimmed) + assert trimmed_tokens <= budget, trimmed + + +@pytest.mark.parametrize( + ("model", "forwards_tools", "forwards_reasoning_effort"), + [ + ("gpt-5.6-luna", True, True), + ("gpt-oss-120b", True, True), + ("qwen3.6-35b-a3b", True, True), + ("LFM2.5-1.2B-Instruct", False, False), + ], +) +def test_tools_and_reasoning_effort_reach_zerogpu_only_for_models_that_support_them( + model: str, forwards_tools: bool, forwards_reasoning_effort: bool +): + zerogpu: Final = _RecordingZeroGPU({"prompt_tokens": 3, "completion_tokens": 1, "total_tokens": 4}) + + with zerogpu.client() as client: + litellm.completion( + model=f"zerogpu/{model}", + messages=[{"role": "user", "content": "weather in Paris?"}], + tools=[WEATHER_TOOL], + tool_choice="auto", + reasoning_effort="high", + drop_params=True, + client=client, + ) + + tools: Final = {"tools": [WEATHER_TOOL], "tool_choice": "auto"} if forwards_tools else {} + reasoning: Final = {"reasoning_effort": "high"} if forwards_reasoning_effort else {} + assert json.loads(zerogpu.requests[0].content) == { + "model": model, + "messages": [{"role": "user", "content": "weather in Paris?"}], + **tools, + **reasoning, + } + + +@pytest.mark.parametrize("model", ZEROGPU_MODELS) +def test_model_info_reports_the_reasoning_support_litellm_routes_with(model: str): + model_info: Final = litellm.get_model_info(f"zerogpu/{model}") + + assert bool(model_info.get("supports_reasoning")) == litellm.utils.supports_reasoning(f"zerogpu/{model}") + + +def test_zerogpu_key_and_base_come_from_zerogpu_env_vars(monkeypatch: pytest.MonkeyPatch): + monkeypatch.setenv("ZEROGPU_API_KEY", "zerogpu-env-key") + monkeypatch.delenv("ZEROGPU_API_BASE", raising=False) + + assert get_llm_provider(model="zerogpu/gpt-oss-120b") == ( + "gpt-oss-120b", + "zerogpu", + "zerogpu-env-key", + ZEROGPU_API_BASE, + ) + + monkeypatch.setenv("ZEROGPU_API_BASE", "https://zerogpu.internal.example/v1") + + assert get_llm_provider(model="zerogpu/gpt-oss-120b")[3] == "https://zerogpu.internal.example/v1" + + +def test_zerogpu_wildcard_lists_every_zerogpu_catalog_model(): + catalog_models: Final = sorted( + key for key, row in litellm.model_cost.items() if row.get("litellm_provider") == "zerogpu" + ) + + assert catalog_models + assert sorted(get_provider_models("zerogpu") or []) == catalog_models diff --git a/tests/unit/proxy/public_endpoints/test_public_endpoints.py b/tests/unit/proxy/public_endpoints/test_public_endpoints.py index f60c2254b41..d55ce25d878 100644 --- a/tests/unit/proxy/public_endpoints/test_public_endpoints.py +++ b/tests/unit/proxy/public_endpoints/test_public_endpoints.py @@ -363,6 +363,26 @@ def test_cognition_provider_fields(): assert fields_by_key["api_base"]["required"] is False +def test_zerogpu_add_model_form_defaults_resolve_to_a_priced_zerogpu_model(): + import litellm + from litellm.litellm_core_utils.get_llm_provider_logic import get_llm_provider + + app_instance: Final = FastAPI() + app_instance.include_router(router) + providers: Final = TestClient(app_instance).get("/public/providers/fields").json() + + zerogpu: Final = next(p for p in providers if p["litellm_provider"] == "zerogpu") + fields_by_key: Final = {f["key"]: f for f in zerogpu["credential_fields"]} + placeholder: Final = zerogpu["default_model_placeholder"] + _, provider, _, api_base = get_llm_provider(model=placeholder) + + assert provider == "zerogpu" + assert api_base == fields_by_key["api_base"]["placeholder"] + assert litellm.model_cost[placeholder]["litellm_provider"] == "zerogpu" + assert fields_by_key["api_key"]["required"] is True + assert fields_by_key["api_key"]["field_type"] == "password" + + def test_qwen_mainland_provider_fields_carry_the_qianwen_brand(): app_instance = FastAPI() app_instance.include_router(router) diff --git a/tests/unit/test_model_prices_and_context_window.py b/tests/unit/test_model_prices_and_context_window.py index bbfe304429d..2aceed493a7 100644 --- a/tests/unit/test_model_prices_and_context_window.py +++ b/tests/unit/test_model_prices_and_context_window.py @@ -115,6 +115,23 @@ def test_github_copilot_backup_rows_match_the_main_catalog() -> None: assert _github_copilot_catalog_rows(PRICES_PATH) == _github_copilot_catalog_rows(BACKUP_PRICES_PATH) +class _TokenLimits(BaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + max_input_tokens: int + max_output_tokens: int + + +@pytest.mark.parametrize("path", (PRICES_PATH, BACKUP_PRICES_PATH), ids=("main", "backup")) +def test_zerogpu_rows_have_an_output_limit_that_fits_the_context_window(path: Path) -> None: + prices: Final[dict[str, JsonValue]] = TypeAdapter(dict[str, JsonValue]).validate_json(path.read_bytes()) + rows: Final = MappingProxyType( + {key: _TokenLimits.model_validate(value) for key, value in prices.items() if key.startswith("zerogpu/")} + ) + assert rows, f"No ZeroGPU rows found in {path}" + assert tuple(key for key, row in rows.items() if row.max_output_tokens > row.max_input_tokens) == () + + def test_github_copilot_rows_resolve_through_get_model_info() -> None: rows: Final[Mapping[str, _CopilotEndpointRow]] = _copilot_endpoint_rows(PRICES_PATH) assert { diff --git a/ui/litellm-dashboard/src/components/provider_info_helpers.test.tsx b/ui/litellm-dashboard/src/components/provider_info_helpers.test.tsx index 0ebd2cdff17..83e80911cff 100644 --- a/ui/litellm-dashboard/src/components/provider_info_helpers.test.tsx +++ b/ui/litellm-dashboard/src/components/provider_info_helpers.test.tsx @@ -224,6 +224,7 @@ describe("provider_info_helpers", () => { Providers.Sail, Providers.WANDB, Providers.ZAI, + Providers.ZeroGPU, ]; const logolessProviders = Object.values(Providers).filter((provider) => !providerLogoMap[provider]); expect([...logolessProviders].sort()).toEqual([...knownLogolessProviders].sort()); @@ -357,6 +358,10 @@ describe("provider_info_helpers", () => { expect(getPlaceholder(Providers.Cognition)).toBe("cognition/swe-1.7"); }); + it("should return zerogpu/gpt-oss-120b placeholder for ZeroGPU provider", () => { + expect(getPlaceholder("ZeroGPU")).toBe("zerogpu/gpt-oss-120b"); + }); + it("should return a chatgpt/ placeholder for the CHATGPT dropdown key", () => { expect(getPlaceholder("CHATGPT")).toBe("chatgpt/gpt-5.4"); }); diff --git a/ui/litellm-dashboard/src/components/provider_info_helpers.tsx b/ui/litellm-dashboard/src/components/provider_info_helpers.tsx index 2541aa45050..bfb578b218a 100644 --- a/ui/litellm-dashboard/src/components/provider_info_helpers.tsx +++ b/ui/litellm-dashboard/src/components/provider_info_helpers.tsx @@ -197,6 +197,7 @@ export enum Providers { xAI = "xAI", XINFERENCE = "Xinference", ZAI = "Z.AI (Zhipu AI)", + ZeroGPU = "ZeroGPU", } export const provider_map: Record = { @@ -325,6 +326,7 @@ export const provider_map: Record = { xAI: "xai", XINFERENCE: "xinference", ZAI: "zai", + ZeroGPU: "zerogpu", }; const standaloneSubproviderSlugs = new Set(["bedrock_mantle"]); @@ -497,6 +499,7 @@ const providerPlaceholderMap: Partial> = { [Providers.Voyage]: "voyage/", [Providers.WATSONX]: "watsonx/ibm/granite-3-3-8b-instruct", [Providers.ZAI]: "zai/glm-4.5", + [Providers.ZeroGPU]: "zerogpu/gpt-oss-120b", }; export const getPlaceholder = (selectedProvider: string): string => {