From 0fd8a42972de6413c1c4467e80c582fa689885e8 Mon Sep 17 00:00:00 2001 From: Gaurav Pandey <112387553+gaurav-pandey-zocdoc@users.noreply.github.com> Date: Tue, 1 Sep 2026 16:22:03 +0530 Subject: [PATCH 01/37] fix(alerting): clarify budget threshold messages Generated with AI Co-Authored-By: Claude Code --- litellm/integrations/SlackAlerting/slack_alerting.py | 4 ++-- .../SlackAlerting/test_slack_alerting.py | 12 ++++++------ 2 files changed, 8 insertions(+), 8 deletions(-) diff --git a/litellm/integrations/SlackAlerting/slack_alerting.py b/litellm/integrations/SlackAlerting/slack_alerting.py index 94d734546be..432cbd0917b 100644 --- a/litellm/integrations/SlackAlerting/slack_alerting.py +++ b/litellm/integrations/SlackAlerting/slack_alerting.py @@ -632,10 +632,10 @@ class SlackAlerting(CustomBatchLogger): event_message += f"Budget Crossed\n Total Budget:`{user_info.max_budget}`" elif percent_left <= SLACK_ALERTING_THRESHOLD_5_PERCENT: event = "threshold_crossed" - event_message += "5% Threshold Crossed " + event_message += "5% or less of budget remaining" elif percent_left <= SLACK_ALERTING_THRESHOLD_15_PERCENT: event = "threshold_crossed" - event_message += "15% Threshold Crossed" + event_message += "15% or less of budget remaining" return event, event_message diff --git a/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting.py b/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting.py index 55e2dcdc270..6eaa3147dc3 100644 --- a/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting.py +++ b/tests/test_litellm/integrations/SlackAlerting/test_slack_alerting.py @@ -10,6 +10,7 @@ import pytest import litellm from litellm.caching.caching import DualCache +from litellm.integrations.SlackAlerting.budget_alert_types import get_budget_alert_type from litellm.integrations.SlackAlerting.slack_alerting import SlackAlerting from litellm.proxy._types import CallInfo, Litellm_EntityType from litellm.types.integrations.slack_alerting import AlertType, SlackAlertingCacheKeys @@ -46,9 +47,8 @@ class TestSlackAlerting(unittest.TestCase): self.assertEqual(result, -0.2) def test_get_event_and_event_message_max_budget(self): - # Initial setup with no event event = None - event_message = "Test Message: " + event_message = get_budget_alert_type("user_budget").get_event_message() # Test case 1: When spend exceeds max_budget user_info = CallInfo( @@ -63,7 +63,7 @@ class TestSlackAlerting(unittest.TestCase): self.assertEqual(event, "budget_crossed") self.assertTrue("Budget Crossed" in event_message) - # Test case 2: When 5% of max_budget is left + event_message = get_budget_alert_type("user_budget").get_event_message() user_info = CallInfo( max_budget=100.0, spend=95.0, @@ -74,9 +74,9 @@ class TestSlackAlerting(unittest.TestCase): user_info=user_info, event=event, event_message=event_message ) self.assertEqual(event, "threshold_crossed") - self.assertTrue("5% Threshold Crossed" in event_message) + self.assertEqual(event_message, "User Budget: 5% or less of budget remaining") - # Test case 3: When 15% of max_budget is left + event_message = get_budget_alert_type("user_budget").get_event_message() user_info = CallInfo( max_budget=100.0, spend=85.0, @@ -87,7 +87,7 @@ class TestSlackAlerting(unittest.TestCase): user_info=user_info, event=event, event_message=event_message ) self.assertEqual(event, "threshold_crossed") - self.assertTrue("15% Threshold Crossed" in event_message) + self.assertEqual(event_message, "User Budget: 15% or less of budget remaining") def test_get_event_and_event_message_soft_budget(self): # Initial setup with no event From 233337628f4b6f1c9ec0527d5d442cfe512ac08b Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Wed, 9 Sep 2026 18:06:25 -0400 Subject: [PATCH 02/37] feat(batches): support Mistral files/batches and per-page OCR batch cost tracking Adds MistralFilesConfig and MistralBatchesConfig so Mistral can be used as a Files and Batches provider through the shared BaseLLMHTTPHandler path, the same way Bedrock plugs in. /v1/ocr is now an accepted batch endpoint, and completed OCR batches are billed per page (ocr_cost_per_page_batches, half the synchronous rate) instead of per token. Resolves #29914 --- litellm/batches/batch_utils.py | 41 ++- litellm/batches/main.py | 20 +- litellm/cost_calculator.py | 61 ++++ litellm/files/main.py | 9 +- litellm/files/types.py | 2 +- litellm/llms/mistral/batches/__init__.py | 0 .../llms/mistral/batches/transformation.py | 186 +++++++++++++ litellm/llms/mistral/common_utils.py | 36 +++ litellm/llms/mistral/files/__init__.py | 0 litellm/llms/mistral/files/transformation.py | 226 +++++++++++++++ ...odel_prices_and_context_window_backup.json | 40 ++- litellm/types/llms/openai.py | 2 +- litellm/types/utils.py | 4 + litellm/utils.py | 10 + model_prices_and_context_window.json | 40 ++- .../test_litellm/batches/test_batch_utils.py | 83 ++++++ tests/test_litellm/batches/test_main.py | 42 +++ .../llms/mistral/batches/__init__.py | 0 .../test_mistral_batches_transformation.py | 260 ++++++++++++++++++ .../llms/mistral/files/__init__.py | 0 .../test_mistral_files_transformation.py | 189 +++++++++++++ .../llms/mistral/ocr/test_mistral_ocr_cost.py | 4 +- tests/test_litellm/test_utils.py | 2 + 23 files changed, 1216 insertions(+), 41 deletions(-) create mode 100644 litellm/llms/mistral/batches/__init__.py create mode 100644 litellm/llms/mistral/batches/transformation.py create mode 100644 litellm/llms/mistral/common_utils.py create mode 100644 litellm/llms/mistral/files/__init__.py create mode 100644 litellm/llms/mistral/files/transformation.py create mode 100644 tests/test_litellm/llms/mistral/batches/__init__.py create mode 100644 tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py create mode 100644 tests/test_litellm/llms/mistral/files/__init__.py create mode 100644 tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py diff --git a/litellm/batches/batch_utils.py b/litellm/batches/batch_utils.py index 87f8fd3946e..077d6e72fd7 100644 --- a/litellm/batches/batch_utils.py +++ b/litellm/batches/batch_utils.py @@ -9,6 +9,7 @@ import litellm from litellm._logging import verbose_logger from litellm.litellm_core_utils.get_litellm_params import AWS_CREDENTIAL_KWARGS_KEYS from litellm.litellm_core_utils.llm_cost_calc.utils import parse_prompt_tokens_details +from litellm.llms.base_llm.ocr.transformation import OCRUsageInfo from litellm.types.llms.openai import Batch from litellm.types.utils import ModelInfo, Usage from litellm.utils import token_counter @@ -50,7 +51,7 @@ def batch_cost_is_final(batch: Batch) -> bool: async def calculate_batch_cost_and_usage( file_content_dictionary: list[dict], - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic"], + custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "mistral"], model_name: str | None = None, model_info: ModelInfo | None = None, ) -> BatchCostUsageResult: @@ -80,7 +81,7 @@ async def calculate_batch_cost_and_usage( async def _handle_completed_batch( batch: Batch, - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic"], + custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "mistral"], model_name: str | None = None, litellm_params: dict | None = None, model_info: ModelInfo | None = None, @@ -166,7 +167,7 @@ class _BatchOutputLineStats: def _classify_output_line_stats( entries: Iterable[dict], - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock"], + custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock", "mistral"], model_name: str | None, model_info: ModelInfo | None, ) -> Iterator[_BatchOutputLineStats | _LineOutcome]: @@ -185,7 +186,7 @@ def _classify_output_line_stats( def _safe_output_line_stats( entry: Mapping[str, object], - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock"], + custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock", "mistral"], model_name: str | None, model_info: ModelInfo | None, ) -> _BatchOutputLineStats | None: @@ -207,7 +208,7 @@ def _safe_output_line_stats( def _compute_output_line_stats( entry: Mapping[str, object], - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock"], + custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock", "mistral"], model_name: str | None, model_info: ModelInfo | None, ) -> _BatchOutputLineStats: @@ -218,6 +219,7 @@ def _compute_output_line_stats( response_model: Final = raw_model if isinstance(raw_model, str) and raw_model else None completion_details: Final = usage.completion_tokens_details line_prompt_cost, line_completion_cost = _output_line_cost( + response_body=response_body, usage=usage, custom_llm_provider=custom_llm_provider, model_name=model_name, @@ -237,19 +239,36 @@ def _compute_output_line_stats( ) +def _ocr_usage_info_from_response_body(response_body: Mapping[str, object]) -> OCRUsageInfo | None: + """OCR results report ``usage_info`` (pages) instead of ``usage`` (tokens); None for non-OCR lines.""" + raw_usage_info: Final = response_body.get("usage_info") + if not isinstance(raw_usage_info, Mapping): + return None + return OCRUsageInfo.model_validate(raw_usage_info) + + def _output_line_cost( + response_body: Mapping[str, object], usage: Usage, - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock"], + custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock", "mistral"], model_name: str | None, response_model: str | None, model_info: ModelInfo | None, ) -> tuple[float, float]: """(prompt_cost, completion_cost) for one output line, priced at batch rates.""" - from litellm.cost_calculator import batch_cost_calculator + from litellm.cost_calculator import batch_cost_calculator, ocr_batch_cost cost_model: Final = ( model_name if custom_llm_provider == "bedrock" and model_name else response_model or model_name or "" ) + ocr_usage: Final = _ocr_usage_info_from_response_body(response_body) + if ocr_usage is not None: + return ocr_batch_cost( + model=cost_model, + custom_llm_provider=custom_llm_provider, + usage_info=ocr_usage, + model_info=model_info, + ) return batch_cost_calculator( usage=usage, model=cost_model, @@ -260,7 +279,7 @@ def _output_line_cost( def _aggregate_batch_cost_usage_models( entries: Iterable[dict], - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock"], + custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "bedrock", "mistral"], model_name: str | None = None, model_info: ModelInfo | None = None, ) -> BatchCostUsageResult: @@ -427,7 +446,7 @@ def _provider_output_file_id(output_file_id: str) -> str: async def _fetch_batch_managed_file_content( file_id: str, - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic"] = "openai", + custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "mistral"] = "openai", litellm_params: dict | None = None, ) -> bytes: """ @@ -457,7 +476,7 @@ async def _fetch_batch_managed_file_content( async def _fetch_batch_output_file_content( batch: Batch, - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic"] = "openai", + custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "mistral"] = "openai", litellm_params: dict | None = None, ) -> bytes: """ @@ -479,7 +498,7 @@ async def _fetch_batch_output_file_content( async def count_error_file_failed_requests( batch: Batch, - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic"], + custom_llm_provider: Literal["openai", "azure", "vertex_ai", "hosted_vllm", "anthropic", "mistral"], litellm_params: dict | None, ) -> int: """Count failed requests reported only in the batch's separate error file. diff --git a/litellm/batches/main.py b/litellm/batches/main.py index 77a4fdebf16..76b6c73b375 100644 --- a/litellm/batches/main.py +++ b/litellm/batches/main.py @@ -105,9 +105,11 @@ def _resolve_timeout( @client async def acreate_batch( completion_window: Literal["24h"], - endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses"], + endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses", "/v1/ocr"], input_file_id: str, - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy"] = "openai", + custom_llm_provider: Literal[ + "openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "mistral" + ] = "openai", metadata: dict[str, str] | None = None, extra_headers: dict[str, str] | None = None, extra_body: dict[str, str] | None = None, @@ -155,9 +157,11 @@ async def acreate_batch( @client def create_batch( completion_window: Literal["24h"], - endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses"], + endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses", "/v1/ocr"], input_file_id: str, - custom_llm_provider: Literal["openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy"] = "openai", + custom_llm_provider: Literal[ + "openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "mistral" + ] = "openai", metadata: dict[str, str] | None = None, extra_headers: dict[str, str] | None = None, extra_body: dict[str, str] | None = None, @@ -341,7 +345,7 @@ def create_batch( async def aretrieve_batch( batch_id: str, custom_llm_provider: Literal[ - "openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "anthropic" + "openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "anthropic", "mistral" ] = "openai", metadata: dict[str, str] | None = None, extra_headers: dict[str, str] | None = None, @@ -389,7 +393,7 @@ def _handle_retrieve_batch_providers_without_provider_config( _retrieve_batch_request: RetrieveBatchRequest, _is_async: bool, custom_llm_provider: Literal[ - "openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "anthropic" + "openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "anthropic", "mistral" ] = "openai", logging_obj: LiteLLMLoggingObj | None = None, ): @@ -497,7 +501,7 @@ def _handle_retrieve_batch_providers_without_provider_config( message=( f"LiteLLM doesn't support custom_llm_provider={custom_llm_provider} for 'retrieve_batch' without a `model` kwarg. " "Supported via this path: 'openai', 'azure', 'vertex_ai', 'anthropic'. " - "'bedrock' is supported but requires `model` to be passed so the provider config can be loaded." + "'bedrock' and 'mistral' are supported but require `model` to be passed so the provider config can be loaded." ), model="n/a", llm_provider=custom_llm_provider, @@ -514,7 +518,7 @@ def _handle_retrieve_batch_providers_without_provider_config( def retrieve_batch( batch_id: str, custom_llm_provider: Literal[ - "openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "anthropic" + "openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "anthropic", "mistral" ] = "openai", metadata: dict[str, str] | None = None, extra_headers: dict[str, str] | None = None, diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 814eaaf76f7..0c0c6b05df8 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -139,6 +139,7 @@ if TYPE_CHECKING: from litellm.litellm_core_utils.litellm_logging import ( Logging as LitellmLoggingObject, ) + from litellm.llms.base_llm.ocr.transformation import OCRUsageInfo else: LitellmLoggingObject = Any @@ -1982,6 +1983,66 @@ def ocr_cost( return ocr_pages_cost + annotation_pages_cost, 0.0 +_OCR_PRICING_KEYS: Final = ( + "ocr_cost_per_page", + "ocr_cost_per_page_batches", + "annotation_cost_per_page", + "annotation_cost_per_page_batches", +) + + +def ocr_batch_cost( + model: str, + custom_llm_provider: str | None, + usage_info: "OCRUsageInfo", + model_info: ModelInfo | None = None, +) -> tuple[float, float]: + """Per-page cost of one OCR result inside a batch output file. + + Batch OCR is billed per page at the ``*_batches`` rate, falling back to the + synchronous per-page rate when a model has no batch price recorded, the same + fallback ``batch_cost_calculator`` applies to per-token batch pricing. Returns + ``(prompt_cost, completion_cost)`` with the whole cost in the first slot, like + ``ocr_cost``. + """ + has_ocr_pricing: Final = model_info is not None and any(model_info.get(k) is not None for k in _OCR_PRICING_KEYS) + if has_ocr_pricing: + resolved_info: ModelInfo | None = model_info + else: + try: + resolved_info = litellm.get_model_info(model=model, custom_llm_provider=custom_llm_provider) + except Exception: + resolved_info = None + if resolved_info is None: + verbose_logger.warning( + "OCR batch cost: model=%s custom_llm_provider=%s has no pricing entry; returning 0.0 cost.", + model, + custom_llm_provider, + ) + return 0.0, 0.0 + + page_rate: Final = _first_price(resolved_info, "ocr_cost_per_page_batches", "ocr_cost_per_page") + annotation_rate: Final = _first_price( + resolved_info, "annotation_cost_per_page_batches", "annotation_cost_per_page" + ) + pages_processed: Final = usage_info.pages_processed or 0 + annotation_pages: Final = usage_info.pages_processed_annotation or 0 + if page_rate is None and pages_processed > 0: + verbose_logger.warning( + "OCR batch cost: model=%s custom_llm_provider=%s reported pages_processed=%s but no " + "ocr_cost_per_page is configured; returning 0.0 cost for those pages.", + model, + custom_llm_provider, + pages_processed, + ) + effective_annotation_rate: Final = annotation_rate if annotation_rate is not None else page_rate + return (page_rate or 0.0) * pages_processed + (effective_annotation_rate or 0.0) * annotation_pages, 0.0 + + +def _first_price(model_info: ModelInfo, *keys: str) -> float | None: + return next((price for price in (model_info.get(k) for k in keys) if isinstance(price, (int, float))), None) + + def vector_store_search_cost( model: str | None, custom_llm_provider: str, diff --git a/litellm/files/main.py b/litellm/files/main.py index 218518eb3cd..3d90bf4f299 100644 --- a/litellm/files/main.py +++ b/litellm/files/main.py @@ -27,12 +27,15 @@ FileCreateProvider = Literal[ "litellm_proxy", "manus", "anthropic", + "mistral", ] FileRetrieveProvider = Literal[ - "openai", "azure", "gemini", "vertex_ai", "hosted_vllm", "litellm_proxy", "manus", "anthropic" + "openai", "azure", "gemini", "vertex_ai", "hosted_vllm", "litellm_proxy", "manus", "anthropic", "mistral" ] -FileDeleteProvider = Literal["openai", "azure", "gemini", "bedrock", "litellm_proxy", "manus", "anthropic"] -FileListProvider = Literal["openai", "azure", "litellm_proxy", "manus", "anthropic"] +FileDeleteProvider = Literal[ + "openai", "azure", "gemini", "bedrock", "litellm_proxy", "manus", "anthropic", "mistral" +] +FileListProvider = Literal["openai", "azure", "litellm_proxy", "manus", "anthropic", "mistral"] import litellm from litellm import get_secret_str from litellm.files.streaming import FileContentStreamingResponse diff --git a/litellm/files/types.py b/litellm/files/types.py index b4ec9996f37..01c7970144b 100644 --- a/litellm/files/types.py +++ b/litellm/files/types.py @@ -2,7 +2,7 @@ from collections.abc import AsyncIterator, Iterator from typing import Literal, NamedTuple FileContentProvider = Literal[ - "openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "anthropic", "manus" + "openai", "azure", "vertex_ai", "bedrock", "hosted_vllm", "litellm_proxy", "anthropic", "manus", "mistral" ] diff --git a/litellm/llms/mistral/batches/__init__.py b/litellm/llms/mistral/batches/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/litellm/llms/mistral/batches/transformation.py b/litellm/llms/mistral/batches/transformation.py new file mode 100644 index 00000000000..399319e590b --- /dev/null +++ b/litellm/llms/mistral/batches/transformation.py @@ -0,0 +1,186 @@ +""" +Mistral Batch API. Reference: https://docs.mistral.ai/api/#tag/batch + +Mistral runs one model per job (set on the job, not per input line) and accepts +``/v1/ocr`` as a batch endpoint, which is how OCR gets its 50% batch discount. +Output and error files are OpenAI-shaped JSONL (``{custom_id, response: {status_code, body}}``), +so the shared batch cost accounting reads them without a provider branch. +""" + +from types import MappingProxyType +from typing import Final, Literal + +import httpx +from openai.types.batch import BatchRequestCounts +from openai.types.batch import Errors as BatchErrors +from openai.types.batch_error import BatchError +from pydantic import BaseModel, ConfigDict + +from litellm.litellm_core_utils.url_utils import encode_url_path_segment +from litellm.llms.base_llm.batches.transformation import BaseBatchesConfig +from litellm.llms.base_llm.chat.transformation import BaseLLMException +from litellm.types.llms.openai import AllMessageValues, CreateBatchRequest +from litellm.types.utils import LiteLLMBatch, LlmProviders + +from ..common_utils import get_mistral_api_base, get_mistral_auth_headers, mistral_error + +MistralBatchStatus = Literal[ + "QUEUED", "RUNNING", "SUCCESS", "FAILED", "TIMEOUT_EXCEEDED", "CANCELLATION_REQUESTED", "CANCELLED" +] +OpenAIBatchStatus = Literal[ + "validating", "failed", "in_progress", "finalizing", "completed", "expired", "cancelling", "cancelled" +] + +_STATUS_MAP: Final[MappingProxyType[MistralBatchStatus, OpenAIBatchStatus]] = MappingProxyType( + { + "QUEUED": "validating", + "RUNNING": "in_progress", + "SUCCESS": "completed", + "FAILED": "failed", + "TIMEOUT_EXCEEDED": "expired", + "CANCELLATION_REQUESTED": "cancelling", + "CANCELLED": "cancelled", + } +) + + +class MistralBatchError(BaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + message: str + count: int = 1 + + +class MistralBatchJob(BaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + id: str + input_files: tuple[str, ...] = () + endpoint: str + model: str | None = None + status: MistralBatchStatus + created_at: int + started_at: int | None = None + completed_at: int | None = None + total_requests: int = 0 + completed_requests: int = 0 + succeeded_requests: int = 0 + failed_requests: int = 0 + output_file: str | None = None + error_file: str | None = None + errors: tuple[MistralBatchError, ...] = () + metadata: dict[str, str] | None = None + + +def _to_litellm_batch(job: MistralBatchJob) -> LiteLLMBatch: + status: Final = _STATUS_MAP[job.status] + terminal_at: Final = job.completed_at + return LiteLLMBatch( + id=job.id, + object="batch", + endpoint=job.endpoint, + input_file_id=job.input_files[0] if job.input_files else "", + completion_window="24h", + status=status, + created_at=job.created_at, + in_progress_at=job.started_at, + completed_at=terminal_at if status == "completed" else None, + failed_at=terminal_at if status == "failed" else None, + expired_at=terminal_at if status == "expired" else None, + cancelled_at=terminal_at if status == "cancelled" else None, + output_file_id=job.output_file, + error_file_id=job.error_file, + errors=( + BatchErrors( + object="list", + data=[BatchError(message=f"{e.message} (x{e.count})" if e.count > 1 else e.message) for e in job.errors], + ) + if job.errors + else None + ), + request_counts=BatchRequestCounts( + total=job.total_requests, + completed=job.succeeded_requests, + failed=job.failed_requests, + ), + metadata=job.metadata, + ) + + +class MistralBatchesConfig(BaseBatchesConfig): + @property + def custom_llm_provider(self) -> LlmProviders: + return LlmProviders.MISTRAL + + def validate_environment( + self, + headers: dict, + model: str, + messages: list[AllMessageValues], + optional_params: dict, + litellm_params: dict, + api_key: str | None = None, + api_base: str | None = None, + ) -> dict: + return get_mistral_auth_headers(headers, api_key) + + def get_complete_batch_url( + self, + api_base: str | None, + api_key: str | None, + model: str, + optional_params: dict, + litellm_params: dict, + data: CreateBatchRequest, + ) -> str: + return f"{get_mistral_api_base(api_base)}/v1/batch/jobs" + + def transform_create_batch_request( + self, + model: str, + create_batch_data: CreateBatchRequest, + optional_params: dict, + litellm_params: dict, + ) -> dict[str, object]: + metadata: Final = create_batch_data.get("metadata") + return { + "input_files": [create_batch_data["input_file_id"]], + "endpoint": create_batch_data["endpoint"], + "model": model, + **({"metadata": metadata} if metadata else {}), + **(create_batch_data.get("extra_body") or {}), + } + + def transform_create_batch_response( + self, + model: str | None, + raw_response: httpx.Response, + logging_obj: object, + litellm_params: dict, + ) -> LiteLLMBatch: + return _to_litellm_batch(MistralBatchJob.model_validate(raw_response.json())) + + def transform_retrieve_batch_request( + self, + batch_id: str, + optional_params: dict, + litellm_params: dict, + ) -> dict[str, object]: + encoded_batch_id: Final = encode_url_path_segment(batch_id, field_name="batch_id") + return { + "method": "GET", + "url": f"{get_mistral_api_base(litellm_params.get('api_base'))}/v1/batch/jobs/{encoded_batch_id}", + "headers": get_mistral_auth_headers({}, litellm_params.get("api_key")), + } + + def transform_retrieve_batch_response( + self, + model: str | None, + raw_response: httpx.Response, + logging_obj: object, + litellm_params: dict, + ) -> LiteLLMBatch: + return _to_litellm_batch(MistralBatchJob.model_validate(raw_response.json())) + + def get_error_class(self, error_message: str, status_code: int, headers: dict | httpx.Headers) -> BaseLLMException: + return mistral_error(error_message, status_code, headers) diff --git a/litellm/llms/mistral/common_utils.py b/litellm/llms/mistral/common_utils.py new file mode 100644 index 00000000000..9ea501c860d --- /dev/null +++ b/litellm/llms/mistral/common_utils.py @@ -0,0 +1,36 @@ +from typing import Final + +import httpx + +from litellm.llms.base_llm.chat.transformation import BaseLLMException +from litellm.secret_managers.main import get_secret_str + +MISTRAL_API_BASE: Final = "https://api.mistral.ai" +MISTRAL_API_KEY_ENV_VAR: Final = "MISTRAL_API_KEY" + + +class MistralError(BaseLLMException): + pass + + +def get_mistral_api_base(api_base: str | None) -> str: + """Return the Mistral origin without a trailing ``/v1``, so callers can append ``/v1/``.""" + resolved: Final = (api_base or get_secret_str("MISTRAL_API_BASE") or MISTRAL_API_BASE).rstrip("/") + return resolved.removesuffix("/v1") + + +def get_mistral_auth_headers(headers: dict, api_key: str | None) -> dict: + resolved_key: Final = api_key or get_secret_str(MISTRAL_API_KEY_ENV_VAR) + if resolved_key is None: + raise ValueError( + "Missing Mistral API Key - A call is being made to Mistral but no key is set either in the environment variables or via params" + ) + return {**headers, "Authorization": f"Bearer {resolved_key}"} + + +def mistral_error(error_message: str, status_code: int, headers: dict | httpx.Headers) -> MistralError: + return MistralError( + status_code=status_code, + message=error_message, + headers=headers if isinstance(headers, httpx.Headers) else httpx.Headers(headers), + ) diff --git a/litellm/llms/mistral/files/__init__.py b/litellm/llms/mistral/files/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/litellm/llms/mistral/files/transformation.py b/litellm/llms/mistral/files/transformation.py new file mode 100644 index 00000000000..071b6f58569 --- /dev/null +++ b/litellm/llms/mistral/files/transformation.py @@ -0,0 +1,226 @@ +""" +Mistral Files API. Reference: https://docs.mistral.ai/api/#tag/files + +Mistral's file objects already carry the OpenAI field names (id, bytes, created_at, +filename, purpose), so this config is URL routing, auth, and a purpose mapping: +Mistral only accepts ``fine-tune``, ``batch`` and ``ocr`` as upload purposes. +""" + +import time +from typing import Final, Literal + +import httpx +from openai.types.file_deleted import FileDeleted +from pydantic import BaseModel, ConfigDict + +from litellm.litellm_core_utils.prompt_templates.common_utils import extract_file_data +from litellm.litellm_core_utils.url_utils import encode_url_path_segment +from litellm.llms.base_llm.chat.transformation import BaseLLMException +from litellm.llms.base_llm.files.transformation import BaseFilesConfig, LiteLLMLoggingObj +from litellm.types.llms.openai import ( + CreateFileRequest, + FileContentRequest, + HttpxBinaryResponseContent, + OpenAICreateFileRequestOptionalParams, + OpenAIFileObject, + OpenAIFilesPurpose, +) +from litellm.types.utils import LlmProviders + +from ..common_utils import get_mistral_api_base, get_mistral_auth_headers, mistral_error + +MistralFilePurpose = Literal["fine-tune", "batch", "ocr"] + + +class MistralFile(BaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + id: str + bytes: int = 0 + created_at: int | None = None + filename: str = "" + purpose: MistralFilePurpose = "batch" + expires_at: int | None = None + + +class MistralFileList(BaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + data: tuple[MistralFile, ...] = () + + +class MistralFileDeleted(BaseModel): + model_config = ConfigDict(frozen=True, extra="ignore") + + id: str + deleted: bool = True + + +def _to_openai_file_object(file: MistralFile) -> OpenAIFileObject: + return OpenAIFileObject( + id=file.id, + bytes=file.bytes, + created_at=file.created_at if file.created_at is not None else int(time.time()), + filename=file.filename, + object="file", + purpose=_to_openai_purpose(file.purpose), + status="uploaded", + expires_at=file.expires_at, + ) + + +def _to_openai_purpose(purpose: MistralFilePurpose) -> OpenAIFilesPurpose: + match purpose: + case "fine-tune" | "batch": + return purpose + case "ocr": + return "user_data" + + +def _to_mistral_purpose(purpose: str) -> MistralFilePurpose: + match purpose: + case "fine-tune" | "ocr": + return purpose + case _: + return "batch" + + +class MistralFilesConfig(BaseFilesConfig): + @property + def custom_llm_provider(self) -> LlmProviders: + return LlmProviders.MISTRAL + + def get_complete_url( + self, + api_base: str | None, + api_key: str | None, + model: str, + optional_params: dict, + litellm_params: dict, + stream: bool | None = None, + ) -> str: + return f"{get_mistral_api_base(api_base)}/v1/files" + + def _file_url(self, file_id: str, litellm_params: dict, suffix: str = "") -> str: + encoded_file_id: Final = encode_url_path_segment(file_id, field_name="file_id") + return f"{get_mistral_api_base(litellm_params.get('api_base'))}/v1/files/{encoded_file_id}{suffix}" + + def get_error_class(self, error_message: str, status_code: int, headers: dict | httpx.Headers) -> BaseLLMException: + return mistral_error(error_message, status_code, headers) + + def validate_environment( + self, + headers: dict, + model: str, + messages: list, + optional_params: dict, + litellm_params: dict, + api_key: str | None = None, + api_base: str | None = None, + ) -> dict: + return get_mistral_auth_headers(headers, api_key) + + def get_supported_openai_params(self, model: str) -> list[OpenAICreateFileRequestOptionalParams]: + return ["purpose"] + + def map_openai_params( + self, + non_default_params: dict, + optional_params: dict, + model: str, + drop_params: bool, + ) -> dict: + return optional_params + + def transform_create_file_request( + self, + model: str, + create_file_data: CreateFileRequest, + optional_params: dict, + litellm_params: dict, + ) -> dict: + file_data: Final = create_file_data.get("file") + if file_data is None: + raise ValueError("File data is required") + extracted: Final = extract_file_data(file_data) + filename: Final = extracted["filename"] or f"file_{int(time.time())}.jsonl" + content_type: Final = extracted.get("content_type") or "application/octet-stream" + return { + "file": (filename, extracted["content"], content_type), + "purpose": (None, _to_mistral_purpose(create_file_data.get("purpose", "batch"))), + } + + def transform_create_file_response( + self, + model: str | None, + raw_response: httpx.Response, + logging_obj: LiteLLMLoggingObj, + litellm_params: dict, + ) -> OpenAIFileObject: + return _to_openai_file_object(MistralFile.model_validate(raw_response.json())) + + def transform_retrieve_file_request( + self, + file_id: str, + optional_params: dict, + litellm_params: dict, + ) -> tuple[str, dict]: + return self._file_url(file_id, litellm_params), {} + + def transform_retrieve_file_response( + self, + raw_response: httpx.Response, + logging_obj: LiteLLMLoggingObj, + litellm_params: dict, + ) -> OpenAIFileObject: + return _to_openai_file_object(MistralFile.model_validate(raw_response.json())) + + def transform_delete_file_request( + self, + file_id: str, + optional_params: dict, + litellm_params: dict, + ) -> tuple[str, dict]: + return self._file_url(file_id, litellm_params), {} + + def transform_delete_file_response( + self, + raw_response: httpx.Response, + logging_obj: LiteLLMLoggingObj, + litellm_params: dict, + ) -> FileDeleted: + deleted: Final = MistralFileDeleted.model_validate(raw_response.json()) + return FileDeleted(id=deleted.id, deleted=deleted.deleted, object="file") + + def transform_list_files_request( + self, + purpose: str | None, + optional_params: dict, + litellm_params: dict, + ) -> tuple[str, dict]: + params: Final = {"purpose": _to_mistral_purpose(purpose)} if purpose else {} + return f"{get_mistral_api_base(litellm_params.get('api_base'))}/v1/files", params + + def transform_list_files_response( + self, + raw_response: httpx.Response, + logging_obj: LiteLLMLoggingObj, + litellm_params: dict, + ) -> list[OpenAIFileObject]: + return [_to_openai_file_object(f) for f in MistralFileList.model_validate(raw_response.json()).data] + + def transform_file_content_request( + self, + file_content_request: FileContentRequest, + optional_params: dict, + litellm_params: dict, + ) -> tuple[str, dict]: + return self._file_url(file_content_request["file_id"], litellm_params, suffix="/content"), {} + + def transform_file_content_response( + self, + raw_response: httpx.Response, + logging_obj: LiteLLMLoggingObj, + litellm_params: dict, + ) -> HttpxBinaryResponseContent: + return HttpxBinaryResponseContent(response=raw_response) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 54ebdc85be9..5a934301edd 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -35262,51 +35262,66 @@ "mistral/mistral-ocr-latest": { "litellm_provider": "mistral", "ocr_cost_per_page": 0.004, + "ocr_cost_per_page_batches": 0.002, "annotation_cost_per_page": 0.005, + "annotation_cost_per_page_batches": 0.0025, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, "mistral/mistral-ocr-4-0": { "litellm_provider": "mistral", "ocr_cost_per_page": 0.004, + "ocr_cost_per_page_batches": 0.002, "annotation_cost_per_page": 0.005, + "annotation_cost_per_page_batches": 0.0025, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, "mistral/mistral-ocr-4-1": { "annotation_cost_per_page": 0.005, + "annotation_cost_per_page_batches": 0.0025, "litellm_provider": "mistral", "mode": "ocr", "ocr_cost_per_page": 0.004, + "ocr_cost_per_page_batches": 0.002, "source": "https://docs.mistral.ai/models/model-cards/ocr-4-1", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ] }, "mistral/mistral-ocr-2505-completion": { "deprecation_date": "2026-05-31", "litellm_provider": "mistral", "ocr_cost_per_page": 0.001, + "ocr_cost_per_page_batches": 0.0005, "annotation_cost_per_page": 0.003, + "annotation_cost_per_page_batches": 0.0015, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, "mistral/mistral-ocr-2512": { "litellm_provider": "mistral", "ocr_cost_per_page": 0.002, + "ocr_cost_per_page_batches": 0.001, "annotation_cost_per_page": 0.003, + "annotation_cost_per_page_batches": 0.0015, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, @@ -59822,31 +59837,40 @@ "mistral/mistral-ocr-3": { "litellm_provider": "mistral", "ocr_cost_per_page": 0.002, + "ocr_cost_per_page_batches": 0.001, "annotation_cost_per_page": 0.003, + "annotation_cost_per_page_batches": 0.0015, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, "mistral/mistral-ocr-3-0": { "litellm_provider": "mistral", "ocr_cost_per_page": 0.002, + "ocr_cost_per_page_batches": 0.001, "annotation_cost_per_page": 0.003, + "annotation_cost_per_page_batches": 0.0015, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, "mistral/mistral-ocr-4": { "annotation_cost_per_page": 0.005, + "annotation_cost_per_page_batches": 0.0025, "litellm_provider": "mistral", "mode": "ocr", "ocr_cost_per_page": 0.004, + "ocr_cost_per_page_batches": 0.002, "source": "https://docs.mistral.ai/models/model-cards/ocr-4-1", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ] }, "mistral/voxtral-mini-latest": { diff --git a/litellm/types/llms/openai.py b/litellm/types/llms/openai.py index b6da9490e01..defd59f2be8 100644 --- a/litellm/types/llms/openai.py +++ b/litellm/types/llms/openai.py @@ -498,7 +498,7 @@ class CreateBatchRequest(TypedDict, total=False): """ completion_window: Literal["24h"] - endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses"] + endpoint: Literal["/v1/chat/completions", "/v1/embeddings", "/v1/completions", "/v1/responses", "/v1/ocr"] input_file_id: str metadata: dict[str, str] | None output_expires_after: FileExpiresAfter diff --git a/litellm/types/utils.py b/litellm/types/utils.py index d62f00f3676..ea23e00d2bb 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -320,8 +320,10 @@ class ModelInfoBase(ProviderSpecificModelInfo, total=False): output_cost_per_second_480p: ReadOnly[float | None] output_cost_per_second_4k: ReadOnly[float | None] ocr_cost_per_page: float | None # for OCR models + ocr_cost_per_page_batches: ReadOnly[float | None] ocr_cost_per_credit: float | None # for OCR models priced by credit annotation_cost_per_page: float | None # for OCR models + annotation_cost_per_page_batches: ReadOnly[float | None] search_context_cost_per_query: SearchContextCostPerQuery | None # Cost for using web search tool web_search_billing_unit: ( Literal["per_query", "per_prompt"] | None @@ -3598,8 +3600,10 @@ class CustomPricingLiteLLMParams(MirroredPricingParams): output_cost_per_token_above_512k_tokens: float | None = None output_vector_size: int | None = None ocr_cost_per_page: float | None = None + ocr_cost_per_page_batches: float | None = None ocr_cost_per_credit: float | None = None annotation_cost_per_page: float | None = None + annotation_cost_per_page_batches: float | None = None regional_processing_uplift_multiplier_eu: float | None = None regional_processing_uplift_multiplier_us: float | None = None regional_endpoint_uplift_multiplier: float | None = None diff --git a/litellm/utils.py b/litellm/utils.py index 8df28870544..3ca81605a95 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -5963,8 +5963,10 @@ def _get_model_info_helper( tpm=_model_info.get("tpm", None), rpm=_model_info.get("rpm", None), ocr_cost_per_page=_model_info.get("ocr_cost_per_page", None), + ocr_cost_per_page_batches=_model_info.get("ocr_cost_per_page_batches", None), ocr_cost_per_credit=_model_info.get("ocr_cost_per_credit", None), annotation_cost_per_page=_model_info.get("annotation_cost_per_page", None), + annotation_cost_per_page_batches=_model_info.get("annotation_cost_per_page_batches", None), provider_specific_entry=_model_info.get("provider_specific_entry", None), uses_embed_content=_model_info.get("uses_embed_content", None), supports_image_size=_model_info.get("supports_image_size", None), @@ -8909,6 +8911,10 @@ class ProviderConfigManager: from litellm.llms.anthropic.files.transformation import AnthropicFilesConfig return AnthropicFilesConfig() + elif LlmProviders.MISTRAL == provider: + from litellm.llms.mistral.files.transformation import MistralFilesConfig + + return MistralFilesConfig() return None @staticmethod @@ -8920,6 +8926,10 @@ class ProviderConfigManager: from litellm.llms.bedrock.batches.transformation import BedrockBatchesConfig return BedrockBatchesConfig() + elif LlmProviders.MISTRAL == provider: + from litellm.llms.mistral.batches.transformation import MistralBatchesConfig + + return MistralBatchesConfig() return None @staticmethod diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 54ebdc85be9..5a934301edd 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -35262,51 +35262,66 @@ "mistral/mistral-ocr-latest": { "litellm_provider": "mistral", "ocr_cost_per_page": 0.004, + "ocr_cost_per_page_batches": 0.002, "annotation_cost_per_page": 0.005, + "annotation_cost_per_page_batches": 0.0025, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, "mistral/mistral-ocr-4-0": { "litellm_provider": "mistral", "ocr_cost_per_page": 0.004, + "ocr_cost_per_page_batches": 0.002, "annotation_cost_per_page": 0.005, + "annotation_cost_per_page_batches": 0.0025, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, "mistral/mistral-ocr-4-1": { "annotation_cost_per_page": 0.005, + "annotation_cost_per_page_batches": 0.0025, "litellm_provider": "mistral", "mode": "ocr", "ocr_cost_per_page": 0.004, + "ocr_cost_per_page_batches": 0.002, "source": "https://docs.mistral.ai/models/model-cards/ocr-4-1", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ] }, "mistral/mistral-ocr-2505-completion": { "deprecation_date": "2026-05-31", "litellm_provider": "mistral", "ocr_cost_per_page": 0.001, + "ocr_cost_per_page_batches": 0.0005, "annotation_cost_per_page": 0.003, + "annotation_cost_per_page_batches": 0.0015, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, "mistral/mistral-ocr-2512": { "litellm_provider": "mistral", "ocr_cost_per_page": 0.002, + "ocr_cost_per_page_batches": 0.001, "annotation_cost_per_page": 0.003, + "annotation_cost_per_page_batches": 0.0015, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, @@ -59822,31 +59837,40 @@ "mistral/mistral-ocr-3": { "litellm_provider": "mistral", "ocr_cost_per_page": 0.002, + "ocr_cost_per_page_batches": 0.001, "annotation_cost_per_page": 0.003, + "annotation_cost_per_page_batches": 0.0015, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, "mistral/mistral-ocr-3-0": { "litellm_provider": "mistral", "ocr_cost_per_page": 0.002, + "ocr_cost_per_page_batches": 0.001, "annotation_cost_per_page": 0.003, + "annotation_cost_per_page_batches": 0.0015, "mode": "ocr", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ], "source": "https://mistral.ai/pricing#api-pricing" }, "mistral/mistral-ocr-4": { "annotation_cost_per_page": 0.005, + "annotation_cost_per_page_batches": 0.0025, "litellm_provider": "mistral", "mode": "ocr", "ocr_cost_per_page": 0.004, + "ocr_cost_per_page_batches": 0.002, "source": "https://docs.mistral.ai/models/model-cards/ocr-4-1", "supported_endpoints": [ - "/v1/ocr" + "/v1/ocr", + "/v1/batch" ] }, "mistral/voxtral-mini-latest": { diff --git a/tests/test_litellm/batches/test_batch_utils.py b/tests/test_litellm/batches/test_batch_utils.py index 768ea332677..85f267c2db0 100644 --- a/tests/test_litellm/batches/test_batch_utils.py +++ b/tests/test_litellm/batches/test_batch_utils.py @@ -1787,3 +1787,86 @@ class TestBatchCostIsFinal: @pytest.mark.parametrize("status", ["failed", "expired", "cancelled"]) def test_other_terminal_statuses_are_final(self, status): assert bu.batch_cost_is_final(_retrieved_batch(status)) is True + + +# =========================================================================== # +# OCR batch output lines (Mistral /v1/ocr batches) are billed per page, not per token +# =========================================================================== # + + +def _ocr_row(pages_processed, annotation_pages=None, model="mistral-ocr-latest"): + usage_info = {"pages_processed": pages_processed, "doc_size_bytes": 4096} + if annotation_pages is not None: + usage_info["pages_processed_annotation"] = annotation_pages + return _success_row(model=model, pages=[{"index": i, "markdown": "x"} for i in range(pages_processed)], usage_info=usage_info) + + +def test_ocr_rows_are_priced_per_page_at_batch_rate(monkeypatch): + monkeypatch.setattr( + litellm, + "get_model_info", + lambda model, custom_llm_provider=None: {"ocr_cost_per_page": 0.004, "ocr_cost_per_page_batches": 0.002}, + ) + result = bu._aggregate_batch_cost_usage_models( + entries=[_ocr_row(3), _ocr_row(5), _failed_row(model="mistral-ocr-latest")], + custom_llm_provider="mistral", + model_name="mistral/mistral-ocr-latest", + ) + assert result.cost == pytest.approx(8 * 0.002) + assert result.prompt_cost == pytest.approx(8 * 0.002) + assert result.completion_cost == 0.0 + assert (result.successful_requests, result.failed_requests) == (2, 1) + assert result.usage.total_tokens == 0 + assert result.models == ["mistral/mistral-ocr-latest"] + + +def test_ocr_rows_fall_back_to_sync_page_rate_without_batch_price(monkeypatch): + monkeypatch.setattr(litellm, "get_model_info", lambda model, custom_llm_provider=None: {"ocr_cost_per_page": 0.004}) + result = bu._aggregate_batch_cost_usage_models(entries=[_ocr_row(2)], custom_llm_provider="mistral") + assert result.cost == pytest.approx(2 * 0.004) + + +def test_ocr_rows_bill_annotation_pages_separately(monkeypatch): + monkeypatch.setattr( + litellm, + "get_model_info", + lambda model, custom_llm_provider=None: { + "ocr_cost_per_page_batches": 0.002, + "annotation_cost_per_page_batches": 0.0025, + }, + ) + result = bu._aggregate_batch_cost_usage_models(entries=[_ocr_row(4, annotation_pages=4)], custom_llm_provider="mistral") + assert result.cost == pytest.approx(4 * 0.002 + 4 * 0.0025) + + +def test_ocr_rows_use_deployment_model_info_pricing_over_cost_map(monkeypatch): + monkeypatch.setattr( + litellm, "get_model_info", lambda model, custom_llm_provider=None: pytest.fail("cost map must not be consulted") + ) + result = bu._aggregate_batch_cost_usage_models( + entries=[_ocr_row(10)], + custom_llm_provider="mistral", + model_info={"ocr_cost_per_page_batches": 0.001}, + ) + assert result.cost == pytest.approx(0.01) + + +def test_ocr_rows_without_pricing_bill_zero_but_count_as_successful(monkeypatch): + monkeypatch.setattr(litellm, "get_model_info", lambda model, custom_llm_provider=None: {"mode": "ocr"}) + result = bu._aggregate_batch_cost_usage_models(entries=[_ocr_row(3)], custom_llm_provider="mistral") + assert result.cost == 0.0 + assert (result.successful_requests, result.failed_requests) == (1, 0) + + +def test_chat_rows_from_mistral_still_use_token_pricing(monkeypatch): + monkeypatch.setattr( + litellm, + "get_model_info", + lambda model, custom_llm_provider=None: {"input_cost_per_token": 0.001, "output_cost_per_token": 0.002}, + ) + result = bu._aggregate_batch_cost_usage_models( + entries=[_success_row(model="mistral-small-latest", usage=_usage(10, 5))], + custom_llm_provider="mistral", + ) + assert result.cost == pytest.approx((10 * 0.001 + 5 * 0.002) / 2) + assert result.usage.total_tokens == 15 diff --git a/tests/test_litellm/batches/test_main.py b/tests/test_litellm/batches/test_main.py index b87f9489250..c5a33dd6508 100644 --- a/tests/test_litellm/batches/test_main.py +++ b/tests/test_litellm/batches/test_main.py @@ -778,3 +778,45 @@ def test_retrieve__omits_trusted_model_credentials_when_not_supplied(seams): litellm_params = logging_obj.update_from_kwargs.call_args.kwargs["litellm_params"] assert "_litellm_internal_model_credentials" not in litellm_params + + +# =========================================================================== # +# mistral - a provider-config provider, like bedrock, so it requires `model` +# =========================================================================== # + + +def test_create__mistral_ocr_routes_to_base_http_handler_with_mistral_config(seams): + with patch.object(bm.ProviderConfigManager, "get_provider_batches_config", wraps=bm.ProviderConfigManager.get_provider_batches_config) as get_cfg: + result = bm.create_batch( + completion_window="24h", + endpoint="/v1/ocr", + input_file_id="file-abc", + custom_llm_provider="mistral", + model="mistral/mistral-ocr-latest", + ) + + assert result is seams.base_http.create_batch.return_value + _assert_only(seams.base_http.create_batch, seams, "create_batch") + get_cfg.assert_called_once() + forwarded = seams.base_http.create_batch.call_args.kwargs + assert type(forwarded["provider_config"]).__name__ == "MistralBatchesConfig" + assert forwarded["model"] == "mistral-ocr-latest" + assert forwarded["create_batch_data"]["endpoint"] == "/v1/ocr" + + +def test_create__mistral_without_model_raises_badrequest(seams): + with pytest.raises(litellm.exceptions.BadRequestError): + bm.create_batch(**CREATE_KW, custom_llm_provider="mistral") + + for m in _all_seam_methods(seams, "create_batch"): + m.assert_not_called() + + +def test_retrieve__mistral_routes_to_base_http_handler_with_mistral_config(seams): + result = bm.retrieve_batch(batch_id="job-1", custom_llm_provider="mistral", model="mistral/mistral-ocr-latest") + + assert result is seams.base_http.retrieve_batch.return_value + _assert_only(seams.base_http.retrieve_batch, seams, "retrieve_batch") + forwarded = seams.base_http.retrieve_batch.call_args.kwargs + assert type(forwarded["provider_config"]).__name__ == "MistralBatchesConfig" + assert forwarded["batch_id"] == "job-1" diff --git a/tests/test_litellm/llms/mistral/batches/__init__.py b/tests/test_litellm/llms/mistral/batches/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py b/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py new file mode 100644 index 00000000000..03e9c351a30 --- /dev/null +++ b/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py @@ -0,0 +1,260 @@ +""" +Regression tests for ``MistralBatchesConfig``, the BaseBatchesConfig implementation +behind ``custom_llm_provider="mistral"`` on /v1/batches. + +Locks the request shape Mistral's ``POST /v1/batch/jobs`` accepts (input_files list, +model set on the job, endpoint passed through untouched so ``/v1/ocr`` batches work), +the Mistral -> OpenAI status mapping, request-count and file-id mapping, and auth. +Everything runs for real against canned httpx responses; only the API key env var is +set. +""" + +import json + +import httpx +import pytest + +from litellm.llms.mistral.batches.transformation import MistralBatchesConfig +from litellm.llms.mistral.common_utils import MistralError +from litellm.types.llms.openai import CreateBatchRequest +from litellm.types.utils import LiteLLMBatch, LlmProviders + +STATUS_MAP = { + "QUEUED": "validating", + "RUNNING": "in_progress", + "SUCCESS": "completed", + "FAILED": "failed", + "TIMEOUT_EXCEEDED": "expired", + "CANCELLATION_REQUESTED": "cancelling", + "CANCELLED": "cancelled", +} + + +def _job(**overrides): + base = { + "id": "8ff5e0d1-6bc2-4c3a-9f7d-0d1c2e3f4a5b", + "object": "batch", + "input_files": ["c1a2b3d4-0000-4000-8000-000000000001"], + "endpoint": "/v1/ocr", + "model": "mistral-ocr-latest", + "status": "SUCCESS", + "created_at": 1_757_400_000, + "started_at": 1_757_400_010, + "completed_at": 1_757_400_500, + "total_requests": 3, + "completed_requests": 3, + "succeeded_requests": 2, + "failed_requests": 1, + "output_file": "out-0000-4000-8000-000000000002", + "error_file": "err-0000-4000-8000-000000000003", + "errors": [], + "metadata": {"job_type": "testing"}, + } + return {**base, **overrides} + + +def _response(payload: dict, status_code: int = 200) -> httpx.Response: + return httpx.Response( + status_code=status_code, + content=json.dumps(payload).encode(), + request=httpx.Request("GET", "https://api.mistral.ai/v1/batch/jobs/x"), + ) + + +@pytest.fixture +def config() -> MistralBatchesConfig: + return MistralBatchesConfig() + + +@pytest.fixture +def api_key(monkeypatch) -> str: + monkeypatch.setenv("MISTRAL_API_KEY", "sk-mistral-test") + return "sk-mistral-test" + + +def test_custom_llm_provider(config): + assert config.custom_llm_provider == LlmProviders.MISTRAL + + +# --------------------------------------------------------------------------- # +# create +# --------------------------------------------------------------------------- # + + +def test_create_request_maps_openai_fields_onto_mistral_job(config): + data = CreateBatchRequest( + completion_window="24h", + endpoint="/v1/ocr", + input_file_id="file-123", + metadata={"team": "docs"}, + ) + body = config.transform_create_batch_request( + model="mistral-ocr-latest", create_batch_data=data, optional_params={}, litellm_params={} + ) + assert body == { + "input_files": ["file-123"], + "endpoint": "/v1/ocr", + "model": "mistral-ocr-latest", + "metadata": {"team": "docs"}, + } + + +def test_create_request_omits_empty_metadata_and_forwards_extra_body(config): + data = CreateBatchRequest( + completion_window="24h", + endpoint="/v1/chat/completions", + input_file_id="file-123", + metadata=None, + extra_body={"timeout_hours": 48}, + ) + body = config.transform_create_batch_request( + model="mistral-small-latest", create_batch_data=data, optional_params={}, litellm_params={} + ) + assert "metadata" not in body + assert body["timeout_hours"] == 48 + + +@pytest.mark.parametrize( + "api_base,expected", + [ + (None, "https://api.mistral.ai/v1/batch/jobs"), + ("https://api.mistral.ai/v1", "https://api.mistral.ai/v1/batch/jobs"), + ("https://proxy.example.com/", "https://proxy.example.com/v1/batch/jobs"), + ], +) +def test_create_url(config, api_base, expected): + url = config.get_complete_batch_url( + api_base=api_base, api_key="k", model="m", optional_params={}, litellm_params={}, data={} + ) + assert url == expected + + +def test_validate_environment_uses_bearer_auth(config, api_key): + headers = config.validate_environment( + headers={"x-extra": "1"}, model="m", messages=[], optional_params={}, litellm_params={} + ) + assert headers == {"x-extra": "1", "Authorization": f"Bearer {api_key}"} + + +def test_validate_environment_explicit_key_wins(config, api_key): + headers = config.validate_environment( + headers={}, model="m", messages=[], optional_params={}, litellm_params={}, api_key="sk-explicit" + ) + assert headers["Authorization"] == "Bearer sk-explicit" + + +def test_validate_environment_without_key_raises(config, monkeypatch): + monkeypatch.delenv("MISTRAL_API_KEY", raising=False) + with pytest.raises(ValueError, match="Missing Mistral API Key"): + config.validate_environment(headers={}, model="m", messages=[], optional_params={}, litellm_params={}) + + +def test_create_response_maps_job_onto_openai_batch(config): + batch = config.transform_create_batch_response( + model="mistral-ocr-latest", + raw_response=_response(_job(status="QUEUED", started_at=None, completed_at=None)), + logging_obj=None, + litellm_params={}, + ) + assert isinstance(batch, LiteLLMBatch) + assert batch.id == "8ff5e0d1-6bc2-4c3a-9f7d-0d1c2e3f4a5b" + assert batch.endpoint == "/v1/ocr" + assert batch.input_file_id == "c1a2b3d4-0000-4000-8000-000000000001" + assert batch.status == "validating" + assert batch.created_at == 1_757_400_000 + assert batch.in_progress_at is None + assert batch.completed_at is None + assert batch.metadata == {"job_type": "testing"} + + +# --------------------------------------------------------------------------- # +# retrieve +# --------------------------------------------------------------------------- # + + +def test_retrieve_request_is_presigned_get_with_auth(config, api_key): + req = config.transform_retrieve_batch_request( + batch_id="job/with slash", optional_params={}, litellm_params={"api_base": "https://api.mistral.ai"} + ) + assert req["method"] == "GET" + assert req["url"] == "https://api.mistral.ai/v1/batch/jobs/job%2Fwith%20slash" + assert req["headers"] == {"Authorization": f"Bearer {api_key}"} + + +def test_retrieve_request_prefers_litellm_params_api_key(config, api_key): + req = config.transform_retrieve_batch_request( + batch_id="job-1", optional_params={}, litellm_params={"api_key": "sk-from-deployment"} + ) + assert req["headers"]["Authorization"] == "Bearer sk-from-deployment" + + +@pytest.mark.parametrize("mistral_status,openai_status", sorted(STATUS_MAP.items())) +def test_retrieve_response_status_mapping(config, mistral_status, openai_status): + batch = config.transform_retrieve_batch_response( + model=None, raw_response=_response(_job(status=mistral_status)), logging_obj=None, litellm_params={} + ) + assert batch.status == openai_status + + +@pytest.mark.parametrize( + "mistral_status,populated_field", + [ + ("SUCCESS", "completed_at"), + ("FAILED", "failed_at"), + ("TIMEOUT_EXCEEDED", "expired_at"), + ("CANCELLED", "cancelled_at"), + ], +) +def test_retrieve_response_terminal_timestamp_lands_on_matching_field(config, mistral_status, populated_field): + batch = config.transform_retrieve_batch_response( + model=None, raw_response=_response(_job(status=mistral_status)), logging_obj=None, litellm_params={} + ) + terminal_fields = {"completed_at", "failed_at", "expired_at", "cancelled_at"} + assert getattr(batch, populated_field) == 1_757_400_500 + for other in terminal_fields - {populated_field}: + assert getattr(batch, other) is None + assert batch.in_progress_at == 1_757_400_010 + + +def test_retrieve_response_maps_counts_and_files(config): + batch = config.transform_retrieve_batch_response( + model=None, raw_response=_response(_job()), logging_obj=None, litellm_params={} + ) + assert batch.request_counts.total == 3 + assert batch.request_counts.completed == 2 + assert batch.request_counts.failed == 1 + assert batch.output_file_id == "out-0000-4000-8000-000000000002" + assert batch.error_file_id == "err-0000-4000-8000-000000000003" + assert batch.errors is None + + +def test_retrieve_response_surfaces_job_errors(config): + batch = config.transform_retrieve_batch_response( + model=None, + raw_response=_response( + _job(status="FAILED", errors=[{"message": "invalid document", "count": 2}, {"message": "timeout"}]) + ), + logging_obj=None, + litellm_params={}, + ) + assert [e.message for e in batch.errors.data] == ["invalid document (x2)", "timeout"] + + +def test_retrieve_response_without_files_or_input(config): + batch = config.transform_retrieve_batch_response( + model=None, + raw_response=_response(_job(input_files=[], output_file=None, error_file=None, metadata=None)), + logging_obj=None, + litellm_params={}, + ) + assert batch.input_file_id == "" + assert batch.output_file_id is None + assert batch.error_file_id is None + assert batch.metadata is None + + +def test_get_error_class(config): + err = config.get_error_class("nope", 401, {"x-request-id": "r1"}) + assert isinstance(err, MistralError) + assert err.status_code == 401 + assert err.message == "nope" diff --git a/tests/test_litellm/llms/mistral/files/__init__.py b/tests/test_litellm/llms/mistral/files/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py new file mode 100644 index 00000000000..d6ad8a34b35 --- /dev/null +++ b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py @@ -0,0 +1,189 @@ +""" +Regression tests for ``MistralFilesConfig``, the BaseFilesConfig implementation behind +``custom_llm_provider="mistral"`` on /v1/files. + +Locks the URL routing for each file operation, the multipart upload shape Mistral's +``POST /v1/files`` accepts (purpose restricted to fine-tune/batch/ocr), and the +Mistral -> OpenAI file object mapping. Runs against canned httpx responses. +""" + +import json + +import httpx +import pytest +from openai.types.file_deleted import FileDeleted + +from litellm.llms.mistral.files.transformation import MistralFilesConfig +from litellm.types.llms.openai import CreateFileRequest, FileContentRequest, OpenAIFileObject +from litellm.types.utils import LlmProviders + +FILE_ID = "497f6eca-6276-4993-bfeb-53cbbbba6f09" + + +def _file(**overrides): + base = { + "id": FILE_ID, + "object": "file", + "bytes": 13000, + "created_at": 1_716_963_433, + "filename": "batch_input.jsonl", + "purpose": "batch", + "sample_type": "batch_request", + "num_lines": 3, + "source": "upload", + } + return {**base, **overrides} + + +def _response(payload) -> httpx.Response: + return httpx.Response( + status_code=200, + content=json.dumps(payload).encode(), + request=httpx.Request("GET", "https://api.mistral.ai/v1/files"), + ) + + +@pytest.fixture +def config() -> MistralFilesConfig: + return MistralFilesConfig() + + +@pytest.fixture +def api_key(monkeypatch) -> str: + monkeypatch.setenv("MISTRAL_API_KEY", "sk-mistral-test") + return "sk-mistral-test" + + +def test_custom_llm_provider(config): + assert config.custom_llm_provider == LlmProviders.MISTRAL + + +@pytest.mark.parametrize( + "api_base,expected", + [ + (None, "https://api.mistral.ai/v1/files"), + ("https://api.mistral.ai/v1/", "https://api.mistral.ai/v1/files"), + ("https://proxy.example.com", "https://proxy.example.com/v1/files"), + ], +) +def test_upload_url(config, api_base, expected): + url = config.get_complete_url(api_base=api_base, api_key="k", model="", optional_params={}, litellm_params={}) + assert url == expected + + +def test_validate_environment_uses_bearer_auth(config, api_key): + headers = config.validate_environment(headers={}, model="", messages=[], optional_params={}, litellm_params={}) + assert headers == {"Authorization": f"Bearer {api_key}"} + + +def test_upload_request_is_multipart_with_batch_purpose(config): + body = config.transform_create_file_request( + model="", + create_file_data=CreateFileRequest(file=("in.jsonl", b'{"custom_id":"0"}\n', "application/jsonl"), purpose="batch"), + optional_params={}, + litellm_params={}, + ) + assert body == { + "file": ("in.jsonl", b'{"custom_id":"0"}\n', "application/jsonl"), + "purpose": (None, "batch"), + } + + +@pytest.mark.parametrize( + "openai_purpose,mistral_purpose", + [("batch", "batch"), ("fine-tune", "fine-tune"), ("ocr", "ocr"), ("assistants", "batch"), ("user_data", "batch")], +) +def test_upload_request_maps_purpose_onto_mistral_enum(config, openai_purpose, mistral_purpose): + body = config.transform_create_file_request( + model="", + create_file_data=CreateFileRequest(file=("f.bin", b"x"), purpose=openai_purpose), + optional_params={}, + litellm_params={}, + ) + assert body["purpose"] == (None, mistral_purpose) + + +def test_upload_request_requires_file(config): + with pytest.raises(ValueError, match="File data is required"): + config.transform_create_file_request( + model="", create_file_data=CreateFileRequest(purpose="batch"), optional_params={}, litellm_params={} + ) + + +def test_upload_response_maps_onto_openai_file_object(config): + obj = config.transform_create_file_response( + model=None, raw_response=_response(_file()), logging_obj=None, litellm_params={} + ) + assert obj == OpenAIFileObject( + id=FILE_ID, + bytes=13000, + created_at=1_716_963_433, + filename="batch_input.jsonl", + object="file", + purpose="batch", + status="uploaded", + ) + + +def test_file_response_with_ocr_purpose_maps_onto_user_data(config): + obj = config.transform_retrieve_file_response( + raw_response=_response(_file(purpose="ocr", expires_at=1_800_000_000)), logging_obj=None, litellm_params={} + ) + assert obj.purpose == "user_data" + assert obj.expires_at == 1_800_000_000 + + +@pytest.mark.parametrize( + "method,suffix", + [ + ("transform_retrieve_file_request", ""), + ("transform_delete_file_request", ""), + ], +) +def test_single_file_urls_encode_id_and_honor_api_base(config, method, suffix): + url, params = getattr(config, method)( + file_id="id/with slash", optional_params={}, litellm_params={"api_base": "https://mistral.internal/v1"} + ) + assert url == f"https://mistral.internal/v1/files/id%2Fwith%20slash{suffix}" + assert params == {} + + +def test_file_content_url(config): + url, params = config.transform_file_content_request( + file_content_request=FileContentRequest(file_id=FILE_ID), optional_params={}, litellm_params={} + ) + assert url == f"https://api.mistral.ai/v1/files/{FILE_ID}/content" + assert params == {} + + +def test_file_content_response_is_binary_passthrough(config): + raw = httpx.Response( + 200, content=b'{"custom_id":"0","response":{"status_code":200}}\n', request=httpx.Request("GET", "https://x") + ) + out = config.transform_file_content_response(raw_response=raw, logging_obj=None, litellm_params={}) + assert out.content == b'{"custom_id":"0","response":{"status_code":200}}\n' + + +def test_delete_response(config): + out = config.transform_delete_file_response( + raw_response=_response({"id": FILE_ID, "object": "file", "deleted": True}), logging_obj=None, litellm_params={} + ) + assert out == FileDeleted(id=FILE_ID, deleted=True, object="file") + + +def test_list_request_filters_by_mapped_purpose(config): + url, params = config.transform_list_files_request(purpose="batch", optional_params={}, litellm_params={}) + assert url == "https://api.mistral.ai/v1/files" + assert params == {"purpose": "batch"} + _, no_params = config.transform_list_files_request(purpose=None, optional_params={}, litellm_params={}) + assert no_params == {} + + +def test_list_response(config): + out = config.transform_list_files_response( + raw_response=_response({"data": [_file(), _file(id="second", filename="b.jsonl")], "object": "list", "total": 2}), + logging_obj=None, + litellm_params={}, + ) + assert [f.id for f in out] == [FILE_ID, "second"] + assert out[1].filename == "b.jsonl" diff --git a/tests/test_litellm/llms/mistral/ocr/test_mistral_ocr_cost.py b/tests/test_litellm/llms/mistral/ocr/test_mistral_ocr_cost.py index 40e54f71eeb..9fe6f38003f 100644 --- a/tests/test_litellm/llms/mistral/ocr/test_mistral_ocr_cost.py +++ b/tests/test_litellm/llms/mistral/ocr/test_mistral_ocr_cost.py @@ -72,9 +72,11 @@ def test_ocr3_pricing_entry(cost_map_path: Path) -> None: assert info is not None, f"{OCR3_MODEL} missing from {cost_map_path.name}" assert info["litellm_provider"] == "mistral" assert info["mode"] == "ocr" - assert info["supported_endpoints"] == ["/v1/ocr"] + assert info["supported_endpoints"] == ["/v1/ocr", "/v1/batch"] assert info["ocr_cost_per_page"] == OCR3_COST_PER_PAGE assert info["annotation_cost_per_page"] == OCR3_ANNOTATION_COST_PER_PAGE + assert info["ocr_cost_per_page_batches"] == OCR3_COST_PER_PAGE / 2 + assert info["annotation_cost_per_page_batches"] == OCR3_ANNOTATION_COST_PER_PAGE / 2 def test_ocr3_model_info_price(local_model_cost_map) -> None: diff --git a/tests/test_litellm/test_utils.py b/tests/test_litellm/test_utils.py index e42608c9904..61739846706 100644 --- a/tests/test_litellm/test_utils.py +++ b/tests/test_litellm/test_utils.py @@ -992,7 +992,9 @@ def test_aaamodel_prices_and_context_window_json_is_valid(): "input_cost_per_video_per_second_above_128k_tokens": {"type": "number"}, "input_dbu_cost_per_token": {"type": "number"}, "annotation_cost_per_page": {"type": "number"}, + "annotation_cost_per_page_batches": {"type": "number"}, "ocr_cost_per_page": {"type": "number"}, + "ocr_cost_per_page_batches": {"type": "number"}, "ocr_cost_per_credit": {"type": "number"}, "code_interpreter_cost_per_session": {"type": "number"}, "inference_geo": {"type": "string"}, From 2e5f5a95c813b940dc7f654c10b2ce2036c6fb2a Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Wed, 9 Sep 2026 18:14:16 -0400 Subject: [PATCH 03/37] fix(proxy): retrieve model-routed file ids from the deployment's provider GET /v1/files/{id} for an id encoded with a non-OpenAI deployment forwarded the deployment credentials but let custom_llm_provider default to openai, so a Mistral file was fetched from api.openai.com with the Mistral key and 401'd. Delete and content already passed the provider through; retrieve now does too. --- .../openai_files_endpoints/files_endpoints.py | 5 +- .../test_files_endpoint.py | 59 +++++++++++++++++++ 2 files changed, 63 insertions(+), 1 deletion(-) diff --git a/litellm/proxy/openai_files_endpoints/files_endpoints.py b/litellm/proxy/openai_files_endpoints/files_endpoints.py index c315d30b8f3..49a01495d65 100644 --- a/litellm/proxy/openai_files_endpoints/files_endpoints.py +++ b/litellm/proxy/openai_files_endpoints/files_endpoints.py @@ -1139,7 +1139,10 @@ async def get_file( include_internal_credentials=True, ) - response = await litellm.afile_retrieve(**data) + response = await litellm.afile_retrieve( + custom_llm_provider=credentials["custom_llm_provider"], + **data, + ) # Keep the encoded ID in response if it was originally encoded if original_file_id and response and hasattr(response, "id") and response.id: diff --git a/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py b/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py index 5f1e7e1fe0c..5faae166fca 100644 --- a/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py +++ b/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py @@ -4819,3 +4819,62 @@ def test_get_file_content_keeps_the_status_of_a_rejection_raised_inside_the_rout error = response.json()["error"] assert error["message"].startswith("Storage backend error") assert (error["type"], error["param"], error["code"]) == ("invalid_request_error", "file_id", "400") + + +def test_get_file_model_routed_id_forwards_deployment_provider(mocker: MockerFixture, monkeypatch): + """ + Regression: a file id encoded with a non-OpenAI deployment (here Mistral) must be + retrieved from that deployment's provider. Before the fix the retrieve path only + forwarded the credentials and let ``custom_llm_provider`` default to openai, so a + Mistral file id was sent to api.openai.com with the Mistral key and 401'd. + """ + import litellm.proxy.proxy_server as ps + from litellm.proxy._types import LitellmUserRoles + from litellm.proxy.openai_files_endpoints.common_utils import encode_file_id_with_model + + router = Router( + model_list=[ + { + "model_name": "mistral-ocr", + "litellm_params": {"model": "mistral/mistral-ocr-latest", "api_key": "mistral-key"}, + "model_info": {"id": "mistral-ocr-id"}, + } + ] + ) + proxy_logging_obj = setup_proxy_logging_object(monkeypatch, router) + monkeypatch.setattr("litellm.proxy.proxy_server.master_key", None) + monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", None) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", router) + proxy_logging_obj.update_request_status = mocker.AsyncMock() + proxy_logging_obj.post_call_failure_hook = mocker.AsyncMock() + + captured_kwargs: dict = {} + + async def _mock_afile_retrieve(**kwargs): + captured_kwargs.update(kwargs) + return OpenAIFileObject( + id="7a13fa8e-fcf8-42c5-aa61-c93c10e2c7df", + object="file", + bytes=2, + created_at=1234567890, + filename="batch.jsonl", + purpose="batch", + status="uploaded", + ) + + monkeypatch.setattr(litellm, "afile_retrieve", _mock_afile_retrieve) + app.dependency_overrides[ps.user_api_key_auth] = lambda: UserAPIKeyAuth( + api_key="test-key", user_role=LitellmUserRoles.PROXY_ADMIN, user_id="test-user" + ) + encoded_id = encode_file_id_with_model("7a13fa8e-fcf8-42c5-aa61-c93c10e2c7df", "mistral-ocr") + + try: + response = client.get(f"/v1/files/{encoded_id}", headers={"Authorization": "Bearer test-key"}) + finally: + app.dependency_overrides.pop(ps.user_api_key_auth, None) + + assert response.status_code == 200, response.text + assert captured_kwargs["custom_llm_provider"] == "mistral" + assert captured_kwargs["api_key"] == "mistral-key" + assert captured_kwargs["file_id"] == "7a13fa8e-fcf8-42c5-aa61-c93c10e2c7df" + assert response.json()["id"] == encoded_id From c246f75e3ec93192f95e0cb2fc3f50485bf13ebc Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Wed, 9 Sep 2026 18:37:01 -0400 Subject: [PATCH 04/37] refactor(mistral): satisfy type-discipline and basedpyright gates for files/batches configs --- litellm/cost_calculator.py | 23 ++-- litellm/files/main.py | 4 +- .../llms/mistral/batches/transformation.py | 114 ++++++++++------ litellm/llms/mistral/common_utils.py | 13 +- litellm/llms/mistral/files/transformation.py | 129 +++++++++++------- .../openai_files_endpoints/files_endpoints.py | 2 +- .../test_litellm/batches/test_batch_utils.py | 46 +++++-- tests/test_litellm/batches/test_main.py | 76 +++-------- .../test_mistral_batches_transformation.py | 16 ++- .../test_mistral_files_transformation.py | 8 +- 10 files changed, 251 insertions(+), 180 deletions(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 0c0c6b05df8..7c95941d77d 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -2006,13 +2006,11 @@ def ocr_batch_cost( ``ocr_cost``. """ has_ocr_pricing: Final = model_info is not None and any(model_info.get(k) is not None for k in _OCR_PRICING_KEYS) - if has_ocr_pricing: - resolved_info: ModelInfo | None = model_info - else: - try: - resolved_info = litellm.get_model_info(model=model, custom_llm_provider=custom_llm_provider) - except Exception: - resolved_info = None + resolved_info: Final = ( + model_info + if has_ocr_pricing + else _lookup_model_info_or_none(model=model, custom_llm_provider=custom_llm_provider) + ) if resolved_info is None: verbose_logger.warning( "OCR batch cost: model=%s custom_llm_provider=%s has no pricing entry; returning 0.0 cost.", @@ -2022,9 +2020,7 @@ def ocr_batch_cost( return 0.0, 0.0 page_rate: Final = _first_price(resolved_info, "ocr_cost_per_page_batches", "ocr_cost_per_page") - annotation_rate: Final = _first_price( - resolved_info, "annotation_cost_per_page_batches", "annotation_cost_per_page" - ) + annotation_rate: Final = _first_price(resolved_info, "annotation_cost_per_page_batches", "annotation_cost_per_page") pages_processed: Final = usage_info.pages_processed or 0 annotation_pages: Final = usage_info.pages_processed_annotation or 0 if page_rate is None and pages_processed > 0: @@ -2039,6 +2035,13 @@ def ocr_batch_cost( return (page_rate or 0.0) * pages_processed + (effective_annotation_rate or 0.0) * annotation_pages, 0.0 +def _lookup_model_info_or_none(model: str, custom_llm_provider: str | None) -> ModelInfo | None: + try: + return litellm.get_model_info(model=model, custom_llm_provider=custom_llm_provider) + except Exception: + return None + + def _first_price(model_info: ModelInfo, *keys: str) -> float | None: return next((price for price in (model_info.get(k) for k in keys) if isinstance(price, (int, float))), None) diff --git a/litellm/files/main.py b/litellm/files/main.py index 3d90bf4f299..5cf0f8e576a 100644 --- a/litellm/files/main.py +++ b/litellm/files/main.py @@ -32,9 +32,7 @@ FileCreateProvider = Literal[ FileRetrieveProvider = Literal[ "openai", "azure", "gemini", "vertex_ai", "hosted_vllm", "litellm_proxy", "manus", "anthropic", "mistral" ] -FileDeleteProvider = Literal[ - "openai", "azure", "gemini", "bedrock", "litellm_proxy", "manus", "anthropic", "mistral" -] +FileDeleteProvider = Literal["openai", "azure", "gemini", "bedrock", "litellm_proxy", "manus", "anthropic", "mistral"] FileListProvider = Literal["openai", "azure", "litellm_proxy", "manus", "anthropic", "mistral"] import litellm from litellm import get_secret_str diff --git a/litellm/llms/mistral/batches/transformation.py b/litellm/llms/mistral/batches/transformation.py index 399319e590b..ef9ee5ff503 100644 --- a/litellm/llms/mistral/batches/transformation.py +++ b/litellm/llms/mistral/batches/transformation.py @@ -7,14 +7,16 @@ Output and error files are OpenAI-shaped JSONL (``{custom_id, response: {status_ so the shared batch cost accounting reads them without a provider branch. """ +from collections.abc import Mapping, Sequence from types import MappingProxyType -from typing import Final, Literal +from typing import Final, Literal, TypeAlias import httpx from openai.types.batch import BatchRequestCounts from openai.types.batch import Errors as BatchErrors from openai.types.batch_error import BatchError from pydantic import BaseModel, ConfigDict +from typing_extensions import NotRequired, ReadOnly, TypedDict from litellm.litellm_core_utils.url_utils import encode_url_path_segment from litellm.llms.base_llm.batches.transformation import BaseBatchesConfig @@ -24,13 +26,14 @@ from litellm.types.utils import LiteLLMBatch, LlmProviders from ..common_utils import get_mistral_api_base, get_mistral_auth_headers, mistral_error -MistralBatchStatus = Literal[ +MistralBatchStatus: TypeAlias = Literal[ "QUEUED", "RUNNING", "SUCCESS", "FAILED", "TIMEOUT_EXCEEDED", "CANCELLATION_REQUESTED", "CANCELLED" ] -OpenAIBatchStatus = Literal[ +OpenAIBatchStatus: TypeAlias = Literal[ "validating", "failed", "in_progress", "finalizing", "completed", "expired", "cancelling", "cancelled" ] +_NO_HEADERS: Final[Mapping[str, str]] = MappingProxyType({}) # mutable-ok: frozen at module scope _STATUS_MAP: Final[MappingProxyType[MistralBatchStatus, OpenAIBatchStatus]] = MappingProxyType( { "QUEUED": "validating", @@ -44,6 +47,23 @@ _STATUS_MAP: Final[MappingProxyType[MistralBatchStatus, OpenAIBatchStatus]] = Ma ) +class MistralCreateBatchJobRequest(TypedDict): + """Body of ``POST /v1/batch/jobs``.""" + + input_files: ReadOnly[tuple[str, ...]] + endpoint: ReadOnly[str] + model: ReadOnly[str] + metadata: NotRequired[ReadOnly[Mapping[str, str]]] + + +class MistralPresignedRequest(TypedDict): + """A fully-formed request the shared HTTP handler sends as-is (its ``method`` branch).""" + + method: ReadOnly[Literal["GET"]] + url: ReadOnly[str] + headers: ReadOnly[Mapping[str, str]] + + class MistralBatchError(BaseModel): model_config = ConfigDict(frozen=True, extra="ignore") @@ -69,7 +89,18 @@ class MistralBatchJob(BaseModel): output_file: str | None = None error_file: str | None = None errors: tuple[MistralBatchError, ...] = () - metadata: dict[str, str] | None = None + metadata: dict[str, str] | None = None # mutable-ok: LiteLLMBatch.metadata is typed as dict + + +def _to_batch_errors(errors: Sequence[MistralBatchError]) -> BatchErrors | None: + if not errors: + return None + return BatchErrors( + object="list", + data=[ # mutable-ok: openai Batch.Errors.data is typed as list + BatchError(message=f"{e.message} (x{e.count})" if e.count > 1 else e.message) for e in errors + ], + ) def _to_litellm_batch(job: MistralBatchJob) -> LiteLLMBatch: @@ -90,14 +121,7 @@ def _to_litellm_batch(job: MistralBatchJob) -> LiteLLMBatch: cancelled_at=terminal_at if status == "cancelled" else None, output_file_id=job.output_file, error_file_id=job.error_file, - errors=( - BatchErrors( - object="list", - data=[BatchError(message=f"{e.message} (x{e.count})" if e.count > 1 else e.message) for e in job.errors], - ) - if job.errors - else None - ), + errors=_to_batch_errors(job.errors), request_counts=BatchRequestCounts( total=job.total_requests, completed=job.succeeded_requests, @@ -114,14 +138,14 @@ class MistralBatchesConfig(BaseBatchesConfig): def validate_environment( self, - headers: dict, + headers: Mapping[str, str], model: str, - messages: list[AllMessageValues], - optional_params: dict, - litellm_params: dict, + messages: Sequence[AllMessageValues], + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], api_key: str | None = None, api_base: str | None = None, - ) -> dict: + ) -> dict[str, str]: # mutable-ok: BaseBatchesConfig signature return get_mistral_auth_headers(headers, api_key) def get_complete_batch_url( @@ -129,8 +153,8 @@ class MistralBatchesConfig(BaseBatchesConfig): api_base: str | None, api_key: str | None, model: str, - optional_params: dict, - litellm_params: dict, + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], data: CreateBatchRequest, ) -> str: return f"{get_mistral_api_base(api_base)}/v1/batch/jobs" @@ -139,48 +163,58 @@ class MistralBatchesConfig(BaseBatchesConfig): self, model: str, create_batch_data: CreateBatchRequest, - optional_params: dict, - litellm_params: dict, - ) -> dict[str, object]: + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], + ) -> dict[str, object]: # mutable-ok: BaseBatchesConfig signature + input_file_id: Final = create_batch_data.get("input_file_id") + endpoint: Final = create_batch_data.get("endpoint") + if input_file_id is None or endpoint is None: + raise ValueError("input_file_id and endpoint are required to create a Mistral batch job") metadata: Final = create_batch_data.get("metadata") - return { - "input_files": [create_batch_data["input_file_id"]], - "endpoint": create_batch_data["endpoint"], - "model": model, - **({"metadata": metadata} if metadata else {}), - **(create_batch_data.get("extra_body") or {}), - } + body: Final = ( + MistralCreateBatchJobRequest( + input_files=(input_file_id,), endpoint=endpoint, model=model, metadata=metadata + ) + if metadata + else MistralCreateBatchJobRequest(input_files=(input_file_id,), endpoint=endpoint, model=model) + ) + return dict(body) # mutable-ok: BaseBatchesConfig signature def transform_create_batch_response( self, model: str | None, raw_response: httpx.Response, logging_obj: object, - litellm_params: dict, + litellm_params: Mapping[str, object], ) -> LiteLLMBatch: return _to_litellm_batch(MistralBatchJob.model_validate(raw_response.json())) def transform_retrieve_batch_request( self, batch_id: str, - optional_params: dict, - litellm_params: dict, - ) -> dict[str, object]: + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], + ) -> dict[str, object]: # mutable-ok: BaseBatchesConfig signature encoded_batch_id: Final = encode_url_path_segment(batch_id, field_name="batch_id") - return { - "method": "GET", - "url": f"{get_mistral_api_base(litellm_params.get('api_base'))}/v1/batch/jobs/{encoded_batch_id}", - "headers": get_mistral_auth_headers({}, litellm_params.get("api_key")), - } + api_base: Final = litellm_params.get("api_base") + api_key: Final = litellm_params.get("api_key") + request: Final = MistralPresignedRequest( + method="GET", + url=f"{get_mistral_api_base(api_base if isinstance(api_base, str) else None)}/v1/batch/jobs/{encoded_batch_id}", + headers=get_mistral_auth_headers(_NO_HEADERS, api_key if isinstance(api_key, str) else None), + ) + return dict(request) # mutable-ok: BaseBatchesConfig signature def transform_retrieve_batch_response( self, model: str | None, raw_response: httpx.Response, logging_obj: object, - litellm_params: dict, + litellm_params: Mapping[str, object], ) -> LiteLLMBatch: return _to_litellm_batch(MistralBatchJob.model_validate(raw_response.json())) - def get_error_class(self, error_message: str, status_code: int, headers: dict | httpx.Headers) -> BaseLLMException: + def get_error_class( + self, error_message: str, status_code: int, headers: Mapping[str, str] | httpx.Headers + ) -> BaseLLMException: return mistral_error(error_message, status_code, headers) diff --git a/litellm/llms/mistral/common_utils.py b/litellm/llms/mistral/common_utils.py index 9ea501c860d..2f14328afdf 100644 --- a/litellm/llms/mistral/common_utils.py +++ b/litellm/llms/mistral/common_utils.py @@ -1,3 +1,4 @@ +from collections.abc import Mapping from typing import Final import httpx @@ -19,18 +20,22 @@ def get_mistral_api_base(api_base: str | None) -> str: return resolved.removesuffix("/v1") -def get_mistral_auth_headers(headers: dict, api_key: str | None) -> dict: +def get_mistral_auth_headers( + headers: Mapping[str, str], api_key: str | None +) -> dict[str, str]: # mutable-ok: BaseConfig.validate_environment contract returns dict resolved_key: Final = api_key or get_secret_str(MISTRAL_API_KEY_ENV_VAR) if resolved_key is None: raise ValueError( "Missing Mistral API Key - A call is being made to Mistral but no key is set either in the environment variables or via params" ) - return {**headers, "Authorization": f"Bearer {resolved_key}"} + return dict(headers, Authorization=f"Bearer {resolved_key}") # mutable-ok: BaseConfig contract returns dict -def mistral_error(error_message: str, status_code: int, headers: dict | httpx.Headers) -> MistralError: +def mistral_error(error_message: str, status_code: int, headers: Mapping[str, str] | httpx.Headers) -> MistralError: return MistralError( status_code=status_code, message=error_message, - headers=headers if isinstance(headers, httpx.Headers) else httpx.Headers(headers), + headers=headers + if isinstance(headers, httpx.Headers) + else httpx.Headers(dict(headers)), # mutable-ok: httpx.Headers takes a dict ) diff --git a/litellm/llms/mistral/files/transformation.py b/litellm/llms/mistral/files/transformation.py index 071b6f58569..6d58311813c 100644 --- a/litellm/llms/mistral/files/transformation.py +++ b/litellm/llms/mistral/files/transformation.py @@ -7,11 +7,13 @@ Mistral only accepts ``fine-tune``, ``batch`` and ``ocr`` as upload purposes. """ import time -from typing import Final, Literal +from collections.abc import Mapping, Sequence +from typing import Final, Literal, TypeAlias import httpx from openai.types.file_deleted import FileDeleted from pydantic import BaseModel, ConfigDict +from typing_extensions import ReadOnly, TypedDict from litellm.litellm_core_utils.prompt_templates.common_utils import extract_file_data from litellm.litellm_core_utils.url_utils import encode_url_path_segment @@ -29,7 +31,16 @@ from litellm.types.utils import LlmProviders from ..common_utils import get_mistral_api_base, get_mistral_auth_headers, mistral_error -MistralFilePurpose = Literal["fine-tune", "batch", "ocr"] +MistralFilePurpose: TypeAlias = Literal["fine-tune", "batch", "ocr"] + +_NO_QUERY_PARAMS: Final[dict[str, str]] = {} # mutable-ok: BaseFilesConfig request transforms return tuple[str, dict] + + +class MistralMultipartUpload(TypedDict): + """``files=`` payload for ``POST /v1/files``: each value is an httpx multipart tuple.""" + + file: ReadOnly[tuple[str, object, str]] + purpose: ReadOnly[tuple[None, MistralFilePurpose]] class MistralFile(BaseModel): @@ -85,6 +96,11 @@ def _to_mistral_purpose(purpose: str) -> MistralFilePurpose: return "batch" +def _api_base_from(litellm_params: Mapping[str, object]) -> str: + api_base: Final = litellm_params.get("api_base") + return get_mistral_api_base(api_base if isinstance(api_base, str) else None) + + class MistralFilesConfig(BaseFilesConfig): @property def custom_llm_provider(self) -> LlmProviders: @@ -95,99 +111,103 @@ class MistralFilesConfig(BaseFilesConfig): api_base: str | None, api_key: str | None, model: str, - optional_params: dict, - litellm_params: dict, + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], stream: bool | None = None, ) -> str: return f"{get_mistral_api_base(api_base)}/v1/files" - def _file_url(self, file_id: str, litellm_params: dict, suffix: str = "") -> str: + def _file_url(self, file_id: str, litellm_params: Mapping[str, object], suffix: str = "") -> str: encoded_file_id: Final = encode_url_path_segment(file_id, field_name="file_id") - return f"{get_mistral_api_base(litellm_params.get('api_base'))}/v1/files/{encoded_file_id}{suffix}" + return f"{_api_base_from(litellm_params)}/v1/files/{encoded_file_id}{suffix}" - def get_error_class(self, error_message: str, status_code: int, headers: dict | httpx.Headers) -> BaseLLMException: + def get_error_class( + self, error_message: str, status_code: int, headers: Mapping[str, str] | httpx.Headers + ) -> BaseLLMException: return mistral_error(error_message, status_code, headers) def validate_environment( self, - headers: dict, + headers: Mapping[str, str], model: str, - messages: list, - optional_params: dict, - litellm_params: dict, + messages: Sequence[object], + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], api_key: str | None = None, api_base: str | None = None, - ) -> dict: + ) -> dict[str, str]: # mutable-ok: BaseFilesConfig signature return get_mistral_auth_headers(headers, api_key) - def get_supported_openai_params(self, model: str) -> list[OpenAICreateFileRequestOptionalParams]: - return ["purpose"] + def get_supported_openai_params( + self, model: str + ) -> list[OpenAICreateFileRequestOptionalParams]: # mutable-ok: BaseFilesConfig signature + return ["purpose"] # mutable-ok: BaseFilesConfig signature def map_openai_params( self, - non_default_params: dict, - optional_params: dict, + non_default_params: Mapping[str, object], + optional_params: dict[str, object], # mutable-ok: BaseConfig signature, returned as-is model: str, drop_params: bool, - ) -> dict: + ) -> dict[str, object]: # mutable-ok: BaseConfig signature return optional_params def transform_create_file_request( self, model: str, create_file_data: CreateFileRequest, - optional_params: dict, - litellm_params: dict, - ) -> dict: - file_data: Final = create_file_data.get("file") - if file_data is None: + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], + ) -> dict[str, object]: # mutable-ok: BaseFilesConfig signature + if "file" not in create_file_data: raise ValueError("File data is required") - extracted: Final = extract_file_data(file_data) + extracted: Final = extract_file_data(create_file_data["file"]) filename: Final = extracted["filename"] or f"file_{int(time.time())}.jsonl" content_type: Final = extracted.get("content_type") or "application/octet-stream" - return { - "file": (filename, extracted["content"], content_type), - "purpose": (None, _to_mistral_purpose(create_file_data.get("purpose", "batch"))), - } + upload: Final = MistralMultipartUpload( + file=(filename, extracted["content"], content_type), + purpose=(None, _to_mistral_purpose(create_file_data.get("purpose", "batch"))), + ) + return dict(upload) # mutable-ok: BaseFilesConfig signature def transform_create_file_response( self, model: str | None, raw_response: httpx.Response, logging_obj: LiteLLMLoggingObj, - litellm_params: dict, + litellm_params: Mapping[str, object], ) -> OpenAIFileObject: return _to_openai_file_object(MistralFile.model_validate(raw_response.json())) def transform_retrieve_file_request( self, file_id: str, - optional_params: dict, - litellm_params: dict, - ) -> tuple[str, dict]: - return self._file_url(file_id, litellm_params), {} + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], + ) -> tuple[str, dict[str, str]]: # mutable-ok: BaseFilesConfig signature + return self._file_url(file_id, litellm_params), _NO_QUERY_PARAMS def transform_retrieve_file_response( self, raw_response: httpx.Response, logging_obj: LiteLLMLoggingObj, - litellm_params: dict, + litellm_params: Mapping[str, object], ) -> OpenAIFileObject: return _to_openai_file_object(MistralFile.model_validate(raw_response.json())) def transform_delete_file_request( self, file_id: str, - optional_params: dict, - litellm_params: dict, - ) -> tuple[str, dict]: - return self._file_url(file_id, litellm_params), {} + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], + ) -> tuple[str, dict[str, str]]: # mutable-ok: BaseFilesConfig signature + return self._file_url(file_id, litellm_params), _NO_QUERY_PARAMS def transform_delete_file_response( self, raw_response: httpx.Response, logging_obj: LiteLLMLoggingObj, - litellm_params: dict, + litellm_params: Mapping[str, object], ) -> FileDeleted: deleted: Final = MistralFileDeleted.model_validate(raw_response.json()) return FileDeleted(id=deleted.id, deleted=deleted.deleted, object="file") @@ -195,32 +215,39 @@ class MistralFilesConfig(BaseFilesConfig): def transform_list_files_request( self, purpose: str | None, - optional_params: dict, - litellm_params: dict, - ) -> tuple[str, dict]: - params: Final = {"purpose": _to_mistral_purpose(purpose)} if purpose else {} - return f"{get_mistral_api_base(litellm_params.get('api_base'))}/v1/files", params + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], + ) -> tuple[str, dict[str, str]]: # mutable-ok: BaseFilesConfig signature + url: Final = f"{_api_base_from(litellm_params)}/v1/files" + if not purpose: + return url, _NO_QUERY_PARAMS + return url, {"purpose": _to_mistral_purpose(purpose)} # mutable-ok: BaseFilesConfig signature returns dict def transform_list_files_response( self, raw_response: httpx.Response, logging_obj: LiteLLMLoggingObj, - litellm_params: dict, - ) -> list[OpenAIFileObject]: - return [_to_openai_file_object(f) for f in MistralFileList.model_validate(raw_response.json()).data] + litellm_params: Mapping[str, object], + ) -> list[OpenAIFileObject]: # mutable-ok: BaseFilesConfig signature + return [ # mutable-ok: BaseFilesConfig signature + _to_openai_file_object(f) for f in MistralFileList.model_validate(raw_response.json()).data + ] def transform_file_content_request( self, file_content_request: FileContentRequest, - optional_params: dict, - litellm_params: dict, - ) -> tuple[str, dict]: - return self._file_url(file_content_request["file_id"], litellm_params, suffix="/content"), {} + optional_params: Mapping[str, object], + litellm_params: Mapping[str, object], + ) -> tuple[str, dict[str, str]]: # mutable-ok: BaseFilesConfig signature + file_id: Final = file_content_request.get("file_id") + if file_id is None: + raise ValueError("file_id is required to download file content") + return self._file_url(file_id, litellm_params, suffix="/content"), _NO_QUERY_PARAMS def transform_file_content_response( self, raw_response: httpx.Response, logging_obj: LiteLLMLoggingObj, - litellm_params: dict, + litellm_params: Mapping[str, object], ) -> HttpxBinaryResponseContent: return HttpxBinaryResponseContent(response=raw_response) diff --git a/litellm/proxy/openai_files_endpoints/files_endpoints.py b/litellm/proxy/openai_files_endpoints/files_endpoints.py index 49a01495d65..b3bb1fa9a01 100644 --- a/litellm/proxy/openai_files_endpoints/files_endpoints.py +++ b/litellm/proxy/openai_files_endpoints/files_endpoints.py @@ -1130,7 +1130,7 @@ async def get_file( check_file_id_encoding=True, ) - if should_route: + if should_route and credentials is not None: # Use model-based routing with credentials from config prepare_data_with_credentials( data=data, diff --git a/tests/test_litellm/batches/test_batch_utils.py b/tests/test_litellm/batches/test_batch_utils.py index 85f267c2db0..56cd3298db6 100644 --- a/tests/test_litellm/batches/test_batch_utils.py +++ b/tests/test_litellm/batches/test_batch_utils.py @@ -645,9 +645,7 @@ async def test_calculate_vertex_disable_transform_needs_model_name(monkeypatch): lambda content, model: pytest.fail("raw vertex path should not run"), ) - result = await bu.calculate_batch_cost_and_usage( - file_content_dictionary=[], custom_llm_provider="vertex_ai" - ) + result = await bu.calculate_batch_cost_and_usage(file_content_dictionary=[], custom_llm_provider="vertex_ai") assert result.cost == 0.0 assert result.usage.total_tokens == 0 assert result.models == [] @@ -1250,6 +1248,7 @@ async def test_handle_completed_batch_no_output_file_is_zero(monkeypatch): result set - zero cost, zero usage, no models - instead of letting the file fetch raise "Output file id is None" on every aretrieve_batch logging poll. """ + # The output-file fetch must not even be attempted when there is no output file. async def _must_not_fetch(*args, **kwargs): pytest.fail("_fetch_batch_output_file_content should not be called") @@ -1376,7 +1375,10 @@ def test_anthropic_response_body_is_result_message(): def test_anthropic_usage_conversion_includes_cache_tokens(): - body = {"model": "claude-sonnet-4-5-20250929", "usage": _anthropic_usage(1000, 200, cache_creation=2000, cache_read=8000)} + body = { + "model": "claude-sonnet-4-5-20250929", + "usage": _anthropic_usage(1000, 200, cache_creation=2000, cache_read=8000), + } usage = bu._get_batch_job_usage_from_response_body(body, custom_llm_provider="anthropic") assert usage.prompt_tokens == 11000 assert usage.completion_tokens == 200 @@ -1391,7 +1393,9 @@ def test_bedrock_model_output_line_success_check(): "modelOutput": {"model": "claude-sonnet-4-6", "usage": {"input_tokens": 13, "output_tokens": 5}}, } assert bu._batch_response_was_successful(row, custom_llm_provider="bedrock") is True - assert bu._get_response_from_batch_job_output_file(row, custom_llm_provider="bedrock")["model"] == "claude-sonnet-4-6" + assert ( + bu._get_response_from_batch_job_output_file(row, custom_llm_provider="bedrock")["model"] == "claude-sonnet-4-6" + ) def test_bedrock_cost_uses_deployment_model_name(): @@ -1445,7 +1449,13 @@ def test_total_usage_without_cache_tokens_has_no_prompt_details(monkeypatch): rows = [ { "custom_id": "req-1", - "response": {"status_code": 200, "body": {"model": "gpt-5.2", "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}}}, + "response": { + "status_code": 200, + "body": { + "model": "gpt-5.2", + "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15}, + }, + }, } ] result = bu._aggregate_batch_cost_usage_models(entries=rows, custom_llm_provider="openai") @@ -1487,7 +1497,9 @@ def test_anthropic_cost_without_model_info_uses_batch_cost_calculator(monkeypatc lambda **kw: pytest.fail("anthropic rows must not go through completion_cost"), ) - result = bu._aggregate_batch_cost_usage_models(entries=[_anthropic_succeeded_row()], custom_llm_provider="anthropic") + result = bu._aggregate_batch_cost_usage_models( + entries=[_anthropic_succeeded_row()], custom_llm_provider="anthropic" + ) assert result.cost == pytest.approx(0.3) assert seen[0]["model"] == "claude-sonnet-4-5-20250929" @@ -1522,7 +1534,11 @@ async def test_calculate_batch_cost_and_usage_anthropic_end_to_end(): ) assert result.cost == pytest.approx(1000 * 3e-6 / 2 + 8000 * 3e-7 / 2 + 2000 * 3.75e-6 / 2 + 200 * 15e-6 / 2) - assert (result.usage.prompt_tokens, result.usage.completion_tokens, result.usage.total_tokens) == (11000, 200, 11200) + assert (result.usage.prompt_tokens, result.usage.completion_tokens, result.usage.total_tokens) == ( + 11000, + 200, + 11200, + ) assert result.models == ["claude-sonnet-4-5"] @@ -1689,7 +1705,10 @@ async def test_handle_completed_batch_honors_deployment_pricing(monkeypatch) -> def test_bedrock_converse_shaped_batch_usage_is_parsed(): - body = {"model": "us.amazon.nova-lite-v1:0", "usage": {"inputTokens": 2202, "outputTokens": 540, "totalTokens": 2742}} + body = { + "model": "us.amazon.nova-lite-v1:0", + "usage": {"inputTokens": 2202, "outputTokens": 540, "totalTokens": 2742}, + } usage = bu._get_batch_job_usage_from_response_body(body, custom_llm_provider="bedrock") assert (usage.prompt_tokens, usage.completion_tokens, usage.total_tokens) == (2202, 540, 2742) @@ -1739,6 +1758,7 @@ def test_unparsable_bedrock_batch_usage_warns(caplog): # batch_cost_is_final # --------------------------------------------------------------------------- # + def _retrieved_batch( status: str, output_file_id: str | None = None, counts: BatchRequestCounts | None = None ) -> LiteLLMBatch: @@ -1798,7 +1818,9 @@ def _ocr_row(pages_processed, annotation_pages=None, model="mistral-ocr-latest") usage_info = {"pages_processed": pages_processed, "doc_size_bytes": 4096} if annotation_pages is not None: usage_info["pages_processed_annotation"] = annotation_pages - return _success_row(model=model, pages=[{"index": i, "markdown": "x"} for i in range(pages_processed)], usage_info=usage_info) + return _success_row( + model=model, pages=[{"index": i, "markdown": "x"} for i in range(pages_processed)], usage_info=usage_info + ) def test_ocr_rows_are_priced_per_page_at_batch_rate(monkeypatch): @@ -1835,7 +1857,9 @@ def test_ocr_rows_bill_annotation_pages_separately(monkeypatch): "annotation_cost_per_page_batches": 0.0025, }, ) - result = bu._aggregate_batch_cost_usage_models(entries=[_ocr_row(4, annotation_pages=4)], custom_llm_provider="mistral") + result = bu._aggregate_batch_cost_usage_models( + entries=[_ocr_row(4, annotation_pages=4)], custom_llm_provider="mistral" + ) assert result.cost == pytest.approx(4 * 0.002 + 4 * 0.0025) diff --git a/tests/test_litellm/batches/test_main.py b/tests/test_litellm/batches/test_main.py index c5a33dd6508..26dc4083b0b 100644 --- a/tests/test_litellm/batches/test_main.py +++ b/tests/test_litellm/batches/test_main.py @@ -66,9 +66,7 @@ def seams(): stack.enter_context(patch.object(bm, "openai_batches_instance", openai_i)) stack.enter_context(patch.object(bm, "azure_batches_instance", azure_i)) stack.enter_context(patch.object(bm, "vertex_ai_batches_instance", vertex_i)) - stack.enter_context( - patch.object(bm, "anthropic_batches_instance", anthropic_i) - ) + stack.enter_context(patch.object(bm, "anthropic_batches_instance", anthropic_i)) stack.enter_context(patch.object(bm, "base_llm_http_handler", base_http)) stack.enter_context(patch.object(bm, "BedrockBatchesHandler", bedrock_arn)) yield Seams( @@ -174,9 +172,7 @@ def test_create__provider_config_routes_to_base_http_handler(seams): "get_provider_batches_config", return_value=MagicMock(name="provider_config"), ): - result = bm.create_batch( - **CREATE_KW, custom_llm_provider="bedrock", model="bedrock/my-batch-model" - ) + result = bm.create_batch(**CREATE_KW, custom_llm_provider="bedrock", model="bedrock/my-batch-model") assert result is seams.base_http.create_batch.return_value _assert_only(seams.base_http.create_batch, seams, "create_batch") @@ -281,9 +277,7 @@ def test_retrieve__bedrock_model_invocation_job_arn(seams): result = bm.retrieve_batch(batch_id=arn, custom_llm_provider="bedrock") seams.bedrock_arn._handle_model_invocation_job_status.assert_called_once() - assert ( - result is seams.bedrock_arn._handle_model_invocation_job_status.return_value - ) + assert result is seams.bedrock_arn._handle_model_invocation_job_status.return_value seams.bedrock_arn._handle_async_invoke_status.assert_not_called() @@ -385,9 +379,7 @@ def test_cancel__unsupported_provider_raises_badrequest(seams): def test_cancel__async_flag_propagates_is_async(seams): - bm.cancel_batch( - batch_id="batch-1", custom_llm_provider="openai", acancel_batch=True - ) + bm.cancel_batch(batch_id="batch-1", custom_llm_provider="openai", acancel_batch=True) assert seams.openai.cancel_batch.call_args.kwargs["_is_async"] is True @@ -415,9 +407,7 @@ async def test_acreate_batch_delegates_to_create_batch(): @pytest.mark.asyncio async def test_aretrieve_batch_delegates_to_retrieve_batch(): with patch.object(bm, "retrieve_batch", MagicMock(return_value="SENTINEL")) as m: - result = await bm.aretrieve_batch( - batch_id="batch-1", custom_llm_provider="azure" - ) + result = await bm.aretrieve_batch(batch_id="batch-1", custom_llm_provider="azure") assert result == "SENTINEL" assert m.call_count == 1 @@ -429,9 +419,7 @@ async def test_aretrieve_batch_delegates_to_retrieve_batch(): @pytest.mark.asyncio async def test_alist_batches_delegates_to_list_batches(): with patch.object(bm, "list_batches", MagicMock(return_value="SENTINEL")) as m: - result = await bm.alist_batches( - after="cur", limit=3, custom_llm_provider="vertex_ai" - ) + result = await bm.alist_batches(after="cur", limit=3, custom_llm_provider="vertex_ai") assert result == "SENTINEL" assert m.call_count == 1 @@ -444,9 +432,7 @@ async def test_alist_batches_delegates_to_list_batches(): @pytest.mark.asyncio async def test_acancel_batch_delegates_to_cancel_batch(): with patch.object(bm, "cancel_batch", MagicMock(return_value="SENTINEL")) as m: - result = await bm.acancel_batch( - batch_id="batch-1", custom_llm_provider="openai" - ) + result = await bm.acancel_batch(batch_id="batch-1", custom_llm_provider="openai") assert result == "SENTINEL" assert m.call_count == 1 @@ -499,9 +485,7 @@ def _sent(mock_method, *keys): def test_create__openai_credentials_passthrough(seams): bm.create_batch(**CREATE_KW, custom_llm_provider="openai", **OPENAI_CREDS) - assert _sent( - seams.openai.create_batch, "api_key", "api_base", "organization", "max_retries" - ) == { + assert _sent(seams.openai.create_batch, "api_key", "api_base", "organization", "max_retries") == { "api_key": "sk-user-openai", "api_base": "https://openai.user.test", "organization": "org-user-123", @@ -512,9 +496,7 @@ def test_create__openai_credentials_passthrough(seams): def test_create__azure_credentials_passthrough(seams): bm.create_batch(**CREATE_KW, custom_llm_provider="azure", **AZURE_CREDS) - assert _sent( - seams.azure.create_batch, "api_key", "api_base", "api_version" - ) == { + assert _sent(seams.azure.create_batch, "api_key", "api_base", "api_version") == { "api_key": "sk-user-azure", "api_base": "https://azure.user.test", "api_version": "2024-12-99", @@ -564,9 +546,7 @@ def test_create__provider_config_credentials_passthrough(seams): def test_retrieve__openai_credentials_passthrough(seams): bm.retrieve_batch(batch_id="b1", custom_llm_provider="openai", **OPENAI_CREDS) - assert _sent( - seams.openai.retrieve_batch, "api_key", "api_base", "organization" - ) == { + assert _sent(seams.openai.retrieve_batch, "api_key", "api_base", "organization") == { "api_key": "sk-user-openai", "api_base": "https://openai.user.test", "organization": "org-user-123", @@ -576,9 +556,7 @@ def test_retrieve__openai_credentials_passthrough(seams): def test_retrieve__azure_credentials_passthrough(seams): bm.retrieve_batch(batch_id="b1", custom_llm_provider="azure", **AZURE_CREDS) - assert _sent( - seams.azure.retrieve_batch, "api_key", "api_base", "api_version" - ) == { + assert _sent(seams.azure.retrieve_batch, "api_key", "api_base", "api_version") == { "api_key": "sk-user-azure", "api_base": "https://azure.user.test", "api_version": "2024-12-99", @@ -640,9 +618,7 @@ def test_retrieve__provider_config_credentials_passthrough(seams): def test_list__openai_credentials_passthrough(seams): bm.list_batches(custom_llm_provider="openai", **OPENAI_CREDS) - assert _sent( - seams.openai.list_batches, "api_key", "api_base", "organization" - ) == { + assert _sent(seams.openai.list_batches, "api_key", "api_base", "organization") == { "api_key": "sk-user-openai", "api_base": "https://openai.user.test", "organization": "org-user-123", @@ -652,9 +628,7 @@ def test_list__openai_credentials_passthrough(seams): def test_list__azure_credentials_passthrough(seams): bm.list_batches(custom_llm_provider="azure", **AZURE_CREDS) - assert _sent( - seams.azure.list_batches, "api_key", "api_base", "api_version" - ) == { + assert _sent(seams.azure.list_batches, "api_key", "api_base", "api_version") == { "api_key": "sk-user-azure", "api_base": "https://azure.user.test", "api_version": "2024-12-99", @@ -682,9 +656,7 @@ def test_list__vertex_credentials_passthrough(seams): def test_cancel__openai_credentials_passthrough(seams): bm.cancel_batch(batch_id="b1", custom_llm_provider="openai", **OPENAI_CREDS) - assert _sent( - seams.openai.cancel_batch, "api_key", "api_base", "organization" - ) == { + assert _sent(seams.openai.cancel_batch, "api_key", "api_base", "organization") == { "api_key": "sk-user-openai", "api_base": "https://openai.user.test", "organization": "org-user-123", @@ -694,9 +666,7 @@ def test_cancel__openai_credentials_passthrough(seams): def test_cancel__azure_credentials_passthrough(seams): bm.cancel_batch(batch_id="b1", custom_llm_provider="azure", **AZURE_CREDS) - assert _sent( - seams.azure.cancel_batch, "api_key", "api_base", "api_version" - ) == { + assert _sent(seams.azure.cancel_batch, "api_key", "api_base", "api_version") == { "api_key": "sk-user-azure", "api_base": "https://azure.user.test", "api_version": "2024-12-99", @@ -786,18 +756,16 @@ def test_retrieve__omits_trusted_model_credentials_when_not_supplied(seams): def test_create__mistral_ocr_routes_to_base_http_handler_with_mistral_config(seams): - with patch.object(bm.ProviderConfigManager, "get_provider_batches_config", wraps=bm.ProviderConfigManager.get_provider_batches_config) as get_cfg: - result = bm.create_batch( - completion_window="24h", - endpoint="/v1/ocr", - input_file_id="file-abc", - custom_llm_provider="mistral", - model="mistral/mistral-ocr-latest", - ) + result = bm.create_batch( + completion_window="24h", + endpoint="/v1/ocr", + input_file_id="file-abc", + custom_llm_provider="mistral", + model="mistral/mistral-ocr-latest", + ) assert result is seams.base_http.create_batch.return_value _assert_only(seams.base_http.create_batch, seams, "create_batch") - get_cfg.assert_called_once() forwarded = seams.base_http.create_batch.call_args.kwargs assert type(forwarded["provider_config"]).__name__ == "MistralBatchesConfig" assert forwarded["model"] == "mistral-ocr-latest" diff --git a/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py b/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py index 03e9c351a30..03cfeedece2 100644 --- a/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py +++ b/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py @@ -92,26 +92,34 @@ def test_create_request_maps_openai_fields_onto_mistral_job(config): model="mistral-ocr-latest", create_batch_data=data, optional_params={}, litellm_params={} ) assert body == { - "input_files": ["file-123"], + "input_files": ("file-123",), "endpoint": "/v1/ocr", "model": "mistral-ocr-latest", "metadata": {"team": "docs"}, } -def test_create_request_omits_empty_metadata_and_forwards_extra_body(config): +def test_create_request_omits_empty_metadata(config): data = CreateBatchRequest( completion_window="24h", endpoint="/v1/chat/completions", input_file_id="file-123", metadata=None, - extra_body={"timeout_hours": 48}, ) body = config.transform_create_batch_request( model="mistral-small-latest", create_batch_data=data, optional_params={}, litellm_params={} ) assert "metadata" not in body - assert body["timeout_hours"] == 48 + + +def test_create_request_requires_input_file_and_endpoint(config): + with pytest.raises(ValueError, match="input_file_id and endpoint are required"): + config.transform_create_batch_request( + model="m", + create_batch_data=CreateBatchRequest(completion_window="24h"), + optional_params={}, + litellm_params={}, + ) @pytest.mark.parametrize( diff --git a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py index d6ad8a34b35..f62645be7ee 100644 --- a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py +++ b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py @@ -79,7 +79,9 @@ def test_validate_environment_uses_bearer_auth(config, api_key): def test_upload_request_is_multipart_with_batch_purpose(config): body = config.transform_create_file_request( model="", - create_file_data=CreateFileRequest(file=("in.jsonl", b'{"custom_id":"0"}\n', "application/jsonl"), purpose="batch"), + create_file_data=CreateFileRequest( + file=("in.jsonl", b'{"custom_id":"0"}\n', "application/jsonl"), purpose="batch" + ), optional_params={}, litellm_params={}, ) @@ -181,7 +183,9 @@ def test_list_request_filters_by_mapped_purpose(config): def test_list_response(config): out = config.transform_list_files_response( - raw_response=_response({"data": [_file(), _file(id="second", filename="b.jsonl")], "object": "list", "total": 2}), + raw_response=_response( + {"data": [_file(), _file(id="second", filename="b.jsonl")], "object": "list", "total": 2} + ), logging_obj=None, litellm_params={}, ) From 91e7df3d8fe83c37268b61080a37e48ef0e63e3f Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Thu, 10 Sep 2026 13:35:35 -0400 Subject: [PATCH 05/37] chore: regenerate cost-map schema and UI API types, drop test banner comments --- model_prices_and_context_window.schema.json | 8 ++++++++ tests/test_litellm/batches/test_batch_utils.py | 5 ----- .../batches/test_mistral_batches_transformation.py | 10 ---------- .../llms/mistral/ocr/test_mistral_ocr_cost.py | 1 - ui/litellm-dashboard/src/lib/http/schema.d.ts | 8 ++++++++ 5 files changed, 16 insertions(+), 16 deletions(-) diff --git a/model_prices_and_context_window.schema.json b/model_prices_and_context_window.schema.json index 47a1934a703..09385db728b 100644 --- a/model_prices_and_context_window.schema.json +++ b/model_prices_and_context_window.schema.json @@ -53,6 +53,10 @@ "type": "number", "minimum": 0 }, + "annotation_cost_per_page_batches": { + "type": "number", + "minimum": 0 + }, "audio_transcription_config": { "type": "string" }, @@ -432,6 +436,10 @@ "type": "number", "minimum": 0 }, + "ocr_cost_per_page_batches": { + "type": "number", + "minimum": 0 + }, "output_cost_per_audio_token": { "type": "number", "minimum": 0 diff --git a/tests/test_litellm/batches/test_batch_utils.py b/tests/test_litellm/batches/test_batch_utils.py index 56cd3298db6..0d66c0eb5ec 100644 --- a/tests/test_litellm/batches/test_batch_utils.py +++ b/tests/test_litellm/batches/test_batch_utils.py @@ -1809,11 +1809,6 @@ class TestBatchCostIsFinal: assert bu.batch_cost_is_final(_retrieved_batch(status)) is True -# =========================================================================== # -# OCR batch output lines (Mistral /v1/ocr batches) are billed per page, not per token -# =========================================================================== # - - def _ocr_row(pages_processed, annotation_pages=None, model="mistral-ocr-latest"): usage_info = {"pages_processed": pages_processed, "doc_size_bytes": 4096} if annotation_pages is not None: diff --git a/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py b/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py index 03cfeedece2..4073879e3b8 100644 --- a/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py +++ b/tests/test_litellm/llms/mistral/batches/test_mistral_batches_transformation.py @@ -76,11 +76,6 @@ def test_custom_llm_provider(config): assert config.custom_llm_provider == LlmProviders.MISTRAL -# --------------------------------------------------------------------------- # -# create -# --------------------------------------------------------------------------- # - - def test_create_request_maps_openai_fields_onto_mistral_job(config): data = CreateBatchRequest( completion_window="24h", @@ -175,11 +170,6 @@ def test_create_response_maps_job_onto_openai_batch(config): assert batch.metadata == {"job_type": "testing"} -# --------------------------------------------------------------------------- # -# retrieve -# --------------------------------------------------------------------------- # - - def test_retrieve_request_is_presigned_get_with_auth(config, api_key): req = config.transform_retrieve_batch_request( batch_id="job/with slash", optional_params={}, litellm_params={"api_base": "https://api.mistral.ai"} diff --git a/tests/test_litellm/llms/mistral/ocr/test_mistral_ocr_cost.py b/tests/test_litellm/llms/mistral/ocr/test_mistral_ocr_cost.py index 9fe6f38003f..d72e866949f 100644 --- a/tests/test_litellm/llms/mistral/ocr/test_mistral_ocr_cost.py +++ b/tests/test_litellm/llms/mistral/ocr/test_mistral_ocr_cost.py @@ -63,7 +63,6 @@ def test_ocr4_cost_scales_with_pages(model: str, pages_processed: int) -> None: assert cost == pytest.approx(OCR4_COST_PER_PAGE * pages_processed) - @pytest.mark.parametrize("cost_map_path", [MAIN_COST_MAP, BACKUP_COST_MAP]) def test_ocr3_pricing_entry(cost_map_path: Path) -> None: with open(cost_map_path) as f: diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index 0f0c1fc9af4..5d1ed79098e 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -29493,6 +29493,8 @@ export interface components { allow_client_keepalive_override: boolean | null; /** Annotation Cost Per Page */ annotation_cost_per_page?: number | null; + /** Annotation Cost Per Page Batches */ + annotation_cost_per_page_batches?: number | null; /** Api Base */ api_base?: string | null; /** Api Key */ @@ -29704,6 +29706,8 @@ export interface components { ocr_cost_per_credit?: number | null; /** Ocr Cost Per Page */ ocr_cost_per_page?: number | null; + /** Ocr Cost Per Page Batches */ + ocr_cost_per_page_batches?: number | null; /** Organization */ organization?: string | null; /** Otpm */ @@ -39684,6 +39688,8 @@ export interface components { allow_client_keepalive_override: boolean | null; /** Annotation Cost Per Page */ annotation_cost_per_page?: number | null; + /** Annotation Cost Per Page Batches */ + annotation_cost_per_page_batches?: number | null; /** Api Base */ api_base?: string | null; /** Api Key */ @@ -39895,6 +39901,8 @@ export interface components { ocr_cost_per_credit?: number | null; /** Ocr Cost Per Page */ ocr_cost_per_page?: number | null; + /** Ocr Cost Per Page Batches */ + ocr_cost_per_page_batches?: number | null; /** Organization */ organization?: string | null; /** Otpm */ From edd5727f3c9077f449eec37367ace43945e649fa Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Thu, 10 Sep 2026 17:28:27 -0400 Subject: [PATCH 06/37] fix(proxy): enforce key/team/org/project model grants on model-routed file and batch credentials Files and batches routes take their model from a header, query param or a model-encoded resource id, which the auth layer never sees, so any key could name any deployment and act on that provider account with its server-side key. Every caller-supplied model now goes through can_key_call_resolved_model before deployment credentials are resolved, covering file create/retrieve/content/ delete/list, batch create/retrieve/list/cancel, and vector store files. --- litellm/proxy/batches_endpoints/endpoints.py | 17 +- .../openai_files_endpoints/common_utils.py | 61 +++++- .../openai_files_endpoints/files_endpoints.py | 20 +- .../vector_store_files_endpoints/endpoints.py | 6 +- .../proxy/batches_endpoints/test_endpoints.py | 58 +++++- .../test_files_endpoint.py | 184 ++++++++++++++++-- .../test_batch_x_litellm_model_encoding.py | 53 ++--- 7 files changed, 322 insertions(+), 77 deletions(-) diff --git a/litellm/proxy/batches_endpoints/endpoints.py b/litellm/proxy/batches_endpoints/endpoints.py index 5c4bacd757c..c99f66d032e 100644 --- a/litellm/proxy/batches_endpoints/endpoints.py +++ b/litellm/proxy/batches_endpoints/endpoints.py @@ -34,9 +34,9 @@ from litellm.proxy.openai_files_endpoints.common_utils import ( encode_batch_response_ids, encode_file_id_with_model, ensure_batch_response_managed_file_ids, + get_authorized_credentials_for_model, get_batch_from_database, get_batch_id_from_unified_batch_id, - get_credentials_for_model, get_model_id_from_unified_batch_id, get_models_from_unified_file_id, get_original_file_id, @@ -218,9 +218,10 @@ async def create_batch( # SCENARIO 1: File ID is encoded with model info if model_from_file_id is not None and input_file_id: - credentials = get_credentials_for_model( + credentials = await get_authorized_credentials_for_model( llm_router=llm_router, model_id=model_from_file_id, + user_api_key_dict=user_api_key_dict, operation_context="batch creation (file created with model)", ) @@ -310,9 +311,10 @@ async def create_batch( # SCENARIO 2 & 3: Model from header/query OR custom_llm_provider fallback if model_param: # SCENARIO 2: Use model-based routing from header/query/body - credentials = get_credentials_for_model( + credentials = await get_authorized_credentials_for_model( llm_router=llm_router, model_id=model_param, + user_api_key_dict=user_api_key_dict, operation_context="batch creation", ) @@ -540,9 +542,10 @@ async def retrieve_batch( # Retrieve from provider (for non-terminal states or if DB lookup failed) # SCENARIO 1: Batch ID is encoded with model info if model_from_id is not None: - credentials: Final = get_credentials_for_model( + credentials: Final = await get_authorized_credentials_for_model( llm_router=llm_router, model_id=model_from_id, + user_api_key_dict=user_api_key_dict, operation_context="batch retrieval (batch created with model)", ) @@ -764,9 +767,10 @@ async def list_batches( data.get("model") or request.query_params.get("model") or request.headers.get("x-litellm-model") ): # SCENARIO 2: Use model-based routing from header/query/body - credentials: Final = get_credentials_for_model( + credentials: Final = await get_authorized_credentials_for_model( llm_router=llm_router, model_id=model_param, + user_api_key_dict=user_api_key_dict, operation_context="batch listing", ) @@ -952,9 +956,10 @@ async def cancel_batch( # SCENARIO 1: Batch ID is encoded with model info if model_from_id is not None: - credentials: Final = get_credentials_for_model( + credentials: Final = await get_authorized_credentials_for_model( llm_router=llm_router, model_id=model_from_id, + user_api_key_dict=user_api_key_dict, operation_context="batch cancellation (batch created with model)", ) diff --git a/litellm/proxy/openai_files_endpoints/common_utils.py b/litellm/proxy/openai_files_endpoints/common_utils.py index b1f282a0978..4202a6d1689 100644 --- a/litellm/proxy/openai_files_endpoints/common_utils.py +++ b/litellm/proxy/openai_files_endpoints/common_utils.py @@ -351,6 +351,10 @@ def get_credentials_for_model( """ Retrieve API credentials for a model from the LLM Router. + Does not check whether the caller may use ``model_id``; use + ``get_authorized_credentials_for_model`` for anything driven by a caller-supplied + model name (request body, header, query param, or a model-encoded resource id). + Args: llm_router: LiteLLM Router instance model_id: Model name or deployment ID @@ -381,6 +385,48 @@ def get_credentials_for_model( return credentials +async def authorize_model_for_key( + model_id: str, + llm_router: Optional["Router"], + user_api_key_dict: "UserAPIKeyAuth", +) -> None: + """ + Enforce the caller's model grants on a model name the auth layer never saw. + + The files and batches routes carry their model in a header, query param, or a + model-encoded resource id rather than the request body, so ``user_api_key_auth`` + cannot check it. Run the same key, team (incl. team-member and access-group + fallbacks), org and project allowlist checks a chat request would get, so a + restricted key cannot borrow another deployment's server-side credentials. + + Raises: + ProxyException (403): the caller is not allowed to use ``model_id`` + """ + from litellm.proxy.auth.auth_checks import can_key_call_resolved_model + + await can_key_call_resolved_model( + model=model_id, + llm_model_list=None, + valid_token=user_api_key_dict, + llm_router=llm_router, + ) + + +async def get_authorized_credentials_for_model( + llm_router: Optional["Router"], + model_id: str, + user_api_key_dict: "UserAPIKeyAuth", + operation_context: str = "file operation", +) -> dict: # mutable-ok: same contract as get_credentials_for_model, callers merge it into request data + """``get_credentials_for_model`` gated by ``authorize_model_for_key``.""" + await authorize_model_for_key(model_id=model_id, llm_router=llm_router, user_api_key_dict=user_api_key_dict) + return get_credentials_for_model( + llm_router=llm_router, + model_id=model_id, + operation_context=operation_context, + ) + + def get_team_provider_credentials( llm_router: Optional["Router"], user_api_key_dict: "UserAPIKeyAuth", @@ -573,21 +619,27 @@ def prepare_data_with_credentials( data["file_id"] = file_id -def handle_model_based_routing( +async def handle_model_based_routing( file_id: str, request, # FastAPI Request object llm_router, # Router instance data: dict, + user_api_key_dict: "UserAPIKeyAuth", check_file_id_encoding: bool = True, ) -> tuple[bool, str | None, str | None, dict | None]: """ Orchestrate model-based credential routing for file operations. + The model name comes from the caller (embedded in the file id, or a header, query + param or body field), so it is authorized against the caller's key, team, org and + project grants before any deployment credentials are resolved. + Args: file_id: File ID (may contain embedded model info) request: FastAPI request object llm_router: LiteLLM Router instance data: Request data dictionary + user_api_key_dict: The authenticated caller check_file_id_encoding: Whether to check for embedded model in file_id Returns: @@ -599,6 +651,7 @@ def handle_model_based_routing( Raises: HTTPException: If router unavailable or model not found + ProxyException: If the caller is not allowed to use the model """ model_from_id, model_from_param = extract_model_from_sources( file_id=file_id, @@ -608,9 +661,10 @@ def handle_model_based_routing( # Priority 1: Model embedded in file_id if check_file_id_encoding and model_from_id is not None: - credentials = get_credentials_for_model( + credentials = await get_authorized_credentials_for_model( llm_router=llm_router, model_id=model_from_id, + user_api_key_dict=user_api_key_dict, operation_context=f"file operation (file created with model '{model_from_id}')", ) original_file_id: Final = get_original_file_id(file_id) @@ -618,9 +672,10 @@ def handle_model_based_routing( # Priority 2: Model from header/query/body elif model_from_param is not None: - credentials = get_credentials_for_model( + credentials = await get_authorized_credentials_for_model( llm_router=llm_router, model_id=model_from_param, + user_api_key_dict=user_api_key_dict, operation_context="file operation", ) return True, model_from_param, None, credentials diff --git a/litellm/proxy/openai_files_endpoints/files_endpoints.py b/litellm/proxy/openai_files_endpoints/files_endpoints.py index b3bb1fa9a01..0efd618e171 100644 --- a/litellm/proxy/openai_files_endpoints/files_endpoints.py +++ b/litellm/proxy/openai_files_endpoints/files_endpoints.py @@ -68,7 +68,7 @@ from litellm.proxy.openai_files_endpoints.common_utils import ( apply_team_provider_credentials, encode_file_id_with_model, extract_file_creation_params, - get_credentials_for_model, + get_authorized_credentials_for_model, handle_model_based_routing, prepare_data_with_credentials, validate_file_list_limit, @@ -267,9 +267,10 @@ async def route_create_file( # NEW: Handle model-based routing (no DB required) if model is not None: # Get credentials from model_list via router - credentials: Final = get_credentials_for_model( + credentials: Final = await get_authorized_credentials_for_model( llm_router=llm_router, model_id=model, + user_api_key_dict=user_api_key_dict, operation_context="file upload", ) @@ -907,11 +908,12 @@ async def get_file_content( model_used, original_file_id, credentials, - ) = handle_model_based_routing( + ) = await handle_model_based_routing( file_id=file_id, request=request, llm_router=llm_router, data=data, + user_api_key_dict=user_api_key_dict, check_file_id_encoding=True, ) @@ -1122,11 +1124,12 @@ async def get_file( model_used, original_file_id, credentials, - ) = handle_model_based_routing( + ) = await handle_model_based_routing( file_id=file_id, request=request, llm_router=llm_router, data=data, + user_api_key_dict=user_api_key_dict, check_file_id_encoding=True, ) @@ -1330,11 +1333,12 @@ async def delete_file( model_used, original_file_id, credentials, - ) = handle_model_based_routing( + ) = await handle_model_based_routing( file_id=file_id, request=request, llm_router=llm_router, data=data, + user_api_key_dict=user_api_key_dict, check_file_id_encoding=True, ) @@ -1517,11 +1521,12 @@ async def list_files( response: Any | None = None # Check for model-based credential routing (no file_id encoding check for list) - should_route, model_used, _, credentials = handle_model_based_routing( + should_route, model_used, _, credentials = await handle_model_based_routing( file_id="", # No file_id for list endpoint request=request, llm_router=llm_router, data=data, + user_api_key_dict=user_api_key_dict, check_file_id_encoding=False, ) @@ -1548,9 +1553,10 @@ async def list_files( status_code=500, detail="LLM Router not initialized. Ensure models added to proxy.", ) - credentials = get_credentials_for_model( + credentials = await get_authorized_credentials_for_model( llm_router=llm_router, model_id=target_model_names_list[0], + user_api_key_dict=user_api_key_dict, operation_context="file list", ) prepare_data_with_credentials(data=data, credentials=credentials) diff --git a/litellm/proxy/vector_store_files_endpoints/endpoints.py b/litellm/proxy/vector_store_files_endpoints/endpoints.py index 957ed9fd0b9..11ef8efb598 100644 --- a/litellm/proxy/vector_store_files_endpoints/endpoints.py +++ b/litellm/proxy/vector_store_files_endpoints/endpoints.py @@ -144,11 +144,12 @@ async def _update_request_data_with_managed_file_id( model_used, original_file_id, credentials, - ) = handle_model_based_routing( + ) = await handle_model_based_routing( file_id=file_id, request=request, llm_router=llm_router, data=data, + user_api_key_dict=user_api_key_dict, check_file_id_encoding=True, ) @@ -273,11 +274,12 @@ async def _update_request_data_with_model_routing_hint( _model_used, _original_file_id, credentials, - ) = handle_model_based_routing( + ) = await handle_model_based_routing( file_id="", request=request, llm_router=llm_router, data=data, + user_api_key_dict=user_api_key_dict, check_file_id_encoding=False, ) diff --git a/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py b/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py index a37c8ff2bb4..57e42e79a42 100644 --- a/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py +++ b/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py @@ -177,6 +177,8 @@ def harness(): logging.get_proxy_hook = MagicMock(return_value=None) router = MagicMock(spec=Router) + router.model_group_alias = {} + router.get_model_access_groups = MagicMock(return_value={}) router.acreate_batch = AsyncMock(return_value=make_batch()) router.get_deployment_credentials_with_provider = MagicMock(side_effect=_creds_lookup) @@ -1161,6 +1163,8 @@ def retrieve_harness(): logging.get_proxy_hook = MagicMock(return_value=None) router = MagicMock(spec=Router) + router.model_group_alias = {} + router.get_model_access_groups = MagicMock(return_value={}) router.aretrieve_batch = AsyncMock(return_value=make_batch()) router.get_deployment_credentials_with_provider = MagicMock(side_effect=_creds_lookup) @@ -1616,6 +1620,8 @@ def list_harness(): logging.get_proxy_hook = MagicMock(return_value=None) router = MagicMock(spec=Router) + router.model_group_alias = {} + router.get_model_access_groups = MagicMock(return_value={}) router.alist_batches = AsyncMock(return_value=FakeListPage([])) router.get_deployment_credentials_with_provider = MagicMock(side_effect=_creds_lookup) @@ -2012,6 +2018,8 @@ def cancel_harness(): logging.get_proxy_hook = MagicMock(return_value=None) router = MagicMock(spec=Router) + router.model_group_alias = {} + router.get_model_access_groups = MagicMock(return_value={}) router.acancel_batch = AsyncMock(return_value=make_batch()) router.get_deployment_credentials_with_provider = MagicMock(side_effect=_creds_lookup) @@ -2733,8 +2741,6 @@ async def test_cancel__unified_batch_id_allowed_when_managed_files_required(canc assert cancel_harness.router_acancel.call_count == 1 - - @pytest.mark.asyncio async def test_retrieve__managed_batch_defers_cost_to_the_poller_when_it_is_running(retrieve_harness): with patch.object(endpoints, "batch_cost_poller_is_active", MagicMock(return_value=True)): @@ -2762,3 +2768,51 @@ async def test_retrieve__raw_batch_id_is_untouched_by_the_poller_handoff(retriev metadata = retrieve_harness.litellm_aretrieve.await_args.kwargs.get("litellm_metadata") or {} assert metadata.get("batch_ignore_default_logging") is None + + +def _key_restricted_to(*models: str) -> UserAPIKeyAuth: + return UserAPIKeyAuth(api_key="sk-restricted", team_id="team-a", team_models=list(models), models=list(models)) + + +@pytest.mark.asyncio +async def test_create__header_model_rejects_key_without_model_grant(harness): + """A key not granted the model named in x-litellm-model must not receive that deployment's credentials.""" + set_body(harness, {"input_file_id": "file-plain", "endpoint": "/v1/chat/completions", "completion_window": "24h"}) + + with pytest.raises(ProxyException) as exc_info: + await call_create(harness, user=_key_restricted_to("azure/gpt-4o"), headers={"x-litellm-model": "vertex-model"}) + + assert exc_info.value.code == "403" + harness.creds_resolver.assert_not_called() + harness.litellm_acreate.assert_not_called() + + +@pytest.mark.asyncio +async def test_create__header_model_allows_key_with_model_grant(harness): + set_body(harness, {"input_file_id": "file-plain", "endpoint": "/v1/chat/completions", "completion_window": "24h"}) + + await call_create(harness, user=_key_restricted_to("vertex-model"), headers={"x-litellm-model": "vertex-model"}) + + harness.creds_resolver.assert_called_once_with(model_id="vertex-model") + assert harness.acreate_kwargs()["custom_llm_provider"] == "vertex_ai" + + +@pytest.mark.asyncio +async def test_retrieve__model_encoded_id_rejects_key_without_model_grant(retrieve_harness): + """The model embedded in a batch id is caller-controlled, so it is checked against the key's grants too.""" + with pytest.raises(ProxyException) as exc_info: + await call_retrieve(retrieve_harness, AZURE_BATCH_ID, user=_key_restricted_to("vertex-model")) + + assert exc_info.value.code == "403" + retrieve_harness.creds_resolver.assert_not_called() + retrieve_harness.litellm_aretrieve.assert_not_called() + + +@pytest.mark.asyncio +async def test_cancel__model_encoded_id_rejects_key_without_model_grant(cancel_harness): + with pytest.raises(ProxyException) as exc_info: + await call_cancel(cancel_harness, AZURE_BATCH_ID, user=_key_restricted_to("vertex-model")) + + assert exc_info.value.code == "403" + cancel_harness.creds_resolver.assert_not_called() + cancel_harness.litellm_acancel.assert_not_called() diff --git a/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py b/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py index 5faae166fca..548c0eb0d91 100644 --- a/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py +++ b/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py @@ -1877,7 +1877,7 @@ def test_get_file_content_streams_openai_direct_path( monkeypatch.setattr(litellm, "afile_content", _mock_afile_content) monkeypatch.setattr( "litellm.proxy.openai_files_endpoints.files_endpoints.handle_model_based_routing", - lambda **kwargs: (False, None, None, None), + AsyncMock(return_value=(False, None, None, None)), ) app.dependency_overrides[ps.user_api_key_auth] = lambda: UserAPIKeyAuth( @@ -1942,15 +1942,17 @@ def test_get_file_content_routed_provider_skips_streaming_when_resolved_provider ) monkeypatch.setattr( "litellm.proxy.openai_files_endpoints.files_endpoints.handle_model_based_routing", - lambda **kwargs: ( - True, - "azure-gpt-3-5-turbo", - "file-original-123", - { - "custom_llm_provider": "azure", - "api_key": "azure-key", - "api_base": "https://azure.example.com", - }, + AsyncMock( + return_value=( + True, + "azure-gpt-3-5-turbo", + "file-original-123", + { + "custom_llm_provider": "azure", + "api_key": "azure-key", + "api_base": "https://azure.example.com", + }, + ) ), ) @@ -2015,7 +2017,7 @@ def test_get_file_content_non_openai_provider_skips_streaming_handler( ) monkeypatch.setattr( "litellm.proxy.openai_files_endpoints.files_endpoints.handle_model_based_routing", - lambda **kwargs: (False, None, None, None), + AsyncMock(return_value=(False, None, None, None)), ) app.dependency_overrides[ps.user_api_key_auth] = lambda: UserAPIKeyAuth( @@ -2463,14 +2465,16 @@ def test_list_files_model_routing_does_not_forward_custom_llm_provider_twice( monkeypatch.setattr(litellm, "afile_list", _mock_afile_list) monkeypatch.setattr( "litellm.proxy.openai_files_endpoints.files_endpoints.handle_model_based_routing", - lambda **kwargs: ( - True, - "azure-gpt-4o", - None, - { - "custom_llm_provider": "azure", - "api_key": "azure-key", - }, + AsyncMock( + return_value=( + True, + "azure-gpt-4o", + None, + { + "custom_llm_provider": "azure", + "api_key": "azure-key", + }, + ) ), ) @@ -4878,3 +4882,145 @@ def test_get_file_model_routed_id_forwards_deployment_provider(mocker: MockerFix assert captured_kwargs["api_key"] == "mistral-key" assert captured_kwargs["file_id"] == "7a13fa8e-fcf8-42c5-aa61-c93c10e2c7df" assert response.json()["id"] == encoded_id + + +def _mistral_plus_anthropic_router() -> Router: + return Router( + model_list=[ + { + "model_name": "mistral-ocr", + "litellm_params": {"model": "mistral/mistral-ocr-latest", "api_key": "mistral-key"}, + "model_info": {"id": "mistral-ocr-id"}, + }, + { + "model_name": "claude-opus-4-6", + "litellm_params": {"model": "anthropic/claude-opus-4-6", "api_key": "anthropic-key"}, + "model_info": {"id": "claude-id"}, + }, + ] + ) + + +def _restricted_key(key_models: list[str]) -> UserAPIKeyAuth: + from litellm.proxy._types import LitellmUserRoles + + return UserAPIKeyAuth( + api_key="test-key", + user_role=LitellmUserRoles.INTERNAL_USER, + user_id="test-user", + team_id="team-a", + team_models=["claude-opus-4-6", "mistral-ocr"], + models=key_models, + ) + + +@pytest.mark.parametrize( + "http_method, path_suffix, litellm_fn", + [ + ("get", "", "afile_retrieve"), + ("get", "/content", "afile_content"), + ("delete", "", "afile_delete"), + ], +) +def test_model_routed_file_ops_reject_key_without_model_grant( + mocker: MockerFixture, monkeypatch, http_method: str, path_suffix: str, litellm_fn: str +): + """ + Regression: a key whose allowlist does not include the deployment named in a + model-encoded file id must be refused before that deployment's server-side + credentials are resolved. Previously any key could name any deployment via the + id (or the x-litellm-model header) and act on that provider account's files. + """ + import litellm.proxy.proxy_server as ps + from litellm.proxy.openai_files_endpoints.common_utils import encode_file_id_with_model + + router = _mistral_plus_anthropic_router() + proxy_logging_obj = setup_proxy_logging_object(monkeypatch, router) + monkeypatch.setattr("litellm.proxy.proxy_server.master_key", None) + monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", None) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", router) + proxy_logging_obj.update_request_status = mocker.AsyncMock() + proxy_logging_obj.post_call_failure_hook = mocker.AsyncMock() + + upstream = mocker.AsyncMock(side_effect=AssertionError("provider must not be called")) + monkeypatch.setattr(litellm, litellm_fn, upstream) + app.dependency_overrides[ps.user_api_key_auth] = lambda: _restricted_key(["claude-opus-4-6"]) + encoded_id = encode_file_id_with_model("7a13fa8e-fcf8-42c5-aa61-c93c10e2c7df", "mistral-ocr") + + try: + response = getattr(client, http_method)( + f"/v1/files/{encoded_id}{path_suffix}", headers={"Authorization": "Bearer test-key"} + ) + finally: + app.dependency_overrides.pop(ps.user_api_key_auth, None) + + assert response.status_code == 403, response.text + assert "not allowed to access model" in response.text + upstream.assert_not_called() + + +def test_list_files_header_model_rejects_key_without_model_grant(mocker: MockerFixture, monkeypatch): + import litellm.proxy.proxy_server as ps + + router = _mistral_plus_anthropic_router() + proxy_logging_obj = setup_proxy_logging_object(monkeypatch, router) + monkeypatch.setattr("litellm.proxy.proxy_server.master_key", None) + monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", None) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", router) + proxy_logging_obj.update_request_status = mocker.AsyncMock() + proxy_logging_obj.post_call_failure_hook = mocker.AsyncMock() + + upstream = mocker.AsyncMock(side_effect=AssertionError("provider must not be called")) + monkeypatch.setattr(litellm, "afile_list", upstream) + app.dependency_overrides[ps.user_api_key_auth] = lambda: _restricted_key(["claude-opus-4-6"]) + + try: + response = client.get( + "/v1/files", headers={"Authorization": "Bearer test-key", "x-litellm-model": "mistral-ocr"} + ) + finally: + app.dependency_overrides.pop(ps.user_api_key_auth, None) + + assert response.status_code == 403, response.text + upstream.assert_not_called() + + +def test_model_routed_file_retrieve_allows_key_with_model_grant(mocker: MockerFixture, monkeypatch): + """The grant check must not break the happy path: a key allowed the deployment still resolves its credentials.""" + import litellm.proxy.proxy_server as ps + from litellm.proxy.openai_files_endpoints.common_utils import encode_file_id_with_model + + router = _mistral_plus_anthropic_router() + proxy_logging_obj = setup_proxy_logging_object(monkeypatch, router) + monkeypatch.setattr("litellm.proxy.proxy_server.master_key", None) + monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", None) + monkeypatch.setattr("litellm.proxy.proxy_server.llm_router", router) + proxy_logging_obj.update_request_status = mocker.AsyncMock() + proxy_logging_obj.post_call_failure_hook = mocker.AsyncMock() + + captured_kwargs: dict = {} + + async def _mock_afile_retrieve(**kwargs): + captured_kwargs.update(kwargs) + return OpenAIFileObject( + id="7a13fa8e-fcf8-42c5-aa61-c93c10e2c7df", + object="file", + bytes=2, + created_at=1234567890, + filename="batch.jsonl", + purpose="batch", + status="uploaded", + ) + + monkeypatch.setattr(litellm, "afile_retrieve", _mock_afile_retrieve) + app.dependency_overrides[ps.user_api_key_auth] = lambda: _restricted_key(["mistral-ocr"]) + encoded_id = encode_file_id_with_model("7a13fa8e-fcf8-42c5-aa61-c93c10e2c7df", "mistral-ocr") + + try: + response = client.get(f"/v1/files/{encoded_id}", headers={"Authorization": "Bearer test-key"}) + finally: + app.dependency_overrides.pop(ps.user_api_key_auth, None) + + assert response.status_code == 200, response.text + assert captured_kwargs["api_key"] == "mistral-key" + assert captured_kwargs["custom_llm_provider"] == "mistral" diff --git a/tests/test_litellm/proxy/test_batch_x_litellm_model_encoding.py b/tests/test_litellm/proxy/test_batch_x_litellm_model_encoding.py index 3d1831bb4cd..fe4903b547c 100644 --- a/tests/test_litellm/proxy/test_batch_x_litellm_model_encoding.py +++ b/tests/test_litellm/proxy/test_batch_x_litellm_model_encoding.py @@ -58,10 +58,7 @@ def _make_batch_response( def test_get_batch_id_from_unified_batch_id_handles_appended_fields(): - decoded_id = ( - "litellm_proxy;model_id:deployment-123;" - "llm_batch_id:batch_openai_123;llm_output_file_id:file-output" - ) + decoded_id = "litellm_proxy;model_id:deployment-123;llm_batch_id:batch_openai_123;llm_output_file_id:file-output" assert get_batch_id_from_unified_batch_id(decoded_id) == "batch_openai_123" @@ -107,12 +104,10 @@ async def test_create_batch_with_x_litellm_model_encodes_batch_id(): } ), ), + patch("litellm.proxy.batches_endpoints.endpoints.ProxyBaseLLMRequestProcessing") as mock_processor_cls, patch( - "litellm.proxy.batches_endpoints.endpoints.ProxyBaseLLMRequestProcessing" - ) as mock_processor_cls, - patch( - "litellm.proxy.batches_endpoints.endpoints.get_credentials_for_model", - return_value=mock_credentials, + "litellm.proxy.batches_endpoints.endpoints.get_authorized_credentials_for_model", + new=AsyncMock(return_value=mock_credentials), ), patch( "litellm.proxy.batches_endpoints.endpoints.prepare_data_with_credentials", @@ -165,23 +160,15 @@ async def test_create_batch_with_x_litellm_model_encodes_batch_id(): ) # The batch_id should be encoded with model info - assert ( - response.id != raw_batch_id - ), f"Expected batch_id to be encoded, but got raw ID: {response.id}" - assert response.id.startswith( - "batch_" - ), f"Encoded batch_id should keep batch_ prefix, got: {response.id}" + assert response.id != raw_batch_id, f"Expected batch_id to be encoded, but got raw ID: {response.id}" + assert response.id.startswith("batch_"), f"Encoded batch_id should keep batch_ prefix, got: {response.id}" # Should be decodable back to the original decoded_model = decode_model_from_file_id(response.id) - assert ( - decoded_model == model_name - ), f"Expected model '{model_name}' from decoded batch_id, got: {decoded_model}" + assert decoded_model == model_name, f"Expected model '{model_name}' from decoded batch_id, got: {decoded_model}" original_id = get_original_file_id(response.id) - assert ( - original_id == raw_batch_id - ), f"Expected original ID '{raw_batch_id}', got: {original_id}" + assert original_id == raw_batch_id, f"Expected original ID '{raw_batch_id}', got: {original_id}" assert mock_create_batch.call_args.kwargs["metadata"] == {"customer_id": "cust-123"} @@ -227,12 +214,10 @@ async def test_create_batch_with_x_litellm_model_encodes_output_and_error_file_i } ), ), + patch("litellm.proxy.batches_endpoints.endpoints.ProxyBaseLLMRequestProcessing") as mock_processor_cls, patch( - "litellm.proxy.batches_endpoints.endpoints.ProxyBaseLLMRequestProcessing" - ) as mock_processor_cls, - patch( - "litellm.proxy.batches_endpoints.endpoints.get_credentials_for_model", - return_value=mock_credentials, + "litellm.proxy.batches_endpoints.endpoints.get_authorized_credentials_for_model", + new=AsyncMock(return_value=mock_credentials), ), patch( "litellm.proxy.batches_endpoints.endpoints.prepare_data_with_credentials", @@ -316,9 +301,7 @@ async def test_create_batch_without_x_litellm_model_returns_raw_ids(monkeypatch) } ), ), - patch( - "litellm.proxy.batches_endpoints.endpoints.ProxyBaseLLMRequestProcessing" - ) as mock_processor_cls, + patch("litellm.proxy.batches_endpoints.endpoints.ProxyBaseLLMRequestProcessing") as mock_processor_cls, patch( "litellm.acreate_batch", new=AsyncMock(return_value=mock_response), @@ -383,9 +366,7 @@ class TestBatchIdRoundTripWithRetrieve: raw_batch_id = "batch_vllm_12345" # What create_batch does: - encoded_id = encode_file_id_with_model( - file_id=raw_batch_id, model=model_name, id_type="batch" - ) + encoded_id = encode_file_id_with_model(file_id=raw_batch_id, model=model_name, id_type="batch") # What retrieve_batch does: decoded_model = decode_model_from_file_id(encoded_id) @@ -410,9 +391,7 @@ class TestBatchIdRoundTripWithRetrieve: ] for raw_id, model in test_cases: - encoded = encode_file_id_with_model( - file_id=raw_id, model=model, id_type="batch" - ) + encoded = encode_file_id_with_model(file_id=raw_id, model=model, id_type="batch") assert encoded.startswith("batch_") assert decode_model_from_file_id(encoded) == model assert get_original_file_id(encoded) == raw_id @@ -440,9 +419,7 @@ async def test_cancel_batch_with_unified_id_routes_with_decoded_model_and_batch_ mock_user_api_key_dict.team_metadata = {} with ( - patch( - "litellm.proxy.batches_endpoints.endpoints.ProxyBaseLLMRequestProcessing" - ) as mock_processor_cls, + patch("litellm.proxy.batches_endpoints.endpoints.ProxyBaseLLMRequestProcessing") as mock_processor_cls, patch( "litellm.proxy.batches_endpoints.endpoints.update_batch_in_database", new=AsyncMock(), From bae731ddfc394b23d3c6f44a85cfaa58472897e7 Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Thu, 10 Sep 2026 17:54:57 -0400 Subject: [PATCH 07/37] fix(proxy): apply model grants to unified file and batch ids on batch routes Unified ids carry the deployment model inside the id, so a restricted key could create, retrieve or cancel a batch on a deployment it is not granted. The model parsed from a unified id now goes through the same grant check as header, query and model-encoded id sources before the router is called. --- litellm/proxy/batches_endpoints/endpoints.py | 16 ++++- .../proxy/batches_endpoints/test_endpoints.py | 58 +++++++++++++++++++ .../test_batch_x_litellm_model_encoding.py | 7 +-- 3 files changed, 75 insertions(+), 6 deletions(-) diff --git a/litellm/proxy/batches_endpoints/endpoints.py b/litellm/proxy/batches_endpoints/endpoints.py index c99f66d032e..c2489ce52ea 100644 --- a/litellm/proxy/batches_endpoints/endpoints.py +++ b/litellm/proxy/batches_endpoints/endpoints.py @@ -29,6 +29,7 @@ from litellm.proxy.openai_files_endpoints.common_utils import ( _is_base64_encoded_unified_file_id, add_internal_model_credentials, apply_team_provider_credentials, + authorize_model_for_key, batch_cost_poller_is_active, decode_model_from_file_id, encode_batch_response_ids, @@ -286,6 +287,7 @@ async def create_batch( detail={"error": f"Expected 1 model, got {len(target_model_names)}"}, ) model: Final = target_model_names[0] + await authorize_model_for_key(model_id=model, llm_router=llm_router, user_api_key_dict=user_api_key_dict) _create_batch_data["model"] = model resolved_storage_url: Final = await _resolve_managed_input_file_storage_url(input_file_id) @@ -582,10 +584,17 @@ async def retrieve_batch( ) if unified_batch_id: + unified_model_id: Final = get_model_id_from_unified_batch_id(unified_batch_id) + if unified_model_id is not None: + await authorize_model_for_key( + model_id=llm_router.resolve_model_name_from_model_id(unified_model_id) or unified_model_id, + llm_router=llm_router, + user_api_key_dict=user_api_key_dict, + ) add_internal_model_credentials( data=data, llm_router=llm_router, - model_id=get_model_id_from_unified_batch_id(unified_batch_id), + model_id=unified_model_id, ) response = await llm_router.aretrieve_batch(**data) @@ -998,6 +1007,11 @@ async def cancel_batch( status_code=400, detail={"error": "Invalid LiteLLM managed batch ID. Missing model_id."}, ) + await authorize_model_for_key( + model_id=llm_router.resolve_model_name_from_model_id(model_id_from_batch) or model_id_from_batch, + llm_router=llm_router, + user_api_key_dict=user_api_key_dict, + ) data["model"] = model_id_from_batch data["batch_id"] = get_batch_id_from_unified_batch_id(unified_batch_id) response = await llm_router.acancel_batch(**data) diff --git a/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py b/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py index 57e42e79a42..5be64ca0847 100644 --- a/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py +++ b/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py @@ -179,6 +179,8 @@ def harness(): router = MagicMock(spec=Router) router.model_group_alias = {} router.get_model_access_groups = MagicMock(return_value={}) + router.resolve_model_name_from_model_id = MagicMock(side_effect=lambda model_id: model_id) + router.model_list = [] router.acreate_batch = AsyncMock(return_value=make_batch()) router.get_deployment_credentials_with_provider = MagicMock(side_effect=_creds_lookup) @@ -1165,6 +1167,8 @@ def retrieve_harness(): router = MagicMock(spec=Router) router.model_group_alias = {} router.get_model_access_groups = MagicMock(return_value={}) + router.resolve_model_name_from_model_id = MagicMock(side_effect=lambda model_id: model_id) + router.model_list = [] router.aretrieve_batch = AsyncMock(return_value=make_batch()) router.get_deployment_credentials_with_provider = MagicMock(side_effect=_creds_lookup) @@ -1622,6 +1626,8 @@ def list_harness(): router = MagicMock(spec=Router) router.model_group_alias = {} router.get_model_access_groups = MagicMock(return_value={}) + router.resolve_model_name_from_model_id = MagicMock(side_effect=lambda model_id: model_id) + router.model_list = [] router.alist_batches = AsyncMock(return_value=FakeListPage([])) router.get_deployment_credentials_with_provider = MagicMock(side_effect=_creds_lookup) @@ -2020,6 +2026,8 @@ def cancel_harness(): router = MagicMock(spec=Router) router.model_group_alias = {} router.get_model_access_groups = MagicMock(return_value={}) + router.resolve_model_name_from_model_id = MagicMock(side_effect=lambda model_id: model_id) + router.model_list = [] router.acancel_batch = AsyncMock(return_value=make_batch()) router.get_deployment_credentials_with_provider = MagicMock(side_effect=_creds_lookup) @@ -2816,3 +2824,53 @@ async def test_cancel__model_encoded_id_rejects_key_without_model_grant(cancel_h assert exc_info.value.code == "403" cancel_harness.creds_resolver.assert_not_called() cancel_harness.litellm_acancel.assert_not_called() + + +def _b64_unified_id(decoded: str) -> str: + return base64.urlsafe_b64encode(decoded.encode()).decode().rstrip("=") + + +UNIFIED_FILE_ID_FOR_GPT4O_MINI = _b64_unified_id( + "litellm_proxy:application/octet-stream;unified_id,c4843482-b176-4901-8292-7523fd0f2c6e;" + "target_model_names,gpt-4o-mini;llm_output_file_id,file-provider;llm_output_file_model_id,dep-1" +) +UNIFIED_BATCH_ID_FOR_GPT4O_MINI = _b64_unified_id(UNIFIED_BATCH_ID) + + +@pytest.mark.asyncio +async def test_create__unified_file_id_rejects_key_without_model_grant(harness): + """The model carried inside a unified file id is caller-controlled too, so it is checked against the key's grants.""" + set_body( + harness, + { + "input_file_id": UNIFIED_FILE_ID_FOR_GPT4O_MINI, + "endpoint": "/v1/chat/completions", + "completion_window": "24h", + }, + ) + + with pytest.raises(ProxyException) as exc_info: + await call_create(harness, user=_key_restricted_to("vertex-model")) + + assert exc_info.value.code == "403" + harness.router_acreate.assert_not_called() + harness.litellm_acreate.assert_not_called() + + +@pytest.mark.asyncio +async def test_retrieve__unified_batch_id_rejects_key_without_model_grant(retrieve_harness): + with pytest.raises(ProxyException) as exc_info: + await call_retrieve(retrieve_harness, UNIFIED_BATCH_ID_FOR_GPT4O_MINI, user=_key_restricted_to("vertex-model")) + + assert exc_info.value.code == "403" + retrieve_harness.router_aretrieve.assert_not_called() + retrieve_harness.creds_resolver.assert_not_called() + + +@pytest.mark.asyncio +async def test_cancel__unified_batch_id_rejects_key_without_model_grant(cancel_harness): + with pytest.raises(ProxyException) as exc_info: + await call_cancel(cancel_harness, UNIFIED_BATCH_ID_FOR_GPT4O_MINI, user=_key_restricted_to("vertex-model")) + + assert exc_info.value.code == "403" + cancel_harness.router_acancel.assert_not_called() diff --git a/tests/test_litellm/proxy/test_batch_x_litellm_model_encoding.py b/tests/test_litellm/proxy/test_batch_x_litellm_model_encoding.py index fe4903b547c..3161fe99e68 100644 --- a/tests/test_litellm/proxy/test_batch_x_litellm_model_encoding.py +++ b/tests/test_litellm/proxy/test_batch_x_litellm_model_encoding.py @@ -11,6 +11,7 @@ from unittest.mock import AsyncMock, MagicMock, patch import pytest +from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.openai_files_endpoints.common_utils import ( decode_model_from_file_id, get_batch_id_from_unified_batch_id, @@ -412,11 +413,7 @@ async def test_cancel_batch_with_unified_id_routes_with_decoded_model_and_batch_ mock_request.url.path = f"/v1/batches/{unified_batch_id}/cancel" mock_fastapi_response = MagicMock() mock_fastapi_response.headers = {} - mock_user_api_key_dict = MagicMock() - mock_user_api_key_dict.parent_otel_span = None - mock_user_api_key_dict.user_id = "test_user" - mock_user_api_key_dict.allowed_model_region = None - mock_user_api_key_dict.team_metadata = {} + mock_user_api_key_dict = UserAPIKeyAuth(api_key="sk-test", user_id="test_user", team_metadata={}) with ( patch("litellm.proxy.batches_endpoints.endpoints.ProxyBaseLLMRequestProcessing") as mock_processor_cls, From 95fdefa390af6586affb5ff825955b0ccb3bce17 Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Thu, 10 Sep 2026 19:14:45 -0400 Subject: [PATCH 08/37] fix(logging): tolerate a missing api_base in pre_call for presigned batch retrieves Provider batch configs that build their own request URL (Mistral, Bedrock) hand pre_call api_base=None, and mask_api_base_credentials raised TypeError on it, so every such retrieve logged a non-blocking LoggingError and lost its pre-call logging. --- litellm/litellm_core_utils/litellm_logging.py | 4 +- .../test_litellm_logging.py | 63 ++++++++++--------- 2 files changed, 35 insertions(+), 32 deletions(-) diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py index cb9209be267..38e88493986 100644 --- a/litellm/litellm_core_utils/litellm_logging.py +++ b/litellm/litellm_core_utils/litellm_logging.py @@ -1199,8 +1199,8 @@ class Logging(LiteLLMLoggingBaseClass): return {"error": f"Unable to parse raw request body. Got - {data}"} return data - def _get_masked_api_base(self, api_base: str) -> str: - return str(mask_api_base_credentials(api_base)) + def _get_masked_api_base(self, api_base: str | None) -> str: + return str(mask_api_base_credentials(api_base or "")) def _pre_call(self, input, api_key, model=None, additional_args={}): """ diff --git a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py index 6aa77745e3d..570f4339cd4 100644 --- a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py +++ b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py @@ -58,6 +58,16 @@ def test_get_masked_api_base(logging_obj): assert type(masked_api_base) == str +def test_pre_call_tolerates_missing_api_base(logging_obj): + """Presigned batch retrieves (Mistral, Bedrock) build their own URL and pass api_base=None + to pre_call; masking must not raise or the request's pre-call logging is silently lost.""" + logging_obj.update_environment_variables(litellm_params={}, optional_params={}) + + logging_obj.pre_call(input="", api_key="", additional_args={"api_base": None, "headers": {}}) + + assert logging_obj.model_call_details["litellm_params"]["api_base"] == "" + + def test_post_call_serializes_dict_with_datetime(logging_obj): import datetime @@ -3976,9 +3986,7 @@ def test_get_standard_logging_object_payload_carries_matched_access_groups(loggi "model": "gpt-4o", "messages": [], "litellm_params": { - "metadata": { - "user_api_key_matched_model_access_groups": ["premium-pool", "shared-pool"] - }, + "metadata": {"user_api_key_matched_model_access_groups": ["premium-pool", "shared-pool"]}, "proxy_server_request": {"body": {}}, }, }, @@ -4062,9 +4070,7 @@ def _model_router_response(selected_model: str, stamp: bool): from litellm.types.utils import ModelResponse response = ModelResponse(model=selected_model) - response._hidden_params = ( - {AZURE_MODEL_ROUTER_SELECTED_MODEL_KEY: selected_model} if stamp else {} - ) + response._hidden_params = {AZURE_MODEL_ROUTER_SELECTED_MODEL_KEY: selected_model} if stamp else {} return response @@ -4088,9 +4094,7 @@ def test_standard_logging_payload_uses_stamped_model_router_model(logging_obj): "messages": [], "litellm_params": {"metadata": {}}, }, - init_response_obj=_model_router_response( - "azure_ai/grok-4-1-fast-reasoning", stamp=True - ), + init_response_obj=_model_router_response("azure_ai/grok-4-1-fast-reasoning", stamp=True), start_time=now, end_time=now, logging_obj=logging_obj, @@ -4122,9 +4126,7 @@ def test_standard_logging_payload_keeps_requested_model_without_router_stamp( "messages": [], "litellm_params": {"metadata": {}}, }, - init_response_obj=_model_router_response( - "azure_ai/grok-4-1-fast-reasoning", stamp=False - ), + init_response_obj=_model_router_response("azure_ai/grok-4-1-fast-reasoning", stamp=False), start_time=now, end_time=now, logging_obj=logging_obj, @@ -5536,9 +5538,7 @@ class TestNonInferenceCallTypesAreNotBilled: init_response_obj=self._retrieved_response(), start_time=now, end_time=now, - logging_obj=self._logging_obj( - "aget_responses", litellm_metadata=self.BACKGROUND_POLL_METADATA - ), + logging_obj=self._logging_obj("aget_responses", litellm_metadata=self.BACKGROUND_POLL_METADATA), status="success", ) @@ -5784,9 +5784,7 @@ async def test_streaming_success_callbacks_survive_cost_calculation_failure(): releasing.async_log_success_event = AsyncMock() patcher, logging_obj = _streaming_logging_obj_with_callbacks([releasing]) - with patcher, patch.object( - logging_obj, "_response_cost_calculator", side_effect=ValueError("bad usage block") - ): + with patcher, patch.object(logging_obj, "_response_cost_calculator", side_effect=ValueError("bad usage block")): await logging_obj.async_success_handler(result=_assembled_stream_result()) assert logging_obj.model_call_details["response_cost"] is None @@ -5799,8 +5797,9 @@ async def test_streaming_success_callbacks_survive_standard_logging_payload_fail releasing.async_log_success_event = AsyncMock() patcher, logging_obj = _streaming_logging_obj_with_callbacks([releasing]) - with patcher, patch.object( - logging_obj, "_build_standard_logging_payload", side_effect=ValueError("incomplete stream") + with ( + patcher, + patch.object(logging_obj, "_build_standard_logging_payload", side_effect=ValueError("incomplete stream")), ): await logging_obj.async_success_handler(result=_assembled_stream_result()) @@ -6073,6 +6072,8 @@ def test_prompt_hooks_skip_prompt_managers_when_no_prompt_id(logging_obj, tmp_pa ) for hook in [cb for cb in litellm.callbacks if isinstance(cb, VectorStorePreCallHook)]: litellm.logging_callback_manager.remove_callback_from_list_by_object(litellm.callbacks, hook) + + def test_newrelic_dispatch_prefers_otel_v2_when_flag_on(monkeypatch): """With LITELLM_OTEL_V2 on and operator credentials present, the "newrelic" callback builds the OTel v2 logger (per-team credential routing); with the @@ -6228,7 +6229,9 @@ def test_get_error_information_skips_traceback_for_budget_rejection_with_provide from litellm.litellm_core_utils.litellm_logging import StandardLoggingPayloadSetup assert litellm.log_client_error_tracebacks is False - over_budget = _raise_and_catch(litellm.BudgetExceededError(current_cost=0.01, max_budget=0.0, llm_provider="anthropic")) + over_budget = _raise_and_catch( + litellm.BudgetExceededError(current_cost=0.01, max_budget=0.0, llm_provider="anthropic") + ) result = StandardLoggingPayloadSetup.get_error_information(over_budget) assert result["error_code"] == "429" assert result["llm_provider"] == "anthropic" @@ -6745,9 +6748,7 @@ def test_passthrough_embeddings_result_swapped_for_callbacks(): ], "model": "EmbeddingsGigaR", }, - request=httpx.Request( - "POST", "https://gigachat.devices.sberbank.ru/api/v1/embeddings" - ), + request=httpx.Request("POST", "https://gigachat.devices.sberbank.ru/api/v1/embeddings"), ) _, _, swapped_result = logging_obj._success_handler_helper_fn( @@ -6766,12 +6767,14 @@ def test_get_status_fields_ranks_guardrail_flagged_between_success_and_intervene request-level guardrail_status but never mask an intervention.""" flagged = {"guardrail_status": "guardrail_flagged"} - assert _get_status_fields( - "success", [{"guardrail_status": "success"}, flagged], None - )["guardrail_status"] == "guardrail_flagged" - assert _get_status_fields( - "success", [flagged, {"guardrail_status": "guardrail_intervened"}], None - )["guardrail_status"] == "guardrail_intervened" + assert ( + _get_status_fields("success", [{"guardrail_status": "success"}, flagged], None)["guardrail_status"] + == "guardrail_flagged" + ) + assert ( + _get_status_fields("success", [flagged, {"guardrail_status": "guardrail_intervened"}], None)["guardrail_status"] + == "guardrail_intervened" + ) def test_get_error_information_redacts_provider_key_from_upstream_url(): From e6bc4e47c7a63e8540a856f1b2f66710a4b47142 Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Thu, 10 Sep 2026 19:46:55 -0400 Subject: [PATCH 09/37] fix(mistral): reject file purposes Mistral lacks instead of mapping them to batch The proxy runs batch-file validation and guardrails only for purpose=batch, so a purpose such as assistants that was silently rewritten to batch on the way to Mistral let an upload skip both. Only batch, fine-tune and ocr pass through now; anything else is a 400. --- litellm/llms/mistral/files/transformation.py | 9 ++++-- .../test_mistral_files_transformation.py | 29 ++++++++++++++----- 2 files changed, 28 insertions(+), 10 deletions(-) diff --git a/litellm/llms/mistral/files/transformation.py b/litellm/llms/mistral/files/transformation.py index 6d58311813c..bf1ef7cb69f 100644 --- a/litellm/llms/mistral/files/transformation.py +++ b/litellm/llms/mistral/files/transformation.py @@ -89,11 +89,14 @@ def _to_openai_purpose(purpose: MistralFilePurpose) -> OpenAIFilesPurpose: def _to_mistral_purpose(purpose: str) -> MistralFilePurpose: + """Only Mistral's own purposes pass through. Silently mapping anything else to ``batch`` + would let an upload skip the proxy's batch-file validation and guardrails, which only + run when the caller says ``purpose=batch``.""" match purpose: - case "fine-tune" | "ocr": + case "batch" | "fine-tune" | "ocr": return purpose case _: - return "batch" + raise ValueError(f"Mistral does not support purpose={purpose!r}. Use one of: batch, fine-tune, ocr") def _api_base_from(litellm_params: Mapping[str, object]) -> str: @@ -166,7 +169,7 @@ class MistralFilesConfig(BaseFilesConfig): content_type: Final = extracted.get("content_type") or "application/octet-stream" upload: Final = MistralMultipartUpload( file=(filename, extracted["content"], content_type), - purpose=(None, _to_mistral_purpose(create_file_data.get("purpose", "batch"))), + purpose=(None, _to_mistral_purpose(create_file_data.get("purpose") or "batch")), ) return dict(upload) # mutable-ok: BaseFilesConfig signature diff --git a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py index f62645be7ee..b81740c0429 100644 --- a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py +++ b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py @@ -91,18 +91,28 @@ def test_upload_request_is_multipart_with_batch_purpose(config): } -@pytest.mark.parametrize( - "openai_purpose,mistral_purpose", - [("batch", "batch"), ("fine-tune", "fine-tune"), ("ocr", "ocr"), ("assistants", "batch"), ("user_data", "batch")], -) -def test_upload_request_maps_purpose_onto_mistral_enum(config, openai_purpose, mistral_purpose): +@pytest.mark.parametrize("purpose", ["batch", "fine-tune", "ocr"]) +def test_upload_request_passes_mistral_purposes_through(config, purpose): body = config.transform_create_file_request( model="", - create_file_data=CreateFileRequest(file=("f.bin", b"x"), purpose=openai_purpose), + create_file_data=CreateFileRequest(file=("f.bin", b"x"), purpose=purpose), optional_params={}, litellm_params={}, ) - assert body["purpose"] == (None, mistral_purpose) + assert body["purpose"] == (None, purpose) + + +@pytest.mark.parametrize("purpose", ["assistants", "user_data", "vision", "evals"]) +def test_upload_request_rejects_purposes_mistral_lacks(config, purpose): + """Regression: these used to be silently rewritten to ``batch``, so an upload that skipped the + proxy's batch-only validation and guardrails still landed on Mistral as a batch input file.""" + with pytest.raises(ValueError, match=f"purpose={purpose!r}"): + config.transform_create_file_request( + model="", + create_file_data=CreateFileRequest(file=("f.bin", b"x"), purpose=purpose), + optional_params={}, + litellm_params={}, + ) def test_upload_request_requires_file(config): @@ -181,6 +191,11 @@ def test_list_request_filters_by_mapped_purpose(config): assert no_params == {} +def test_list_request_rejects_purposes_mistral_lacks(config): + with pytest.raises(ValueError, match="purpose='assistants'"): + config.transform_list_files_request(purpose="assistants", optional_params={}, litellm_params={}) + + def test_list_response(config): out = config.transform_list_files_response( raw_response=_response( From 785c6cffc4826ef44c73a797981e28a294a45668 Mon Sep 17 00:00:00 2001 From: Marty Sullivan Date: Thu, 13 Aug 2026 23:34:40 -0400 Subject: [PATCH 10/37] fix(cost): carry image and video input tokens through the Responses usage bridge Realtime cost is computed from *_tokens_details after the usage round-trips through the Responses shape, and the input half of that shape carried audio only, so image and video prompt tokens stopped being billable as themselves. Vertex splits prompt tokens by modality, so a session sending camera frames arrives with image_tokens set. Those were folded into text_tokens and lost their attribution. The amount happens not to move today, because the calculator falls back to input_cost_per_token when no per-modality rate is set, but the tokens have to survive before any such rate can ever apply. InputTokensDetails now declares image_tokens and video_tokens instead of leaning on pydantic extras, the repeated per-field copying is a loop over the modality names so adding a modality no longer adds a branch, and the read-back in ResponseAPILoggingUtils picks up video_tokens, which PromptTokensDetailsWrapper already declared. The output half of the original change is dropped: 449c091391 landed the same OutputTokensDetails.audio_tokens fix upstream, with its own coverage in test_gemini_realtime_transformation.py, and it always sets output_tokens_details rather than only when non-empty. That structure is kept as upstream wrote it. --- .../transformation.py | 2 ++ litellm/responses/utils.py | 1 + litellm/types/llms/openai.py | 2 ++ .../test_litellm_completion_responses.py | 33 +++++++++++++++++++ 4 files changed, 38 insertions(+) diff --git a/litellm/responses/litellm_completion_transformation/transformation.py b/litellm/responses/litellm_completion_transformation/transformation.py index 01fb6cb483d..ffd7ce491b1 100644 --- a/litellm/responses/litellm_completion_transformation/transformation.py +++ b/litellm/responses/litellm_completion_transformation/transformation.py @@ -2851,6 +2851,8 @@ class LiteLLMCompletionResponsesConfig: cached_tokens=prompt_details.cached_tokens if prompt_details.cached_tokens is not None else 0, text_tokens=prompt_details.text_tokens, audio_tokens=prompt_details.audio_tokens, + image_tokens=prompt_details.image_tokens, + video_tokens=prompt_details.video_tokens, cached_tokens_details=( cached_tokens_details if isinstance(cached_tokens_details, CachedTokensDetails) else None ), diff --git a/litellm/responses/utils.py b/litellm/responses/utils.py index 41a3ded7022..f50c17aff85 100644 --- a/litellm/responses/utils.py +++ b/litellm/responses/utils.py @@ -1182,6 +1182,7 @@ class ResponseAPILoggingUtils: cached_tokens_details=getattr( response_api_usage.input_tokens_details, "cached_tokens_details", None ), + video_tokens=getattr(response_api_usage.input_tokens_details, "video_tokens", None), cache_write_tokens=getattr(response_api_usage.input_tokens_details, "cache_write_tokens", None), web_search_requests=getattr(response_api_usage.input_tokens_details, "web_search_requests", None), google_maps_grounding_requests=getattr( diff --git a/litellm/types/llms/openai.py b/litellm/types/llms/openai.py index e3eac9b9205..98548705979 100644 --- a/litellm/types/llms/openai.py +++ b/litellm/types/llms/openai.py @@ -1291,7 +1291,9 @@ class InputTokensDetails(BaseLiteLLMOpenAIResponseObject): audio_tokens: int | None = None cached_tokens: int = 0 cached_tokens_details: CachedTokensDetails | None = None + image_tokens: int | None = None text_tokens: int | None = None + video_tokens: int | None = None model_config = {"extra": "allow"} diff --git a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py index 0ed101952be..2f9f7adfcf1 100644 --- a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py +++ b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py @@ -2885,6 +2885,39 @@ class TestUsageTransformation: assert getattr(response_usage.input_tokens_details, "cache_write_tokens", None) == 800 assert response_usage.input_tokens_details.model_dump()["cache_write_tokens"] == 800 + def test_transform_usage_preserves_input_modality_tokens(self): + """Regression: the bridge dropped image and video input tokens. + + Vertex reports prompt tokens split by modality, so a Live session that sends + camera frames arrives with image_tokens set. InputTokensDetails declared only + audio/cached/text, so those tokens were folded into text and lost their + attribution, and any per-modality rate could never apply to them. + """ + usage = Usage( + prompt_tokens=300, + completion_tokens=10, + total_tokens=310, + prompt_tokens_details=PromptTokensDetailsWrapper( + text_tokens=20, audio_tokens=80, image_tokens=150, video_tokens=50, cached_tokens=0 + ), + completion_tokens_details=CompletionTokensDetailsWrapper(text_tokens=10), + ) + + response_usage = LiteLLMCompletionResponsesConfig._transform_chat_completion_usage_to_responses_usage( + chat_completion_response=usage + ) + details = response_usage.input_tokens_details + assert details is not None + assert getattr(details, "image_tokens", None) == 150 + assert getattr(details, "video_tokens", None) == 50 + assert getattr(details, "audio_tokens", None) == 80 + + from litellm.responses.utils import ResponseAPILoggingUtils + + back = ResponseAPILoggingUtils._transform_response_api_usage_to_chat_usage(response_usage.model_dump()) + assert back.prompt_tokens_details.image_tokens == 150 + assert back.prompt_tokens_details.video_tokens == 50 + def test_transform_usage_with_reasoning_tokens_gemini(self): """Test that reasoning_tokens from Gemini are properly transformed to output_tokens_details""" # Setup: Simulate Gemini usage with thoughtsTokenCount From 76488beaf8d1a44e0f07b6a4a66c06b9b4390222 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Tue, 15 Sep 2026 05:55:25 -0700 Subject: [PATCH 11/37] fix(utils): reject an untranslatable tool_choice with a 400 instead of a 500 --- litellm/main.py | 2 +- litellm/utils.py | 16 +++- tests/litellm_utils_tests/test_utils.py | 6 +- .../test_validate_tool_choice.py | 74 ++++++++++--------- .../test_litellm_completion_responses.py | 15 ++++ tests/test_litellm/test_main.py | 14 ++++ 6 files changed, 85 insertions(+), 42 deletions(-) diff --git a/litellm/main.py b/litellm/main.py index f6f4ec1bf63..9eea6779abf 100644 --- a/litellm/main.py +++ b/litellm/main.py @@ -5102,7 +5102,7 @@ def completion( messages = validate_and_fix_openai_messages(messages=messages) tools = validate_and_fix_openai_tools(tools=tools) # validate tool_choice - tool_choice = validate_chat_completion_tool_choice(tool_choice=tool_choice) + tool_choice = validate_chat_completion_tool_choice(tool_choice=tool_choice, model=model) # validate optional params stop = validate_openai_optional_params(stop=stop) thinking = validate_and_fix_thinking_param(thinking=thinking) diff --git a/litellm/utils.py b/litellm/utils.py index 18df5e2abf7..17088f0475b 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -8038,6 +8038,7 @@ def validate_chat_completion_user_messages(messages: list[AllMessageValues]): def validate_chat_completion_tool_choice( tool_choice: dict | str | None, + model: str, ) -> dict | str | None: """ Confirm the tool choice is passed in the OpenAI format. @@ -8053,12 +8054,19 @@ def validate_chat_completion_tool_choice( # Standard OpenAI format: {"type": "function", "function": {...}} if tool_choice.get("type") is None or tool_choice.get("function") is None: - raise Exception( - f"Invalid tool choice, tool_choice={tool_choice}. Please ensure tool_choice follows the OpenAI spec" + raise BadRequestError( + message=f"Invalid tool choice, tool_choice={tool_choice}. Please ensure tool_choice follows the OpenAI spec", + model=model, + llm_provider="", ) return tool_choice - raise Exception( - f"Invalid tool choice, tool_choice={tool_choice}. Got={type(tool_choice)}. Expecting str, or dict. Please ensure tool_choice follows the OpenAI tool_choice spec" + raise BadRequestError( + message=( + f"Invalid tool choice, tool_choice={tool_choice}. Got={type(tool_choice)}. Expecting str, or dict. " + "Please ensure tool_choice follows the OpenAI tool_choice spec" + ), + model=model, + llm_provider="", ) diff --git a/tests/litellm_utils_tests/test_utils.py b/tests/litellm_utils_tests/test_utils.py index 0ccfae55290..11d089719c6 100644 --- a/tests/litellm_utils_tests/test_utils.py +++ b/tests/litellm_utils_tests/test_utils.py @@ -1334,10 +1334,10 @@ def test_validate_chat_completion_tool_choice(tool_choice, expected_bool): from litellm.utils import validate_chat_completion_tool_choice if expected_bool: - validate_chat_completion_tool_choice(tool_choice=tool_choice) + validate_chat_completion_tool_choice(tool_choice=tool_choice, model="gpt-5.6-sol") else: - with pytest.raises(Exception, match="Invalid tool choice"): - validate_chat_completion_tool_choice(tool_choice=tool_choice) + with pytest.raises(litellm.BadRequestError, match="Invalid tool choice"): + validate_chat_completion_tool_choice(tool_choice=tool_choice, model="gpt-5.6-sol") def test_models_by_provider(): diff --git a/tests/litellm_utils_tests/test_validate_tool_choice.py b/tests/litellm_utils_tests/test_validate_tool_choice.py index b8246fe0deb..b4272af7b90 100644 --- a/tests/litellm_utils_tests/test_validate_tool_choice.py +++ b/tests/litellm_utils_tests/test_validate_tool_choice.py @@ -1,60 +1,66 @@ +import re +from typing import Final + import pytest - +import litellm from litellm.utils import validate_chat_completion_tool_choice +MODEL: Final = "anthropic/claude-haiku-4-5" + def test_validate_tool_choice_none(): """Test that None is returned as-is.""" - result = validate_chat_completion_tool_choice(None) + result = validate_chat_completion_tool_choice(None, model=MODEL) assert result is None def test_validate_tool_choice_string(): """Test that string values are returned as-is.""" - assert validate_chat_completion_tool_choice("auto") == "auto" - assert validate_chat_completion_tool_choice("none") == "none" - assert validate_chat_completion_tool_choice("required") == "required" + assert validate_chat_completion_tool_choice("auto", model=MODEL) == "auto" + assert validate_chat_completion_tool_choice("none", model=MODEL) == "none" + assert validate_chat_completion_tool_choice("required", model=MODEL) == "required" def test_validate_tool_choice_standard_dict(): """Test standard OpenAI format with function.""" tool_choice = {"type": "function", "function": {"name": "my_function"}} - result = validate_chat_completion_tool_choice(tool_choice) + result = validate_chat_completion_tool_choice(tool_choice, model=MODEL) assert result == tool_choice def test_validate_tool_choice_cursor_format(): """Cursor IDE format {"type": "auto"} is unwrapped to the bare string.""" - assert validate_chat_completion_tool_choice({"type": "auto"}) == "auto" - assert validate_chat_completion_tool_choice({"type": "none"}) == "none" - assert validate_chat_completion_tool_choice({"type": "required"}) == "required" + assert validate_chat_completion_tool_choice({"type": "auto"}, model=MODEL) == "auto" + assert validate_chat_completion_tool_choice({"type": "none"}, model=MODEL) == "none" + assert validate_chat_completion_tool_choice({"type": "required"}, model=MODEL) == "required" -def test_validate_tool_choice_invalid_dict(): - """Test that invalid dict formats raise exceptions.""" - # Missing both type and function - with pytest.raises(Exception, match='Invalid tool choice, tool_choice=\\{\\}\\. Please ensure') as exc_info: - validate_chat_completion_tool_choice({}) - assert "Invalid tool choice" in str(exc_info.value) - - # Invalid type value - with pytest.raises(Exception, match="Invalid tool choice, tool_choice=\\{'type': 'invalid'\\}\\.") as exc_info: - validate_chat_completion_tool_choice({"type": "invalid"}) - assert "Invalid tool choice" in str(exc_info.value) - - # Has type but missing function when type is "function" - with pytest.raises(Exception, match="Invalid tool choice, tool_choice=\\{'type': 'function'\\}\\.") as exc_info: - validate_chat_completion_tool_choice({"type": "function"}) - assert "Invalid tool choice" in str(exc_info.value) +@pytest.mark.parametrize( + "tool_choice", + [ + {}, + {"type": "invalid"}, + {"type": "function"}, + {"name": "lookup_fruit"}, + {"type": "file_search"}, + ], +) +def test_validate_tool_choice_invalid_dict_is_a_400(tool_choice): + """A dict shape chat completions cannot carry is the caller's mistake: a 400 that names the field, never a 500.""" + with pytest.raises( + litellm.BadRequestError, match=f"Invalid tool choice, tool_choice={re.escape(str(tool_choice))}\\. Please ensure" + ) as exc_info: + validate_chat_completion_tool_choice(tool_choice, model=MODEL) + assert exc_info.value.status_code == 400 + assert exc_info.value.model == MODEL -def test_validate_tool_choice_invalid_type(): - """Test that invalid types raise exceptions.""" - with pytest.raises(Exception, match="\\. Expecting str, or dict\\. Please ensure") as exc_info: - validate_chat_completion_tool_choice(123) - assert "Got=" in str(exc_info.value) - - with pytest.raises(Exception, match="Invalid tool choice, tool_choice=\\[\\]\\. Got=\\.") as exc_info: - validate_chat_completion_tool_choice([]) - assert "Got=" in str(exc_info.value) +@pytest.mark.parametrize("tool_choice", [123, []]) +def test_validate_tool_choice_invalid_type_is_a_400(tool_choice): + """A non-str, non-dict tool_choice is rejected as a 400 that names the type it got.""" + with pytest.raises( + litellm.BadRequestError, match=f"Got={re.escape(str(type(tool_choice)))}\\. Expecting str, or dict\\." + ) as exc_info: + validate_chat_completion_tool_choice(tool_choice, model=MODEL) + assert exc_info.value.status_code == 400 diff --git a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py index 0ed101952be..3950fd549ef 100644 --- a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py +++ b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py @@ -4906,3 +4906,18 @@ class TestStreamingSnapshotItemIds: reasoning_items = _bridged_output_items(completed_event.response, "reasoning") assert len(reasoning_items) == 1 assert reasoning_items[0].id == streamed_event.item_id + + +@pytest.mark.parametrize("stream", [True, False]) +async def test_bridge_rejects_untranslatable_tool_choice_with_a_400(stream: bool): + with pytest.raises(litellm.BadRequestError) as exc_info: + await litellm.aresponses( + model="anthropic/claude-haiku-4-5", + input="Which fruit is red?", + tools=[{"type": "function", "name": "lookup_fruit", "parameters": {"type": "object"}}], + tool_choice={"type": "file_search"}, + stream=stream, + api_key="sk-unused", + ) + assert exc_info.value.status_code == 400 + assert "tool_choice={'type': 'file_search'}" in str(exc_info.value) diff --git a/tests/test_litellm/test_main.py b/tests/test_litellm/test_main.py index 3dccb2b35bf..81ad161772e 100644 --- a/tests/test_litellm/test_main.py +++ b/tests/test_litellm/test_main.py @@ -3850,3 +3850,17 @@ def test_bridged_responses_with_openai_http_handler_keeps_forwarded_headers_out_ assert "extra_headers" not in body assert body["model"] == "gpt-5.4" assert {k: request.headers[k] for k in FORWARDED_CLIENT_HEADERS} == FORWARDED_CLIENT_HEADERS + + +@pytest.mark.parametrize("tool_choice", [{"type": "bogus"}, {"name": "lookup_fruit"}, {"type": "file_search"}]) +def test_completion_rejects_untranslatable_tool_choice_with_a_400(tool_choice): + with pytest.raises(litellm.BadRequestError) as exc_info: + litellm.completion( + model="anthropic/claude-haiku-4-5", + messages=[{"role": "user", "content": "Which fruit is red?"}], + tools=[{"type": "function", "function": {"name": "lookup_fruit", "parameters": {"type": "object"}}}], + tool_choice=tool_choice, + api_key="sk-unused", + ) + assert exc_info.value.status_code == 400 + assert f"tool_choice={tool_choice}" in str(exc_info.value) From 2bbf34c6520ef3266a62121510470868f73e499e Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Tue, 15 Sep 2026 07:00:33 -0700 Subject: [PATCH 12/37] fix(utils): keep the tool_choice validator's model argument optional --- litellm/utils.py | 2 +- tests/litellm_utils_tests/test_validate_tool_choice.py | 8 ++++++++ 2 files changed, 9 insertions(+), 1 deletion(-) diff --git a/litellm/utils.py b/litellm/utils.py index 17088f0475b..c2dbbda2b68 100644 --- a/litellm/utils.py +++ b/litellm/utils.py @@ -8038,7 +8038,7 @@ def validate_chat_completion_user_messages(messages: list[AllMessageValues]): def validate_chat_completion_tool_choice( tool_choice: dict | str | None, - model: str, + model: str = "", ) -> dict | str | None: """ Confirm the tool choice is passed in the OpenAI format. diff --git a/tests/litellm_utils_tests/test_validate_tool_choice.py b/tests/litellm_utils_tests/test_validate_tool_choice.py index b4272af7b90..a9dacf9fa15 100644 --- a/tests/litellm_utils_tests/test_validate_tool_choice.py +++ b/tests/litellm_utils_tests/test_validate_tool_choice.py @@ -64,3 +64,11 @@ def test_validate_tool_choice_invalid_type_is_a_400(tool_choice): ) as exc_info: validate_chat_completion_tool_choice(tool_choice, model=MODEL) assert exc_info.value.status_code == 400 + + +def test_validate_tool_choice_without_model_is_still_a_400(): + """Callers that predate the model argument keep getting a 400, with an empty model on the error.""" + with pytest.raises(litellm.BadRequestError, match="Invalid tool choice") as exc_info: + validate_chat_completion_tool_choice({"type": "bogus"}) + assert exc_info.value.status_code == 400 + assert exc_info.value.model == "" From d2af0577d535a68f438e39273c79b3b77cdf9987 Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Thu, 17 Sep 2026 15:39:07 -0400 Subject: [PATCH 13/37] test(files): assert key_model_access_denied error type instead of message text main (15f2e25e8a) replaced the configurable model-access-denied message with a fixed client message, so match on the stable error type. --- .../proxy/openai_files_endpoint/test_files_endpoint.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py b/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py index cf80be6f655..c48fc572e40 100644 --- a/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py +++ b/tests/test_litellm/proxy/openai_files_endpoint/test_files_endpoint.py @@ -5357,7 +5357,7 @@ def test_model_routed_file_ops_reject_key_without_model_grant( app.dependency_overrides.pop(ps.user_api_key_auth, None) assert response.status_code == 403, response.text - assert "not allowed to access model" in response.text + assert response.json()["error"]["type"] == "key_model_access_denied" upstream.assert_not_called() From 4e4008cea93291b72b0785014e25746be747b2a6 Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Thu, 17 Sep 2026 15:43:46 -0400 Subject: [PATCH 14/37] lint(cost): justify blind except in _lookup_model_info_or_none get_model_info raises a bare Exception for unmapped models, so BLE001 cannot be narrowed; mark it noqa with the reason to stay within the strict budget. --- litellm/cost_calculator.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 6cb1ef8228e..352d95a4fdf 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -2170,7 +2170,7 @@ def ocr_batch_cost( def _lookup_model_info_or_none(model: str, custom_llm_provider: str | None) -> ModelInfo | None: try: return litellm.get_model_info(model=model, custom_llm_provider=custom_llm_provider) - except Exception: + except Exception: # noqa: BLE001 # get_model_info raises bare Exception for unmapped models; caller logs and bills 0.0 return None From 8cf2606e2dd7184ed1ff27a29940d17b75d69e95 Mon Sep 17 00:00:00 2001 From: mubashir1osmani Date: Fri, 18 Sep 2026 22:50:19 -0400 Subject: [PATCH 15/37] fix(batches): mask pre-signed request auth headers before raw-request logging A pre-signed batch/file request (Mistral, Bedrock) carries its auth header inside the transformed request body, which pre_call logs verbatim into raw_request_typed_dict and raw-request callbacks, leaking the provider key. Mask the nested headers channel before handing the request to pre_call. Co-Authored-By: Claude Fable 5 --- litellm/llms/custom_httpx/llm_http_handler.py | 23 ++++++- .../custom_httpx/test_llm_http_handler.py | 62 +++++++++++++++++++ 2 files changed, 82 insertions(+), 3 deletions(-) diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py index 857adf5b9f1..f1c8add7152 100644 --- a/litellm/llms/custom_httpx/llm_http_handler.py +++ b/litellm/llms/custom_httpx/llm_http_handler.py @@ -278,6 +278,23 @@ def _has_pre_call_deployment_hook(logging_obj: LiteLLMLoggingObj) -> bool: return False +def _mask_presigned_request_headers(transformed_request: bytes | str | dict) -> bytes | str | dict: + """A pre-signed request carries its auth inside its own ``headers`` key, which + logging treats as request body (only the top-level headers channel gets masked), + so mask it here before the request is handed to ``pre_call``.""" + if not isinstance(transformed_request, dict): + return transformed_request + request_headers: Final = transformed_request.get("headers") + if not isinstance(request_headers, dict): + return transformed_request + + from litellm.litellm_core_utils.litellm_logging import ( + _get_masked_values, # pyright: ignore[reportPrivateUsage] # the shared header-masking helper has no public name + ) + + return {**transformed_request, "headers": _get_masked_values(request_headers)} + + def _aws_signing_overrides(optional_params: Mapping[str, Any], litellm_params: Mapping[str, Any]) -> Mapping[str, Any]: return MappingProxyType( { @@ -3692,7 +3709,7 @@ class BaseLLMHTTPHandler: "complete_input_dict": ( "" if isinstance(transformed_request, dict) and "streaming_media_upload" in transformed_request - else transformed_request + else _mask_presigned_request_headers(transformed_request) ), "api_base": api_base, "headers": headers, @@ -4115,7 +4132,7 @@ class BaseLLMHTTPHandler: input="", api_key="", additional_args={ - "complete_input_dict": transformed_request, + "complete_input_dict": _mask_presigned_request_headers(transformed_request), "api_base": api_base, "headers": headers, }, @@ -4194,7 +4211,7 @@ class BaseLLMHTTPHandler: input="", api_key="", additional_args={ - "complete_input_dict": transformed_request, + "complete_input_dict": _mask_presigned_request_headers(transformed_request), "api_base": api_base, "headers": headers, "batch_id": batch_id, diff --git a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py index 95dceccb2f5..6252947ef56 100644 --- a/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py +++ b/tests/test_litellm/llms/custom_httpx/test_llm_http_handler.py @@ -2761,6 +2761,68 @@ def test_direct_vector_store_search_debug_log_omits_stored_credentials(caplog, i assert "sk-embedding-s3cret" not in logged +@pytest.mark.asyncio +async def test_async_retrieve_batch_masks_presigned_auth_header_in_raw_request_log(): + """Regression: a pre-signed retrieve-batch request (Mistral, Bedrock) embeds its auth + header inside the transformed request, which pre_call logs verbatim as the raw request + body, so the provider key landed unmasked in raw_request_typed_dict and every + raw-request callback.""" + from litellm.litellm_core_utils.litellm_logging import Logging as LitellmLogging + from litellm.llms.mistral.batches.transformation import MistralBatchesConfig + + provider_key = "mistral-s3cret-provider-key-123456" + job_payload = { + "id": "batch-1", + "input_files": ["file-1"], + "endpoint": "/v1/ocr", + "model": "mistral-ocr-latest", + "status": "SUCCESS", + "created_at": 1_757_400_000, + } + sent_requests = [] + + def _capture(request: httpx.Request) -> httpx.Response: + sent_requests.append(request) + return httpx.Response(200, json=job_payload) + + client = AsyncHTTPHandler() + client.client = httpx.AsyncClient(transport=httpx.MockTransport(_capture)) + + logging_obj = LitellmLogging( + model="mistral/mistral-ocr-latest", + messages=[], + stream=False, + call_type="batch_retrieve", + start_time=time.time(), + litellm_call_id="batch-retrieve-call-id", + function_id="batch-retrieve-function-id", + log_raw_request_response=True, + ) + logging_obj.update_environment_variables( + model="mistral/mistral-ocr-latest", + optional_params={}, + litellm_params={"litellm_call_id": "batch-retrieve-call-id", "metadata": {}}, + ) + + result = await BaseLLMHTTPHandler().retrieve_batch( + batch_id="batch-1", + litellm_params={"api_key": provider_key}, + provider_config=MistralBatchesConfig(), + headers={}, + api_base=None, + api_key=provider_key, + logging_obj=logging_obj, + _is_async=True, + client=client, + model="mistral/mistral-ocr-latest", + ) + + assert result.id == "batch-1" + assert sent_requests[0].headers["Authorization"] == f"Bearer {provider_key}" + raw_request_body = logging_obj.model_call_details["raw_request_typed_dict"]["raw_request_body"] + assert provider_key not in json.dumps(raw_request_body) + + @pytest.mark.asyncio async def test_async_anthropic_messages_handler_carries_deployment_vertex_location_for_pricing(monkeypatch): """ From 044f88ee91942d60cdd4cc8b060b1e95fc60cb0c Mon Sep 17 00:00:00 2001 From: Yuneng Jiang Date: Fri, 18 Sep 2026 21:04:25 -0700 Subject: [PATCH 16/37] fix(proxy): register transcribe as a known provider for model grants #41515 added the cost map entry transcribe/StartTranscriptionJob under a new litellm_provider value "transcribe" without registering that provider anywhere else, so litellm.models_by_provider had no "transcribe" key. test_models_by_provider derives its provider set from the cost map itself, so it went red on main. The user-visible half is that get_provider_models returned None for the provider, which get_known_models_from_wildcard turns into an empty list, leaving a transcribe/* key or team grant resolving to no models. Mirror the aws_polly registration: an enum member, a model set, an ingestion branch, and a models_by_provider entry. Amazon Transcribe is reached through the pass-through route rather than the Add Model form, so it joins the frozen unlisted set the Add Model drift test tracks. --- litellm/__init__.py | 4 ++++ litellm/types/utils.py | 1 + .../test_litellm/proxy/auth/test_model_checks.py | 15 +++++++++++++++ .../public_endpoints/test_public_endpoints.py | 1 + 4 files changed, 21 insertions(+) diff --git a/litellm/__init__.py b/litellm/__init__.py index fcfc4768ff3..e17ab613dac 100644 --- a/litellm/__init__.py +++ b/litellm/__init__.py @@ -701,6 +701,7 @@ github_copilot_models: Set = set() chatgpt_models: Set = set() minimax_models: Set = set() aws_polly_models: Set = set() +transcribe_models: Set = set() gigachat_models: Set = set() llamagate_models: Set = set() reducto_models: Set = set() @@ -980,6 +981,8 @@ def _populate_provider_model_sets(model_cost_map: Dict) -> None: minimax_models.add(key) elif value.get("litellm_provider") == "aws_polly": aws_polly_models.add(key) + elif value.get("litellm_provider") == "transcribe": + transcribe_models.add(key) elif value.get("litellm_provider") == "gigachat": gigachat_models.add(key) elif value.get("litellm_provider") == "llamagate": @@ -1227,6 +1230,7 @@ def _build_models_by_provider() -> dict: "chatgpt": chatgpt_models, "minimax": minimax_models, "aws_polly": aws_polly_models, + "transcribe": transcribe_models, "gigachat": gigachat_models, "llamagate": llamagate_models, "reducto": reducto_models, diff --git a/litellm/types/utils.py b/litellm/types/utils.py index c63d971b89b..dbd293a7dc8 100644 --- a/litellm/types/utils.py +++ b/litellm/types/utils.py @@ -3989,6 +3989,7 @@ class LlmProviders(str, Enum): REDUCTO = "reducto" RUNWAYML = "runwayml" AWS_POLLY = "aws_polly" + TRANSCRIBE = "transcribe" HUGGINGFACE = "huggingface" TOGETHER_AI = "together_ai" OPENROUTER = "openrouter" diff --git a/tests/test_litellm/proxy/auth/test_model_checks.py b/tests/test_litellm/proxy/auth/test_model_checks.py index 3c6733cb86d..f10622e954b 100644 --- a/tests/test_litellm/proxy/auth/test_model_checks.py +++ b/tests/test_litellm/proxy/auth/test_model_checks.py @@ -879,3 +879,18 @@ def test_get_complete_model_list_sentinel_only_grants_nothing(): infer_model_from_keys=False, ) assert result == [] + + +def test_transcribe_is_a_known_provider_for_wildcard_expansion(): + import litellm + from litellm.proxy.auth.model_checks import ( + get_known_models_from_wildcard, + get_provider_models, + ) + + assert "transcribe" in litellm.models_by_provider + assert "transcribe/StartTranscriptionJob" in litellm.models_by_provider["transcribe"] + assert get_provider_models("transcribe") == ["transcribe/StartTranscriptionJob"] + assert get_known_models_from_wildcard("transcribe/*") == [ + "transcribe/StartTranscriptionJob" + ] diff --git a/tests/test_litellm/proxy/public_endpoints/test_public_endpoints.py b/tests/test_litellm/proxy/public_endpoints/test_public_endpoints.py index 0d82ed778f5..aaa3b205312 100644 --- a/tests/test_litellm/proxy/public_endpoints/test_public_endpoints.py +++ b/tests/test_litellm/proxy/public_endpoints/test_public_endpoints.py @@ -384,6 +384,7 @@ ADD_MODEL_UNLISTED_PROVIDERS: Final = frozenset( "tencent", "tensormesh", "text-completion-inception", + "transcribe", "valkey", "xiaomi_mimo", "zai", From f3b198c1b7aaa890c6a17b1bb02e65750370231f Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 21:56:43 -0700 Subject: [PATCH 17/37] fix(mistral): accept user_data as the OCR file purpose and keep OCR cost warnings single-line --- litellm/cost_calculator.py | 12 ++++--- litellm/llms/custom_httpx/llm_http_handler.py | 5 ++- litellm/llms/mistral/files/transformation.py | 31 +++++++++++-------- .../test_mistral_files_transformation.py | 22 ++++++++++++- 4 files changed, 51 insertions(+), 19 deletions(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 352d95a4fdf..30f2cb8489c 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -2146,8 +2146,8 @@ def ocr_batch_cost( if resolved_info is None: verbose_logger.warning( "OCR batch cost: model=%s custom_llm_provider=%s has no pricing entry; returning 0.0 cost.", - model, - custom_llm_provider, + _single_log_line(model), + _single_log_line(custom_llm_provider), ) return 0.0, 0.0 @@ -2159,14 +2159,18 @@ def ocr_batch_cost( verbose_logger.warning( "OCR batch cost: model=%s custom_llm_provider=%s reported pages_processed=%s but no " "ocr_cost_per_page is configured; returning 0.0 cost for those pages.", - model, - custom_llm_provider, + _single_log_line(model), + _single_log_line(custom_llm_provider), pages_processed, ) effective_annotation_rate: Final = annotation_rate if annotation_rate is not None else page_rate return (page_rate or 0.0) * pages_processed + (effective_annotation_rate or 0.0) * annotation_pages, 0.0 +def _single_log_line(value: str | None) -> str: + return str(value).replace("\n", "").replace("\r", "") + + def _lookup_model_info_or_none(model: str, custom_llm_provider: str | None) -> ModelInfo | None: try: return litellm.get_model_info(model=model, custom_llm_provider=custom_llm_provider) diff --git a/litellm/llms/custom_httpx/llm_http_handler.py b/litellm/llms/custom_httpx/llm_http_handler.py index 196b437c66b..8e0cdf547a2 100644 --- a/litellm/llms/custom_httpx/llm_http_handler.py +++ b/litellm/llms/custom_httpx/llm_http_handler.py @@ -307,7 +307,10 @@ def _mask_presigned_request_headers(transformed_request: bytes | str | dict) -> _get_masked_values, # pyright: ignore[reportPrivateUsage] # the shared header-masking helper has no public name ) - return {**transformed_request, "headers": _get_masked_values(request_headers)} + return { # mutable-ok: logging's curl and raw-request builders take dict + **transformed_request, + "headers": _get_masked_values(request_headers), + } def _aws_signing_overrides(optional_params: Mapping[str, Any], litellm_params: Mapping[str, Any]) -> Mapping[str, Any]: diff --git a/litellm/llms/mistral/files/transformation.py b/litellm/llms/mistral/files/transformation.py index bf1ef7cb69f..6e64961485a 100644 --- a/litellm/llms/mistral/files/transformation.py +++ b/litellm/llms/mistral/files/transformation.py @@ -8,6 +8,7 @@ Mistral only accepts ``fine-tune``, ``batch`` and ``ocr`` as upload purposes. import time from collections.abc import Mapping, Sequence +from types import MappingProxyType from typing import Final, Literal, TypeAlias import httpx @@ -33,6 +34,14 @@ from ..common_utils import get_mistral_api_base, get_mistral_auth_headers, mistr MistralFilePurpose: TypeAlias = Literal["fine-tune", "batch", "ocr"] +_OPENAI_PURPOSE_BY_MISTRAL: Final[Mapping[MistralFilePurpose, OpenAIFilesPurpose]] = MappingProxyType( + {"fine-tune": "fine-tune", "batch": "batch", "ocr": "user_data"} +) +_MISTRAL_PURPOSE_BY_OPENAI: Final[Mapping[str, MistralFilePurpose]] = MappingProxyType( + {"fine-tune": "fine-tune", "batch": "batch", "ocr": "ocr", "user_data": "ocr"} +) +_SUPPORTED_PURPOSES: Final = ", ".join(_MISTRAL_PURPOSE_BY_OPENAI) + _NO_QUERY_PARAMS: Final[dict[str, str]] = {} # mutable-ok: BaseFilesConfig request transforms return tuple[str, dict] @@ -81,22 +90,18 @@ def _to_openai_file_object(file: MistralFile) -> OpenAIFileObject: def _to_openai_purpose(purpose: MistralFilePurpose) -> OpenAIFilesPurpose: - match purpose: - case "fine-tune" | "batch": - return purpose - case "ocr": - return "user_data" + return _OPENAI_PURPOSE_BY_MISTRAL[purpose] def _to_mistral_purpose(purpose: str) -> MistralFilePurpose: - """Only Mistral's own purposes pass through. Silently mapping anything else to ``batch`` - would let an upload skip the proxy's batch-file validation and guardrails, which only - run when the caller says ``purpose=batch``.""" - match purpose: - case "batch" | "fine-tune" | "ocr": - return purpose - case _: - raise ValueError(f"Mistral does not support purpose={purpose!r}. Use one of: batch, fine-tune, ocr") + """``user_data`` is what an OCR file reads back as, since OpenAI's purpose literal has no ``ocr``, + so it maps back onto ``ocr``. Every other purpose Mistral lacks is rejected: silently rewriting + it to ``batch`` would let an upload skip the proxy's batch-file validation and guardrails, which + only run when the caller says ``purpose=batch``.""" + mistral_purpose: Final = _MISTRAL_PURPOSE_BY_OPENAI.get(purpose) + if mistral_purpose is None: + raise ValueError(f"Mistral does not support purpose={purpose!r}. Use one of: {_SUPPORTED_PURPOSES}") + return mistral_purpose def _api_base_from(litellm_params: Mapping[str, object]) -> str: diff --git a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py index b81740c0429..1dcd6d92a4b 100644 --- a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py +++ b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py @@ -102,7 +102,17 @@ def test_upload_request_passes_mistral_purposes_through(config, purpose): assert body["purpose"] == (None, purpose) -@pytest.mark.parametrize("purpose", ["assistants", "user_data", "vision", "evals"]) +def test_upload_request_maps_user_data_onto_ocr(config): + body = config.transform_create_file_request( + model="", + create_file_data=CreateFileRequest(file=("scan.pdf", b"%PDF"), purpose="user_data"), + optional_params={}, + litellm_params={}, + ) + assert body["purpose"] == (None, "ocr") + + +@pytest.mark.parametrize("purpose", ["assistants", "vision", "evals"]) def test_upload_request_rejects_purposes_mistral_lacks(config, purpose): """Regression: these used to be silently rewritten to ``batch``, so an upload that skipped the proxy's batch-only validation and guardrails still landed on Mistral as a batch input file.""" @@ -191,6 +201,16 @@ def test_list_request_filters_by_mapped_purpose(config): assert no_params == {} +def test_list_request_accepts_the_purpose_an_ocr_file_reads_back_as(config): + """Regression: an OCR file reads back as ``purpose=user_data``, and listing with that purpose + used to raise, so ``files.list(purpose=file.purpose)`` could never find OCR files.""" + ocr_file = config.transform_retrieve_file_response( + raw_response=_response(_file(purpose="ocr")), logging_obj=None, litellm_params={} + ) + _, params = config.transform_list_files_request(purpose=ocr_file.purpose, optional_params={}, litellm_params={}) + assert params == {"purpose": "ocr"} + + def test_list_request_rejects_purposes_mistral_lacks(config): with pytest.raises(ValueError, match="purpose='assistants'"): config.transform_list_files_request(purpose="assistants", optional_params={}, litellm_params={}) From 5783a38e27374919a003a6a8265098c7e676640e Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 22:30:07 -0700 Subject: [PATCH 18/37] fix(proxy): enforce the unified batch model grant before the DB shortcut and skip it for registry-routed vector store models retrieve_batch returned a terminal batch from the DB before checking that the key may use the model encoded in a unified batch id; the grant check now runs right after pre-call processing. The vector store file list helper authorized data["model"] through handle_model_based_routing even when the vector store registry set it server-side and even with no caller, which crashed on a None key; it now authorizes only a caller-supplied hint and resolves credentials directly. --- litellm/proxy/batches_endpoints/endpoints.py | 18 ++++++---- .../vector_store_files_endpoints/endpoints.py | 20 +++-------- .../proxy/batches_endpoints/test_endpoints.py | 15 ++++++++ .../test_vector_store_endpoints.py | 34 +++++++++++++++++++ 4 files changed, 65 insertions(+), 22 deletions(-) diff --git a/litellm/proxy/batches_endpoints/endpoints.py b/litellm/proxy/batches_endpoints/endpoints.py index 7bc38cdb33d..b8a485310c3 100644 --- a/litellm/proxy/batches_endpoints/endpoints.py +++ b/litellm/proxy/batches_endpoints/endpoints.py @@ -470,6 +470,17 @@ async def retrieve_batch( route_type="aretrieve_batch", ) + unified_model_id: Final = get_model_id_from_unified_batch_id(unified_batch_id) if unified_batch_id else None + if unified_model_id is not None: + resolved_unified_model: Final = ( + llm_router.resolve_model_name_from_model_id(unified_model_id) if llm_router is not None else None + ) + await authorize_model_for_key( + model_id=resolved_unified_model or unified_model_id, + llm_router=llm_router, + user_api_key_dict=user_api_key_dict, + ) + # FIX: First, try to read from ManagedObjectTable for consistent state managed_files_obj: Final = proxy_logging_obj.get_proxy_hook("managed_files") from litellm.proxy.proxy_server import prisma_client @@ -590,13 +601,6 @@ async def retrieve_batch( ) if unified_batch_id: - unified_model_id: Final = get_model_id_from_unified_batch_id(unified_batch_id) - if unified_model_id is not None: - await authorize_model_for_key( - model_id=llm_router.resolve_model_name_from_model_id(unified_model_id) or unified_model_id, - llm_router=llm_router, - user_api_key_dict=user_api_key_dict, - ) add_internal_model_credentials( data=data, llm_router=llm_router, diff --git a/litellm/proxy/vector_store_files_endpoints/endpoints.py b/litellm/proxy/vector_store_files_endpoints/endpoints.py index 11ef8efb598..50a98d01625 100644 --- a/litellm/proxy/vector_store_files_endpoints/endpoints.py +++ b/litellm/proxy/vector_store_files_endpoints/endpoints.py @@ -14,6 +14,7 @@ from litellm.proxy.common_utils.openai_endpoint_utils import ( get_custom_llm_provider_from_request_query, ) from litellm.proxy.openai_files_endpoints.common_utils import ( + get_credentials_for_model, handle_model_based_routing, prepare_data_with_credentials, ) @@ -262,26 +263,15 @@ async def _update_request_data_with_model_routing_hint( model_id=model_hint, team_id=caller_team_id ) should_route = credentials is not None - else: - if isinstance(model_hint, str) and should_authorize_model_hint: + elif isinstance(model_hint, str): + if should_authorize_model_hint: await _authorize_model_routing_hint( model=model_hint, llm_router=llm_router, user_api_key_dict=user_api_key_dict, ) - ( - should_route, - _model_used, - _original_file_id, - credentials, - ) = await handle_model_based_routing( - file_id="", - request=request, - llm_router=llm_router, - data=data, - user_api_key_dict=user_api_key_dict, - check_file_id_encoding=False, - ) + credentials = get_credentials_for_model(llm_router=llm_router, model_id=model_hint) + should_route = True if should_route and credentials is not None: prepare_data_with_credentials( diff --git a/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py b/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py index 46aec7a7031..cfbe48a241d 100644 --- a/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py +++ b/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py @@ -2908,6 +2908,21 @@ async def test_retrieve__unified_batch_id_rejects_key_without_model_grant(retrie retrieve_harness.creds_resolver.assert_not_called() +@pytest.mark.asyncio +async def test_retrieve__unified_batch_id_rejects_key_without_model_grant_before_db_terminal_shortcut( + retrieve_harness, +): + retrieve_harness.get_batch_from_db.return_value = (MagicMock(), make_batch(id="batch-from-db", status="completed")) + + with pytest.raises(ProxyException) as exc_info: + await call_retrieve(retrieve_harness, UNIFIED_BATCH_ID_FOR_GPT4O_MINI, user=_key_restricted_to("vertex-model")) + + assert exc_info.value.code == "403" + retrieve_harness.logging.post_call_success_hook.assert_not_called() + retrieve_harness.ensure_managed_files.assert_not_called() + retrieve_harness.router_aretrieve.assert_not_called() + + @pytest.mark.asyncio async def test_cancel__unified_batch_id_rejects_key_without_model_grant(cancel_harness): with pytest.raises(ProxyException) as exc_info: diff --git a/tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py b/tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py index dc445ec007c..52672b596ea 100644 --- a/tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py +++ b/tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py @@ -306,6 +306,40 @@ async def test_vector_store_file_list_resolves_credentials_from_model_query_para ) +@pytest.mark.asyncio +async def test_vector_store_file_list_registry_routed_model_skips_key_model_grant(): + request = MagicMock(spec=Request) + request.query_params = {} + request.headers = {} + + llm_router = MagicMock() + llm_router.get_deployment_credentials_with_provider.return_value = { + "api_key": "sk-team-openai", + "api_base": "https://api.openai.com/v1", + "custom_llm_provider": "openai", + "model": "openai/gpt-4o-mini", + } + + data = {"vector_store_id": "vs_123", "model": "team-openai"} + user_api_key_dict = UserAPIKeyAuth( + models=["restricted-deployment"], + team_models=["restricted-deployment"], + ) + + result = await _update_request_data_with_model_routing_hint( + data=data, + request=request, + llm_router=llm_router, + user_api_key_dict=user_api_key_dict, + ) + + assert result["api_key"] == "sk-team-openai" + assert result["model"] == "openai/gpt-4o-mini" + llm_router.get_deployment_credentials_with_provider.assert_called_once_with( + model_id="team-openai" + ) + + @pytest.mark.asyncio async def test_vector_store_file_list_resolves_single_openai_team_deployment(): request = MagicMock(spec=Request) From f89ca64481d6064770eeb49ee8795e7dbcc10432 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Fri, 18 Sep 2026 23:57:41 -0700 Subject: [PATCH 19/37] fix(batches): honor deployment OCR page pricing in batch cost and answer 400 for unsupported Mistral file purposes --- litellm/cost_calculator.py | 54 ++++++++++++------- litellm/litellm_core_utils/litellm_logging.py | 8 ++- litellm/llms/mistral/files/transformation.py | 6 ++- .../provider_endpoints_support_backup.json | 2 +- provider_endpoints_support.json | 2 +- .../test_litellm/batches/test_batch_utils.py | 29 ++++++++++ .../test_litellm_logging.py | 32 +++++++++++ .../test_mistral_files_transformation.py | 10 ++-- 8 files changed, 116 insertions(+), 27 deletions(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 30f2cb8489c..04c9675cc83 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -2115,12 +2115,8 @@ def ocr_cost( return ocr_pages_cost + annotation_pages_cost, 0.0 -_OCR_PRICING_KEYS: Final = ( - "ocr_cost_per_page", - "ocr_cost_per_page_batches", - "annotation_cost_per_page", - "annotation_cost_per_page_batches", -) +_OCR_BATCH_PAGE_RATE_KEYS: Final = ("ocr_cost_per_page_batches", "ocr_cost_per_page") +_OCR_BATCH_ANNOTATION_RATE_KEYS: Final = ("annotation_cost_per_page_batches", "annotation_cost_per_page") def ocr_batch_cost( @@ -2133,17 +2129,27 @@ def ocr_batch_cost( Batch OCR is billed per page at the ``*_batches`` rate, falling back to the synchronous per-page rate when a model has no batch price recorded, the same - fallback ``batch_cost_calculator`` applies to per-token batch pricing. Returns - ``(prompt_cost, completion_cost)`` with the whole cost in the first slot, like - ``ocr_cost``. + fallback ``batch_cost_calculator`` applies to per-token batch pricing. Each + per-page family (OCR pages, annotation pages) belongs to the deployment's + ``model_info`` when it prices that family at either rate and to the published + cost map otherwise, so a deployment overriding one family keeps the model's + published rate for the other, and the cost map is only consulted for a family + the deployment leaves out. Returns ``(prompt_cost, completion_cost)`` with the + whole cost in the first slot, like ``ocr_cost``. """ - has_ocr_pricing: Final = model_info is not None and any(model_info.get(k) is not None for k in _OCR_PRICING_KEYS) - resolved_info: Final = ( - model_info - if has_ocr_pricing - else _lookup_model_info_or_none(model=model, custom_llm_provider=custom_llm_provider) + pages_processed: Final = usage_info.pages_processed or 0 + annotation_pages: Final = usage_info.pages_processed_annotation or 0 + deployment_page_rate: Final = _first_price(model_info, *_OCR_BATCH_PAGE_RATE_KEYS) + deployment_annotation_rate: Final = _first_price(model_info, *_OCR_BATCH_ANNOTATION_RATE_KEYS) + needs_published_pricing: Final = (pages_processed > 0 and deployment_page_rate is None) or ( + annotation_pages > 0 and deployment_annotation_rate is None ) - if resolved_info is None: + published: Final = ( + _lookup_model_info_or_none(model=model, custom_llm_provider=custom_llm_provider) + if needs_published_pricing + else None + ) + if needs_published_pricing and published is None: verbose_logger.warning( "OCR batch cost: model=%s custom_llm_provider=%s has no pricing entry; returning 0.0 cost.", _single_log_line(model), @@ -2151,10 +2157,16 @@ def ocr_batch_cost( ) return 0.0, 0.0 - page_rate: Final = _first_price(resolved_info, "ocr_cost_per_page_batches", "ocr_cost_per_page") - annotation_rate: Final = _first_price(resolved_info, "annotation_cost_per_page_batches", "annotation_cost_per_page") - pages_processed: Final = usage_info.pages_processed or 0 - annotation_pages: Final = usage_info.pages_processed_annotation or 0 + page_rate: Final = ( + deployment_page_rate + if deployment_page_rate is not None + else _first_price(published, *_OCR_BATCH_PAGE_RATE_KEYS) + ) + annotation_rate: Final = ( + deployment_annotation_rate + if deployment_annotation_rate is not None + else _first_price(published, *_OCR_BATCH_ANNOTATION_RATE_KEYS) + ) if page_rate is None and pages_processed > 0: verbose_logger.warning( "OCR batch cost: model=%s custom_llm_provider=%s reported pages_processed=%s but no " @@ -2178,7 +2190,9 @@ def _lookup_model_info_or_none(model: str, custom_llm_provider: str | None) -> M return None -def _first_price(model_info: ModelInfo, *keys: str) -> float | None: +def _first_price(model_info: ModelInfo | None, *keys: str) -> float | None: + if model_info is None: + return None return next((price for price in (model_info.get(k) for k in keys) if isinstance(price, (int, float))), None) diff --git a/litellm/litellm_core_utils/litellm_logging.py b/litellm/litellm_core_utils/litellm_logging.py index eeebcec50b1..f7679b31f69 100644 --- a/litellm/litellm_core_utils/litellm_logging.py +++ b/litellm/litellm_core_utils/litellm_logging.py @@ -371,6 +371,10 @@ _DEPLOYMENT_PRICING_KEYS: Final = ( "output_cost_per_token", "input_cost_per_token_batches", "output_cost_per_token_batches", + "ocr_cost_per_page", + "ocr_cost_per_page_batches", + "annotation_cost_per_page", + "annotation_cost_per_page_batches", ) @@ -386,7 +390,9 @@ def deployment_pricing_model_info(model_id: str | None, deployment_model: str | the model's published rates instead of billing as zero. Ownership is per token direction: declaring either rate for a direction takes that whole direction, so a published batch rate can never displace a standard rate - the deployment configured itself. + the deployment configured itself. OCR per-page rates count as declared + pricing too; they pass through as registered and ``ocr_batch_cost`` layers + the published rate under each per-page family the deployment leaves out. """ if model_id is None: return None diff --git a/litellm/llms/mistral/files/transformation.py b/litellm/llms/mistral/files/transformation.py index 6e64961485a..88867ea4802 100644 --- a/litellm/llms/mistral/files/transformation.py +++ b/litellm/llms/mistral/files/transformation.py @@ -100,7 +100,11 @@ def _to_mistral_purpose(purpose: str) -> MistralFilePurpose: only run when the caller says ``purpose=batch``.""" mistral_purpose: Final = _MISTRAL_PURPOSE_BY_OPENAI.get(purpose) if mistral_purpose is None: - raise ValueError(f"Mistral does not support purpose={purpose!r}. Use one of: {_SUPPORTED_PURPOSES}") + raise mistral_error( + f"Mistral does not support purpose={purpose!r}. Use one of: {_SUPPORTED_PURPOSES}", + status_code=400, + headers=httpx.Headers(), + ) return mistral_purpose diff --git a/litellm/provider_endpoints_support_backup.json b/litellm/provider_endpoints_support_backup.json index dbeaccdda2d..30c1e0b894e 100644 --- a/litellm/provider_endpoints_support_backup.json +++ b/litellm/provider_endpoints_support_backup.json @@ -1462,7 +1462,7 @@ "audio_transcriptions": false, "audio_speech": false, "moderations": false, - "batches": false, + "batches": true, "rerank": false, "ocr": true, "a2a": true, diff --git a/provider_endpoints_support.json b/provider_endpoints_support.json index c71f4a82a4a..af9b194bbee 100644 --- a/provider_endpoints_support.json +++ b/provider_endpoints_support.json @@ -1577,7 +1577,7 @@ "audio_transcriptions": false, "audio_speech": false, "moderations": false, - "batches": false, + "batches": true, "rerank": false, "ocr": true, "a2a": true, diff --git a/tests/test_litellm/batches/test_batch_utils.py b/tests/test_litellm/batches/test_batch_utils.py index 4cad45ed809..1e7e0200754 100644 --- a/tests/test_litellm/batches/test_batch_utils.py +++ b/tests/test_litellm/batches/test_batch_utils.py @@ -1940,6 +1940,35 @@ def test_ocr_rows_use_deployment_model_info_pricing_over_cost_map(monkeypatch): assert result.cost == pytest.approx(0.01) +def test_ocr_rows_keep_the_published_page_rate_when_the_deployment_prices_only_annotations(monkeypatch): + monkeypatch.setattr( + litellm, + "get_model_info", + lambda model, custom_llm_provider=None: { + "ocr_cost_per_page_batches": 0.002, + "annotation_cost_per_page_batches": 0.0025, + }, + ) + result = bu._aggregate_batch_cost_usage_models( + entries=[_ocr_row(4, annotation_pages=4)], + custom_llm_provider="mistral", + model_info={"annotation_cost_per_page_batches": 0.01}, + ) + assert result.cost == pytest.approx(4 * 0.002 + 4 * 0.01) + + +def test_ocr_rows_bill_the_deployment_sync_page_rate_over_the_published_batch_rate(monkeypatch): + monkeypatch.setattr( + litellm, "get_model_info", lambda model, custom_llm_provider=None: pytest.fail("cost map must not be consulted") + ) + result = bu._aggregate_batch_cost_usage_models( + entries=[_ocr_row(3)], + custom_llm_provider="mistral", + model_info={"ocr_cost_per_page": 0.0912}, + ) + assert result.cost == pytest.approx(3 * 0.0912) + + def test_ocr_rows_without_pricing_bill_zero_but_count_as_successful(monkeypatch): monkeypatch.setattr(litellm, "get_model_info", lambda model, custom_llm_provider=None: {"mode": "ocr"}) result = bu._aggregate_batch_cost_usage_models(entries=[_ocr_row(3)], custom_llm_provider="mistral") diff --git a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py index 473d16a43f2..999adbdd935 100644 --- a/tests/test_litellm/litellm_core_utils/test_litellm_logging.py +++ b/tests/test_litellm/litellm_core_utils/test_litellm_logging.py @@ -17,12 +17,14 @@ from openai._legacy_response import HttpxBinaryResponseContent import litellm from litellm._logging import session_id_var, trace_id_var from litellm.constants import SENTRY_DENYLIST, SENTRY_PII_DENYLIST +from litellm.cost_calculator import ocr_batch_cost from litellm.integrations.custom_logger import CustomLogger from litellm.litellm_core_utils.litellm_logging import Logging as LitellmLogging from litellm.litellm_core_utils.litellm_logging import ( _get_status_fields, set_callbacks, ) +from litellm.llms.base_llm.ocr.transformation import OCRUsageInfo from litellm.types.llms.openai import ResponseAPIUsage, ResponsesAPIResponse from litellm.types.utils import ( CallTypes, @@ -529,6 +531,36 @@ class TestGetRouterDeploymentModelInfo: finally: litellm.model_cost.pop(deployment_id, None) + def test_ocr_only_deployment_pricing_reaches_batch_ocr_cost(self, logging_obj) -> None: + """Regression: a deployment priced only per page was treated as unpriced, so a retrieved OCR batch + billed at the published rate while the same deployment's synchronous OCR calls billed at its own.""" + deployment_id = "deploy-ocr-only-pricing-1" + litellm.model_cost[deployment_id] = { + "id": deployment_id, + "litellm_provider": "mistral", + "mode": "ocr", + "ocr_cost_per_page": 0.0456, + "ocr_cost_per_page_batches": 0.0123, + } + logging_obj.litellm_params = { + "litellm_metadata": {"model_info": {"id": deployment_id}}, + "model": "mistral/mistral-ocr-latest", + } + logging_obj.model_call_details["model"] = "mistral/mistral-ocr-latest" + published_annotation_rate = litellm.model_cost["mistral/mistral-ocr-latest"]["annotation_cost_per_page_batches"] + try: + info = logging_obj.get_router_deployment_model_info() + assert info is not None + assert info["ocr_cost_per_page_batches"] == 0.0123 + pages_only = OCRUsageInfo(pages_processed=3) + assert ocr_batch_cost("mistral-ocr-latest", "mistral", pages_only, info)[0] == pytest.approx(3 * 0.0123) + with_annotations = OCRUsageInfo(pages_processed=3, pages_processed_annotation=2) + assert ocr_batch_cost("mistral-ocr-latest", "mistral", with_annotations, info)[0] == pytest.approx( + 3 * 0.0123 + 2 * published_annotation_rate + ) + finally: + litellm.model_cost.pop(deployment_id, None) + class TestRetrieveBatchCostPassesModelIdentity: """Regression: retrieving a batch priced it with no model identity at all. diff --git a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py index 1dcd6d92a4b..5043cc583ef 100644 --- a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py +++ b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py @@ -13,6 +13,7 @@ import httpx import pytest from openai.types.file_deleted import FileDeleted +from litellm.llms.base_llm.chat.transformation import BaseLLMException from litellm.llms.mistral.files.transformation import MistralFilesConfig from litellm.types.llms.openai import CreateFileRequest, FileContentRequest, OpenAIFileObject from litellm.types.utils import LlmProviders @@ -115,14 +116,16 @@ def test_upload_request_maps_user_data_onto_ocr(config): @pytest.mark.parametrize("purpose", ["assistants", "vision", "evals"]) def test_upload_request_rejects_purposes_mistral_lacks(config, purpose): """Regression: these used to be silently rewritten to ``batch``, so an upload that skipped the - proxy's batch-only validation and guardrails still landed on Mistral as a batch input file.""" - with pytest.raises(ValueError, match=f"purpose={purpose!r}"): + proxy's batch-only validation and guardrails still landed on Mistral as a batch input file. The + rejection is a 400 provider error, so the proxy answers invalid_request_error instead of a 500.""" + with pytest.raises(BaseLLMException, match=f"purpose={purpose!r}") as exc_info: config.transform_create_file_request( model="", create_file_data=CreateFileRequest(file=("f.bin", b"x"), purpose=purpose), optional_params={}, litellm_params={}, ) + assert exc_info.value.status_code == 400 def test_upload_request_requires_file(config): @@ -212,8 +215,9 @@ def test_list_request_accepts_the_purpose_an_ocr_file_reads_back_as(config): def test_list_request_rejects_purposes_mistral_lacks(config): - with pytest.raises(ValueError, match="purpose='assistants'"): + with pytest.raises(BaseLLMException, match="purpose='assistants'") as exc_info: config.transform_list_files_request(purpose="assistants", optional_params={}, litellm_params={}) + assert exc_info.value.status_code == 400 def test_list_response(config): From 0feca8641f88a3c9674633243ecb17a8649a0fb7 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 00:30:15 -0700 Subject: [PATCH 20/37] fix(batches): price model-encoded batch retrievals by their deployment A batch retrieved by its model-encoded id takes the direct (non-router) path, which resolved credentials without stamping the deployment's model_info, so a completed batch on a deployment with its own per-page pricing was billed at the published rate with an empty model_id on the spend row. Extract the router's credential lookup into get_credential_deployment and stamp the resolved deployment's model_info onto the retrieve call the way the router does for routed calls. --- litellm/proxy/batches_endpoints/endpoints.py | 2 + .../openai_files_endpoints/common_utils.py | 19 ++++ litellm/router.py | 90 +++++++++++-------- .../proxy/batches_endpoints/test_endpoints.py | 26 ++++++ tests/test_litellm/test_router.py | 39 ++++++++ 5 files changed, 138 insertions(+), 38 deletions(-) diff --git a/litellm/proxy/batches_endpoints/endpoints.py b/litellm/proxy/batches_endpoints/endpoints.py index b8a485310c3..0e348fa6e06 100644 --- a/litellm/proxy/batches_endpoints/endpoints.py +++ b/litellm/proxy/batches_endpoints/endpoints.py @@ -32,6 +32,7 @@ from litellm.proxy.common_utils.openai_endpoint_utils import ( from litellm.proxy.openai_files_endpoints.common_utils import ( BATCH_CREATE_HIDDEN_PARAM, _is_base64_encoded_unified_file_id, + add_deployment_model_info, add_internal_model_credentials, apply_team_provider_credentials, authorize_model_for_key, @@ -580,6 +581,7 @@ async def retrieve_batch( # so litellm.aretrieve_batch can load BedrockBatchesConfig. Without # it the call falls into the legacy provider switch and 400s. data["model"] = model_from_id + add_deployment_model_info(data=data, llm_router=llm_router, model_id=model_from_id) # Retrieve batch using model credentials response = await litellm.aretrieve_batch( diff --git a/litellm/proxy/openai_files_endpoints/common_utils.py b/litellm/proxy/openai_files_endpoints/common_utils.py index 2d7e9f221b5..3d6c72de09f 100644 --- a/litellm/proxy/openai_files_endpoints/common_utils.py +++ b/litellm/proxy/openai_files_endpoints/common_utils.py @@ -593,6 +593,25 @@ def add_internal_model_credentials( data["_litellm_internal_model_credentials"] = MappingProxyType(dict(credentials)) +def add_deployment_model_info( + data: dict, + llm_router: Optional["Router"], + model_id: str, +) -> None: + """ + Stamp the resolved deployment's `model_info` onto a direct (non-router) batch call + (in-place), the way the router does for routed calls, so the completed batch is + priced by its deployment id instead of the published model rate. + """ + deployment: Final = llm_router.get_credential_deployment(model_id=model_id) if llm_router is not None else None + if deployment is None: + return + data["litellm_metadata"] = { + **(data.get("litellm_metadata") or {}), + "model_info": deployment.model_info.model_dump(), + } + + def prepare_data_with_credentials( data: dict, credentials: dict, diff --git a/litellm/router.py b/litellm/router.py index 74adf6f909d..3f3daaf2eae 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -10313,6 +10313,55 @@ class Router: return display_name return None + def get_credential_deployment(self, model_id: str, team_id: str | None = None) -> Deployment | None: + """ + The deployment a passthrough endpoint (files, batches, etc.) resolves for a + model id or model name: by deployment id first, then by model_name, then by + the team's exact public model name, then by wildcard pattern (team wildcards + before global ones, so a global "openai/*" never shadows the team's own + entry). Name and wildcard lookups never resolve another team's deployment. + + Returns None when nothing matches or the match is paused via + `LiteLLM_ProxyModelTable.blocked`, so callers cannot bypass an admin pause + by resolving the deployment directly. + """ + deployment: Final = ( + self.get_deployment(model_id=model_id) + or self._get_model_group_deployment_usable_by_team(model_group_name=model_id, team_id=team_id) + or self._get_team_public_name_deployment(model_id=model_id, team_id=team_id) + or self._get_wildcard_deployment_usable_by_team(model_id=model_id, team_id=team_id) + ) + if deployment is None or self._is_deployment_blocked(deployment): + return None + return deployment + + def _get_team_public_name_deployment(self, model_id: str, team_id: str | None) -> Deployment | None: + if team_id is None: + return None + team_indices: Final = self.team_model_to_deployment_indices.get((team_id, model_id)) + if not team_indices: + return None + team_model: Final = self.model_list[team_indices[0]] + return Deployment(**team_model) if isinstance(team_model, dict) else team_model + + def _get_wildcard_deployment_usable_by_team(self, model_id: str, team_id: str | None) -> Deployment | None: + team_pattern_router: Final = self.team_pattern_routers.get(team_id) if team_id is not None else None + team_wildcard_models: Final = team_pattern_router.route(model_id) if team_pattern_router else None + global_wildcard_models: Final = tuple( + wildcard_model + for wildcard_model in (self.pattern_router.route(model_id) or ()) + if self._deployment_usable_by_team(wildcard_model, team_id) + ) + potential_wildcard_models: Final = team_wildcard_models or global_wildcard_models + if not potential_wildcard_models: + return None + wildcard_deployment: Final = potential_wildcard_models[0] + if isinstance(wildcard_deployment, dict): + return Deployment(**wildcard_deployment) + if isinstance(wildcard_deployment, Deployment): + return wildcard_deployment + return None + def get_deployment_credentials_with_provider( self, model_id: str, team_id: str | None = None ) -> dict[str, Any] | None: @@ -10320,8 +10369,8 @@ class Router: Get API credentials and provider info from a model name in model_list. Useful for passthrough endpoints (files, batches, etc.) that need credentials. - This method tries to find a deployment by model_id first, and if not found, - it tries to find by model_group_name (model_name). + Resolves the deployment with `get_credential_deployment` (by deployment id, + then model_name, team public model name, and wildcard pattern). Args: model_id: Model ID or model name from model_list (e.g., "gpt-4o-litellm") @@ -10342,43 +10391,8 @@ class Router: credentials = router.get_deployment_credentials_with_provider("gpt-4o-litellm") # Returns: {"api_key": "sk-...", "custom_llm_provider": "openai", "model": "gpt-4o", ...} """ - # Try to get deployment by model_id first - deployment = self.get_deployment(model_id=model_id) - - # If not found, try by model_group_name + deployment: Final = self.get_credential_deployment(model_id=model_id, team_id=team_id) if deployment is None: - deployment = self._get_model_group_deployment_usable_by_team(model_group_name=model_id, team_id=team_id) - - # If not found, check team-scoped deployments whose team public model - # name exactly matches model_id (wildcard team names are matched via - # team_pattern_routers below). - if deployment is None and team_id is not None: - team_indices: Final = self.team_model_to_deployment_indices.get((team_id, model_id), []) - if team_indices: - team_model: Final = self.model_list[team_indices[0]] - deployment = Deployment(**team_model) if isinstance(team_model, dict) else team_model - - # If still not found, check for wildcard pattern matches. Team wildcard - # matches take priority so a global pattern (e.g. "openai/*") doesn't - # shadow the team's own entry. - if deployment is None: - team_pattern_router: Final = self.team_pattern_routers.get(team_id) if team_id is not None else None - team_wildcard_models: Final = (team_pattern_router.route(model_id) or []) if team_pattern_router else [] - global_wildcard_models: Final = [ - wildcard_model - for wildcard_model in (self.pattern_router.route(model_id) or []) - if self._deployment_usable_by_team(wildcard_model, team_id) - ] - potential_wildcard_models: Final = team_wildcard_models or global_wildcard_models - if potential_wildcard_models: - # Use the first matching wildcard deployment - deployment_dict: Final = potential_wildcard_models[0] - if isinstance(deployment_dict, dict): - deployment = Deployment(**deployment_dict) - elif isinstance(deployment_dict, Deployment): - deployment = deployment_dict - - if deployment is None or self._is_deployment_blocked(deployment): return None # Get basic credentials diff --git a/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py b/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py index cfbe48a241d..1b3f3806d79 100644 --- a/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py +++ b/tests/test_litellm/proxy/batches_endpoints/test_endpoints.py @@ -51,6 +51,7 @@ from litellm.proxy.openai_files_endpoints.common_utils import ( from litellm.proxy.utils import ProxyLogging from litellm.router import Router from litellm.types.llms.openai import BatchJobStatus +from litellm.types.router import Deployment, LiteLLM_Params, ModelInfo from litellm.types.utils import CredentialItem, LiteLLMBatch from fastapi import Request, Response @@ -1194,6 +1195,7 @@ def retrieve_harness(): router.model_list = [] router.aretrieve_batch = AsyncMock(return_value=make_batch()) router.get_deployment_credentials_with_provider = MagicMock(side_effect=_creds_lookup) + router.get_credential_deployment = MagicMock(return_value=None) pre_call = AsyncMock(side_effect=lambda **kw: (data_holder["data"], MagicMock())) get_headers = MagicMock(return_value={}) @@ -1312,6 +1314,30 @@ async def test_retrieve__model_encoded_id(retrieve_harness): assert retrieve_harness.update_batch_in_db.call_args.kwargs["operation"] == "retrieve" +@pytest.mark.asyncio +async def test_retrieve__model_encoded_id__stamps_deployment_model_info_for_cost(retrieve_harness): + """Regression: this path calls litellm.aretrieve_batch directly, so nothing stamped the + deployment's model_info the way the router does for routed calls. Cost tracking then never + saw the deployment id, and a completed batch on a deployment with its own per-page pricing + was billed at the published rate with an empty model_id on the spend row.""" + retrieve_harness.router.get_credential_deployment.return_value = Deployment( + model_name="azure-gpt", + litellm_params=LiteLLM_Params(model="azure/gpt-4o"), + model_info=ModelInfo(id="dep-123"), + ) + retrieve_harness.pre_call.side_effect = lambda **kw: ( + {**retrieve_harness.data["data"], "litellm_metadata": {"user_api_key_alias": "qa-key"}}, + MagicMock(), + ) + + await call_retrieve(retrieve_harness, AZURE_BATCH_ID) + + retrieve_harness.router.get_credential_deployment.assert_called_once_with(model_id="azure/gpt-4o") + litellm_metadata = retrieve_harness.aretrieve_kwargs()["litellm_metadata"] + assert litellm_metadata["model_info"]["id"] == "dep-123" + assert litellm_metadata["user_api_key_alias"] == "qa-key" + + @pytest.mark.asyncio async def test_retrieve__model_encoded_id__forwards_decoded_model_not_deployment( retrieve_harness, diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 7cc2a9e4c82..85954c7eb5f 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -5745,6 +5745,45 @@ async def test_router_unknown_model_error_message_renders_model_name_literally() assert " " not in message # no padding run from an expanded format field +def test_get_credential_deployment_is_the_deployment_credentials_resolve_to(): + """Regression: a batch retrieved with credentials resolved by model name was priced + without its deployment id, so per-deployment pricing never applied. The deployment + behind the credentials must be reachable by name and by id, carrying its model_info.""" + router = litellm.Router( + model_list=[ + { + "model_name": "mistral-ocr", + "litellm_params": {"model": "mistral/mistral-ocr-latest", "api_key": "sk-ocr"}, + "model_info": {"id": "ocr-dep", "ocr_cost_per_page_batches": 0.0123}, + } + ] + ) + + by_name = router.get_credential_deployment(model_id="mistral-ocr") + by_id = router.get_credential_deployment(model_id="ocr-dep") + + assert by_name is not None and by_id is not None + assert by_name.model_info.id == by_id.model_info.id == "ocr-dep" + assert by_name.model_info.model_dump()["ocr_cost_per_page_batches"] == 0.0123 + assert router.get_deployment_credentials_with_provider(model_id="mistral-ocr")["api_key"] == "sk-ocr" + assert router.get_credential_deployment(model_id="no-such-model") is None + + +def test_get_credential_deployment_skips_a_paused_deployment(): + router = litellm.Router( + model_list=[ + { + "model_name": "paused-ocr", + "litellm_params": {"model": "mistral/mistral-ocr-latest", "api_key": "sk-ocr"}, + "model_info": {"id": "paused-dep", "blocked": True}, + } + ] + ) + + assert router.get_credential_deployment(model_id="paused-ocr") is None + assert router.get_credential_deployment(model_id="paused-dep") is None + + def test_get_deployment_credentials_with_provider_aws_bedrock_runtime_endpoint(): """ Test that get_deployment_credentials_with_provider correctly copies From 7133baa7775c50134864883b4d6f7bb82cbc7c0f Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 00:54:41 -0700 Subject: [PATCH 21/37] fix(vector_stores): run the full model grant check on caller-supplied model hints The vector-store file routes accept a model hint through the ?model= query param and the x-litellm-model header. That hint was authorized with a hand rolled check that covered only the key allowlist and the team allowlist, so a key restricted by its project's model grant, a team-member restriction, or a key config still routed through the hinted deployment. Greptile flagged the gap as a P1 on the replacement PR. The hint now goes through the same authorize_model_for_key path the batches and files routes use, which runs can_key_call_resolved_model with every rule the proxy enforces elsewhere. Keys those extra rules deny now get a 403 on these routes. The two remaining behavioral differences are edge cases the old check tolerated: a key whose team_models is set without a team_id no longer runs the team allowlist, and a key with a config set skips the key allowlist, both matching the rest of the proxy. The regression test caches a project whose grant excludes the hinted model and asserts the request is refused before any deployment lookup. The two patch() calls on litellm.proxy.proxy_server carry a test-quality-ok reason because can_key_call_resolved_model reads prisma_client and user_api_key_cache through a lazy module import with no injection seam. --- .../vector_store_files_endpoints/endpoints.py | 23 +---------- .../test_vector_store_endpoints.py | 38 +++++++++++++++++++ 2 files changed, 40 insertions(+), 21 deletions(-) diff --git a/litellm/proxy/vector_store_files_endpoints/endpoints.py b/litellm/proxy/vector_store_files_endpoints/endpoints.py index 50a98d01625..97367e59023 100644 --- a/litellm/proxy/vector_store_files_endpoints/endpoints.py +++ b/litellm/proxy/vector_store_files_endpoints/endpoints.py @@ -5,7 +5,6 @@ from fastapi.responses import ORJSONResponse import litellm from litellm.proxy._types import UserAPIKeyAuth -from litellm.proxy.auth.auth_checks import _can_object_call_model, can_key_call_model from litellm.proxy.auth.user_api_key_auth import user_api_key_auth from litellm.proxy.common_request_processing import ProxyBaseLLMRequestProcessing from litellm.proxy.common_utils.openai_endpoint_utils import ( @@ -14,6 +13,7 @@ from litellm.proxy.common_utils.openai_endpoint_utils import ( get_custom_llm_provider_from_request_query, ) from litellm.proxy.openai_files_endpoints.common_utils import ( + authorize_model_for_key, get_credentials_for_model, handle_model_based_routing, prepare_data_with_credentials, @@ -212,26 +212,7 @@ async def _authorize_model_routing_hint( ) -> None: if user_api_key_dict is None: return - - key_models: Final = getattr(user_api_key_dict, "models", None) - if not (isinstance(key_models, list) and "all-team-models" in key_models): - await can_key_call_model( - model=model, - llm_model_list=None, - valid_token=user_api_key_dict, - llm_router=llm_router, - ) - - team_models: Final = getattr(user_api_key_dict, "team_models", None) - if isinstance(team_models, list) and len(team_models) > 0: - _can_object_call_model( - model=model, - llm_router=llm_router, - models=team_models, - team_model_aliases=user_api_key_dict.team_model_aliases, - team_id=user_api_key_dict.team_id, - object_type="team", - ) + await authorize_model_for_key(model_id=model, llm_router=llm_router, user_api_key_dict=user_api_key_dict) async def _update_request_data_with_model_routing_hint( diff --git a/tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py b/tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py index 52672b596ea..20484e787bd 100644 --- a/tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py +++ b/tests/test_litellm/proxy/vector_store_endpoints/test_vector_store_endpoints.py @@ -609,6 +609,44 @@ async def test_vector_store_file_list_authorizes_model_query_param_before_creden llm_router.get_deployment_credentials_with_provider.assert_not_called() +@pytest.mark.asyncio +async def test_vector_store_file_list_model_query_param_enforces_project_model_grant(): + from litellm.proxy._types import LiteLLM_ProjectTableCachedObj, LiteLLM_TeamTableCachedObj + from litellm.proxy.auth.auth_checks import ProxyException + from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache, project_cache_key + + request = MagicMock(spec=Request) + request.query_params = {"model": "team-openai"} + request.headers = {} + + llm_router = MagicMock() + llm_router.model_group_alias = {} + cache = UserApiKeyCache() + await cache.async_set_cache( + key="team_id:team-123", + value=LiteLLM_TeamTableCachedObj(team_id="team-123", models=["team-openai"]), + ) + await cache.async_set_cache( + key=project_cache_key("proj-1"), + value=LiteLLM_ProjectTableCachedObj(project_id="proj-1", models=["other-deployment"]), + ) + user_api_key_dict = UserAPIKeyAuth(team_id="team-123", team_models=["team-openai"], project_id="proj-1") + + with ( + patch("litellm.proxy.proxy_server.prisma_client", MagicMock()), # test-quality-ok: proxy_server global, no seam + patch("litellm.proxy.proxy_server.user_api_key_cache", cache), # test-quality-ok: proxy_server global, no seam + ): + with pytest.raises(ProxyException): + await _update_request_data_with_model_routing_hint( + data={"vector_store_id": "vs_123"}, + request=request, + llm_router=llm_router, + user_api_key_dict=user_api_key_dict, + ) + + llm_router.get_deployment_credentials_with_provider.assert_not_called() + + @pytest.mark.asyncio async def test_update_request_data_with_litellm_managed_vector_store_registry(): """ From 00214ac371470dd5a57162fe412992eb1e956d64 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 01:05:01 -0700 Subject: [PATCH 22/37] test(router): cover the team-scoped credential deployment lookups The router coverage gate in code-quality flags every router.py function no router test calls by name, and the two helpers get_credential_deployment gained (the team public-name lookup and the team-aware wildcard lookup) were only reached through it. Each now has a test of its own: the public-name lookup resolves only for the owning team, and the wildcard lookup prefers the team's own pattern over the shared one and never hands another team's wildcard deployment to a caller outside that team. --- tests/test_litellm/test_router.py | 48 +++++++++++++++++++++++++++++++ 1 file changed, 48 insertions(+) diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 85954c7eb5f..c5ae5d4b151 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -5784,6 +5784,54 @@ def test_get_credential_deployment_skips_a_paused_deployment(): assert router.get_credential_deployment(model_id="paused-dep") is None +def test_get_team_public_name_deployment_only_resolves_the_owning_team(): + router = litellm.Router( + model_list=[ + { + "model_name": "mistral/mistral-ocr-latest", + "litellm_params": {"model": "mistral/mistral-ocr-latest", "api_key": "sk-team-a"}, + "model_info": {"id": "team-a-ocr", "team_id": "team-a", "team_public_model_name": "ocr"}, + } + ] + ) + + owning_team = router._get_team_public_name_deployment(model_id="ocr", team_id="team-a") + + assert owning_team is not None and owning_team.model_info.id == "team-a-ocr" + assert router._get_team_public_name_deployment(model_id="ocr", team_id="team-b") is None + assert router._get_team_public_name_deployment(model_id="ocr", team_id=None) is None + assert router.get_credential_deployment(model_id="ocr", team_id="team-a").model_info.id == "team-a-ocr" + assert router.get_credential_deployment(model_id="ocr", team_id="team-b") is None + + +def test_get_wildcard_deployment_usable_by_team_prefers_the_team_pattern(): + router = litellm.Router( + model_list=[ + { + "model_name": "mistral/*", + "litellm_params": {"model": "mistral/*", "api_key": "sk-shared"}, + "model_info": {"id": "shared-wildcard"}, + }, + { + "model_name": "mistral/*", + "litellm_params": {"model": "mistral/*", "api_key": "sk-team-a"}, + "model_info": {"id": "team-a-wildcard", "team_id": "team-a", "team_public_model_name": "mistral/*"}, + }, + ] + ) + ocr = "mistral/mistral-ocr-latest" + + team_match = router._get_wildcard_deployment_usable_by_team(model_id=ocr, team_id="team-a") + other_team_match = router._get_wildcard_deployment_usable_by_team(model_id=ocr, team_id="team-b") + anonymous_match = router._get_wildcard_deployment_usable_by_team(model_id=ocr, team_id=None) + + assert team_match is not None and team_match.model_info.id == "team-a-wildcard" + assert other_team_match is not None and other_team_match.model_info.id == "shared-wildcard" + assert anonymous_match is not None and anonymous_match.model_info.id == "shared-wildcard" + assert router._get_wildcard_deployment_usable_by_team(model_id="openai/gpt-5.6", team_id="team-a") is None + assert router.get_credential_deployment(model_id=ocr, team_id="team-b").model_info.id == "shared-wildcard" + + def test_get_deployment_credentials_with_provider_aws_bedrock_runtime_endpoint(): """ Test that get_deployment_credentials_with_provider correctly copies From 6f4d1c5911253c866586007026c7604bc6a95fb2 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 01:41:35 -0700 Subject: [PATCH 23/37] fix(guardrails): stop the Javelin api_version default leaking into Azure Content Safety LitellmParams mixes every provider config model into one class, so the Javelin api_version default of "v1" reached the Azure Content Safety guardrails whenever config.yaml omitted api_version and Azure answered 404. The shared field now defaults to None, Javelin keeps filling in "v1" itself, and the Azure guardrails fall back to the documented 2024-09-01 at request time so a DB update that omits api_version stays on the default too. --- litellm/proxy/_lazy_openapi_snapshot.json | 3 +- .../guardrails/guardrail_hooks/azure/base.py | 7 ++- litellm/types/guardrails.py | 2 +- .../azure/test_azure_prompt_shield.py | 55 +++++++++++++++++++ .../azure/test_azure_text_moderation.py | 31 +++++++++++ .../guardrail_hooks/test_javelin.py | 42 ++++++++++++++ ui/litellm-dashboard/src/lib/http/schema.d.ts | 3 +- 7 files changed, 136 insertions(+), 7 deletions(-) create mode 100644 tests/test_litellm/proxy/guardrails/guardrail_hooks/test_javelin.py diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index 73ea8cf1991..3ddf0def821 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -11155,7 +11155,6 @@ "type": "null" } ], - "default": "v1", "description": "API version for Javelin service", "title": "Api Version" }, @@ -19622,7 +19621,7 @@ } } }, - "description": "\n Unified rate-limit error.\n\n Every rate-limit condition surfaced by litellm \u2014 whether it originated from\n an upstream LLM provider, a vendor batch endpoint, or one of litellm's own\n proxy-side limiters (parallel-requests, dynamic-rate, batch-rate, budget,\n max-iterations, etc.) \u2014 is raised as an instance of this class.\n\n The :attr:`category` attribute lets callers distinguish the source. See\n :class:`RateLimitErrorCategory` for the available values.\n " + "description": "\nUnified rate-limit error.\n\nEvery rate-limit condition surfaced by litellm \u2014 whether it originated from\nan upstream LLM provider, a vendor batch endpoint, or one of litellm's own\nproxy-side limiters (parallel-requests, dynamic-rate, batch-rate, budget,\nmax-iterations, etc.) \u2014 is raised as an instance of this class.\n\nThe :attr:`category` attribute lets callers distinguish the source. See\n:class:`RateLimitErrorCategory` for the available values.\n" }, "500": { "content": { diff --git a/litellm/proxy/guardrails/guardrail_hooks/azure/base.py b/litellm/proxy/guardrails/guardrail_hooks/azure/base.py index 42f0220cc4d..2338ed2e30d 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/azure/base.py +++ b/litellm/proxy/guardrails/guardrail_hooks/azure/base.py @@ -20,6 +20,8 @@ AZURE_CONTENT_SAFETY_MAX_TEXT_LENGTH: Final = 10000 # chunk of N characters consumes ceil(N / 1000) text records. AZURE_CONTENT_SAFETY_TEXT_RECORD_LENGTH: Final = 1000 +AZURE_CONTENT_SAFETY_DEFAULT_API_VERSION: Final = "2024-09-01" + class AzureGuardrailBase: """ @@ -43,7 +45,7 @@ class AzureGuardrailBase: self.async_handler = get_async_httpx_client(llm_provider=httpxSpecialProvider.GuardrailCallback) self.api_key = api_key self.api_base = api_base - self.api_version: str = kwargs.get("api_version") or "2024-09-01" + self.api_version: str | None = kwargs.get("api_version") async def _post_to_content_safety(self, endpoint_path: str, request_body: dict[str, object]) -> dict[str, Any]: """POST to an Azure Content Safety endpoint with standard auth headers. @@ -56,7 +58,8 @@ class AzureGuardrailBase: Returns: Parsed JSON response dict. """ - url: Final = f"{self.api_base}/contentsafety/{endpoint_path}?api-version={self.api_version}" + api_version: Final = self.api_version or AZURE_CONTENT_SAFETY_DEFAULT_API_VERSION + url: Final = f"{self.api_base}/contentsafety/{endpoint_path}?api-version={api_version}" headers: Final = { "Ocp-Apim-Subscription-Key": self.api_key, "Content-Type": "application/json", diff --git a/litellm/types/guardrails.py b/litellm/types/guardrails.py index 400cadd69e7..172edf136fd 100644 --- a/litellm/types/guardrails.py +++ b/litellm/types/guardrails.py @@ -819,7 +819,7 @@ class JavelinGuardrailConfigModel(BaseModel): """Configuration parameters for the Javelin guardrail""" guard_name: str | None = Field(default=None, description="Name of the Javelin guard to use") - api_version: str | None = Field(default="v1", description="API version for Javelin service") + api_version: str | None = Field(default=None, description="API version for Javelin service") metadata: dict | None = Field(default=None, description="Additional metadata to send with requests") application: str | None = Field(default=None, description="Application name for Javelin service") config: dict | None = Field(default=None, description="Additional configuration for the guardrail") diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_prompt_shield.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_prompt_shield.py index 17e7222fa44..0e7bf72706c 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_prompt_shield.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_prompt_shield.py @@ -7,6 +7,7 @@ from litellm.proxy._types import UserAPIKeyAuth from litellm.proxy.guardrails.guardrail_hooks.azure.prompt_shield import ( AzureContentSafetyPromptShieldGuardrail, ) +from litellm.proxy.guardrails.guardrail_registry import InMemoryGuardrailHandler from litellm.types.guardrails import LitellmParams @@ -635,3 +636,57 @@ def test_update_in_memory_litellm_params_dead_env_credential_rejected_untouched( assert guardrail.api_key == "azure_prompt_shield_api_key" assert guardrail.price_per_1000_text_records == 0.38 + + +@pytest.mark.asyncio +async def test_config_without_api_version_calls_documented_azure_api_version(): + """A config.yaml entry that omits api_version must reach Azure at the documented + default. LitellmParams inherits every provider's config model, so a sibling + provider's api_version default used to leak into the Azure URL and 404.""" + handler = InMemoryGuardrailHandler() + registered = handler.initialize_guardrail( + guardrail={ + "guardrail_name": "azure-prompt-shield-no-api-version", + "litellm_params": { + "guardrail": "azure/prompt_shield", + "mode": "pre_call", + "api_key": "azure_prompt_shield_api_key", + "api_base": "https://example.cognitiveservices.azure.com", + }, + } + ) + assert registered is not None + guardrail = handler.guardrail_id_to_custom_guardrail[registered["guardrail_id"]] + assert isinstance(guardrail, AzureContentSafetyPromptShieldGuardrail) + + with patch.object(guardrail.async_handler, "post", return_value=_shield_response(False)) as mock_post: + result = await guardrail.apply_guardrail(inputs={"texts": ["hello"]}, request_data={}, input_type="request") + + assert result == {"texts": ["hello"]} + assert mock_post.call_args.kwargs["url"] == ( + "https://example.cognitiveservices.azure.com/contentsafety/text:shieldPrompt?api-version=2024-09-01" + ) + + +@pytest.mark.asyncio +async def test_update_without_api_version_keeps_documented_azure_api_version(): + """The DB update path copies every LitellmParams attribute onto the live + instance, api_version included, so an update that omits it must still leave + the request on the documented default rather than a None or leaked value.""" + guardrail = _shield_guardrail() + guardrail.update_in_memory_litellm_params( + LitellmParams( + guardrail="azure/prompt_shield", + mode="pre_call", + api_key="azure_prompt_shield_api_key", + api_base="https://example.cognitiveservices.azure.com", + ) + ) + + with patch.object(guardrail.async_handler, "post", return_value=_shield_response(False)) as mock_post: + result = await guardrail.apply_guardrail(inputs={"texts": ["hello"]}, request_data={}, input_type="request") + + assert result == {"texts": ["hello"]} + assert mock_post.call_args.kwargs["url"] == ( + "https://example.cognitiveservices.azure.com/contentsafety/text:shieldPrompt?api-version=2024-09-01" + ) diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py index a43f95062f9..1798565f383 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py @@ -4,6 +4,7 @@ import pytest from fastapi import HTTPException from litellm.proxy._types import UserAPIKeyAuth +from litellm.proxy.guardrails.guardrail_registry import InMemoryGuardrailHandler from litellm.proxy.guardrails.guardrail_hooks.azure.text_moderation import ( AzureContentSafetyTextModerationGuardrail, ) @@ -463,3 +464,33 @@ async def test_apply_guardrail_handles_missing_texts_key(): mock_post.assert_not_called() assert result == {"images": ["x"]} + + +@pytest.mark.asyncio +async def test_config_without_api_version_calls_documented_azure_api_version(): + """A config.yaml entry that omits api_version must reach Azure at the documented + default. LitellmParams inherits every provider's config model, so a sibling + provider's api_version default used to leak into the Azure URL and 404.""" + handler = InMemoryGuardrailHandler() + registered = handler.initialize_guardrail( + guardrail={ + "guardrail_name": "azure-text-moderation-no-api-version", + "litellm_params": { + "guardrail": "azure/text_moderations", + "mode": "pre_call", + "api_key": "azure_text_moderation_api_key", + "api_base": "https://example.cognitiveservices.azure.com", + }, + } + ) + assert registered is not None + guardrail = handler.guardrail_id_to_custom_guardrail[registered["guardrail_id"]] + assert isinstance(guardrail, AzureContentSafetyTextModerationGuardrail) + + with patch.object(guardrail.async_handler, "post", return_value=_moderation_response(0)) as mock_post: + result = await guardrail.apply_guardrail(inputs={"texts": ["hello"]}, request_data={}, input_type="request") + + assert result == {"texts": ["hello"]} + assert mock_post.call_args.kwargs["url"] == ( + "https://example.cognitiveservices.azure.com/contentsafety/text:analyze?api-version=2024-09-01" + ) diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_javelin.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_javelin.py new file mode 100644 index 00000000000..b5283255eb2 --- /dev/null +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_javelin.py @@ -0,0 +1,42 @@ +from unittest.mock import Mock, patch + +import pytest + +from litellm.proxy.guardrails.guardrail_hooks.javelin.javelin import JavelinGuardrail +from litellm.proxy.guardrails.guardrail_registry import InMemoryGuardrailHandler +from litellm.types.guardrails import GuardrailEventHooks + + +@pytest.mark.asyncio +async def test_config_without_api_version_calls_javelin_v1(): + """Javelin's v1 default no longer lives in the shared LitellmParams model (it + leaked into every other provider), so the Javelin initializer has to supply + it itself when the config omits api_version.""" + handler = InMemoryGuardrailHandler() + registered = handler.initialize_guardrail( + guardrail={ + "guardrail_name": "javelin-no-api-version", + "litellm_params": { + "guardrail": "javelin", + "mode": "pre_call", + "api_key": "javelin_api_key", + "api_base": "https://javelin.example", + "guard_name": "trustsafety", + }, + } + ) + assert registered is not None + guardrail = handler.guardrail_id_to_custom_guardrail[registered["guardrail_id"]] + assert isinstance(guardrail, JavelinGuardrail) + assessments = [{"trustsafety": {"request_reject": False}}] + response = Mock() + response.json.return_value = {"assessments": assessments} + + with patch.object(guardrail.async_handler, "post", return_value=response) as mock_post: + result = await guardrail.call_javelin_guard( + request={"input": {"text": "hello"}, "config": None, "metadata": None}, + event_type=GuardrailEventHooks.pre_call, + ) + + assert result == {"assessments": assessments} + assert mock_post.call_args.kwargs["url"] == "https://javelin.example/v1/guardrail/trustsafety/apply" diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts index a5ede63bf7f..2bb98fa4c66 100644 --- a/ui/litellm-dashboard/src/lib/http/schema.d.ts +++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts @@ -31795,9 +31795,8 @@ export interface components { /** * Api Version * @description API version for Javelin service - * @default v1 */ - api_version: string | null; + api_version?: string | null; /** * Application * @description Application name for Javelin service From dd79c1f77d7f1bb6391e9aac06e11b15e61c17ad Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 01:59:39 -0700 Subject: [PATCH 24/37] fix(cost): keep the deployment's OCR page rate when the model has no published price When a deployment priced one OCR batch family and the other still needed a published rate, a failed cost-map lookup returned zero for the whole line and discarded the deployment rate that was already resolved. Those pages were billed as free. The lookup failure now only logs, and the families the deployment prices are billed at the configured rate --- litellm/cost_calculator.py | 4 ++-- tests/test_litellm/batches/test_batch_utils.py | 13 +++++++++++++ 2 files changed, 15 insertions(+), 2 deletions(-) diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py index 04c9675cc83..38758867a11 100644 --- a/litellm/cost_calculator.py +++ b/litellm/cost_calculator.py @@ -2151,11 +2151,11 @@ def ocr_batch_cost( ) if needs_published_pricing and published is None: verbose_logger.warning( - "OCR batch cost: model=%s custom_llm_provider=%s has no pricing entry; returning 0.0 cost.", + "OCR batch cost: model=%s custom_llm_provider=%s has no pricing entry; " + "billing only the per-page families the deployment prices.", _single_log_line(model), _single_log_line(custom_llm_provider), ) - return 0.0, 0.0 page_rate: Final = ( deployment_page_rate diff --git a/tests/test_litellm/batches/test_batch_utils.py b/tests/test_litellm/batches/test_batch_utils.py index 1e7e0200754..708c472939e 100644 --- a/tests/test_litellm/batches/test_batch_utils.py +++ b/tests/test_litellm/batches/test_batch_utils.py @@ -1957,6 +1957,19 @@ def test_ocr_rows_keep_the_published_page_rate_when_the_deployment_prices_only_a assert result.cost == pytest.approx(4 * 0.002 + 4 * 0.01) +def test_ocr_rows_keep_the_deployment_page_rate_when_the_unmapped_model_has_no_annotation_price(monkeypatch): + def _unmapped(model, custom_llm_provider=None): + raise Exception(f"This model isn't mapped yet: {model}") + + monkeypatch.setattr(litellm, "get_model_info", _unmapped) + result = bu._aggregate_batch_cost_usage_models( + entries=[_ocr_row(4, annotation_pages=4, model="my-private-ocr-model")], + custom_llm_provider="mistral", + model_info={"ocr_cost_per_page_batches": 0.001}, + ) + assert result.cost == pytest.approx(4 * 0.001 + 4 * 0.001) + + def test_ocr_rows_bill_the_deployment_sync_page_rate_over_the_published_batch_rate(monkeypatch): monkeypatch.setattr( litellm, "get_model_info", lambda model, custom_llm_provider=None: pytest.fail("cost map must not be consulted") From f784681bfaf3c4af42c98e1c9c1bd13ca740ac01 Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Sat, 19 Sep 2026 09:28:14 +0000 Subject: [PATCH 25/37] refactor(types): replace Any with proven types in 6 files Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/integrations/otel/logger.py | 9 +++++++-- .../bedrock/chat/agentcore/transformation.py | 5 +++-- .../guardrails/guardrail_hooks/alice/alice.py | 11 +++++++--- .../guardrail_hooks/grayswan/grayswan.py | 10 +++++++--- .../promptguard/promptguard.py | 20 ++++++++++++++----- .../guardrail_hooks/singulr/singulr.py | 11 +++++----- 6 files changed, 46 insertions(+), 20 deletions(-) diff --git a/litellm/integrations/otel/logger.py b/litellm/integrations/otel/logger.py index c3b30f0983e..607611eb971 100644 --- a/litellm/integrations/otel/logger.py +++ b/litellm/integrations/otel/logger.py @@ -5,7 +5,7 @@ from collections.abc import Callable, Iterator, Mapping, Sequence from contextlib import contextmanager from datetime import datetime from types import MappingProxyType -from typing import TYPE_CHECKING, Any, Final, cast +from typing import TYPE_CHECKING, Final, cast from opentelemetry.context import Context, attach, get_current from opentelemetry.sdk._logs import LoggerProvider @@ -21,6 +21,7 @@ from opentelemetry.trace import ( use_span, ) from opentelemetry.trace import TracerProvider as ApiTracerProvider +from typing_extensions import TypedDict, Unpack import litellm from litellm._logging import verbose_logger @@ -140,6 +141,10 @@ def _request_trace_links(context: Context | None) -> tuple[Link, ...] | None: return (Link(anchor),) if anchor.is_valid else None +class _CustomLoggerOptions(TypedDict, total=False, extra_items=object): + """Keyword arguments forwarded untouched to ``CustomLogger`` and ``OpenTelemetryV2Config``.""" + + class _LLMCallSpan: """The state carried from the ``pre_call`` boundary to span close. @@ -179,7 +184,7 @@ class OpenTelemetryV2(CustomLogger): tracer_provider: TracerProvider | None = None, logger_provider: LoggerProvider | None = None, meter_provider: "MeterProvider | None" = None, - **kwargs: Any, + **kwargs: Unpack[_CustomLoggerOptions], ) -> None: super().__init__(**kwargs) self.config: OpenTelemetryV2Config = config or OpenTelemetryV2Config(**kwargs) diff --git a/litellm/llms/bedrock/chat/agentcore/transformation.py b/litellm/llms/bedrock/chat/agentcore/transformation.py index 6aa17372258..e1a9a807abc 100644 --- a/litellm/llms/bedrock/chat/agentcore/transformation.py +++ b/litellm/llms/bedrock/chat/agentcore/transformation.py @@ -6,7 +6,7 @@ https://docs.aws.amazon.com/bedrock/latest/APIReference/API_agentcore_InvokeAgen import json from collections.abc import AsyncGenerator -from typing import TYPE_CHECKING, Any, Final, Optional, Union, cast +from typing import TYPE_CHECKING, Any, Final, Optional, Union from urllib.parse import quote import httpx @@ -31,6 +31,7 @@ from litellm.types.llms.openai import AllMessageValues from litellm.types.utils import ( Choices, Delta, + LlmProviders, Message, ModelResponse, ModelResponseStream, @@ -872,7 +873,7 @@ class AmazonAgentCoreConfig(BaseConfig, BaseAWSLLM): ) if client is None or not isinstance(client, AsyncHTTPHandler): - client = get_async_httpx_client(llm_provider=cast(Any, "bedrock"), params={}) + client = get_async_httpx_client(llm_provider=LlmProviders.BEDROCK, params={}) verbose_logger.debug("Making async streaming request to: %s", api_base) diff --git a/litellm/proxy/guardrails/guardrail_hooks/alice/alice.py b/litellm/proxy/guardrails/guardrail_hooks/alice/alice.py index 9cabac2d0fa..da97359b299 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/alice/alice.py +++ b/litellm/proxy/guardrails/guardrail_hooks/alice/alice.py @@ -11,14 +11,13 @@ from collections.abc import Mapping from itertools import islice from typing import ( TYPE_CHECKING, - Any, # noqa: TID251 # **kwargs forwards verbatim to CustomGuardrail.__init__; see ruff-strict.toml Final, Literal, Optional, ) import httpx -from typing_extensions import NotRequired, ReadOnly, TypedDict +from typing_extensions import NotRequired, ReadOnly, TypedDict, Unpack from litellm._logging import verbose_proxy_logger from litellm.exceptions import GuardrailRaisedException, Timeout @@ -92,6 +91,10 @@ class AliceVerdict(TypedDict): replacements: ReadOnly[NotRequired["tuple[AliceReplacement, ...]"]] +class _CustomGuardrailOptions(TypedDict, total=False, extra_items=object): + """Base-class constructor options this guardrail forwards untouched to CustomGuardrail.""" + + class AliceGuardrailMissingSecrets(Exception): """Raised when the Alice API key is not configured.""" @@ -144,7 +147,9 @@ class AliceGuardrail(CustomGuardrail): api_key: str | None = None, api_base: str | None = None, unreachable_fallback: Literal["fail_closed", "fail_open"] = "fail_closed", - **kwargs: Any, # kwargs-ok: forwarded verbatim to CustomGuardrail.__init__, whose param list is wide and evolving + **kwargs: Unpack[ # kwargs-ok: forwarded verbatim to CustomGuardrail.__init__, whose param list is wide and evolving + _CustomGuardrailOptions + ], ) -> None: self.async_handler = get_async_httpx_client(llm_provider=httpxSpecialProvider.GuardrailCallback) diff --git a/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py b/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py index 48832f8ed5e..14f60ce09a0 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py +++ b/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py @@ -2,10 +2,10 @@ import os import time -from typing import TYPE_CHECKING, Any, Final, Literal, Optional, Protocol +from typing import TYPE_CHECKING, Final, Literal, Optional, Protocol from fastapi import HTTPException -from typing_extensions import NotRequired, ReadOnly, TypedDict +from typing_extensions import NotRequired, ReadOnly, TypedDict, Unpack from litellm._logging import verbose_proxy_logger from litellm.integrations.custom_guardrail import ( @@ -38,6 +38,10 @@ class _GraySwanMonitorResponse(TypedDict): ipi: ReadOnly[NotRequired[bool | None]] +class _CustomGuardrailOptions(TypedDict, total=False, extra_items=object): + """Base-class constructor options this guardrail forwards untouched to CustomGuardrail.""" + + class _GraySwanMonitorHTTPResponse(Protocol): def raise_for_status(self) -> object: ... @@ -103,7 +107,7 @@ class GraySwanGuardrail(CustomGuardrail): streaming_sampling_rate: int = 5, fail_open: bool | None = True, guardrail_timeout: float | None = 30.0, - **kwargs: Any, + **kwargs: Unpack[_CustomGuardrailOptions], ) -> None: self.async_handler: _GraySwanMonitorHTTPClient = get_async_httpx_client( llm_provider=httpxSpecialProvider.GuardrailCallback diff --git a/litellm/proxy/guardrails/guardrail_hooks/promptguard/promptguard.py b/litellm/proxy/guardrails/guardrail_hooks/promptguard/promptguard.py index 2edd6567850..5ab47dfc3e1 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/promptguard/promptguard.py +++ b/litellm/proxy/guardrails/guardrail_hooks/promptguard/promptguard.py @@ -7,9 +7,10 @@ before and after LLM calls. """ import os -from typing import TYPE_CHECKING, Any, Final, Literal, Optional, TypedDict +from typing import TYPE_CHECKING, Final, Literal, Optional, TypedDict -from typing_extensions import ReadOnly +from typing_extensions import ReadOnly, Unpack +from typing_extensions import TypedDict as ExtraItemsTypedDict from litellm._logging import verbose_proxy_logger from litellm.exceptions import GuardrailRaisedException @@ -53,6 +54,12 @@ class PromptGuardHTTPView(TypedDict): guard_response: ReadOnly[PromptGuardGuardAPIResponse] +class _CustomGuardrailOptions(ExtraItemsTypedDict, total=False, extra_items=object): + """Base-class constructor options this guardrail forwards untouched to CustomGuardrail.""" + + supported_event_hooks: ReadOnly[list[GuardrailEventHooks] | None] + + class PromptGuardMissingCredentials(Exception): pass @@ -63,7 +70,7 @@ class PromptGuardGuardrail(CustomGuardrail): api_key: str | None = None, api_base: str | None = None, block_on_error: bool | None = None, - **kwargs: Any, + **kwargs: Unpack[_CustomGuardrailOptions], ) -> None: self.api_key = api_key or os.environ.get( "PROMPTGUARD_API_KEY", @@ -92,9 +99,12 @@ class PromptGuardGuardrail(CustomGuardrail): llm_provider=httpxSpecialProvider.GuardrailCallback, ) - kwargs.setdefault("supported_event_hooks", list(self.get_supported_event_hooks())) + options: Final[_CustomGuardrailOptions] = { + "supported_event_hooks": list(self.get_supported_event_hooks()), + **kwargs, + } - super().__init__(**kwargs) + super().__init__(**options) @staticmethod def get_config_model() -> type["GuardrailConfigModel"] | None: diff --git a/litellm/proxy/guardrails/guardrail_hooks/singulr/singulr.py b/litellm/proxy/guardrails/guardrail_hooks/singulr/singulr.py index a91812bb474..06d4b39f5f6 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/singulr/singulr.py +++ b/litellm/proxy/guardrails/guardrail_hooks/singulr/singulr.py @@ -24,6 +24,7 @@ from litellm.llms.custom_httpx.http_handler import ( ) from litellm.proxy._types import UserAPIKeyAuth from litellm.types.guardrails import GuardrailEventHooks +from litellm.types.llms.openai import AllMessageValues from litellm.types.proxy.guardrails.guardrail_hooks.base import ( GuardrailConfigModel, ) @@ -40,7 +41,7 @@ from litellm.types.utils import CallTypes, GenericGuardrailAPIInputs _DEFAULT_API_BASE: Final = "http://localhost:8003" _GUARD_ENDPOINT: Final = "/api/v1/ai-gateway/litellm-v2" _DEFAULT_TIMEOUT: Final = 30.0 -_EMPTY_MAPPING: Final[Mapping[str, Any]] = MappingProxyType({}) +_EMPTY_MAPPING: Final[Mapping[str, Mapping[str, object]]] = MappingProxyType({}) _MCP_MODEL_PREFIX: Final = "MCP:" @@ -159,7 +160,7 @@ class SingulrGuardrail(CustomGuardrail): return {key: value for key, value in resolved if value} # mutable-ok: short-lived JSON payload dict @staticmethod - def _build_user_message(text: str) -> Mapping[str, Any]: + def _build_user_message(text: str) -> Mapping[str, str]: return {"role": "user", "content": text} # mutable-ok: short-lived JSON payload dict def _build_headers(self) -> Mapping[str, str]: @@ -224,7 +225,7 @@ class SingulrGuardrail(CustomGuardrail): self, inputs: GenericGuardrailAPIInputs, texts: Sequence[str], - structured_messages: Sequence[Any], + structured_messages: Sequence[AllMessageValues], request_data: Mapping[str, Any], ) -> GenericGuardrailAPIInputs: messages: Final = ( @@ -271,12 +272,12 @@ class SingulrGuardrail(CustomGuardrail): return request_data.get("mcp_tool_name") or request_data.get("name") @staticmethod - def _mcp_arguments(request_data: Mapping[str, Any]) -> object: + def _mcp_arguments(request_data: Mapping[str, object]) -> object: arguments: Final = request_data.get("mcp_arguments") return arguments if arguments is not None else request_data.get("arguments") @staticmethod - def _is_mcp_call(request_data: Mapping[str, Any], logging_obj: LiteLLMLoggingObj | None) -> bool: + def _is_mcp_call(request_data: Mapping[str, object], logging_obj: LiteLLMLoggingObj | None) -> bool: call_type: Final = logging_obj.call_type if logging_obj is not None else request_data.get("call_type") if call_type is not None: return call_type == CallTypes.call_mcp_tool.value From 24064e3b3181dd6d65afe10f5e2f51f3d58e0c19 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 02:29:43 -0700 Subject: [PATCH 26/37] fix(guardrails): treat the stored Javelin api_version default as unset for Azure Content Safety Guardrails created through POST /guardrails on older releases have api_version "v1" saved in the database, because the writer persists every default. Azure Content Safety never accepts that value, so those guardrails kept answering 404 after the default moved to None. The Azure base now resolves "v1" to 2024-09-01 the same way it resolves a missing value. Also restores the OpenAPI snapshot line that a Python 3.14 regeneration had dedented --- litellm/proxy/_lazy_openapi_snapshot.json | 2 +- .../guardrails/guardrail_hooks/azure/base.py | 9 ++++- .../azure/test_azure_text_moderation.py | 35 +++++++++++++++++++ 3 files changed, 44 insertions(+), 2 deletions(-) diff --git a/litellm/proxy/_lazy_openapi_snapshot.json b/litellm/proxy/_lazy_openapi_snapshot.json index 3ddf0def821..4cfb2bf8c38 100644 --- a/litellm/proxy/_lazy_openapi_snapshot.json +++ b/litellm/proxy/_lazy_openapi_snapshot.json @@ -19621,7 +19621,7 @@ } } }, - "description": "\nUnified rate-limit error.\n\nEvery rate-limit condition surfaced by litellm \u2014 whether it originated from\nan upstream LLM provider, a vendor batch endpoint, or one of litellm's own\nproxy-side limiters (parallel-requests, dynamic-rate, batch-rate, budget,\nmax-iterations, etc.) \u2014 is raised as an instance of this class.\n\nThe :attr:`category` attribute lets callers distinguish the source. See\n:class:`RateLimitErrorCategory` for the available values.\n" + "description": "\n Unified rate-limit error.\n\n Every rate-limit condition surfaced by litellm \u2014 whether it originated from\n an upstream LLM provider, a vendor batch endpoint, or one of litellm's own\n proxy-side limiters (parallel-requests, dynamic-rate, batch-rate, budget,\n max-iterations, etc.) \u2014 is raised as an instance of this class.\n\n The :attr:`category` attribute lets callers distinguish the source. See\n :class:`RateLimitErrorCategory` for the available values.\n " }, "500": { "content": { diff --git a/litellm/proxy/guardrails/guardrail_hooks/azure/base.py b/litellm/proxy/guardrails/guardrail_hooks/azure/base.py index 2338ed2e30d..d2aa11da7c9 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/azure/base.py +++ b/litellm/proxy/guardrails/guardrail_hooks/azure/base.py @@ -21,6 +21,13 @@ AZURE_CONTENT_SAFETY_MAX_TEXT_LENGTH: Final = 10000 AZURE_CONTENT_SAFETY_TEXT_RECORD_LENGTH: Final = 1000 AZURE_CONTENT_SAFETY_DEFAULT_API_VERSION: Final = "2024-09-01" +JAVELIN_API_VERSION_STORED_BY_OLDER_RELEASES: Final = "v1" + + +def resolve_content_safety_api_version(configured: str | None) -> str: + if not configured or configured == JAVELIN_API_VERSION_STORED_BY_OLDER_RELEASES: + return AZURE_CONTENT_SAFETY_DEFAULT_API_VERSION + return configured class AzureGuardrailBase: @@ -58,7 +65,7 @@ class AzureGuardrailBase: Returns: Parsed JSON response dict. """ - api_version: Final = self.api_version or AZURE_CONTENT_SAFETY_DEFAULT_API_VERSION + api_version: Final = resolve_content_safety_api_version(self.api_version) url: Final = f"{self.api_base}/contentsafety/{endpoint_path}?api-version={api_version}" headers: Final = { "Ocp-Apim-Subscription-Key": self.api_key, diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py index 1798565f383..acb15b4d869 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py @@ -494,3 +494,38 @@ async def test_config_without_api_version_calls_documented_azure_api_version(): assert mock_post.call_args.kwargs["url"] == ( "https://example.cognitiveservices.azure.com/contentsafety/text:analyze?api-version=2024-09-01" ) + + +@pytest.mark.parametrize( + ("stored_api_version", "expected_api_version"), + [("v1", "2024-09-01"), ("2023-10-01", "2023-10-01")], +) +@pytest.mark.asyncio +async def test_guardrail_loaded_with_stored_api_version_calls_azure_at(stored_api_version, expected_api_version): + """Releases before the api_version default fix saved every guardrail created + through the API or dashboard with Javelin's "v1", which Azure always answers + with 404. A row like that must reach Azure at the documented default, while a + real Azure version an admin chose is sent as written.""" + handler = InMemoryGuardrailHandler() + registered = handler.initialize_guardrail( + guardrail={ + "guardrail_name": f"azure-text-moderation-stored-{stored_api_version}", + "litellm_params": { + "guardrail": "azure/text_moderations", + "mode": "pre_call", + "api_key": "azure_text_moderation_api_key", + "api_base": "https://example.cognitiveservices.azure.com", + "api_version": stored_api_version, + }, + } + ) + assert registered is not None + guardrail = handler.guardrail_id_to_custom_guardrail[registered["guardrail_id"]] + + with patch.object(guardrail.async_handler, "post", return_value=_moderation_response(0)) as mock_post: + result = await guardrail.apply_guardrail(inputs={"texts": ["hello"]}, request_data={}, input_type="request") + + assert result == {"texts": ["hello"]} + assert mock_post.call_args.kwargs["url"] == ( + f"https://example.cognitiveservices.azure.com/contentsafety/text:analyze?api-version={expected_api_version}" + ) From 6e0356a7a0fa180f11e1e038022b4dfe204e28a3 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 02:34:27 -0700 Subject: [PATCH 27/37] test: fail a required shard when a cost map provider is unregistered --- .../test_litellm/test_model_prices_schema.py | 49 +++++++++++++++++++ 1 file changed, 49 insertions(+) diff --git a/tests/test_litellm/test_model_prices_schema.py b/tests/test_litellm/test_model_prices_schema.py index a9015397b32..fa42c65fb9a 100644 --- a/tests/test_litellm/test_model_prices_schema.py +++ b/tests/test_litellm/test_model_prices_schema.py @@ -11,7 +11,9 @@ from typing import Final import jsonschema import pytest +import litellm from litellm.llms.openai.chat.gpt_5_transformation import is_gpt_reasoning_series_name +from litellm.llms.openai_like.json_loader import JSONProviderRegistry from litellm.router_utils.reasoning_effort_capability import resolve_supported_reasoning_efforts REPO_ROOT = Path(__file__).parents[2] @@ -363,3 +365,50 @@ def test_active_mistral_chat_rows_price_cache_reads_below_input(path: Path): and not cache_read_is_tenth_of_input(entry) ] assert drifted == [] + + +PROVIDER_LABELS_WITHOUT_A_MODEL_SET: Final = frozenset({"sagemaker", "bedrock_converse"}) +MODES_SERVED_OUTSIDE_THE_LLM_PROVIDER_REGISTRY: Final = frozenset({"search", "evaluation"}) + + +def is_registered_provider(label: str) -> bool: + family_root: Final = label.split("-", 1)[0] + return any( + name in litellm.models_by_provider or JSONProviderRegistry.exists(name) for name in (label, family_root) + ) + + +def unregistered_providers(rows: Mapping[str, object]) -> list[str]: + return sorted( + { + entry["litellm_provider"] + for name, entry in rows.items() + if name != "sample_spec" + and isinstance(entry, dict) + and "litellm_provider" in entry + and entry.get("mode") not in MODES_SERVED_OUTSIDE_THE_LLM_PROVIDER_REGISTRY + and entry["litellm_provider"] not in PROVIDER_LABELS_WITHOUT_A_MODEL_SET + and not is_registered_provider(entry["litellm_provider"]) + } + ) + + +@pytest.mark.parametrize("path", (PRICES_PATH, BACKUP_PRICES_PATH), ids=("main", "backup")) +def test_every_cost_map_provider_is_registered(path: Path): + assert unregistered_providers(json.loads(path.read_text())) == [], ( + f"{path.name} carries a litellm_provider that litellm.models_by_provider does not know, so a `/*` " + "grant expands to no models. Add a `_models` set in litellm/__init__.py, fill it in " + "_populate_provider_model_sets, and list it in _build_models_by_provider" + ) + + +def test_unregistered_provider_guard_flags_only_labels_nobody_registered(): + rows: Final = { + "sample_spec": {"litellm_provider": "one of the supported providers", "mode": "chat"}, + "nobody_registered/StartJob": {"litellm_provider": "nobody_registered", "mode": "audio_transcription"}, + "gpt-4o": {"litellm_provider": "openai", "mode": "chat"}, + "vertex_ai/new-family-model": {"litellm_provider": "vertex_ai-new_family_models", "mode": "chat"}, + "unknown_root/model": {"litellm_provider": "unknown_root-new_family_models", "mode": "chat"}, + "some_search/search": {"litellm_provider": "some_search", "mode": "search"}, + } + assert unregistered_providers(rows) == ["nobody_registered", "unknown_root-new_family_models"] From 988bb65aa236393c81a3b267882d538376cc04b9 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 02:50:31 -0700 Subject: [PATCH 28/37] test: require a provider family's rows to reach its wildcard list --- .../test_litellm/test_model_prices_schema.py | 57 +++++++++++++------ 1 file changed, 41 insertions(+), 16 deletions(-) diff --git a/tests/test_litellm/test_model_prices_schema.py b/tests/test_litellm/test_model_prices_schema.py index fa42c65fb9a..918aff806c1 100644 --- a/tests/test_litellm/test_model_prices_schema.py +++ b/tests/test_litellm/test_model_prices_schema.py @@ -369,46 +369,71 @@ def test_active_mistral_chat_rows_price_cache_reads_below_input(path: Path): PROVIDER_LABELS_WITHOUT_A_MODEL_SET: Final = frozenset({"sagemaker", "bedrock_converse"}) MODES_SERVED_OUTSIDE_THE_LLM_PROVIDER_REGISTRY: Final = frozenset({"search", "evaluation"}) +VERTEX_FAMILIES_A_VERTEX_WILDCARD_GRANT_DOES_NOT_LIST: Final = frozenset( + { + "vertex_ai-ai21_models", + "vertex_ai-embedding-models", + "vertex_ai-image-models", + "vertex_ai-llama_models", + "vertex_ai-mistral_models", + "vertex_ai-openai_models", + "vertex_ai-qwen_models", + "vertex_ai-video-models", + } +) -def is_registered_provider(label: str) -> bool: +def is_registered_provider(label: str, model_names: tuple[str, ...]) -> bool: + if label in litellm.models_by_provider or JSONProviderRegistry.exists(label): + return True family_root: Final = label.split("-", 1)[0] + wildcard_models: Final = litellm.models_by_provider.get(family_root, ()) return any( - name in litellm.models_by_provider or JSONProviderRegistry.exists(name) for name in (label, family_root) + name in wildcard_models or name.removeprefix(f"{family_root}/") in wildcard_models for name in model_names ) def unregistered_providers(rows: Mapping[str, object]) -> list[str]: + labelled_rows: Final = tuple( + (name, entry["litellm_provider"]) + for name, entry in rows.items() + if name != "sample_spec" + and isinstance(entry, dict) + and "litellm_provider" in entry + and entry.get("mode") not in MODES_SERVED_OUTSIDE_THE_LLM_PROVIDER_REGISTRY + and entry["litellm_provider"] not in PROVIDER_LABELS_WITHOUT_A_MODEL_SET + and entry["litellm_provider"] not in VERTEX_FAMILIES_A_VERTEX_WILDCARD_GRANT_DOES_NOT_LIST + ) return sorted( - { - entry["litellm_provider"] - for name, entry in rows.items() - if name != "sample_spec" - and isinstance(entry, dict) - and "litellm_provider" in entry - and entry.get("mode") not in MODES_SERVED_OUTSIDE_THE_LLM_PROVIDER_REGISTRY - and entry["litellm_provider"] not in PROVIDER_LABELS_WITHOUT_A_MODEL_SET - and not is_registered_provider(entry["litellm_provider"]) - } + label + for label in {label for _, label in labelled_rows} + if not is_registered_provider(label, tuple(name for name, row_label in labelled_rows if row_label == label)) ) @pytest.mark.parametrize("path", (PRICES_PATH, BACKUP_PRICES_PATH), ids=("main", "backup")) def test_every_cost_map_provider_is_registered(path: Path): assert unregistered_providers(json.loads(path.read_text())) == [], ( - f"{path.name} carries a litellm_provider that litellm.models_by_provider does not know, so a `/*` " - "grant expands to no models. Add a `_models` set in litellm/__init__.py, fill it in " - "_populate_provider_model_sets, and list it in _build_models_by_provider" + f"{path.name} carries a litellm_provider whose models a `/*` grant does not list. A new provider " + "needs a `_models` set in litellm/__init__.py, filled in _populate_provider_model_sets and listed " + "in _build_models_by_provider. A new `-` label needs its rows added to a set that " + "`models_by_provider[]` includes" ) def test_unregistered_provider_guard_flags_only_labels_nobody_registered(): + wired_vertex_model: Final = sorted(litellm.vertex_language_models)[0] rows: Final = { "sample_spec": {"litellm_provider": "one of the supported providers", "mode": "chat"}, "nobody_registered/StartJob": {"litellm_provider": "nobody_registered", "mode": "audio_transcription"}, "gpt-4o": {"litellm_provider": "openai", "mode": "chat"}, + wired_vertex_model: {"litellm_provider": "vertex_ai-language-models", "mode": "chat"}, "vertex_ai/new-family-model": {"litellm_provider": "vertex_ai-new_family_models", "mode": "chat"}, "unknown_root/model": {"litellm_provider": "unknown_root-new_family_models", "mode": "chat"}, "some_search/search": {"litellm_provider": "some_search", "mode": "search"}, } - assert unregistered_providers(rows) == ["nobody_registered", "unknown_root-new_family_models"] + assert unregistered_providers(rows) == [ + "nobody_registered", + "unknown_root-new_family_models", + "vertex_ai-new_family_models", + ] From e327a6ae7652a4e8c1949af5839abebfdc408e2d Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 02:52:42 -0700 Subject: [PATCH 29/37] test(guardrails): drop regression docstrings from the api_version tests --- .../guardrail_hooks/azure/test_azure_prompt_shield.py | 6 ------ .../guardrail_hooks/azure/test_azure_text_moderation.py | 7 ------- .../proxy/guardrails/guardrail_hooks/test_javelin.py | 3 --- 3 files changed, 16 deletions(-) diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_prompt_shield.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_prompt_shield.py index 0e7bf72706c..f4af4b5ead7 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_prompt_shield.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_prompt_shield.py @@ -640,9 +640,6 @@ def test_update_in_memory_litellm_params_dead_env_credential_rejected_untouched( @pytest.mark.asyncio async def test_config_without_api_version_calls_documented_azure_api_version(): - """A config.yaml entry that omits api_version must reach Azure at the documented - default. LitellmParams inherits every provider's config model, so a sibling - provider's api_version default used to leak into the Azure URL and 404.""" handler = InMemoryGuardrailHandler() registered = handler.initialize_guardrail( guardrail={ @@ -670,9 +667,6 @@ async def test_config_without_api_version_calls_documented_azure_api_version(): @pytest.mark.asyncio async def test_update_without_api_version_keeps_documented_azure_api_version(): - """The DB update path copies every LitellmParams attribute onto the live - instance, api_version included, so an update that omits it must still leave - the request on the documented default rather than a None or leaked value.""" guardrail = _shield_guardrail() guardrail.update_in_memory_litellm_params( LitellmParams( diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py index acb15b4d869..4fbc33edcd6 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/azure/test_azure_text_moderation.py @@ -468,9 +468,6 @@ async def test_apply_guardrail_handles_missing_texts_key(): @pytest.mark.asyncio async def test_config_without_api_version_calls_documented_azure_api_version(): - """A config.yaml entry that omits api_version must reach Azure at the documented - default. LitellmParams inherits every provider's config model, so a sibling - provider's api_version default used to leak into the Azure URL and 404.""" handler = InMemoryGuardrailHandler() registered = handler.initialize_guardrail( guardrail={ @@ -502,10 +499,6 @@ async def test_config_without_api_version_calls_documented_azure_api_version(): ) @pytest.mark.asyncio async def test_guardrail_loaded_with_stored_api_version_calls_azure_at(stored_api_version, expected_api_version): - """Releases before the api_version default fix saved every guardrail created - through the API or dashboard with Javelin's "v1", which Azure always answers - with 404. A row like that must reach Azure at the documented default, while a - real Azure version an admin chose is sent as written.""" handler = InMemoryGuardrailHandler() registered = handler.initialize_guardrail( guardrail={ diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_javelin.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_javelin.py index b5283255eb2..dc58b67e3f8 100644 --- a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_javelin.py +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_javelin.py @@ -9,9 +9,6 @@ from litellm.types.guardrails import GuardrailEventHooks @pytest.mark.asyncio async def test_config_without_api_version_calls_javelin_v1(): - """Javelin's v1 default no longer lives in the shared LitellmParams model (it - leaked into every other provider), so the Javelin initializer has to supply - it itself when the config omits api_version.""" handler = InMemoryGuardrailHandler() registered = handler.initialize_guardrail( guardrail={ From 343e1eeac88ee360f42096c18e131a84a9d67a91 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 02:53:36 -0700 Subject: [PATCH 30/37] ci(unit): fail a hung test in 120s with a traceback instead of idling the shard to its step timeout --- .github/workflows/_test-unit-base.yml | 17 ++++ tests/test_litellm/rerank_api/test_main.py | 1 + .../test_unit_shard_per_test_timeout.py | 87 +++++++++++++++++++ 3 files changed, 105 insertions(+) create mode 100644 tests/test_litellm/test_unit_shard_per_test_timeout.py diff --git a/.github/workflows/_test-unit-base.yml b/.github/workflows/_test-unit-base.yml index bbf0cb4e891..617b09a8075 100644 --- a/.github/workflows/_test-unit-base.yml +++ b/.github/workflows/_test-unit-base.yml @@ -37,6 +37,18 @@ on: required: false type: number default: 60 + test-timeout-seconds: + description: >- + Per-test ceiling enforced by pytest-timeout, covering fixture setup and + teardown as well as the test body. A test that hangs fails with a + traceback of where it was stuck instead of idling the shard until + `timeout-minutes` cancels it. Timed-out tests are excluded from reruns + because pytest-timeout arms its timer once per test and + pytest-rerunfailures reruns inside that same window, so a rerun of a + timed-out test would run with no timer at all. + required: false + type: number + default: 120 max-failures: description: "Stop after this many failures" required: false @@ -137,6 +149,7 @@ jobs: MAX_FAILURES: ${{ inputs.max-failures }} WORKERS: ${{ inputs.workers }} RERUNS: ${{ inputs.reruns }} + TEST_TIMEOUT_SECONDS: ${{ inputs.test-timeout-seconds }} DIST: ${{ inputs.dist }} COVERAGE_CORE: sysmon run: | @@ -146,6 +159,8 @@ jobs: --maxfail="${MAX_FAILURES}" \ --reruns "${RERUNS}" \ --reruns-delay 1 \ + --timeout="${TEST_TIMEOUT_SECONDS}" \ + --rerun-except "from pytest-timeout" \ --durations=20 \ --cov=./litellm --cov=./enterprise/litellm_enterprise \ --cov-report=xml:coverage.xml \ @@ -157,6 +172,8 @@ jobs: -n "${WORKERS}" \ --reruns "${RERUNS}" \ --reruns-delay 1 \ + --timeout="${TEST_TIMEOUT_SECONDS}" \ + --rerun-except "from pytest-timeout" \ --dist="${DIST}" \ --durations=20 \ --cov=./litellm --cov=./enterprise/litellm_enterprise \ diff --git a/tests/test_litellm/rerank_api/test_main.py b/tests/test_litellm/rerank_api/test_main.py index 0992cd9bb37..aca0c970dd5 100644 --- a/tests/test_litellm/rerank_api/test_main.py +++ b/tests/test_litellm/rerank_api/test_main.py @@ -239,6 +239,7 @@ async def test_arerank_error_is_mapped_to_litellm_exception(respx_mock: respx.Mo @pytest.mark.asyncio +@pytest.mark.timeout(300) async def test_arerank_declared_authenticating_provider_skips_resolution(monkeypatch): """Regression for the event-loop hazard in arerank's provider pre-resolution: get_llm_provider runs the blocking OAuth device flow for github_copilot/chatgpt, diff --git a/tests/test_litellm/test_unit_shard_per_test_timeout.py b/tests/test_litellm/test_unit_shard_per_test_timeout.py new file mode 100644 index 00000000000..e6c3db460e1 --- /dev/null +++ b/tests/test_litellm/test_unit_shard_per_test_timeout.py @@ -0,0 +1,87 @@ +import shlex +import subprocess +import sys +from pathlib import Path +from string import Template +from types import MappingProxyType +from typing import Final + +import pytest +import yaml + +_REPO_ROOT: Final = Path(__file__).resolve().parents[2] +_BASE_WORKFLOW: Final = _REPO_ROOT / ".github" / "workflows" / "_test-unit-base.yml" +_SHARD_ENV: Final = MappingProxyType({"WORKERS": "2", "RERUNS": "2", "DIST": "loadscope", "TEST_TIMEOUT_SECONDS": "1"}) +_HANG_GUARD_FLAGS: Final = frozenset(("-n", "--dist", "--reruns", "--reruns-delay", "--timeout", "--rerun-except")) +_HUNG_TEST_MODULE: Final = """ +import threading + +import pytest + + +@pytest.fixture +def hangs_on_teardown(): + yield + threading.Event().wait() + + +def test_body_waits_forever(): + threading.Event().wait() + + +def test_fixture_teardown_waits_forever(hangs_on_teardown): + assert True + + +def test_passes(): + assert True +""" + + +def _run_tests_script() -> str: + workflow: Final = yaml.safe_load(_BASE_WORKFLOW.read_text()) + return next(step["run"] for step in workflow["jobs"]["run"]["steps"] if step.get("name") == "Run tests") + + +def _pytest_invocations(script: str) -> tuple[tuple[str, ...], ...]: + return tuple(tuple(shlex.split(line)) for line in script.replace("\\\n", " ").splitlines() if " pytest " in line) + + +def _hang_guard_args(invocation: tuple[str, ...]) -> tuple[str, ...]: + return tuple( + Template(token).safe_substitute(_SHARD_ENV) + for previous, token in zip(("", *invocation), invocation) + if token.split("=", 1)[0] in _HANG_GUARD_FLAGS or previous in _HANG_GUARD_FLAGS + ) + + +_INVOCATIONS: Final = _pytest_invocations(_run_tests_script()) + + +def test_the_shard_script_runs_pytest_serially_and_under_xdist() -> None: + assert sorted("-n" in invocation for invocation in _INVOCATIONS) == [False, True] + + +@pytest.mark.parametrize( + "invocation", _INVOCATIONS, ids=tuple("xdist" if "-n" in invocation else "serial" for invocation in _INVOCATIONS) +) +def test_a_hung_test_fails_fast_and_names_itself_under_the_shard_flags( + invocation: tuple[str, ...], tmp_path: Path +) -> None: + hung_module: Final = tmp_path / "test_hung.py" + hung_module.write_text(_HUNG_TEST_MODULE) + + result: Final = subprocess.run( + (sys.executable, "-m", "pytest", str(hung_module), "-p", "no:cacheprovider", *_hang_guard_args(invocation)), + cwd=tmp_path, + capture_output=True, + text=True, + timeout=90, + check=False, + ) + + assert result.returncode == 1, result.stdout + assert "FAILED test_hung.py::test_body_waits_forever" in result.stdout + assert "ERROR test_hung.py::test_fixture_teardown_waits_forever" in result.stdout + assert "Timeout (>1.0s) from pytest-timeout" in result.stdout + assert "1 failed, 2 passed, 1 error" in result.stdout From 217ff78ae7ebc2a77f0069a61364d13cc000dfc2 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 03:04:45 -0700 Subject: [PATCH 31/37] fix(mistral): read back files whose purpose Mistral never lets us upload as user_data --- litellm/llms/mistral/files/transformation.py | 12 +++++++----- .../files/test_mistral_files_transformation.py | 17 +++++++++++++++++ 2 files changed, 24 insertions(+), 5 deletions(-) diff --git a/litellm/llms/mistral/files/transformation.py b/litellm/llms/mistral/files/transformation.py index 88867ea4802..6edf188d247 100644 --- a/litellm/llms/mistral/files/transformation.py +++ b/litellm/llms/mistral/files/transformation.py @@ -3,7 +3,8 @@ Mistral Files API. Reference: https://docs.mistral.ai/api/#tag/files Mistral's file objects already carry the OpenAI field names (id, bytes, created_at, filename, purpose), so this config is URL routing, auth, and a purpose mapping: -Mistral only accepts ``fine-tune``, ``batch`` and ``ocr`` as upload purposes. +Mistral only accepts ``fine-tune``, ``batch`` and ``ocr`` as upload purposes, while files +other Mistral products created read back with purposes outside that set and map onto ``user_data``. """ import time @@ -34,9 +35,10 @@ from ..common_utils import get_mistral_api_base, get_mistral_auth_headers, mistr MistralFilePurpose: TypeAlias = Literal["fine-tune", "batch", "ocr"] -_OPENAI_PURPOSE_BY_MISTRAL: Final[Mapping[MistralFilePurpose, OpenAIFilesPurpose]] = MappingProxyType( +_OPENAI_PURPOSE_BY_MISTRAL: Final[Mapping[str, OpenAIFilesPurpose]] = MappingProxyType( {"fine-tune": "fine-tune", "batch": "batch", "ocr": "user_data"} ) +_OPENAI_PURPOSE_FOR_UNMAPPED: Final[OpenAIFilesPurpose] = "user_data" _MISTRAL_PURPOSE_BY_OPENAI: Final[Mapping[str, MistralFilePurpose]] = MappingProxyType( {"fine-tune": "fine-tune", "batch": "batch", "ocr": "ocr", "user_data": "ocr"} ) @@ -59,7 +61,7 @@ class MistralFile(BaseModel): bytes: int = 0 created_at: int | None = None filename: str = "" - purpose: MistralFilePurpose = "batch" + purpose: str = "batch" expires_at: int | None = None @@ -89,8 +91,8 @@ def _to_openai_file_object(file: MistralFile) -> OpenAIFileObject: ) -def _to_openai_purpose(purpose: MistralFilePurpose) -> OpenAIFilesPurpose: - return _OPENAI_PURPOSE_BY_MISTRAL[purpose] +def _to_openai_purpose(purpose: str) -> OpenAIFilesPurpose: + return _OPENAI_PURPOSE_BY_MISTRAL.get(purpose, _OPENAI_PURPOSE_FOR_UNMAPPED) def _to_mistral_purpose(purpose: str) -> MistralFilePurpose: diff --git a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py index 5043cc583ef..303afe99a2c 100644 --- a/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py +++ b/tests/test_litellm/llms/mistral/files/test_mistral_files_transformation.py @@ -158,6 +158,23 @@ def test_file_response_with_ocr_purpose_maps_onto_user_data(config): assert obj.expires_at == 1_800_000_000 +@pytest.mark.parametrize("purpose", ["playground", "audio", "code_interpreter"]) +def test_files_with_purposes_mistral_never_lets_us_upload_still_read_back(config, purpose): + """Regression: Mistral's live API returns purposes its upload endpoint rejects for files + other Mistral products created, and both the unfiltered list and a retrieve of such a file + used to fail validation, so one playground file 500'd ``GET /v1/files`` for the whole key.""" + retrieved = config.transform_retrieve_file_response( + raw_response=_response(_file(purpose=purpose)), logging_obj=None, litellm_params={} + ) + assert retrieved.purpose == "user_data" + listed = config.transform_list_files_response( + raw_response=_response({"data": [_file(purpose=purpose), _file(id="second")], "object": "list", "total": 2}), + logging_obj=None, + litellm_params={}, + ) + assert [(f.id, f.purpose) for f in listed] == [(FILE_ID, "user_data"), ("second", "batch")] + + @pytest.mark.parametrize( "method,suffix", [ From 8ee7591fc8065dfc92d528bab48866c0360e01ba Mon Sep 17 00:00:00 2001 From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> Date: Sat, 19 Sep 2026 10:05:14 +0000 Subject: [PATCH 32/37] refactor(types): drop nonessential TypedDict docstrings Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- litellm/integrations/otel/logger.py | 2 +- litellm/proxy/guardrails/guardrail_hooks/alice/alice.py | 2 +- litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py | 2 +- .../proxy/guardrails/guardrail_hooks/promptguard/promptguard.py | 2 -- 4 files changed, 3 insertions(+), 5 deletions(-) diff --git a/litellm/integrations/otel/logger.py b/litellm/integrations/otel/logger.py index 607611eb971..6b673967427 100644 --- a/litellm/integrations/otel/logger.py +++ b/litellm/integrations/otel/logger.py @@ -142,7 +142,7 @@ def _request_trace_links(context: Context | None) -> tuple[Link, ...] | None: class _CustomLoggerOptions(TypedDict, total=False, extra_items=object): - """Keyword arguments forwarded untouched to ``CustomLogger`` and ``OpenTelemetryV2Config``.""" + pass class _LLMCallSpan: diff --git a/litellm/proxy/guardrails/guardrail_hooks/alice/alice.py b/litellm/proxy/guardrails/guardrail_hooks/alice/alice.py index da97359b299..bcc35e7a22f 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/alice/alice.py +++ b/litellm/proxy/guardrails/guardrail_hooks/alice/alice.py @@ -92,7 +92,7 @@ class AliceVerdict(TypedDict): class _CustomGuardrailOptions(TypedDict, total=False, extra_items=object): - """Base-class constructor options this guardrail forwards untouched to CustomGuardrail.""" + pass class AliceGuardrailMissingSecrets(Exception): diff --git a/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py b/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py index 14f60ce09a0..cc3ed7172b6 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py +++ b/litellm/proxy/guardrails/guardrail_hooks/grayswan/grayswan.py @@ -39,7 +39,7 @@ class _GraySwanMonitorResponse(TypedDict): class _CustomGuardrailOptions(TypedDict, total=False, extra_items=object): - """Base-class constructor options this guardrail forwards untouched to CustomGuardrail.""" + pass class _GraySwanMonitorHTTPResponse(Protocol): diff --git a/litellm/proxy/guardrails/guardrail_hooks/promptguard/promptguard.py b/litellm/proxy/guardrails/guardrail_hooks/promptguard/promptguard.py index 5ab47dfc3e1..f51f59ab0d1 100644 --- a/litellm/proxy/guardrails/guardrail_hooks/promptguard/promptguard.py +++ b/litellm/proxy/guardrails/guardrail_hooks/promptguard/promptguard.py @@ -55,8 +55,6 @@ class PromptGuardHTTPView(TypedDict): class _CustomGuardrailOptions(ExtraItemsTypedDict, total=False, extra_items=object): - """Base-class constructor options this guardrail forwards untouched to CustomGuardrail.""" - supported_event_hooks: ReadOnly[list[GuardrailEventHooks] | None] From 4968e89f3cde7390b411270f0476418ccf70abab Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 03:20:48 -0700 Subject: [PATCH 33/37] test(ci): drop the structure-only assertion on the shard script; the parametrized hang test covers both invocations --- tests/test_litellm/test_unit_shard_per_test_timeout.py | 4 ---- 1 file changed, 4 deletions(-) diff --git a/tests/test_litellm/test_unit_shard_per_test_timeout.py b/tests/test_litellm/test_unit_shard_per_test_timeout.py index e6c3db460e1..8096124ca8c 100644 --- a/tests/test_litellm/test_unit_shard_per_test_timeout.py +++ b/tests/test_litellm/test_unit_shard_per_test_timeout.py @@ -58,10 +58,6 @@ def _hang_guard_args(invocation: tuple[str, ...]) -> tuple[str, ...]: _INVOCATIONS: Final = _pytest_invocations(_run_tests_script()) -def test_the_shard_script_runs_pytest_serially_and_under_xdist() -> None: - assert sorted("-n" in invocation for invocation in _INVOCATIONS) == [False, True] - - @pytest.mark.parametrize( "invocation", _INVOCATIONS, ids=tuple("xdist" if "-n" in invocation else "serial" for invocation in _INVOCATIONS) ) From 9e8a847c5b797457a16e187e538b4a0592946766 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 03:21:28 -0700 Subject: [PATCH 34/37] fix(proxy): keep request metadata out of the cost tracking failure alert The cost tracking callback f-stringed chosen_metadata, litellm_metadata, and old_metadata into the failed_tracking_spend alert on every failure, at every log level, so one 250-byte request produced a 23 KB alert carrying the client's metadata, headers, and key-auth reprs four times over. The alert now carries the exception, the traceback, the model, and the call type; the metadata keys are logged once at debug level through lazy formatting, so nothing is built at warning level --- .../proxy/hooks/proxy_track_cost_callback.py | 33 +++++--- .../hooks/test_proxy_track_cost_callback.py | 78 +++++++++++++++++++ 2 files changed, 102 insertions(+), 9 deletions(-) diff --git a/litellm/proxy/hooks/proxy_track_cost_callback.py b/litellm/proxy/hooks/proxy_track_cost_callback.py index 903255c7b6c..b38fb856215 100644 --- a/litellm/proxy/hooks/proxy_track_cost_callback.py +++ b/litellm/proxy/hooks/proxy_track_cost_callback.py @@ -1,6 +1,6 @@ import asyncio import traceback -from collections.abc import Callable, Sequence +from collections.abc import Callable, Mapping, Sequence from datetime import datetime from typing import TYPE_CHECKING, Any, Final, cast @@ -446,17 +446,26 @@ class _ProxyDBLogger(CustomLogger): f"Cost tracking failed for model={model}.\nDebug info - {cost_tracking_failure_debug_info}\nAdd custom pricing - https://docs.litellm.ai/docs/proxy/custom_pricing" ) except Exception as e: - error_msg = f"Error in tracking cost callback - {e}\n Traceback:{traceback.format_exc()}" - model = kwargs.get("model", "") - metadata = get_litellm_metadata_from_kwargs(kwargs=kwargs) - litellm_metadata: Final = kwargs.get("litellm_params", {}).get("litellm_metadata", {}) - old_metadata: Final = kwargs.get("litellm_params", {}).get("metadata", {}) - call_type = kwargs.get("call_type", "") - error_msg += f"\n Args to _PROXY_track_cost_callback\n model: {model}\n chosen_metadata: {metadata}\n litellm_metadata: {litellm_metadata}\n old_metadata: {old_metadata}\n call_type: {call_type}\n" + failing_model: Final = kwargs.get("model", "") + failing_call_type: Final = kwargs.get("call_type", "") + error_msg: Final = ( + f"Error in tracking cost callback - {e}\n Traceback:{traceback.format_exc()}\n" + f" Args to _PROXY_track_cost_callback\n model: {failing_model}\n call_type: {failing_call_type}\n" + ) + failing_litellm_params: Final = kwargs.get("litellm_params") or {} + verbose_proxy_logger.debug( + "Cost tracking callback failed for model=%s call_type=%s;" + " chosen_metadata keys=%s litellm_metadata keys=%s old_metadata keys=%s", + failing_model, + failing_call_type, + _metadata_keys(get_litellm_metadata_from_kwargs(kwargs=kwargs)), + _metadata_keys(failing_litellm_params.get("litellm_metadata")), + _metadata_keys(failing_litellm_params.get("metadata")), + ) asyncio.create_task( proxy_logging_obj.failed_tracking_alert( error_message=error_msg, - failing_model=model, + failing_model=failing_model, ) ) @@ -614,6 +623,12 @@ def _should_track_cost_callback( return call_type in _UNATTRIBUTED_TRACKABLE_CALL_TYPES +def _metadata_keys(metadata: object) -> tuple[str, ...]: + if not isinstance(metadata, Mapping): + return () + return tuple(sorted(str(key) for key in metadata)) + + def _get_budget_reservation_from_metadata(metadata: dict) -> dict | None: metadata_budget_reservation: Final = metadata.get("user_api_key_budget_reservation") if isinstance(metadata_budget_reservation, dict): diff --git a/tests/test_litellm/proxy/hooks/test_proxy_track_cost_callback.py b/tests/test_litellm/proxy/hooks/test_proxy_track_cost_callback.py index 202495517ad..0e9c336a9eb 100644 --- a/tests/test_litellm/proxy/hooks/test_proxy_track_cost_callback.py +++ b/tests/test_litellm/proxy/hooks/test_proxy_track_cost_callback.py @@ -1,11 +1,13 @@ import asyncio import json +import logging from datetime import datetime from typing import Final from unittest.mock import AsyncMock, MagicMock, patch import pytest +from litellm._logging import verbose_proxy_logger from litellm.litellm_core_utils.internal_call_metadata import MODEL_ACCESS_GROUP_METADATA_KEY from litellm.proxy._types import SpendLogsPayload, UserAPIKeyAuth from litellm.proxy.collector import SpendEventConsumer @@ -2540,3 +2542,79 @@ async def test_async_post_call_failure_hook_persists_no_raw_model_on_an_unknown_ == "/chat/completions: Invalid model name passed in. Call `/v1/models` to view available models for your key." ) assert error_information["error_class"] == "ProxyModelNotFoundError" + + +class _NeverStringifiedMetadataValue: + def __repr__(self) -> str: + raise AssertionError("a request metadata value was stringified by the cost tracking failure path") + + __str__ = __repr__ + + +def _spend_write_kwargs_with_metadata_value(metadata_value: object) -> dict: + return { + "call_type": "acompletion", + "model": "gpt-5.4-mini", + "litellm_call_id": "test-call-id", + "stream": False, + "response_cost": 4.725e-05, + "litellm_params": { + "metadata": { + "user_api_key": "hashed-key", + "user_api_key_user_id": "user-1", + "user_context": metadata_value, + "headers": {"user-agent": metadata_value}, + }, + }, + } + + +@pytest.mark.asyncio +@pytest.mark.parametrize("log_level", [logging.WARNING, logging.DEBUG]) +async def test_track_cost_callback_failure_alert_never_carries_request_metadata_values(log_level): + logger: Final = _ProxyDBLogger() + records: list[logging.LogRecord] = [] + handler: Final = logging.Handler() + handler.emit = records.append + previous_level: Final = verbose_proxy_logger.level + verbose_proxy_logger.setLevel(log_level) + verbose_proxy_logger.addHandler(handler) + try: + with patch( # test-quality-ok: callback imports proxy_logging_obj off proxy_server in its body, no seam + "litellm.proxy.proxy_server.proxy_logging_obj" + ) as mock_proxy_logging: + mock_proxy_logging.failed_tracking_alert = AsyncMock() + mock_proxy_logging.db_spend_update_writer = MagicMock() + mock_proxy_logging.db_spend_update_writer.update_database = AsyncMock( + side_effect=Exception("READONLY You can't write against a read only replica.") + ) + + await logger._PROXY_track_cost_callback( + kwargs=_spend_write_kwargs_with_metadata_value(_NeverStringifiedMetadataValue()), + completion_response=ModelResponse(), + start_time=datetime.now(), + end_time=datetime.now(), + ) + await asyncio.sleep(0) + finally: + verbose_proxy_logger.removeHandler(handler) + verbose_proxy_logger.setLevel(previous_level) + + mock_proxy_logging.failed_tracking_alert.assert_awaited_once() + alert: Final = mock_proxy_logging.failed_tracking_alert.await_args.kwargs + assert alert["failing_model"] == "gpt-5.4-mini" + assert "READONLY You can't write against a read only replica." in alert["error_message"] + assert "model: gpt-5.4-mini" in alert["error_message"] + assert "call_type: acompletion" in alert["error_message"] + + failure_debug_lines: Final = [ + record.getMessage() + for record in records + if record.levelno == logging.DEBUG and "Cost tracking callback failed" in record.getMessage() + ] + if log_level == logging.DEBUG: + assert len(failure_debug_lines) == 1 + assert "user_context" in failure_debug_lines[0] + assert "headers" in failure_debug_lines[0] + else: + assert failure_debug_lines == [] From 7edafd17150a2732b662b69353f90dbdc9419e99 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 03:28:50 -0700 Subject: [PATCH 35/37] fix(masker): memoize shared nodes and fail closed past the depth cap --- .../sensitive_data_masker.py | 52 ++++-- .../test_sensitive_data_masker.py | 156 ++++++++++++++++-- 2 files changed, 178 insertions(+), 30 deletions(-) diff --git a/litellm/litellm_core_utils/sensitive_data_masker.py b/litellm/litellm_core_utils/sensitive_data_masker.py index b4c1beea33e..b68c97e18c9 100644 --- a/litellm/litellm_core_utils/sensitive_data_masker.py +++ b/litellm/litellm_core_utils/sensitive_data_masker.py @@ -1,5 +1,6 @@ from collections.abc import Mapping, Sequence from collections.abc import Set as AbstractSet +from dataclasses import dataclass, field from typing import Any, Final from pydantic import BaseModel @@ -176,26 +177,47 @@ def mask_credentials_in_payload(data: object) -> object: config-dump semantics (``None`` -> ``"None"``, tuples stringified, objects flattened via ``__dict__``) would silently distort the record. + A container referenced from several places in ``data`` is rebuilt once and + referenced from the same places in the copy, so a shared subtree never + fans out into independent copies. A container nested past + ``DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER`` is replaced by + ``REDACTED`` rather than returned unmasked. + Sensitive-key detection is delegated to the shared :class:`SensitiveDataMasker` so pattern updates stay in one place. """ - return _walk_payload(data, key_is_sensitive=False, depth=0) + return _PayloadWalker().walk(data, key_is_sensitive=False, depth=0) -def _walk_payload(node: object, key_is_sensitive: bool, depth: int) -> object: - if depth >= DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER: - return node - if isinstance(node, Mapping): - return {k: _walk_payload(v, _default_masker.is_sensitive_key(k), depth + 1) for k, v in node.items()} - if isinstance(node, list): - return [_walk_payload(item, key_is_sensitive, depth + 1) for item in node] - if isinstance(node, tuple): - return tuple(_walk_payload(item, key_is_sensitive, depth + 1) for item in node) - if isinstance(node, BaseModel): - return _walk_payload(node.model_dump(), key_is_sensitive, depth) - if key_is_sensitive and isinstance(node, str) and node: - return _default_masker._mask_value(node) - return node +@dataclass(frozen=True, slots=True) +class _PayloadWalker: + _memo: dict[tuple[int, bool], tuple[object, object]] = field( # mutable-ok: memo of one walk, pins each keyed node + default_factory=dict + ) + + def walk(self, node: object, key_is_sensitive: bool, depth: int) -> object: + if not isinstance(node, (Mapping, list, tuple, BaseModel)): + return _default_masker._mask_value(node) if key_is_sensitive and isinstance(node, str) and node else node + if depth >= DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER: + return REDACTED + memo_key: Final = (id(node), key_is_sensitive and not isinstance(node, Mapping)) + cached: Final = self._memo.get(memo_key) + if cached is not None: + return cached[1] + rebuilt: Final = self._rebuild(node, key_is_sensitive, depth) + self._memo[memo_key] = (node, rebuilt) + return rebuilt + + def _rebuild( + self, node: Mapping[str, object] | Sequence[object] | BaseModel, key_is_sensitive: bool, depth: int + ) -> object: + if isinstance(node, BaseModel): + return self._rebuild(node.model_dump(), key_is_sensitive, depth) + if isinstance(node, Mapping): + return {k: self.walk(v, _default_masker.is_sensitive_key(k), depth + 1) for k, v in node.items()} + if isinstance(node, tuple): + return tuple(self.walk(item, key_is_sensitive, depth + 1) for item in node) + return [self.walk(item, key_is_sensitive, depth + 1) for item in node] def mask_sensitive_keys(data: dict[str, Any], sensitive_fields: set[str]) -> dict[str, Any]: diff --git a/tests/test_litellm/litellm_core_utils/test_sensitive_data_masker.py b/tests/test_litellm/litellm_core_utils/test_sensitive_data_masker.py index 1551c3fd6e6..de511b0ce11 100644 --- a/tests/test_litellm/litellm_core_utils/test_sensitive_data_masker.py +++ b/tests/test_litellm/litellm_core_utils/test_sensitive_data_masker.py @@ -2,11 +2,11 @@ Unit tests for SensitiveDataMasker - List Preservation """ +from functools import reduce +from typing import Final import pytest -# Add the parent directory to the system path - from litellm.litellm_core_utils.sensitive_data_masker import SensitiveDataMasker @@ -152,9 +152,7 @@ def test_mask_short_values_false_keeps_short_values_readable(): chars of an exception and only masks longer tails), while longer values are still partially masked. """ - masker = SensitiveDataMasker( - visible_prefix=50, visible_suffix=0, mask_short_values=False - ) + masker = SensitiveDataMasker(visible_prefix=50, visible_suffix=0, mask_short_values=False) short = "Test exception for structure validation" assert masker._mask_value(short) == short @@ -202,9 +200,7 @@ def test_mask_sensitive_structure_passes_through_plain_topology_names(): from litellm.litellm_core_utils.sensitive_data_masker import mask_sensitive_structure assert mask_sensitive_structure(["gpt-4", "claude-3-haiku"]) == ["gpt-4", "claude-3-haiku"] - assert mask_sensitive_structure([{"gpt-3.5-turbo": ["claude-3-haiku"]}]) == [ - {"gpt-3.5-turbo": ["claude-3-haiku"]} - ] + assert mask_sensitive_structure([{"gpt-3.5-turbo": ["claude-3-haiku"]}]) == [{"gpt-3.5-turbo": ["claude-3-haiku"]}] assert mask_sensitive_structure(None) is None @@ -233,9 +229,7 @@ def test_mask_sensitive_structure_masks_credentials_nested_in_config_shape(): from litellm.litellm_core_utils.sensitive_data_masker import mask_sensitive_structure secret = "sk-NESTEDINLINESECRET0987654321" - masked = mask_sensitive_structure( - [{"primary-group": [{"model": "gpt-4o", "api_key": secret}]}] - ) + masked = mask_sensitive_structure([{"primary-group": [{"model": "gpt-4o", "api_key": secret}]}]) assert secret not in str(masked) @@ -282,10 +276,7 @@ def test_mask_credentials_in_payload_masks_inside_pydantic_models(): auth_dict = result["user_api_key_auth"] assert isinstance(auth_dict, dict) assert auth_dict["team_alias"] == "acme" - assert ( - auth_dict["token"] - != "1b01552f6e52e0d41963dd6a185bd6b074624e330999534ca7ff5adfdf622dfc" - ) + assert auth_dict["token"] != "1b01552f6e52e0d41963dd6a185bd6b074624e330999534ca7ff5adfdf622dfc" assert "*" in auth_dict["token"] @@ -314,6 +305,140 @@ def test_mask_credentials_in_payload_masks_only_sensitive_string_leaves(): assert masked.endswith(plaintext[-4:]) +def _unique_dict_ids(node: object) -> frozenset[int]: + if isinstance(node, dict): + return frozenset((id(node),)).union(*(_unique_dict_ids(value) for value in node.values())) + if isinstance(node, list): + return frozenset().union(*(_unique_dict_ids(value) for value in node)) + return frozenset() + + +def _nested_under_levels(leaf: object, levels: int) -> object: + return reduce(lambda inner, level: {f"l{level}": inner}, range(levels, 0, -1), leaf) + + +def test_mask_credentials_in_payload_keeps_a_shared_dict_shared(): + """One dict referenced twice comes back as one masked dict referenced + twice. Rebuilding each reference separately is what turned an aliased + retry breadcrumb graph exponential in the v1.100.0 OOM.""" + from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload + + shared: Final = {"api_key": "sk-shared-1234567890abcdef", "model": "gpt-4o-mini"} + result: Final = mask_credentials_in_payload({"first": shared, "second": shared}) + + assert result["first"] is result["second"] + assert result["first"]["model"] == "gpt-4o-mini" + assert result["first"]["api_key"] != "sk-shared-1234567890abcdef" + + +def test_mask_credentials_in_payload_walks_each_dag_node_once(): + """A DAG of 9 dicts where every level references the level below three + times stays 9 dicts after masking, instead of fanning out to 3**8.""" + from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload + + root: Final = reduce( + lambda inner, _: {"a": inner, "b": inner, "c": inner}, range(8), {"api_key": "sk-leaf-1234567890abcdef"} + ) + + result: Final = mask_credentials_in_payload(root) + + assert len(_unique_dict_ids(root)) == 9 + assert len(_unique_dict_ids(result)) == 9 + assert "sk-leaf-1234567890abcdef" not in str(result) + + +def test_mask_credentials_in_payload_masks_a_shared_list_only_under_a_sensitive_key(): + """The same list reached under a plain key and under a sensitive key is + masked in the sensitive spot only, whichever reference the walk meets + first, so the memo can neither leak a secret nor mask a plain value.""" + from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload + + shared: Final = ["sk-list-1234567890abcdef"] + + plain_first: Final = mask_credentials_in_payload({"tags": shared, "api_key": shared}) + assert plain_first["tags"] == ["sk-list-1234567890abcdef"] + assert plain_first["api_key"] != ["sk-list-1234567890abcdef"] + + sensitive_first: Final = mask_credentials_in_payload({"api_key": shared, "tags": shared}) + assert sensitive_first["api_key"] != ["sk-list-1234567890abcdef"] + assert sensitive_first["tags"] == ["sk-list-1234567890abcdef"] + + +def test_mask_credentials_in_payload_masks_a_shared_root_model_list_only_under_a_sensitive_key(): + """A pydantic model that dumps to a list is a list once walked, so the + memo must keep its plain and sensitive rebuilds apart the same way, or + the reference met first decides what the other one shows.""" + from pydantic import RootModel + + from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload + + shared: Final = RootModel[list[str]](["sk-root-1234567890abcdef"]) + + plain_first: Final = mask_credentials_in_payload({"tags": shared, "api_key": shared}) + assert plain_first["tags"] == ["sk-root-1234567890abcdef"] + assert plain_first["api_key"] != ["sk-root-1234567890abcdef"] + + sensitive_first: Final = mask_credentials_in_payload({"api_key": shared, "tags": shared}) + assert sensitive_first["api_key"] != ["sk-root-1234567890abcdef"] + assert sensitive_first["tags"] == ["sk-root-1234567890abcdef"] + + +def test_mask_credentials_in_payload_hides_containers_past_the_depth_cap(): + """A dict nested past DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER is + replaced by the REDACTED marker instead of coming back unmasked, while the + strings sitting exactly at the cap still get the normal per-key treatment: + a sensitive one is masked and a plain one survives verbatim.""" + from litellm.constants import DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER + from litellm.litellm_core_utils.secret_redaction import REDACTED + from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload + + secret: Final = "sk-deep-1234567890abcdef" + cap: Final = DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER + + result: Final = mask_credentials_in_payload(_nested_under_levels({"api_key": secret}, cap)) + + assert secret not in str(result) + at_cap: Final = reduce(lambda node, level: node[f"l{level}"], range(1, cap), result) + assert at_cap == {f"l{cap}": REDACTED} + + strings_at_cap: Final = reduce( + lambda node, level: node[f"l{level}"], + range(1, cap), + mask_credentials_in_payload(_nested_under_levels({"api_key": secret, "model": "gpt-5.4-mini"}, cap - 1)), + ) + assert strings_at_cap["model"] == "gpt-5.4-mini" + assert strings_at_cap["api_key"] != secret + assert strings_at_cap["api_key"].startswith("sk-d") + + +def test_mask_credentials_in_payload_keeps_sibling_models_apart(): + """Two models of the same shape dump into temporaries whose ids CPython + reuses as soon as the first is freed, so an id-keyed memo that does not + pin what it keys hands the second model the first one's masked copy.""" + from pydantic import BaseModel + + from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload + + class Inner(BaseModel): + label: str + api_key: str + + class Outer(BaseModel): + inner: Inner + + result: Final = mask_credentials_in_payload( + { + "first": Outer(inner=Inner(label="one", api_key="sk-first-1234567890abcdef")), + "second": Outer(inner=Inner(label="two", api_key="sk-second-1234567890abcdef")), + } + ) + + assert result["first"]["inner"]["label"] == "one" + assert result["second"]["inner"]["label"] == "two" + assert "sk-second-1234567890abcdef" not in str(result) + assert result["second"]["inner"]["api_key"].startswith("sk-s") + + def test_extra_sensitive_patterns_add_to_the_defaults(): from litellm.litellm_core_utils.sensitive_data_masker import SensitiveDataMasker @@ -344,6 +469,7 @@ def test_the_second_positional_argument_is_still_the_override_set(): assert masker.is_sensitive_key("session_token") is False assert masker.is_sensitive_key("auth_token") is True + def test_redact_credentials_in_payload_leaves_no_fragment_of_the_secret(): """A payload rendered straight to stdout cannot afford the partial reveal mask_credentials_in_payload leaves, so every credential-named value is replaced From 093fb78bafb3c2ef8273e06370f2b072c50341d2 Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 03:58:47 -0700 Subject: [PATCH 36/37] fix(masker): cut cycles at the first back-edge and walk pydantic dumps without self-recursion --- .../sensitive_data_masker.py | 6 +- .../test_sensitive_data_masker.py | 55 +++++++++++++------ 2 files changed, 41 insertions(+), 20 deletions(-) diff --git a/litellm/litellm_core_utils/sensitive_data_masker.py b/litellm/litellm_core_utils/sensitive_data_masker.py index b68c97e18c9..b7bd0a1498b 100644 --- a/litellm/litellm_core_utils/sensitive_data_masker.py +++ b/litellm/litellm_core_utils/sensitive_data_masker.py @@ -179,7 +179,8 @@ def mask_credentials_in_payload(data: object) -> object: A container referenced from several places in ``data`` is rebuilt once and referenced from the same places in the copy, so a shared subtree never - fans out into independent copies. A container nested past + fans out into independent copies, and a reference back into a container + still being rebuilt (a cycle) becomes ``REDACTED``. A container nested past ``DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER`` is replaced by ``REDACTED`` rather than returned unmasked. @@ -204,6 +205,7 @@ class _PayloadWalker: cached: Final = self._memo.get(memo_key) if cached is not None: return cached[1] + self._memo[memo_key] = (node, REDACTED) rebuilt: Final = self._rebuild(node, key_is_sensitive, depth) self._memo[memo_key] = (node, rebuilt) return rebuilt @@ -212,7 +214,7 @@ class _PayloadWalker: self, node: Mapping[str, object] | Sequence[object] | BaseModel, key_is_sensitive: bool, depth: int ) -> object: if isinstance(node, BaseModel): - return self._rebuild(node.model_dump(), key_is_sensitive, depth) + return self.walk(node.model_dump(), key_is_sensitive, depth) if isinstance(node, Mapping): return {k: self.walk(v, _default_masker.is_sensitive_key(k), depth + 1) for k, v in node.items()} if isinstance(node, tuple): diff --git a/tests/test_litellm/litellm_core_utils/test_sensitive_data_masker.py b/tests/test_litellm/litellm_core_utils/test_sensitive_data_masker.py index de511b0ce11..fcdf7fb4798 100644 --- a/tests/test_litellm/litellm_core_utils/test_sensitive_data_masker.py +++ b/tests/test_litellm/litellm_core_utils/test_sensitive_data_masker.py @@ -318,9 +318,6 @@ def _nested_under_levels(leaf: object, levels: int) -> object: def test_mask_credentials_in_payload_keeps_a_shared_dict_shared(): - """One dict referenced twice comes back as one masked dict referenced - twice. Rebuilding each reference separately is what turned an aliased - retry breadcrumb graph exponential in the v1.100.0 OOM.""" from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload shared: Final = {"api_key": "sk-shared-1234567890abcdef", "model": "gpt-4o-mini"} @@ -332,8 +329,6 @@ def test_mask_credentials_in_payload_keeps_a_shared_dict_shared(): def test_mask_credentials_in_payload_walks_each_dag_node_once(): - """A DAG of 9 dicts where every level references the level below three - times stays 9 dicts after masking, instead of fanning out to 3**8.""" from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload root: Final = reduce( @@ -347,10 +342,21 @@ def test_mask_credentials_in_payload_walks_each_dag_node_once(): assert "sk-leaf-1234567890abcdef" not in str(result) +def test_mask_credentials_in_payload_cuts_a_cycle_at_its_first_back_edge(): + from litellm.litellm_core_utils.secret_redaction import REDACTED + from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload + + node: Final[dict[str, object]] = {"api_key": "sk-cycle-1234567890abcdef"} + node["kids"] = [node] * 3 + + result: Final = mask_credentials_in_payload(node) + + assert result["kids"] == [REDACTED, REDACTED, REDACTED] + assert result["api_key"] != "sk-cycle-1234567890abcdef" + assert len(_unique_dict_ids(result)) == 1 + + def test_mask_credentials_in_payload_masks_a_shared_list_only_under_a_sensitive_key(): - """The same list reached under a plain key and under a sensitive key is - masked in the sensitive spot only, whichever reference the walk meets - first, so the memo can neither leak a secret nor mask a plain value.""" from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload shared: Final = ["sk-list-1234567890abcdef"] @@ -365,9 +371,6 @@ def test_mask_credentials_in_payload_masks_a_shared_list_only_under_a_sensitive_ def test_mask_credentials_in_payload_masks_a_shared_root_model_list_only_under_a_sensitive_key(): - """A pydantic model that dumps to a list is a list once walked, so the - memo must keep its plain and sensitive rebuilds apart the same way, or - the reference met first decides what the other one shows.""" from pydantic import RootModel from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload @@ -383,11 +386,21 @@ def test_mask_credentials_in_payload_masks_a_shared_root_model_list_only_under_a assert sensitive_first["tags"] == ["sk-root-1234567890abcdef"] +def test_mask_credentials_in_payload_masks_a_root_model_string_as_one_string(): + from pydantic import RootModel + + from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload + + result: Final = mask_credentials_in_payload( + {"api_key": RootModel[str]("sk-root-1234567890abcdef"), "model": RootModel[str]("gpt-5.4-mini")} + ) + + assert result["model"] == "gpt-5.4-mini" + assert result["api_key"] != "sk-root-1234567890abcdef" + assert result["api_key"].startswith("sk-r") + + def test_mask_credentials_in_payload_hides_containers_past_the_depth_cap(): - """A dict nested past DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER is - replaced by the REDACTED marker instead of coming back unmasked, while the - strings sitting exactly at the cap still get the normal per-key treatment: - a sensitive one is masked and a plain one survives verbatim.""" from litellm.constants import DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER from litellm.litellm_core_utils.secret_redaction import REDACTED from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload @@ -401,6 +414,14 @@ def test_mask_credentials_in_payload_hides_containers_past_the_depth_cap(): at_cap: Final = reduce(lambda node, level: node[f"l{level}"], range(1, cap), result) assert at_cap == {f"l{cap}": REDACTED} + +def test_mask_credentials_in_payload_treats_strings_at_the_depth_cap_per_key(): + from litellm.constants import DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER + from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload + + secret: Final = "sk-deep-1234567890abcdef" + cap: Final = DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER + strings_at_cap: Final = reduce( lambda node, level: node[f"l{level}"], range(1, cap), @@ -412,9 +433,7 @@ def test_mask_credentials_in_payload_hides_containers_past_the_depth_cap(): def test_mask_credentials_in_payload_keeps_sibling_models_apart(): - """Two models of the same shape dump into temporaries whose ids CPython - reuses as soon as the first is freed, so an id-keyed memo that does not - pin what it keys hands the second model the first one's masked copy.""" + """CPython reuses a freed temporary's id, so an id-keyed memo has to pin what it keys.""" from pydantic import BaseModel from litellm.litellm_core_utils.sensitive_data_masker import mask_credentials_in_payload From aceae8e566913faf931a1c13dbbd32698695a20b Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Sat, 19 Sep 2026 04:07:07 -0700 Subject: [PATCH 37/37] test: drop the recursive detector allowlist entry for the removed _walk_payload --- tests/code_coverage_tests/recursive_detector.py | 1 - 1 file changed, 1 deletion(-) diff --git a/tests/code_coverage_tests/recursive_detector.py b/tests/code_coverage_tests/recursive_detector.py index d8e318c61af..3c6a6a58820 100644 --- a/tests/code_coverage_tests/recursive_detector.py +++ b/tests/code_coverage_tests/recursive_detector.py @@ -36,7 +36,6 @@ IGNORE_FUNCTIONS = [ "_collect_argument_paths", # max depth set. "_split_text", # max depth set. "_mask_sequence", # max depth set. - "_walk_payload", # max depth set (DEFAULT_MAX_RECURSE_DEPTH_SENSITIVE_DATA_MASKER). "_delete_nested_value_custom", # max depth set (bounded by number of path segments). "filter_exceptions_from_params", # max depth set (default 20) to prevent infinite recursion. "__getattr__", # lazy loading pattern in litellm/__init__.py with proper caching to prevent infinite recursion.