Merge branch 'main' into litellm_fix_concurrent_single_agent_publish

.agents/skills/database-access/SKILL.md

@@ -1,6 +1,8 @@
 ---
 name: database-access
 description: Add database queries, extract existing Prisma calls into repositories, or explicitly port queries to Rust and SQLx without expanding migration scope
+metadata:
+  internal: true
 ---
 
 # Database access

.github/workflows/spend-logs-guard.yml

@@ -4,10 +4,6 @@ on: # zizmor: ignore[dangerous-triggers] runs the base branch's code only; the P
   pull_request_target:
     branches:
       - main
-  pull_request_review:
-    types:
-      - submitted
-      - dismissed
 
 permissions:
   contents: read

.github/workflows/test-linting.yml

@@ -561,6 +561,9 @@ jobs:
       - name: router_code_coverage
         run: uv run --no-sync python ./tests/code_coverage_tests/router_code_coverage.py
 
+      - name: router_surface_check
+        run: uv run --no-sync python ./tests/code_coverage_tests/router_surface_check.py
+
       - name: test_chat_completion_imports
         run: uv run --no-sync python ./tests/code_coverage_tests/test_chat_completion_imports.py
 

.github/workflows/test-unit.yml

@@ -245,6 +245,7 @@ jobs:
               tests/unit/llms/clarifai
               tests/unit/llms/claude_code
               tests/unit/llms/cloudflare
+              tests/unit/llms/codestral
               tests/unit/llms/codex
               tests/unit/llms/cohere
               tests/unit/llms/cometapi

ci_cd/spend_logs_guard.py

@@ -2,11 +2,12 @@
 needs an approving review from the spend-log owner before it can merge.
 
 CODEOWNERS can only match file paths, and every schema change edits the same schema.prisma copies,
-so a path rule would put every migration on the owner. This gate reads the diff instead: when no
-added line in the watched files mentions the table it passes at once, otherwise it passes only once
-the owner's latest review on the pull request is an approval that still covers every spend-log line
-in the head (approvals survive pushes, so a later commit may not add spend-log lines the owner never
-saw). The workflow re-runs on every review.
+so a path rule would put every migration on the owner. This gate reads the diff instead: a schema.prisma
+hunk counts when it adds or removes a line inside the `model LiteLLM_SpendLogs { ... }` block, a migration
+hunk counts when an added line names the table. Without such a hunk it passes at once, otherwise it passes
+only once the owner's latest review on the pull request is an approval that still covers every spend-log line
+in the head (approvals survive pushes, so a later commit may not add spend-log lines the owner never saw).
+Reviews do not trigger the workflow, so the check is re-run by hand after approving.
 """
 
 from __future__ import annotations
@@ -15,8 +16,10 @@ import argparse
 import re
 import subprocess
 import sys
-from collections.abc import Callable, Sequence
+from collections.abc import Callable, Mapping, Sequence
+from dataclasses import dataclass
 from pathlib import Path
+from types import MappingProxyType
 from typing import Final
 
 from pydantic import BaseModel, ConfigDict, TypeAdapter
@@ -27,7 +30,10 @@ WATCHED: Final = (
     "litellm-proxy-extras/litellm_proxy_extras/schema.prisma",
 )
 MIGRATIONS_PREFIX: Final = "litellm-proxy-extras/litellm_proxy_extras/migrations/"
-TABLE: Final = re.compile(r"(?<![A-Za-z0-9])LiteLLM_SpendLogs", re.IGNORECASE)
+TABLE: Final = re.compile(r"(?<![A-Za-z0-9])LiteLLM_SpendLogs(?![A-Za-z0-9])", re.IGNORECASE)
+DECIDING_STATES: Final = frozenset({"APPROVED", "CHANGES_REQUESTED", "DISMISSED"})
+MODEL_START: Final = re.compile(r"^model\s+LiteLLM_SpendLogs\s*\{")
+HUNK: Final = re.compile(r"^@@ -(\d+)(?:,\d+)? \+(\d+)(?:,\d+)? @@[^\n]*\n", re.MULTILINE)
 
 
 class Reviewer(BaseModel):
@@ -45,26 +51,71 @@ class Review(BaseModel):
 _REVIEWS: Final = TypeAdapter(list[Review])
 
 
+@dataclass(frozen=True, slots=True)
+class Block:
+    """Lines of `model LiteLLM_SpendLogs { ... }` in one schema.prisma copy at the merge base and at the head."""
+
+    base: range
+    head: range
+
+
+@dataclass(frozen=True, slots=True)
+class Hunk:
+    removed: tuple[tuple[int, str], ...]
+    added: tuple[tuple[int, str], ...]
+
+
 def watched(path: str) -> bool:
     return path in WATCHED or (path.startswith(MIGRATIONS_PREFIX) and path.endswith("/migration.sql"))
 
 
-def _added_lines(section: str) -> tuple[str, ...]:
-    body: Final = section.split("\n@@", 1)
-    return tuple(line[1:] for line in body[1].splitlines() if line.startswith("+")) if len(body) == 2 else ()
+def model_block(schema: str) -> range:
+    """1-based line range of `model LiteLLM_SpendLogs { ... }` including both braces, empty when the schema has none."""
+    lines: Final = schema.splitlines()
+    starts: Final = tuple(number for number, line in enumerate(lines, 1) if MODEL_START.match(line))
+    if not starts:
+        return range(0)
+    ends: Final = tuple(number for number, line in enumerate(lines, 1) if number > starts[0] and line.startswith("}"))
+    return range(starts[0], (ends[0] if ends else len(lines)) + 1)
 
 
-def spend_log_hits(diff: str) -> tuple[str, ...]:
-    """`path: added line` for every added line in a watched file that names the table."""
+def _hunks(section: str) -> tuple[Hunk, ...]:
+    headers: Final = tuple(HUNK.finditer(section))
+    ends: Final = tuple(header.start() for header in headers[1:]) + (len(section),)
+    return tuple(_hunk(header, section[header.end() : end]) for header, end in zip(headers, ends))
+
+
+def _hunk(header: re.Match[str], body: str) -> Hunk:
+    lines: Final = body.splitlines()
+    return Hunk(
+        removed=tuple(enumerate((line[1:] for line in lines if line.startswith("-")), int(header.group(1)))),
+        added=tuple(enumerate((line[1:] for line in lines if line.startswith("+")), int(header.group(2)))),
+    )
+
+
+def spend_log_hits(diff: str, blocks: Mapping[str, Block]) -> tuple[str, ...]:
+    """`path: +line` or `path: -line` for every changed line that touches the table.
+
+    A schema.prisma line counts when its number falls inside that copy's `blocks` entry (head for added lines,
+    merge base for removed ones), a migration line when it is added and names the table."""
     sections: Final = re.split(r"^diff --git [^\n]*\n", diff, flags=re.MULTILINE)
-    return tuple(f"{path}: {line.strip()}" for path, line in _watched_added_lines(sections) if TABLE.search(line))
-
-
-def _watched_added_lines(sections: Sequence[str]) -> tuple[tuple[str, str], ...]:
-    pairs: Final = tuple((_path_of(section), section) for section in sections)
+    paths: Final = tuple((_path_of(section), section) for section in sections)
     return tuple(
-        (path, line) for path, section in pairs if watched(path) for line in _added_lines(section)
-    )  # comprehension-ok: path then its lines
+        hit for path, section in paths if watched(path) for hit in _section_hits(path, _hunks(section), blocks)
+    )  # comprehension-ok: path then its hits
+
+
+def _section_hits(path: str, hunks: Sequence[Hunk], blocks: Mapping[str, Block]) -> tuple[str, ...]:
+    if path not in blocks:
+        return tuple(f"{path}: +{line.strip()}" for hunk in hunks for _, line in hunk.added if TABLE.search(line))
+    block: Final = blocks[path]
+    removed: Final = tuple(
+        f"{path}: -{line.strip()}" for hunk in hunks for number, line in hunk.removed if number in block.base
+    )  # comprehension-ok: hunk then its lines
+    added: Final = tuple(
+        f"{path}: +{line.strip()}" for hunk in hunks for number, line in hunk.added if number in block.head
+    )  # comprehension-ok: hunk then its lines
+    return removed + added
 
 
 def _path_of(section: str) -> str:
@@ -73,8 +124,12 @@ def _path_of(section: str) -> str:
 
 
 def approved_by(reviews: Sequence[Review], login: str) -> bool:
-    """True when `login`'s most recent review on the pull request is an approval."""
-    states: Final = tuple(review.state for review in reviews if review.user.login == login)
+    """True when `login`'s most recent review that changes GitHub's approval state is an approval.
+
+    COMMENTED and PENDING reviews leave an earlier approval in force on GitHub, so they are skipped here too.
+    """
+    own: Final = tuple(review.state for review in reviews if review.user.login == login)
+    states: Final = tuple(state for state in own if state in DECIDING_STATES)
     return bool(states) and states[-1] == "APPROVED"
 
 
@@ -106,13 +161,25 @@ def added_diff(base: str, head: str) -> str:
     return result.stdout
 
 
+def file_at(revision: str, path: str) -> str:
+    """Contents of `path` at `revision`, empty when the file does not exist there."""
+    result: Final = subprocess.run(("git", "show", f"{revision}:{path}"), check=False, capture_output=True, text=True)
+    return result.stdout if result.returncode == 0 else ""
+
+
+def schema_blocks(base: str, head: str) -> Mapping[str, Block]:
+    return MappingProxyType(
+        {path: Block(base=model_block(file_at(base, path)), head=model_block(file_at(head, path))) for path in WATCHED}
+    )
+
+
 def fetched(commit: str) -> bool:
     result: Final = subprocess.run(("git", "fetch", "--no-tags", "--depth=1", "origin", commit), check=False)
     return result.returncode == 0
 
 
 def reviewed_hits(base: str, commit: str) -> tuple[str, ...] | None:
-    return spend_log_hits(added_diff(base, commit)) if fetched(commit) else None
+    return spend_log_hits(added_diff(base, commit), schema_blocks(base, commit)) if fetched(commit) else None
 
 
 class Arguments(BaseModel):
@@ -130,9 +197,12 @@ def main(argv: Sequence[str]) -> int:
     parser.add_argument("--approver", required=True, help="GitHub login whose approval unlocks the merge")
     parser.add_argument("--reviews", required=True, help="JSON file with the pull request's reviews")
     args: Final = Arguments.model_validate(vars(parser.parse_args(argv)))
-    hits: Final = spend_log_hits(added_diff(args.base, args.head))
+    hits: Final = spend_log_hits(added_diff(args.base, args.head), schema_blocks(args.base, args.head))
     if not hits:
-        sys.stdout.write("No added line in schema.prisma or a migration names LiteLLM_SpendLogs; nothing to review.\n")
+        sys.stdout.write(
+            "No change inside model LiteLLM_SpendLogs in schema.prisma and no added migration line names it; "
+            "nothing to review.\n"
+        )
         return 0
     reviews: Final = _REVIEWS.validate_json(args.reviews.read_bytes())
     approval: Final = latest_approval(reviews, args.approver) if approved_by(reviews, args.approver) else None
@@ -148,7 +218,7 @@ def main(argv: Sequence[str]) -> int:
         else f"LiteLLM_SpendLogs lines changed after @{args.approver} approved {approval.commit_id}, "
         f"a new approving review is needed:\n"
     )
-    sys.stdout.write(f"{reason}{listed}The check re-runs when a review is submitted.\n")
+    sys.stdout.write(f"{reason}{listed}After approving, re-run this job.\n")
     return 1
 
 

litellm-proxy-extras/litellm_proxy_extras/migration_data_checks.py

@@ -343,11 +343,16 @@ def btree_max_entry_bytes(block_size: int) -> int:
 
 
 def _value_bytes(column: ColumnType, identifier: "sql.Identifier") -> "sql.Composed":
+    """The bytes a value takes in the index entry: its length, or its stored size when
+    Postgres compressed it, since `index_form_tuple` keeps the compressed bytes."""
     from psycopg import sql
 
-    if column.type_name in _OCTET_LENGTH_TYPES:
-        return sql.SQL("octet_length({})").format(identifier)
-    return sql.SQL("octet_length({}::text)").format(identifier)
+    length: Final = (
+        sql.SQL("octet_length({})").format(identifier)
+        if column.type_name in _OCTET_LENGTH_TYPES
+        else sql.SQL("octet_length({}::text)").format(identifier)
+    )
+    return sql.SQL("least({}, pg_column_size({}))").format(length, identifier)
 
 
 def _aligned(offset: "sql.Composable", align: str) -> "sql.Composed":

litellm-rust/.agents/skills/rust-error/SKILL.md

@@ -1,6 +1,8 @@
 ---
 name: rust-error
 description: Define, wrap, convert, classify, report or test Rust error types in litellm-rust with thiserror. Use when adding an error type or variant, mapping one crate's error into another, turning an error into a gateway response or Python exception, deciding retry or fault from an error, or replacing call sites that build errors from strings
+metadata:
+  internal: true
 ---
 
 # Rust errors

litellm-rust/.agents/skills/rust-gateway/SKILL.md

@@ -1,6 +1,8 @@
 ---
 name: rust-gateway
 description: Add or change an HTTP route, extractor, response envelope, stream or router mount in the litellm-rust gateway crates (crates/gateway and crates/gateway-*). Use when adding an endpoint, parsing a request body, choosing an error envelope for an API surface, relaying SSE, or writing a router-level test
+metadata:
+  internal: true
 ---
 
 # Rust gateway

litellm-rust/.agents/skills/rust-string-enums/SKILL.md

@@ -1,6 +1,8 @@
 ---
 name: rust-string-enums
 description: Define or refactor Rust string-valued enums and their Serde adapters in litellm-rust, using Strum and serde_with while preserving parsing, wire values, and schemas
+metadata:
+  internal: true
 ---
 
 # Rust string enums

litellm-rust/.agents/skills/rust-tracing/SKILL.md

@@ -1,6 +1,8 @@
 ---
 name: rust-tracing
 description: Add or change Rust diagnostic tracing in litellm-rust, including route spans, subscriber layers, and Python logger delivery
+metadata:
+  internal: true
 ---
 
 # Rust tracing

litellm-rust/Cargo.lock

@@ -4393,6 +4393,10 @@ dependencies = [
  "rstest",
 ]
 
+[[package]]
+name = "litellm-router-core"
+version = "0.1.0"
+
 [[package]]
 name = "litellm-router-types"
 version = "0.1.0"

litellm-rust/crates/model-catalog/src/lib.rs

@@ -1,3 +1,36 @@
+//! The model catalog: per-model pricing, context limits and capabilities.
+//!
+//! This is the Rust form of `model_prices_and_context_window.json`, the map Python reads as
+//! `litellm.model_cost` and through `litellm.get_model_info`. It answers "what does this model
+//! cost and what can it do"; it knows nothing about configured deployments. The router's view of
+//! deployments (groups, aliases, wildcard patterns, team mappings) is the router registry, defined
+//! in `litellm-router-core`, which reads this catalog through its dynamic config.
+//!
+//! # Structure
+//! - `Catalog`: the parsed map. `Catalog::parse` takes the JSON body and a `Provenance` (where
+//!   the body came from: the bundled copy in the repo or the remote copy, with revision and etag),
+//!   and keeps every entry in file order with its raw fields.
+//! - `ModelEntry`: one model. `ModelEntry::fields` is the raw JSON object, so unknown fields are
+//!   never lost; `ModelEntry::info` is the typed mirror `ModelInfo` (pricing in `pricing`,
+//!   modalities and modes in `capabilities`).
+//! - Aliases: an entry may list `aliases`; each alias resolves to its canonical key. Problems
+//!   found while indexing them (bad lists, collisions) are kept as `AliasIssue`s instead of
+//!   failing the parse.
+//! - `FallbackGeneralizations` / `FallbackRule`: regex rules that map unknown model ids onto known
+//!   families. They are data in the file, not entries.
+//! - Validation: `Catalog::validate` checks a catalog against `IntegrityLimits` (a remote copy that
+//!   lost most of its models is refused), and `validate_registry` checks every entry against the
+//!   checked-in contract.
+//!
+//! # Lookups
+//! `Catalog::lookup` resolves a key exactly first (entry or alias), then case-insensitively, and
+//! returns a `ModelMatch` with both the key that matched and the canonical key of the entry. Provider
+//! prefixes are not stripped here; callers decide which keys to try.
+//!
+//! # Readers
+//! Cost tracking, the router (through `RouterDynamicConfig`) and the gateway. Nothing in this crate
+//! performs I/O; loading and refreshing the file is the caller's job.
+
 mod capabilities;
 mod catalog;
 mod error;

litellm-rust/crates/router-core/Cargo.toml

@@ -0,0 +1,9 @@
+[package]
+name = "litellm-router-core"
+version = "0.1.0"
+edition.workspace = true
+license.workspace = true
+repository.workspace = true
+description = "The router core behind litellm.Router: registry, settings, health, selection and the call loop"
+
+[dependencies]

litellm-rust/crates/router-core/src/lib.rs

@@ -0,0 +1,63 @@
+//! Routing for LiteLLM: decide which deployment serves a request, and drive retries, fallbacks
+//! and cooldowns around the attempts.
+//!
+//! This crate defines the boundary `litellm.Router` is moving behind. Today the Python router in
+//! `litellm/router.py` implements all of it; this crate is where the Rust core will implement the
+//! same interface. Until then it holds this overview and nothing else, so the interface is written
+//! down in code before either side changes.
+//!
+//! # The public surface
+//! Code outside the router uses `litellm.Router` through two kinds of members and nothing else:
+//! - The endpoint methods (`acompletion`, `aembedding`, `aresponses`, `aanthropic_messages`, the
+//!   files, batches, assistants and vector store wrappers, 37 in all). Each is a thin wrapper over
+//!   `call(op, request)`.
+//! - Eight entry points: `registry()`, `settings()`, `health()`, `select()`, `apply()`,
+//!   `observe()`, `register_plugin()` and `retire()`.
+//!
+//! Every other member that outside code uses today has a named replacement among those. The old
+//! names stay available for about a month through a `DeprecatedRouterMembers` mixin, which
+//! implements them over the entry points and warns on use. A CI check
+//! (`tests/code_coverage_tests/router_surface_check.py`) keeps an allowlist of the members outside
+//! code touches; the list may only shrink.
+//!
+//! # Structure
+//! - `Registry`: the configured deployments, indexed by id, model group, alias, wildcard pattern
+//!   and team. Immutable and versioned. Seven queries: `deployment(id)`, `deployments(filter)`,
+//!   `names(filter)`, `resolve(name, scope)`, `group(name)`, `access_groups()` and
+//!   `credentials(id)`. "Registry" is the router's view of what is configured; per-model pricing
+//!   and capabilities are the model catalog (`litellm-model-catalog`), which the router reads
+//!   through its dynamic config.
+//! - `RouterConfig` / `RouterDynamicConfig`: everything the router needs to know, passed in. The
+//!   router never reads process globals. Static config (model list, strategy, retry, timeout and
+//!   cooldown policy, Redis settings, plugins) is fixed for the router's life. Dynamic config
+//!   (pricing, model aliases, fallbacks, retry counts, callbacks) is a versioned snapshot read
+//!   once per call.
+//! - `Settings`: one typed, frozen view of the configuration a caller may read.
+//! - `Health`: cooldowns, blocked deployments and usage, kept in shared cache keys so Python and
+//!   Rust workers see the same state. `report(id, outcome)` records outcomes the router did not
+//!   see itself (a health check, an interceptor that marks a deployment failed).
+//! - `select(model, request)`: resolve a name, filter eligible deployments, pick one with the
+//!   configured strategy, without calling it.
+//! - `apply(changes)`: one `ChangeSet` (upserts, deletes, settings), applied atomically, producing
+//!   a new registry version.
+//! - `call(op, request)`: select, run the attempt, and retry or fall back until it succeeds or the
+//!   policy is exhausted.
+//!
+//! # Execution
+//! Each attempt resolves an `Executor` after a deployment is picked. A Rust-native operation runs
+//! its `inference-*` route machine in-process; anything else goes through `Host::invoke`. A
+//! started attempt is never replayed on a different executor.
+//!
+//! # Hosts
+//! `Host` is what the embedding process provides: `invoke` for non-native operations, `select`
+//! for strategies and sub-routers not yet ported, `allow_fallback` for access and budget checks,
+//! `sleep`, `observe` for routing events, and `cache`, a DualCache-style store that holds all
+//! routing state under the same keys, encodings and TTLs the Python router uses. `PythonHost`
+//! serves them from the existing Python code when the router runs inside Python; `NativeHost`
+//! serves them in the Rust server, where an operation without a Rust implementation makes the
+//! config fail at build time, never partway through a request.
+//!
+//! # Plugins
+//! Routing strategies, pre-call checks and sub-routers (auto, complexity, quality, adaptive)
+//! register through `register_plugin` and see a `RoutingContext` (registry, settings, health and
+//! the current request), never the router itself.

litellm/a2a_protocol/providers/bedrock_agentcore/transformation.py

@@ -19,6 +19,7 @@ from litellm.a2a_protocol.utils import (
 )
 from litellm.exceptions import BadRequestError
 from litellm.llms.bedrock.chat.agentcore.transformation import AmazonAgentCoreConfig
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines
 
 RUNTIME_SESSION_ID_MIN_LENGTH: Final = 33
 RUNTIME_SESSION_ID_MAX_LENGTH: Final = 256
@@ -50,7 +51,7 @@ _RESERVED_PREFIX_HEADERS: Final[tuple[str, ...]] = (
 class _SSELineSource(Protocol):
     """Minimal streaming-response surface used to read SSE lines."""
 
-    def aiter_lines(self) -> AsyncIterator[str]: ...
+    def aiter_text(self) -> AsyncIterator[str]: ...
 
 
 def _filter_reserved_headers(
@@ -244,7 +245,7 @@ class BedrockAgentCoreA2ATransformation:
         Yields:
             Parsed JSON dicts from SSE data lines
         """
-        async for line in response.aiter_lines():
+        async for line in aiter_sse_lines(response.aiter_text()):
             line = line.strip()
             if not line:
                 continue

litellm/a2a_protocol/providers/watsonx_orchestrate/handler.py

@@ -20,6 +20,7 @@ from litellm.llms.custom_httpx.http_handler import (
     AsyncHTTPHandler,
     get_async_httpx_client,
 )
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines
 from litellm.types.llms.custom_http import httpxSpecialProvider
 
 _IBM_CLOUD_IAM_URL: Final = "https://iam.cloud.ibm.com/identity/token"
@@ -74,7 +75,7 @@ class _WXORun(TypedDict, total=False):
 
 
 class _SSELineSource(Protocol):
-    def aiter_lines(self) -> AsyncIterator[str]: ...
+    def aiter_text(self) -> AsyncIterator[str]: ...
 
 
 class _WXOView(TypedDict, total=False):
@@ -231,7 +232,7 @@ class WatsonxOrchestrateHandler:
     async def _accumulate_wxo_sse_text(response: _SSELineSource) -> str:
         source: Final[_WXOView] = {"sse_source": response}
         accumulated_text = ""
-        async for line in source["sse_source"].aiter_lines():
+        async for line in aiter_sse_lines(source["sse_source"].aiter_text()):
             if not line.startswith("data:"):
                 continue
             data_str = line[5:].strip()

litellm/caching/in_memory_cache.py

@@ -274,7 +274,9 @@ class InMemoryCache(BaseCache):
     ) -> list[float] | None:
         results: Final = []
         for increment in increment_list:
-            result = await self.async_increment(increment["key"], increment["increment_value"], **kwargs)
+            result = await self.async_increment(
+                increment["key"], increment["increment_value"], ttl=increment["ttl"], **kwargs
+            )
             results.append(result)
         return results
 

litellm/completion_extras/litellm_responses_transformation/transformation.py

@@ -34,6 +34,7 @@ from litellm.llms.base_llm.base_model_iterator import BaseModelResponseIterator
 from litellm.llms.base_llm.bridges.completion_transformation import (
     CompletionTransformationBridge,
 )
+from litellm.llms.custom_httpx.sse_lines import split_sse_lines
 from litellm.responses.litellm_completion_transformation.tool_search import is_anthropic_tool_search_chat_tool
 from litellm.responses.sse_output_recovery import (
     parse_sse_json_chunk,
@@ -1351,7 +1352,7 @@ class LiteLLMResponsesTransformationHandler(CompletionTransformationBridge):
         recovered_output_items: Final[dict[int, dict[str, object]]] = {}
         recovered_text_only_items: Final[dict[int, dict[str, object]]] = {}
 
-        for chunk in raw_sse.splitlines():
+        for chunk in split_sse_lines(raw_sse):
             parsed_chunk = parse_sse_json_chunk(chunk)
             if parsed_chunk is None:
                 continue

litellm/google_genai/streaming_iterator.py

@@ -4,6 +4,7 @@ from typing import TYPE_CHECKING, Any, Final
 
 import litellm
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines
 from litellm.proxy.pass_through_endpoints.success_handler import (
     PassThroughEndpointLogging,
 )
@@ -131,9 +132,9 @@ class GoogleGenAIGenerateContentStreamingIterator(BaseGoogleGenAIGenerateContent
         self.model = model
         self.generate_content_provider_config = generate_content_provider_config
         self.litellm_metadata = litellm_metadata
-        # Gemini streamGenerateContent uses SSE line framing; iter_lines keeps
+        # Gemini streamGenerateContent uses SSE line framing; iter_sse_lines keeps
         # large inlineData payloads (e.g. image/jpeg) intact within one event.
-        self.stream_iterator = response.iter_lines()
+        self.stream_iterator = iter_sse_lines(response.iter_text())
 
     def __iter__(self):
         return self
@@ -182,9 +183,9 @@ class AsyncGoogleGenAIGenerateContentStreamingIterator(BaseGoogleGenAIGenerateCo
         self.model = model
         self.generate_content_provider_config = generate_content_provider_config
         self.litellm_metadata = litellm_metadata
-        # Gemini streamGenerateContent uses SSE line framing; aiter_lines keeps
+        # Gemini streamGenerateContent uses SSE line framing; aiter_sse_lines keeps
         # large inlineData payloads (e.g. image/jpeg) intact within one event.
-        self.stream_iterator = response.aiter_lines()
+        self.stream_iterator = aiter_sse_lines(response.aiter_text())
 
     def __aiter__(self):
         return self

litellm/images/main.py

@@ -1,9 +1,9 @@
 import asyncio
 import contextvars
 import importlib
-from collections.abc import Coroutine, Mapping
+from collections.abc import Collection, Coroutine, Mapping
 from functools import partial
-from typing import TYPE_CHECKING, Final, Literal, cast, overload
+from typing import TYPE_CHECKING, Final, Literal, cast, get_type_hints, overload
 
 from pydantic import TypeAdapter, ValidationError
 
@@ -61,6 +61,7 @@ from litellm.types.utils import (
     ProviderRoutingHints,
 )
 from litellm.utils import (
+    EXTRA_BODY_ROUTING_KEYS,
     ImageResponse,
     ProviderConfigManager,
     get_llm_provider,
@@ -466,7 +467,7 @@ def image_generation(
             litellm.LlmProviders.QWENCLOUD,
             litellm.LlmProviders.QWEN_AI_PLATFORM,
             litellm.LlmProviders.EDENAI,
-        ):
+        ) or (custom_llm_provider == litellm.LlmProviders.TENCENT and image_generation_config is not None):
             if image_generation_config is None:
                 raise ValueError(f"image generation config is not supported for {custom_llm_provider}")
 
@@ -807,6 +808,32 @@ def image_variation(
     return response
 
 
+_IMAGE_EDIT_OPENAI_PARAMS: Final = frozenset(get_type_hints(ImageEditOptionalRequestParams))
+_DROP_PARAM_KWARGS: Final = frozenset({"drop_params", "additional_drop_params"})
+_ADDITIONAL_DROP_PARAMS: Final = TypeAdapter[tuple[str, ...]](tuple[str, ...])
+
+
+def _listed_drop_params(additional_drop_params: object) -> tuple[str, ...]:
+    try:
+        return _ADDITIONAL_DROP_PARAMS.validate_python(additional_drop_params)
+    except ValidationError:
+        return ()
+
+
+def _tokenhub_passthrough_edit_params(
+    non_default_params: Mapping[str, object],
+    extra_body: Mapping[str, object] | None,
+    mapped_params: Collection[str],
+    additional_drop_params: Collection[str] | None,
+) -> Mapping[str, object]:
+    dropped: Final = frozenset(additional_drop_params or ())
+    unmapped_excluded: Final = _DROP_PARAM_KWARGS | _IMAGE_EDIT_OPENAI_PARAMS | frozenset(mapped_params) | dropped
+    return {
+        **{key: value for key, value in non_default_params.items() if key not in unmapped_excluded},
+        **{key: value for key, value in (extra_body or {}).items() if key not in dropped | EXTRA_BODY_ROUTING_KEYS},
+    }
+
+
 @client
 def image_edit(
     image: FileTypes | list[FileTypes] | None = None,
@@ -935,22 +962,24 @@ def image_edit(
             raise ValueError(f"image edit is not supported for {custom_llm_provider}")
 
         local_vars.update(kwargs)
+        provider_supported_params: Final[frozenset[str]] = frozenset(
+            image_edit_provider_config.get_supported_openai_params(model)
+        )
         # Get ImageEditOptionalRequestParams with only valid parameters
         image_edit_optional_params: Final[ImageEditOptionalRequestParams] = (
             _get_ImageEditRequestUtils().get_requested_image_edit_optional_param(
                 local_vars,
-                provider_supported_params=frozenset(
-                    image_edit_provider_config.get_supported_openai_params(model)
-                ).intersection(non_default_params),
+                provider_supported_params=provider_supported_params.intersection(non_default_params),
             )
         )
         # Get optional parameters for the responses API
+        additional_drop_params: Final = _listed_drop_params(kwargs.get("additional_drop_params"))
         image_edit_request_params: Final[dict] = _get_ImageEditRequestUtils().get_optional_params_image_edit(
             model=model,
             image_edit_provider_config=image_edit_provider_config,
             image_edit_optional_params=image_edit_optional_params,
             drop_params=kwargs.get("drop_params"),
-            additional_drop_params=kwargs.get("additional_drop_params"),
+            additional_drop_params=additional_drop_params,
         )
 
         if image_edit_provider_config.use_multipart_form_data() and (
@@ -1039,7 +1068,19 @@ def image_edit(
             image=images,
             prompt=prompt,
             image_edit_provider_config=image_edit_provider_config,
-            image_edit_optional_request_params=image_edit_request_params,
+            image_edit_optional_request_params={
+                **image_edit_request_params,
+                **(
+                    _tokenhub_passthrough_edit_params(
+                        non_default_params,
+                        extra_body if isinstance(extra_body, dict) else None,
+                        provider_supported_params,
+                        additional_drop_params,
+                    )
+                    if custom_llm_provider == litellm.LlmProviders.TENCENT.value
+                    else {}
+                ),
+            },
             custom_llm_provider=custom_llm_provider,
             litellm_params=litellm_params,
             logging_obj=litellm_logging_obj,

litellm/images/utils.py

@@ -1,6 +1,9 @@
+import base64
+import os
 from collections.abc import Collection, Mapping
 from io import BufferedReader, BytesIO
-from typing import Any, Final, cast, get_type_hints
+from pathlib import Path
+from typing import Any, Final, Protocol, cast, get_type_hints, runtime_checkable
 
 import litellm
 from litellm.litellm_core_utils.token_counter import get_image_type
@@ -16,7 +19,7 @@ class ImageEditRequestUtils:
         image_edit_provider_config: BaseImageEditConfig,
         image_edit_optional_params: ImageEditOptionalRequestParams,
         drop_params: bool | None = None,
-        additional_drop_params: list[str] | None = None,
+        additional_drop_params: Collection[str] | None = None,
     ) -> dict:
         """
         Get optional parameters for the image edit API.
@@ -145,3 +148,36 @@ class ImageEditRequestUtils:
         except Exception:
             # If anything goes wrong, default to PNG
             return FILE_MIME_TYPES[FileType.PNG]
+
+
+@runtime_checkable
+class _SeekableBinaryReader(Protocol):
+    def tell(self) -> int: ...
+
+    def seek(self, offset: int) -> int: ...
+
+    def read(self) -> bytes: ...
+
+
+def _read_image_bytes(image: object) -> bytes:
+    if isinstance(image, bytes):
+        return image
+    if isinstance(image, tuple):
+        return _read_image_bytes(image[1])
+    if isinstance(image, os.PathLike):
+        return Path(image).read_bytes()
+    if isinstance(image, _SeekableBinaryReader):
+        position: Final = image.tell()
+        image.seek(0)
+        data: Final = image.read()
+        image.seek(position)
+        return data
+    raise ValueError(f"Unsupported image type for image edit: {type(image).__name__}")
+
+
+def image_to_data_url(image: object) -> str:
+    if isinstance(image, str):
+        return image
+    image_bytes: Final = _read_image_bytes(image)
+    mime_type: Final = ImageEditRequestUtils.get_image_content_type(image_bytes)
+    return f"data:{mime_type};base64,{base64.b64encode(image_bytes).decode('utf-8')}"

litellm/integrations/otel/model/spans.py

@@ -436,6 +436,7 @@ _POSTGRES_OPERATION_BY_CALL_TYPE: Final[Mapping[str, PostgresOperation]] = Mappi
         "reset_budget_windows": PostgresOperation("select", None),
         "write_budget_windows": PostgresOperation("update", None),
         "roll_window_spend_row": PostgresOperation("update", None),
+        "open_window_spend_row": PostgresOperation("upsert", None),
         "seed_window_spend": PostgresOperation("select", None),
         "select_window_spend_rows": PostgresOperation("select", None),
         "commit_window_spend_updates": PostgresOperation("upsert", None),

litellm/integrations/websearch_interception/handler.py

@@ -97,6 +97,8 @@ WEBSEARCH_NATIVE_BLOCKS_METADATA_KEY: Final = "websearch_native_blocks"
 
 _RESPONSE_CONTENT_FIELD: Final = "content"
 
+_CLAUDE_CODE_WEB_SEARCH_PROMPT_PREFIX: Final = "Perform a web search for the query:"
+
 _ResponseT: Final = TypeVar("_ResponseT")
 
 
@@ -363,7 +365,8 @@ class WebSearchInterceptionLogger(CustomLogger):
             get_last_user_message,
         )
 
-        query: Final = get_last_user_message(cast(list[AllMessageValues], messages))
+        last_user_message: Final = get_last_user_message(cast(list[AllMessageValues], messages)) or ""
+        query: Final = last_user_message.removeprefix(_CLAUDE_CODE_WEB_SEARCH_PROMPT_PREFIX).strip()
         if not query:
             return None
 

litellm/interactions/streaming_iterator.py

@@ -21,6 +21,7 @@ from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLogging
 from litellm.litellm_core_utils.llm_response_utils.get_api_base import get_api_base
 from litellm.litellm_core_utils.thread_pool_executor import executor
 from litellm.llms.base_llm.interactions.transformation import BaseInteractionsAPIConfig
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines
 from litellm.types.interactions import (
     InteractionsAPIStreamingResponse,
 )
@@ -145,7 +146,7 @@ class InteractionsAPIStreamingIterator(BaseInteractionsAPIStreamingIterator):
             litellm_metadata=litellm_metadata,
             custom_llm_provider=custom_llm_provider,
         )
-        self.stream_iterator = response.aiter_lines()
+        self.stream_iterator = aiter_sse_lines(response.aiter_text())
 
     def __aiter__(self):
         return self
@@ -212,7 +213,7 @@ class SyncInteractionsAPIStreamingIterator(BaseInteractionsAPIStreamingIterator)
             litellm_metadata=litellm_metadata,
             custom_llm_provider=custom_llm_provider,
         )
-        self.stream_iterator = response.iter_lines()
+        self.stream_iterator = iter_sse_lines(response.iter_text())
 
     def __iter__(self):
         return self

litellm/litellm_core_utils/get_litellm_params.py

@@ -75,6 +75,8 @@ OPTIONAL_KWARGS_KEYS: Final = (
             "rpm",
             "itpm",
             "otpm",
+            "tpd",
+            "rpd",
             "use_xai_oauth",
             "chatgpt_auth_file",
             "github_copilot_auth_type",

litellm/litellm_core_utils/litellm_logging.py

@@ -2758,7 +2758,7 @@ class Logging(LiteLLMLoggingBaseClass):
         if entry is None or not is_str_keyed_dict(litellm_params):
             return
         metadata: Final[object] = litellm_params.get("metadata")
-        bucket: Final = metadata if is_str_keyed_dict(metadata) else {}
+        bucket: Final = {**metadata} if is_str_keyed_dict(metadata) else {}
         litellm_params["metadata"] = bucket
         with_guardrail_entry(bucket, entry)
 
@@ -3481,7 +3481,7 @@ class Logging(LiteLLMLoggingBaseClass):
                             start_time=start_time,
                             end_time=end_time,
                             print_verbose=print_verbose,
-                            level=LogfireLevel.INFO.value,
+                            level=LogfireLevel.INFO,
                         )
 
                     if callback == "lunary" and lunaryLogger is not None:
@@ -4442,7 +4442,7 @@ class Logging(LiteLLMLoggingBaseClass):
                             response_obj=result,
                             start_time=start_time,
                             end_time=end_time,
-                            level=LogfireLevel.ERROR.value,
+                            level=LogfireLevel.ERROR,
                             print_verbose=print_verbose,
                         )
 

litellm/litellm_core_utils/llm_cost_calc/utils.py

@@ -6,7 +6,7 @@ from collections.abc import Mapping, Sequence
 from dataclasses import dataclass, field, replace
 from datetime import datetime, timezone, tzinfo
 from types import MappingProxyType
-from typing import Final, Literal, TypedDict, cast
+from typing import Final, Literal, Protocol, TypedDict, cast
 from zoneinfo import ZoneInfo, ZoneInfoNotFoundError
 
 from typing_extensions import ReadOnly
@@ -2085,6 +2085,41 @@ def calculate_image_response_web_search_cost(
     )
 
 
+class _ModelPricedImageCostCalculator(Protocol):
+    def __call__(self, model: str, image_response: ImageResponse, model_info: ModelInfo | None = None) -> float: ...
+
+
+def _model_priced_image_cost_calculator(
+    custom_llm_provider: str | None, model: str
+) -> _ModelPricedImageCostCalculator | None:
+    match custom_llm_provider:
+        case litellm.LlmProviders.RECRAFT.value:
+            from litellm.llms.recraft.cost_calculator import cost_calculator as recraft_image_cost_calculator
+
+            return recraft_image_cost_calculator
+        case litellm.LlmProviders.AIML.value:
+            from litellm.llms.aiml.image_generation.cost_calculator import cost_calculator as aiml_image_cost_calculator
+
+            return aiml_image_cost_calculator
+        case litellm.LlmProviders.COMETAPI.value:
+            from litellm.llms.cometapi.image_generation.cost_calculator import (
+                cost_calculator as cometapi_image_cost_calculator,
+            )
+
+            return cometapi_image_cost_calculator
+        case litellm.LlmProviders.RUNWAYML.value:
+            from litellm.llms.runwayml.cost_calculator import cost_calculator as runwayml_image_cost_calculator
+
+            return runwayml_image_cost_calculator
+        case litellm.LlmProviders.TENCENT.value:
+            from litellm.llms.tencent.cost_calculator import image_cost_calculator as tencent_image_cost_calculator
+            from litellm.llms.tencent.image_generation.transformation import is_tokenhub_image_model
+
+            return tencent_image_cost_calculator if is_tokenhub_image_model(model) else None
+        case _:
+            return None
+
+
 class CostCalculatorUtils:
     @staticmethod
     def call_type_has_image_response(call_type: str) -> bool:
@@ -2131,9 +2166,6 @@ class CostCalculatorUtils:
         from litellm.llms.gemini.image_generation.cost_calculator import (
             cost_calculator as gemini_image_cost_calculator,
         )
-        from litellm.llms.recraft.cost_calculator import (
-            cost_calculator as recraft_image_cost_calculator,
-        )
         from litellm.llms.vertex_ai.image_generation.cost_calculator import (
             cost_calculator as vertex_ai_image_cost_calculator,
         )
@@ -2146,7 +2178,10 @@ class CostCalculatorUtils:
         )
         resolved_n: Final = n if n is not None else (len(completion_response.data) if completion_response.data else 0)
         pricing: Final = deployment_pricing(model_info)
+        model_priced_calculator: Final = _model_priced_image_cost_calculator(custom_llm_provider, model)
 
+        if model_priced_calculator is not None:
+            return model_priced_calculator(model=model, image_response=completion_response, model_info=pricing)
         if custom_llm_provider == litellm.LlmProviders.VERTEX_AI.value:
             if isinstance(completion_response, ImageResponse):
                 return vertex_ai_image_cost_calculator(
@@ -2164,36 +2199,6 @@ class CostCalculatorUtils:
                     optional_params=optional_params,
                 )
             raise TypeError("completion_response must be of type ImageResponse for bedrock image cost calculation")
-        elif custom_llm_provider == litellm.LlmProviders.RECRAFT.value:
-            from litellm.llms.recraft.cost_calculator import (
-                cost_calculator as recraft_image_cost_calculator,
-            )
-
-            return recraft_image_cost_calculator(
-                model=model,
-                image_response=completion_response,
-                model_info=pricing,
-            )
-        elif custom_llm_provider == litellm.LlmProviders.AIML.value:
-            from litellm.llms.aiml.image_generation.cost_calculator import (
-                cost_calculator as aiml_image_cost_calculator,
-            )
-
-            return aiml_image_cost_calculator(
-                model=model,
-                image_response=completion_response,
-                model_info=pricing,
-            )
-        elif custom_llm_provider == litellm.LlmProviders.COMETAPI.value:
-            from litellm.llms.cometapi.image_generation.cost_calculator import (
-                cost_calculator as cometapi_image_cost_calculator,
-            )
-
-            return cometapi_image_cost_calculator(
-                model=model,
-                image_response=completion_response,
-                model_info=pricing,
-            )
         elif custom_llm_provider == litellm.LlmProviders.GEMINI.value:
             if call_type in (
                 CallTypes.image_edit.value,
@@ -2237,16 +2242,6 @@ class CostCalculatorUtils:
                 optional_params=optional_params,
                 model_info=pricing,
             )
-        elif custom_llm_provider == litellm.LlmProviders.RUNWAYML.value:
-            from litellm.llms.runwayml.cost_calculator import (
-                cost_calculator as runwayml_image_cost_calculator,
-            )
-
-            return runwayml_image_cost_calculator(
-                model=model,
-                image_response=completion_response,
-                model_info=pricing,
-            )
         elif (
             custom_llm_provider == litellm.LlmProviders.OPENAI.value
             or custom_llm_provider == litellm.LlmProviders.AZURE.value

litellm/litellm_core_utils/prompt_templates/common_utils.py

@@ -21,8 +21,8 @@ from pydantic import TypeAdapter
 
 import litellm
 from litellm import verbose_logger
+from litellm.litellm_core_utils.batch_utils import InMemoryFile
 from litellm.litellm_core_utils.core_helpers import as_str_mapping
-from litellm.router_utils.batch_utils import InMemoryFile
 from litellm.types.llms.openai import (
     AllMessageValues,
     ChatCompletionAssistantMessage,

litellm/litellm_core_utils/safe_json_dumps.py

@@ -34,7 +34,7 @@ def safe_json_structure(
     def _transform(key: str | None, value: str) -> str:
         return value if value_transform is None else value_transform(key, value)
 
-    def _serialize(obj: object, seen: set[int], depth: int, key: str | None = None) -> Any:
+    def _serialize(obj: object, seen: set[int], depth: int, key: str | None = None) -> object:
         # Check for maximum depth.
         if depth > max_depth:
             return "MaxDepthExceeded"
@@ -48,32 +48,32 @@ def safe_json_structure(
         if id(obj) in seen:
             return "CircularReference Detected"
         seen.add(id(obj))
-        result: dict | list | tuple | set | str
         if isinstance(obj, dict):
-            result = {}
-            for k, v in obj.items():
-                if isinstance(k, (str)):
-                    clean_k = k.replace("\x00", "") if "\x00" in k else k
-                    result[clean_k] = _serialize(v, seen, depth + 1, clean_k)
+            mapping_result: Final = {
+                (clean_k := k.replace("\x00", "") if "\x00" in k else k): _serialize(v, seen, depth + 1, clean_k)
+                for k, v in obj.items()
+                if isinstance(k, str)
+            }
             seen.remove(id(obj))
-            return result
+            return mapping_result
         elif isinstance(obj, list):
-            result = [_serialize(item, seen, depth + 1, key) for item in obj]
+            list_result: Final = [_serialize(item, seen, depth + 1, key) for item in obj]
             seen.remove(id(obj))
-            return result
+            return list_result
         elif isinstance(obj, tuple):
-            result = tuple(_serialize(item, seen, depth + 1, key) for item in obj)
+            tuple_result: Final = tuple(_serialize(item, seen, depth + 1, key) for item in obj)
             seen.remove(id(obj))
-            return result
+            return tuple_result
         elif isinstance(obj, set):
-            result = sorted([_serialize(item, seen, depth + 1, key) for item in obj])
+            sortable_values: Final[list[Any]] = [_serialize(item, seen, depth + 1, key) for item in obj]
+            set_result: Final = sorted(sortable_values)
             seen.remove(id(obj))
-            return result
+            return set_result
         elif isinstance(obj, BaseModel):
             dumped: Final = obj.model_dump()
-            result = _serialize(dumped, seen, depth + 1, key)
+            model_result: Final = _serialize(dumped, seen, depth + 1, key)
             seen.remove(id(obj))
-            return result
+            return model_result
         else:
             # Fall back to string conversion for non-serializable objects.
             try:

litellm/litellm_core_utils/streaming_handler.py

@@ -1263,7 +1263,7 @@ class CustomStreamWrapper:
                 )
 
             if _is_delta_empty:
-                model_response.choices[0].delta = Delta(content=None)  # ensure empty delta chunk returned
+                model_response.choices[0].delta = Delta(content=self._close_open_think_block())
                 # get any function call arguments
                 model_response.choices[0].finish_reason = map_finish_reason(
                     finish_reason=self.received_finish_reason
@@ -1292,21 +1292,26 @@ class CustomStreamWrapper:
         if self.merge_reasoning_content_in_choices is not True:
             return
         delta: Final = model_response.choices[0].delta
-        reasoning_content: Final = getattr(delta, "reasoning_content", None)
+        reasoning_content: Final = getattr(delta, "reasoning_content", None) or ""
         content: Final = delta.content or ""
-        if reasoning_content:
-            self._emitted_disqualifying_content = True
-            closes_thinking: Final = bool(content) and not self.sent_last_thinking_block
-            opening: Final = "" if self.sent_first_thinking_block else "<think>"
-            delta.content = opening + reasoning_content + ("</think>" + content if closes_thinking else "")
-            self.sent_first_thinking_block = True
-            self.sent_last_thinking_block = self.sent_last_thinking_block or closes_thinking
-        elif self.sent_first_thinking_block and not self.sent_last_thinking_block and content:
-            delta.content = "</think>" + content
-            self.sent_last_thinking_block = True
-
+        ends_thinking: Final = bool(content or delta.tool_calls or delta.function_call)
+        inside_think_block: Final = self.sent_first_thinking_block and not self.sent_last_thinking_block
+        opening: Final = "<think>" if reasoning_content and not inside_think_block else ""
+        closing: Final = "</think>" if ends_thinking and (reasoning_content or inside_think_block) else ""
         if hasattr(delta, "reasoning_content"):
             del delta.reasoning_content
+        if not reasoning_content and not closing:
+            return
+        delta.content = opening + reasoning_content + closing + content
+        self.sent_first_thinking_block = True
+        self.sent_last_thinking_block = bool(closing)
+        self._emitted_disqualifying_content = self._emitted_disqualifying_content or bool(reasoning_content)
+
+    def _close_open_think_block(self) -> str | None:
+        if not self.sent_first_thinking_block or self.sent_last_thinking_block:
+            return None
+        self.sent_last_thinking_block = True
+        return "</think>"
 
     def _dispatch_provider_chunk(
         self,
@@ -1932,8 +1937,20 @@ class CustomStreamWrapper:
             model_response.choices[0].finish_reason == "stop" and self.tool_call
         ):  # don't overwrite for other - potential error finish reasons
             model_response.choices[0].finish_reason = "tool_calls"
+        close_tag: Final = self._close_open_think_block()
+        if close_tag is not None:
+            model_response.choices[0].delta.content = close_tag
         return model_response
 
+    def _final_chunk_for_logging(self, processed_chunk: "ModelResponseStream") -> "ModelResponseStream | None":
+        if self.received_finish_reason is not None or self.intermittent_finish_reason is not None:
+            return processed_chunk
+        if processed_chunk.choices[0].delta.content is None:
+            return None
+        return processed_chunk.model_copy(
+            update={"choices": [processed_chunk.choices[0].model_copy(update={"finish_reason": None})]}
+        )
+
     def _record_usage_only_chunk(self, model_response: "ModelResponseStream") -> None:
         """
         Keep provider usage-only chunks (e.g. OpenRouter's post-finish chunk, which carries a
@@ -2173,10 +2190,9 @@ class CustomStreamWrapper:
             else:
                 self.sent_last_chunk = True
                 processed_chunk: Final = self.finish_reason_handler()
-                # The logged response is built from self.chunks; keep a finish_reason the provider sent on its
-                # last content chunk (stripped there), but never add the synthetic "stop" used when it sent none.
-                if self.received_finish_reason is not None or self.intermittent_finish_reason is not None:
-                    self.chunks.append(processed_chunk)
+                final_chunk_for_logging: Final = self._final_chunk_for_logging(processed_chunk)
+                if final_chunk_for_logging is not None:
+                    self.chunks.append(final_chunk_for_logging)
                 if self.stream_options is None:
                     self._set_final_chunk_usage(processed_chunk)
                 ## LOGGING
@@ -2406,10 +2422,9 @@ class CustomStreamWrapper:
         else:
             self.sent_last_chunk = True
             processed_chunk: Final = self.finish_reason_handler()
-            # The logged response is built from self.chunks; keep a finish_reason the provider sent on its
-            # last content chunk (stripped there), but never add the synthetic "stop" used when it sent none.
-            if self.received_finish_reason is not None or self.intermittent_finish_reason is not None:
-                self.chunks.append(processed_chunk)
+            final_chunk_for_logging: Final = self._final_chunk_for_logging(processed_chunk)
+            if final_chunk_for_logging is not None:
+                self.chunks.append(final_chunk_for_logging)
             if self.stream_options is None:
                 self._set_final_chunk_usage(processed_chunk)
             # see sync __next__'s sibling branch: deliberately do NOT restore

litellm/llms/anthropic/chat/handler.py

@@ -38,6 +38,7 @@ from litellm.llms.custom_httpx.http_handler import (
     get_async_httpx_client,
     get_httpx_client,
 )
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines
 from litellm.types.llms.anthropic import (
     ContentBlockDelta,
     ContentBlockStart,
@@ -233,7 +234,7 @@ async def make_call(
         raise AnthropicError(status_code=500, message=str(e))
 
     completion_stream: Final = ModelResponseIterator(
-        streaming_response=response.aiter_lines(),
+        streaming_response=aiter_sse_lines(response.aiter_text()),
         sync_stream=False,
         json_mode=json_mode,
         speed=speed,
@@ -304,7 +305,7 @@ def make_sync_call(
         )
 
     completion_stream: Final = ModelResponseIterator(
-        streaming_response=response.iter_lines(),
+        streaming_response=iter_sse_lines(response.iter_text()),
         sync_stream=True,
         json_mode=json_mode,
         speed=speed,

litellm/llms/anthropic/messages/caller_model_checks.py

@@ -268,6 +268,7 @@ async def caller_within_model_budget(
                 user_id=user_id,
                 user_model_max_budget=user_model_max_budget,
                 model=model,
+                team_id=user_api_key_auth.team_id,
             )
         except litellm.BudgetExceededError:
             return "denied"
@@ -309,6 +310,7 @@ async def caller_within_model_budget(
                 end_user_id=end_user_id,
                 end_user_model_max_budget=end_user_model_max_budget,
                 model=model,
+                team_id=user_api_key_auth.team_id,
             )
         except litellm.BudgetExceededError:
             return "denied"

litellm/llms/anthropic/messages/stream_assembly.py

@@ -13,6 +13,7 @@ from litellm.llms.anthropic.chat.handler import (
 )
 from litellm.llms.anthropic.chat.transformation import AnthropicConfig
 from litellm.llms.anthropic.messages.native.utils import MESSAGE_DELTA_OPTIONAL_KEYS
+from litellm.llms.custom_httpx.sse_lines import split_sse_lines
 from litellm.types.utils import (
     Choices,
     Message,
@@ -269,7 +270,7 @@ def build_complete_streaming_response_legacy(
 
 def extract_sse_data(event_str: str) -> dict | None:
     """Parse the JSON object from the ``data:`` line of an Anthropic SSE event."""
-    for line in event_str.splitlines():
+    for line in split_sse_lines(event_str):
         stripped = line.strip()
         if stripped.startswith("data:"):
             payload = stripped[len("data:") :].strip()

litellm/llms/anthropic/prompt_cache_prediction.py

@@ -29,7 +29,7 @@ from litellm.llms.anthropic.messages.native.transformation import (
     AnthropicMessagesConfig,
 )
 from litellm.llms.anthropic.messages.native.utils import AnthropicMessagesRequestUtils, prepare_native_messages
-from litellm.router_utils.baseline_request import (
+from litellm.router_strategy.auto_router.baseline_request import (
     BASELINE_PARAMETERS,
     capture_baseline_parameters,
 )

litellm/llms/azure_ai/agents/handler.py

@@ -35,6 +35,7 @@ from litellm.llms.azure_ai.agents.transformation import (
     AzureAIAgentsConfig,
     AzureAIAgentsError,
 )
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines
 from litellm.types.llms.openai import (
     ChatCompletionAnnotation,
     ChatCompletionAnnotationURLCitation,
@@ -632,7 +633,7 @@ class AzureAIAgentsHandler:
 
         current_event = None
 
-        async for line in response.aiter_lines():
+        async for line in aiter_sse_lines(response.aiter_text()):
             line = line.strip()
 
             if line.startswith("event:"):

litellm/llms/base_llm/base_model_iterator.py

@@ -96,7 +96,7 @@ class BaseModelResponseIterator:
         released (and a backend like vLLM aborts generation) when the stream
         is abandoned before its natural end.
 
-        ``streaming_response`` is usually a bare ``aiter_lines()`` generator
+        ``streaming_response`` is usually a bare ``aiter_sse_lines()`` generator
         that holds no reference to the response, so the handler that owns the
         response attaches it here after construction."""
         if self.http_response is not None:

litellm/llms/base_llm/guardrail_translation/utils.py

@@ -6,6 +6,7 @@ from typing import Final, TypeVar, cast  # noqa: TID251  # a rebuilt chat row ha
 
 from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
 
+from litellm.llms.custom_httpx.sse_lines import split_sse_lines
 from litellm.types.llms.anthropic_messages.anthropic_response import AnthropicUsage
 from litellm.types.llms.openai import AllMessageValues, ResponseAPIUsage
 
@@ -24,7 +25,7 @@ def _anthropic_stream_chunk_events(item: object) -> list[dict]:
 
     events: Final[list[dict]] = []
     for block in chunk.split("\n\n"):
-        for line in block.splitlines():
+        for line in split_sse_lines(block):
             stripped = line.strip()
             if not stripped.startswith("data:"):
                 continue

litellm/llms/base_llm/sandbox/transformation.py

@@ -12,6 +12,7 @@ from typing import Final
 import httpx
 from pydantic import Field, PrivateAttr
 
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines
 from litellm.types.llms.base import LiteLLMPydanticObjectBase
 
 SANDBOX_MAX_OUTPUT_BYTES: Final = 10 * 1024 * 1024
@@ -99,7 +100,7 @@ class BaseSandboxConfig:
     async def _read_capped_lines(self, response: httpx.Response) -> list[str]:
         lines: Final[list[str]] = []
         total = 0
-        async for line in response.aiter_lines():
+        async for line in aiter_sse_lines(response.aiter_text()):
             total += len(line.encode("utf-8"))
             if total > SANDBOX_MAX_OUTPUT_BYTES:
                 raise ValueError(

litellm/llms/bedrock/chat/invoke_transformations/amazon_twelvelabs_pegasus_transformation.py

@@ -10,6 +10,8 @@ import time
 from typing import TYPE_CHECKING, Any, Final
 
 import httpx
+from pydantic import ConfigDict, TypeAdapter, with_config
+from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 import litellm
 from litellm._logging import verbose_logger
@@ -34,6 +36,15 @@ else:
     LiteLLMLoggingObj = Any
 
 
+@with_config(ConfigDict(extra="allow", strict=True))
+class _PegasusResponse(TypedDict):
+    message: ReadOnly[NotRequired[str | None]]
+    finishReason: ReadOnly[NotRequired[str | None]]
+
+
+_PEGASUS_RESPONSE: Final = TypeAdapter(_PegasusResponse)
+
+
 class AmazonTwelveLabsPegasusConfig(AmazonInvokeConfig, BaseConfig):
     """
     Handles transforming OpenAI-style requests into Bedrock InvokeModel requests for
@@ -207,7 +218,7 @@ class AmazonTwelveLabsPegasusConfig(AmazonInvokeConfig, BaseConfig):
         ModelResponse with choices[0].message.content and finish_reason
         """
         try:
-            completion_response: Final = raw_response.json()
+            completion_response: Final = _PEGASUS_RESPONSE.validate_python(raw_response.json())
         except Exception as e:
             raise BedrockError(
                 message=f"Error parsing response: {raw_response.text}, error: {e}",
@@ -224,7 +235,7 @@ class AmazonTwelveLabsPegasusConfig(AmazonInvokeConfig, BaseConfig):
         message_content: Final = completion_response.get("message", "")
 
         # Extract finish reason and map to LiteLLM format
-        finish_reason_raw: Final = completion_response.get("finishReason", "stop")
+        finish_reason_raw: Final = completion_response.get("finishReason") or "stop"
         finish_reason: Final = map_finish_reason(finish_reason_raw)
 
         # Set the response content

litellm/llms/bedrock/chat/invoke_transformations/base_invoke_transformation.py

@@ -1,11 +1,12 @@
 import copy
 import json
 import time
-from collections.abc import Mapping
+from collections.abc import Mapping, Sequence
 from typing import TYPE_CHECKING, Any, Final, cast, get_args
 
 import httpx
-from pydantic import TypeAdapter, ValidationError
+from pydantic import ConfigDict, TypeAdapter, ValidationError, with_config
+from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 import litellm
 from litellm._logging import verbose_logger
@@ -48,6 +49,27 @@ from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 
 _GUARDRAIL_CONFIG_VALIDATOR: Final["TypeAdapter[GuardrailConfigBlock]"] = TypeAdapter(GuardrailConfigBlock)
 
+
+@with_config(ConfigDict(extra="allow", strict=True))
+class _CohereGeneration(TypedDict):
+    text: ReadOnly[str | None]
+    finish_reason: ReadOnly[str | None]
+
+
+@with_config(ConfigDict(extra="allow", strict=True))
+class _CohereInvokeResponse(TypedDict):
+    text: NotRequired[ReadOnly[str | None]]
+    generations: NotRequired[ReadOnly[Sequence[_CohereGeneration]]]
+
+
+@with_config(ConfigDict(extra="allow", strict=True))
+class _LlamaInvokeResponse(TypedDict):
+    generation: ReadOnly[str | None]
+
+
+_COHERE_INVOKE_RESPONSE: Final = TypeAdapter(_CohereInvokeResponse)
+_LLAMA_INVOKE_RESPONSE: Final = TypeAdapter(_LlamaInvokeResponse)
+
 _GUARDRAIL_CONFIG_EXPECTED_FORMAT: Final = (
     "{'guardrailIdentifier': str, 'guardrailVersion': str, 'trace': 'enabled'|'disabled'|'enabled_full'}"
 )
@@ -310,12 +332,13 @@ class AmazonInvokeConfig(BaseConfig, BaseAWSLLM):
         outputText: str | None = None
         try:
             if provider == "cohere":
-                if "text" in completion_response:
-                    outputText = completion_response["text"]
-                elif "generations" in completion_response:
-                    outputText = completion_response["generations"][0]["text"]
+                cohere_response: Final = _COHERE_INVOKE_RESPONSE.validate_python(completion_response)
+                if "text" in cohere_response:
+                    outputText = cohere_response["text"]
+                elif "generations" in cohere_response:
+                    outputText = cohere_response["generations"][0]["text"]
                     model_response.choices[0].finish_reason = map_finish_reason(
-                        completion_response["generations"][0]["finish_reason"]
+                        cohere_response["generations"][0]["finish_reason"] or "stop"
                     )
             elif provider == "anthropic":
                 return litellm.AmazonAnthropicClaudeConfig().transform_response(
@@ -361,7 +384,7 @@ class AmazonInvokeConfig(BaseConfig, BaseAWSLLM):
             elif provider == "ai21":
                 outputText = completion_response.get("completions")[0].get("data").get("text")
             elif provider == "meta" or provider == "llama" or provider == "deepseek_r1":
-                outputText = completion_response["generation"]
+                outputText = _LLAMA_INVOKE_RESPONSE.validate_python(completion_response)["generation"]
             elif provider == "mistral":
                 outputText = litellm.AmazonMistralConfig.get_outputText(completion_response, model_response)
             else:  # amazon titan

litellm/llms/bedrock/files/transformation.py

@@ -49,6 +49,7 @@ from litellm.llms.base_llm.files.transformation import (
     BaseFilesConfig,
     LiteLLMLoggingObj,
 )
+from litellm.llms.custom_httpx.sse_lines import split_sse_lines
 from litellm.types.llms.base import LiteLLMBaseModel
 from litellm.types.llms.bedrock import AwsAuthParams, BedrockBatchRecordKind
 from litellm.types.llms.openai import (
@@ -400,7 +401,7 @@ def _batch_input_line(line_number: int, line: str) -> _BatchInputLine:
 
 
 def _non_blank_lines(file_content: str) -> Iterator[tuple[int, str]]:
-    numbered_lines: Final = enumerate(file_content.splitlines(), start=1)
+    numbered_lines: Final = enumerate(split_sse_lines(file_content), start=1)
     return ((line_number, line) for line_number, line in numbered_lines if line.strip())
 
 

litellm/llms/bedrock/passthrough/guardrail_translation/handler.py

@@ -45,6 +45,7 @@ from litellm.llms.bedrock.passthrough.guardrail_translation.structured_edits imp
     structured_edits,
     tool_edits,
 )
+from litellm.llms.custom_httpx.sse_lines import split_sse_lines
 from litellm.proxy.guardrails.auto_router_compression import is_compression_guardrail
 from litellm.types.utils import GenericGuardrailAPIInputs
 
@@ -947,7 +948,7 @@ def _invoke_frames(items: Sequence[object]) -> tuple[InvokeStreamFrame, ...]:
 def _sse_events(sse: bytes) -> tuple[Mapping[str, object], ...]:
     return tuple(
         _JSON_OBJECT.validate_python(json.loads(line[len("data:") :]))
-        for line in sse.decode().splitlines()
+        for line in split_sse_lines(sse.decode())
         if line.startswith("data:")
     )
 

litellm/llms/bedrock/realtime/transformation.py

@@ -53,6 +53,7 @@ from litellm.types.realtime import (
     RealtimeResponseTypedDict,
 )
 
+_JSON_OBJECT_MESSAGE: Final = TypeAdapter(dict[str, JsonValue], config=ConfigDict(strict=True))
 _JSON_VALUE: Final = TypeAdapter(object)
 _JSON_OBJECT: Final = TypeAdapter(Mapping[str, object])
 
@@ -856,7 +857,7 @@ class BedrockRealtimeConfig(BaseRealtimeConfig):
             List of Bedrock format messages (JSON strings)
         """
         try:
-            json_message: Final[dict[str, JsonValue]] = json.loads(message)
+            json_message: Final = _JSON_OBJECT_MESSAGE.validate_python(json.loads(message))
         except json.JSONDecodeError:
             verbose_logger.warning("Invalid JSON message: %s", message[:200])
             return []

litellm/llms/bedrock/search/transformation.py

@@ -47,6 +47,7 @@ from litellm.llms.base_llm.search.transformation import (
 )
 from litellm.llms.bedrock.base_aws_llm import BaseAWSLLM
 from litellm.llms.bedrock.common_utils import BedrockError
+from litellm.llms.custom_httpx.sse_lines import split_sse_lines
 from litellm.secret_managers.main import get_secret_str
 
 # AgentCore web-search rejects queries longer than 200 characters
@@ -146,7 +147,9 @@ def _iter_sse_events(text: str) -> Iterator[Mapping[str, object]]:
     progress notifications before the JSON-RPC response.
     """
     for chunk in _SSE_EVENT_SEPARATOR.split(text):
-        payload = "\n".join(line[len("data:") :].lstrip() for line in chunk.splitlines() if line.startswith("data:"))
+        payload = "\n".join(
+            line[len("data:") :].lstrip() for line in split_sse_lines(chunk) if line.startswith("data:")
+        )
         if not payload:
             continue
         try:

litellm/llms/chatgpt/responses/transformation.py

@@ -11,6 +11,7 @@ from litellm.litellm_core_utils.llm_response_utils.convert_dict_to_response impo
     safe_convert_created_field,
 )
 from litellm.litellm_core_utils.url_utils import join_api_base
+from litellm.llms.custom_httpx.sse_lines import split_sse_lines
 from litellm.llms.openai.common_utils import OpenAIError
 from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
 from litellm.responses.sse_output_recovery import (
@@ -199,7 +200,7 @@ class ChatGPTResponsesAPIConfig(OpenAIResponsesAPIConfig):
         error_message = None
         streamed_output_items: Final[dict[int, dict[str, object]]] = {}
         text_only_output_items: Final[dict[int, dict[str, object]]] = {}
-        for chunk in body_text.splitlines():
+        for chunk in split_sse_lines(body_text):
             parsed_chunk = parse_sse_json_chunk(chunk)
             if parsed_chunk is None:
                 continue

litellm/llms/codestral/completion/handler.py

@@ -20,6 +20,7 @@ from litellm.llms.custom_httpx.http_handler import (
     AsyncHTTPHandler,
     get_async_httpx_client,
 )
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines
 from litellm.types.utils import TextChoices
 from litellm.utils import CustomStreamWrapper, TextCompletionResponse
 
@@ -109,7 +110,7 @@ async def make_call(
     if response.status_code != 200:
         raise TextCompletionCodestralError(status_code=response.status_code, message=response.text)
 
-    completion_stream: Final = response.aiter_lines()
+    completion_stream: Final = aiter_sse_lines(response.aiter_text())
     # LOGGING
     logging_obj.post_call(
         input=messages,
@@ -339,13 +340,13 @@ class CodestralTextCompletion:
             response = litellm.module_level_client.post(
                 completion_url,
                 headers=headers,
-                data=json.dumps(data),
+                data=json.dumps({**data, "stream": True}),
                 stream=stream,
             )
             _response: Final = CustomStreamWrapper(
-                response.iter_lines(),
+                iter_sse_lines(response.iter_text()),
                 model,
-                custom_llm_provider="codestral",
+                custom_llm_provider="text-completion-codestral",
                 logging_obj=logging_obj,
             )
             return _response

litellm/llms/custom_httpx/aiohttp_handler.py

@@ -21,6 +21,7 @@ from litellm.llms.custom_httpx.http_handler import (
     get_httpx_client,
     get_ssl_configuration,
 )
+from litellm.llms.custom_httpx.sse_lines import iter_sse_lines
 from litellm.types.llms.openai import FileTypes
 from litellm.types.utils import HttpHandlerRequestFields, ImageResponse, LlmProviders
 from litellm.utils import CustomStreamWrapper, ModelResponse, ProviderConfigManager
@@ -466,7 +467,7 @@ class BaseLLMAIOHTTPHandler:
             )
         else:
             completion_stream = provider_config.get_model_response_iterator(
-                streaming_response=response.iter_lines(), sync_stream=True
+                streaming_response=iter_sse_lines(response.iter_text()), sync_stream=True
             )
 
         # LOGGING

litellm/llms/custom_httpx/aiohttp_transport.py

@@ -6,6 +6,7 @@ import ssl
 import sys
 import typing
 import urllib.request
+import weakref
 from collections.abc import Callable, Generator
 from typing import ClassVar, Final
 
@@ -109,6 +110,24 @@ def map_aiohttp_exceptions() -> Generator[None, None, None]:
         raise mapped_exc(message) from exc
 
 
+def _release_transport_anchor(_transport: object) -> None:
+    """Release a transport anchored to a response stream. See ``_anchor_transport_to``."""
+
+
+def _anchor_transport_to(stream: "AiohttpResponseStream", transport: object) -> None:
+    """Keep the transport alive for as long as the stream can still read.
+
+    The transport finalizer closes the session it owns, and nothing on the path from a
+    live response back to that session references the transport, so a client dropped
+    mid-stream would close the session under the body. ``weakref.finalize`` holds the
+    transport in its own registry rather than on the stream, which keeps the transport
+    out of the response's reference cycle (an exception carrying the response, for one)
+    so it is finalized by refcount once the anchor drops and can still schedule an async
+    close, instead of being collected in the same cyclic pass as the session it closes.
+    """
+    weakref.finalize(stream, _release_transport_anchor, transport)
+
+
 class AiohttpResponseStream(httpx.AsyncByteStream):
     CHUNK_SIZE = 1024 * 16
 
@@ -463,10 +482,12 @@ class LiteLLMAiohttpTransport(AiohttpTransport):
                 # Re-raise if it's a different RuntimeError
                 raise
 
+        stream: Final = AiohttpResponseStream(response)
+        _anchor_transport_to(stream, self)
         return httpx.Response(
             status_code=response.status,
             headers=response.headers,
-            stream=AiohttpResponseStream(response),
+            stream=stream,
             request=request,
         )
 

litellm/llms/custom_httpx/llm_http_handler.py

@@ -118,6 +118,7 @@ from litellm.llms.custom_httpx.http_handler import (
     get_httpx_client,
     header_value,
 )
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines
 from litellm.responses.streaming_iterator import (
     BaseResponsesAPIStreamingIterator,
     MockResponsesAPIStreamingIterator,
@@ -1086,7 +1087,7 @@ class BaseLLMHTTPHandler:
             completion_stream: object = MockResponseIterator(model_response=model_response, json_mode=json_mode)
         else:
             completion_stream = provider_config.get_model_response_iterator_for_request(
-                streaming_response=response.iter_lines(),
+                streaming_response=iter_sse_lines(response.iter_text()),
                 sync_stream=True,
                 json_mode=json_mode,
                 optional_params=_OPTIONAL_PARAMS.validate_python(optional_params),
@@ -1229,7 +1230,7 @@ class BaseLLMHTTPHandler:
             completion_stream: object = MockResponseIterator(model_response=model_response, json_mode=json_mode)
         else:
             completion_stream = provider_config.get_model_response_iterator_for_request(
-                streaming_response=response.aiter_lines(),
+                streaming_response=aiter_sse_lines(response.aiter_text()),
                 sync_stream=False,
                 json_mode=json_mode,
                 optional_params=_OPTIONAL_PARAMS.validate_python(optional_params),

litellm/llms/custom_httpx/sse_lines.py

@@ -0,0 +1,74 @@
+"""
+Split a decoded Server-Sent Events body into lines the way the SSE spec does.
+
+httpx's `iter_lines()` / `aiter_lines()` follow `str.splitlines()`, which also breaks on U+0085, U+2028,
+U+2029, \\x0b, \\x0c and \\x1c-\\x1e. JSON leaves U+0085, U+2028 and U+2029 unescaped inside strings, so a
+`data:` line carrying one of them gets cut in half. SSE lines end only at CRLF, LF or CR
+(https://html.spec.whatwg.org/multipage/server-sent-events.html#parsing-an-event-stream)
+"""
+
+import io
+import re
+from collections.abc import AsyncIterable, AsyncIterator, Iterable, Iterator, Sequence
+from typing import Final
+
+_SSE_LINE_BREAK: Final = re.compile(r"\r\n|\r|\n")
+_SSE_LINE_WITH_BREAK: Final = re.compile(r"[^\r\n]*(?:\r\n|\r|\n)|[^\r\n]+")
+
+
+def _split_lines(text: str) -> Sequence[str]:
+    if "\r" not in text:
+        return text.split("\n")
+    crlf_folded: Final = text.replace("\r\n", "\n")
+    if "\r" not in crlf_folded:
+        return crlf_folded.split("\n")
+    return _SSE_LINE_BREAK.split(text)
+
+
+class _SSELineSplitter:
+    def __init__(self) -> None:
+        self._partial_line: io.StringIO = io.StringIO()
+        self._after_cr: bool = False
+
+    def feed(self, text: str) -> tuple[str, ...]:
+        if not text:
+            return ()
+        fresh: Final = text[1:] if self._after_cr and text.startswith("\n") else text
+        self._after_cr = fresh.endswith("\r")
+        *ended, tail = _split_lines(fresh)
+        if not ended:
+            _ = self._partial_line.write(tail)
+            return ()
+        first: Final = self._partial_line.getvalue() + ended[0]
+        self._partial_line = io.StringIO()
+        _ = self._partial_line.write(tail)
+        return (first, *ended[1:])
+
+    def flush(self) -> tuple[str, ...]:
+        rest: Final = self._partial_line.getvalue()
+        self._partial_line = io.StringIO()
+        return (rest,) if rest else ()
+
+
+def iter_sse_lines(text_chunks: Iterable[str]) -> Iterator[str]:
+    splitter: Final = _SSELineSplitter()
+    for text in text_chunks:
+        yield from splitter.feed(text)
+    yield from splitter.flush()
+
+
+def split_sse_lines(text: str) -> tuple[str, ...]:
+    return tuple(iter_sse_lines((text,)))
+
+
+def split_sse_lines_keepends(text: str) -> tuple[str, ...]:
+    return tuple(_SSE_LINE_WITH_BREAK.findall(text))
+
+
+async def aiter_sse_lines(text_chunks: AsyncIterable[str]) -> AsyncIterator[str]:
+    splitter: Final = _SSELineSplitter()
+    async for text in text_chunks:
+        for line in splitter.feed(text):
+            yield line
+    for line in splitter.flush():
+        yield line

litellm/llms/dashscope/embed/transformation.py

@@ -14,6 +14,8 @@ Docs - https://help.aliyun.com/zh/model-studio/text-embedding-synchronous-api
 from typing import Final
 
 import httpx
+from pydantic import ConfigDict, TypeAdapter, with_config
+from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.litellm_core_utils.url_utils import ensure_api_route
@@ -28,6 +30,23 @@ from ..common_utils import DashScopeError
 DEFAULT_API_BASE: Final = "https://dashscope.aliyuncs.com/compatible-mode/v1"
 
 
+@with_config(ConfigDict(extra="allow", strict=True))
+class _DashScopeUsage(TypedDict, total=False):
+    prompt_tokens: NotRequired[ReadOnly[int]]
+    total_tokens: NotRequired[ReadOnly[int]]
+
+
+@with_config(ConfigDict(extra="allow", strict=True))
+class _DashScopeEmbeddingResponse(TypedDict, total=False):
+    data: NotRequired[ReadOnly[list[object]]]
+    model: NotRequired[ReadOnly[str | None]]
+    usage: NotRequired[ReadOnly[_DashScopeUsage | None]]
+    id: NotRequired[ReadOnly[object]]
+
+
+_DASHSCOPE_EMBEDDING_RESPONSE: Final = TypeAdapter(_DashScopeEmbeddingResponse)
+
+
 class DashScopeEmbeddingConfig(BaseEmbeddingConfig):
     """
     Reference: https://help.aliyun.com/zh/model-studio/text-embedding-synchronous-api
@@ -152,11 +171,12 @@ class DashScopeEmbeddingConfig(BaseEmbeddingConfig):
                 message=message,
             )
 
+        parsed_response: Final = _DASHSCOPE_EMBEDDING_RESPONSE.validate_python(response_json)
         model_response.object = "list"
-        model_response.data = response_json.get("data", [])
-        model_response.model = response_json.get("model", model)
+        model_response.data = parsed_response.get("data", [])
+        model_response.model = parsed_response.get("model", model)
 
-        usage: Final = response_json.get("usage") or {}
+        usage: Final = parsed_response.get("usage") or {}
         prompt_tokens: Final = usage.get("prompt_tokens", 0)
         total_tokens: Final = usage.get("total_tokens", prompt_tokens)
         setattr(
@@ -169,8 +189,8 @@ class DashScopeEmbeddingConfig(BaseEmbeddingConfig):
             ),
         )
 
-        if "id" in response_json:
-            setattr(model_response, "id", response_json["id"])
+        if "id" in parsed_response:
+            setattr(model_response, "id", parsed_response["id"])
 
         return model_response
 

litellm/llms/databricks/streaming_utils.py

@@ -1,6 +1,8 @@
 import json
 from typing import Final
 
+from pydantic import ConfigDict, TypeAdapter
+
 import litellm
 from litellm import verbose_logger
 from litellm.types.llms.openai import (
@@ -23,6 +25,9 @@ def _usage_block(usage: Usage) -> ChatCompletionUsageBlock:
     )
 
 
+_JSON_CHUNK: Final = TypeAdapter(dict[str, object], config=ConfigDict(strict=True))
+
+
 class ModelResponseIterator:
     def __init__(self, streaming_response, sync_stream: bool):
         self.streaming_response = streaming_response
@@ -99,7 +104,7 @@ class ModelResponseIterator:
             self.response_iterator = self.streaming_response
         chunk: Final = self._next_sse_data()
         try:
-            json_chunk: Final = json.loads(chunk)
+            json_chunk: Final = _JSON_CHUNK.validate_python(json.loads(chunk))
             return self.chunk_parser(chunk=json_chunk)
         except ValueError as e:
             verbose_logger.debug(
@@ -126,7 +131,7 @@ class ModelResponseIterator:
         if chunk == "[DONE]":
             raise StopAsyncIteration
         try:
-            json_chunk: Final = json.loads(chunk)
+            json_chunk: Final = _JSON_CHUNK.validate_python(json.loads(chunk))
             return self.chunk_parser(chunk=json_chunk)
         except ValueError as e:
             verbose_logger.debug(

litellm/llms/fal_ai/image_edit/flux_lora_depth_transformation.py

@@ -4,13 +4,14 @@ from typing import Final
 
 from httpx._types import RequestFiles
 
+from litellm.images.utils import image_to_data_url
 from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.images.main import ImageEditOptionalRequestParams
 from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import FileTypes
 
-from .transformation import DEFAULT_BASE_URL, FalAIImageEditConfig, to_data_url
+from .transformation import DEFAULT_BASE_URL, FalAIImageEditConfig
 
 FLUX_LORA_DEPTH_ENDPOINT: Final[str] = "fal-ai/flux-lora-depth"
 SUPPORTED_OPENAI_PARAMS: Final[tuple[str, ...]] = ("n", "size")
@@ -68,7 +69,7 @@ class FalAIFluxLoraDepthEditConfig(FalAIImageEditConfig):
         )
         request_body: Final[dict[str, object]] = {  # mutable-ok: base class contract returns a dict
             "prompt": prompt,
-            "image_url": to_data_url(next(iter(images))),
+            "image_url": image_to_data_url(next(iter(images))),
             **provider_params,
         }
         return request_body, ()

litellm/llms/fal_ai/image_edit/transformation.py

@@ -1,15 +1,12 @@
-import base64
-import os
 from collections.abc import Mapping
-from pathlib import Path
 from types import MappingProxyType
-from typing import TYPE_CHECKING, Final, Protocol, runtime_checkable
+from typing import TYPE_CHECKING, Final
 
 import httpx
 from httpx._types import RequestFiles
 from pydantic import ConfigDict, TypeAdapter
 
-from litellm.images.utils import ImageEditRequestUtils
+from litellm.images.utils import image_to_data_url
 from litellm.llms.base_llm.image_edit.transformation import BaseImageEditConfig
 from litellm.llms.fal_ai.image_generation.gpt_image_2_transformation import (
     map_gpt_image_quality,
@@ -39,39 +36,6 @@ PARAM_TRANSLATION: Final[Mapping[str, str]] = MappingProxyType(
 )
 
 
-@runtime_checkable
-class _SeekableBinaryReader(Protocol):
-    def tell(self) -> int: ...
-
-    def seek(self, offset: int) -> int: ...
-
-    def read(self) -> bytes: ...
-
-
-def _read_image_bytes(image: object) -> bytes:
-    if isinstance(image, bytes):
-        return image
-    if isinstance(image, tuple):
-        return _read_image_bytes(image[1])
-    if isinstance(image, os.PathLike):
-        return Path(image).read_bytes()
-    if isinstance(image, _SeekableBinaryReader):
-        position: Final = image.tell()
-        image.seek(0)
-        data: Final = image.read()
-        image.seek(position)
-        return data
-    raise ValueError(f"Unsupported image type for Fal AI image edit: {type(image).__name__}")
-
-
-def to_data_url(image: object) -> str:
-    if isinstance(image, str):
-        return image
-    image_bytes: Final = _read_image_bytes(image)
-    mime_type: Final = ImageEditRequestUtils.get_image_content_type(image_bytes)
-    return f"data:{mime_type};base64,{base64.b64encode(image_bytes).decode('utf-8')}"
-
-
 def _first(value: object) -> object:
     return value[0] if isinstance(value, list) and value else value
 
@@ -146,14 +110,14 @@ class FalAIImageEditConfig(BaseImageEditConfig):
             raise ValueError("Fal AI image edit requires at least one input image")
         mask: Final = _first(image_edit_optional_request_params.get("mask"))
         mask_field: Final[Mapping[str, str]] = (
-            MappingProxyType({"mask_url": to_data_url(mask)}) if mask is not None else MappingProxyType({})
+            MappingProxyType({"mask_url": image_to_data_url(mask)}) if mask is not None else MappingProxyType({})
         )
         provider_params: Final[Mapping[str, object]] = MappingProxyType(
             {key: value for key, value in image_edit_optional_request_params.items() if key != "mask"}
         )
         request_body: Final[dict[str, object]] = {  # mutable-ok: base class contract returns a dict
             "prompt": prompt,
-            "image_urls": tuple(to_data_url(img) for img in images),
+            "image_urls": tuple(image_to_data_url(img) for img in images),
             **mask_field,
             **provider_params,
         }

litellm/llms/gemini/google_genai/guardrail_translation/handler.py

@@ -23,6 +23,7 @@ from litellm.llms.base_llm.guardrail_translation.base_translation import (
     BaseTranslation,
     StreamTransformSink,
 )
+from litellm.llms.custom_httpx.sse_lines import split_sse_lines
 from litellm.types.utils import GenericGuardrailAPIInputs
 
 if TYPE_CHECKING:
@@ -121,7 +122,7 @@ def _parse_json_dict_or_none(payload: str) -> Mapping[str, object] | None:
 def _sse_payload_texts(sse_text: str) -> tuple[str, ...]:
     return tuple(
         text
-        for line in sse_text.splitlines()
+        for line in split_sse_lines(sse_text)
         if line.startswith("data:")
         for payload in (line[len("data:") :].strip(),)
         if payload and payload != "[DONE]"

litellm/llms/gigachat/authenticator.py

@@ -12,6 +12,8 @@ from types import MappingProxyType
 from typing import Final
 
 import httpx
+from pydantic import ConfigDict, TypeAdapter, with_config
+from typing_extensions import NotRequired, ReadOnly, TypedDict
 
 from litellm._logging import verbose_logger
 from litellm.caching.caching import InMemoryCache
@@ -37,6 +39,17 @@ _token_cache: Final = InMemoryCache()
 _NO_LITELLM_PARAMS: Final[Mapping[str, object]] = MappingProxyType({})
 
 
+@with_config(ConfigDict(extra="allow", strict=True))
+class _GigaChatTokenResponse(TypedDict):
+    tok: NotRequired[ReadOnly[str | None]]
+    access_token: NotRequired[ReadOnly[str | None]]
+    exp: NotRequired[ReadOnly[int | float | str | None]]
+    expires_at: NotRequired[ReadOnly[int | float | str | None]]
+
+
+_TOKEN_RESPONSE: Final = TypeAdapter(_GigaChatTokenResponse)
+
+
 class GigaChatAuthError(BaseLLMException):
     """GigaChat authentication error."""
 
@@ -230,7 +243,7 @@ async def _request_token_async(
 
 def _parse_token_response(response: httpx.Response) -> tuple[str, int]:
     """Parse OAuth token response."""
-    data: Final = response.json()
+    data: Final = _TOKEN_RESPONSE.validate_python(response.json())
 
     # GigaChat returns either 'tok'/'exp' or 'access_token'/'expires_at'
     access_token: Final = data.get("tok") or data.get("access_token")

litellm/llms/huggingface/embedding/transformation.py

@@ -17,6 +17,7 @@ from litellm.litellm_core_utils.prompt_templates.factory import (
 )
 from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
 from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
+from litellm.llms.custom_httpx.sse_lines import iter_sse_lines
 from litellm.secret_managers.main import get_secret_str
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.utils import Choices, Message, ModelResponse, Usage
@@ -357,7 +358,7 @@ class HuggingFaceEmbeddingConfig(BaseConfig):
         api_key: str | None = None,
     ) -> list[dict[str, str]]:
         streamed_response: Final = CustomStreamWrapper(
-            completion_stream=response.iter_lines(),
+            completion_stream=iter_sse_lines(response.iter_text()),
             model=model,
             custom_llm_provider="huggingface",
             logging_obj=logging_obj,

litellm/llms/langgraph/chat/sse_iterator.py

@@ -12,6 +12,7 @@ import httpx
 from pydantic import TypeAdapter
 
 from litellm._logging import verbose_logger
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines
 from litellm.types.utils import Delta, ModelResponseStream, StreamingChoices
 
 _DECODED_JSON: Final = TypeAdapter(object)
@@ -36,12 +37,12 @@ class LangGraphSSEStreamIterator:
 
     def __iter__(self):
         """Initialize sync iteration."""
-        self.line_iterator = self.response.iter_lines()
+        self.line_iterator = iter_sse_lines(self.response.iter_text())
         return self
 
     def __aiter__(self):
         """Initialize async iteration."""
-        self.async_line_iterator = self.response.aiter_lines()
+        self.async_line_iterator = aiter_sse_lines(self.response.aiter_text())
         return self
 
     def _parse_sse_line(self, line: str) -> ModelResponseStream | None:

litellm/llms/oobabooga/chat/oobabooga.py

@@ -5,6 +5,7 @@ from typing import TYPE_CHECKING, Final
 import litellm
 from litellm.litellm_core_utils.url_utils import join_api_base
 from litellm.llms.custom_httpx.http_handler import get_httpx_client
+from litellm.llms.custom_httpx.sse_lines import iter_sse_lines
 from litellm.utils import EmbeddingResponse, ModelResponse, Usage
 
 from ..common_utils import OobaboogaError
@@ -74,7 +75,7 @@ def completion(
         stream=optional_params["stream"] if "stream" in optional_params else False,
     )
     if "stream" in optional_params and optional_params["stream"] is True:
-        return response.iter_lines()
+        return iter_sse_lines(response.iter_text())
     else:
         return oobabooga_config.transform_response(
             model=model,

litellm/llms/openai/realtime/handler.py

@@ -5,7 +5,7 @@ This requires websockets, and is currently only supported on LiteLLM Proxy.
 """
 
 import ssl
-from typing import Any, Final, cast
+from typing import TYPE_CHECKING, Final, cast
 
 from litellm._logging import redact_string, verbose_logger
 from litellm.constants import REALTIME_WEBSOCKET_MAX_MESSAGE_SIZE_BYTES
@@ -23,6 +23,9 @@ from ....llms.custom_httpx.http_handler import get_shared_realtime_ssl_context
 from ..common_utils import is_openai_backed_api_base
 from ..openai import OpenAIChatCompletion
 
+if TYPE_CHECKING:
+    from fastapi import WebSocket
+
 
 class OpenAIRealtime(OpenAIChatCompletion):
     """
@@ -131,7 +134,7 @@ class OpenAIRealtime(OpenAIChatCompletion):
     async def async_realtime(
         self,
         model: str,
-        websocket: Any,
+        websocket: "WebSocket",
         logging_obj: LiteLLMLogging,
         api_base: str | None = None,
         api_key: str | None = None,

litellm/llms/openai_like/chat/handler.py

@@ -16,6 +16,7 @@ from litellm import LlmProviders
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.llms.base_llm.base_model_iterator import MockResponseIterator
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler, HTTPHandler
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines
 from litellm.llms.databricks.streaming_utils import ModelResponseIterator
 from litellm.llms.openai.chat.gpt_transformation import OpenAIGPTConfig
 from litellm.llms.openai.openai import OpenAIConfig
@@ -65,7 +66,9 @@ async def make_call(
         model_response: Final = _fake_streamed_model_response(response.json())
         completion_stream = MockResponseIterator(model_response=model_response)
     else:
-        completion_stream = ModelResponseIterator(streaming_response=response.aiter_lines(), sync_stream=False)
+        completion_stream = ModelResponseIterator(
+            streaming_response=aiter_sse_lines(response.aiter_text()), sync_stream=False
+        )
     # LOGGING
     logging_obj.post_call(
         input=messages,
@@ -103,7 +106,9 @@ def make_sync_call(
         model_response: Final = _fake_streamed_model_response(response.json())
         completion_stream = MockResponseIterator(model_response=model_response)
     else:
-        completion_stream = ModelResponseIterator(streaming_response=response.iter_lines(), sync_stream=True)
+        completion_stream = ModelResponseIterator(
+            streaming_response=iter_sse_lines(response.iter_text()), sync_stream=True
+        )
 
     # LOGGING
     logging_obj.post_call(

litellm/llms/perplexity/embedding/transformation.py

@@ -12,12 +12,14 @@ This module decodes them into float arrays for OpenAI-compatible responses.
 """
 
 import base64
+import builtins
 import struct
-from collections.abc import Mapping
-from typing import Final
+from collections.abc import Mapping, Sequence
+from typing import Final, Literal
 
 import httpx
-from pydantic import ConfigDict, TypeAdapter
+from pydantic import ConfigDict, TypeAdapter, with_config
+from typing_extensions import ReadOnly, TypedDict
 
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.litellm_core_utils.url_utils import join_api_base
@@ -31,6 +33,17 @@ _JSON_OBJECT: Final = TypeAdapter(Mapping[str, object], config=ConfigDict(hide_i
 _INT: Final = TypeAdapter(int)
 
 
+@with_config(ConfigDict(extra="allow", strict=True))
+class _PerplexityEmbeddingResponse(TypedDict, total=False):
+    model: ReadOnly[str | None]
+    object: ReadOnly[Literal["list"]]
+    data: ReadOnly[Sequence[builtins.object] | Mapping[str, builtins.object] | Literal[""]]
+    usage: ReadOnly[builtins.object]
+
+
+_EMBEDDING_RESPONSE_ADAPTER: Final = TypeAdapter(_PerplexityEmbeddingResponse)
+
+
 class PerplexityEmbeddingError(BaseLLMException):
     def __init__(
         self,
@@ -153,14 +166,18 @@ class PerplexityEmbeddingConfig(BaseEmbeddingConfig):
         litellm_params: dict = {},
     ) -> EmbeddingResponse:
         try:
-            raw_response_json: Final = raw_response.json()
+            raw_response_json: Final[object] = raw_response.json()
         except Exception:
             raise PerplexityEmbeddingError(message=raw_response.text, status_code=raw_response.status_code)
 
-        model_response.model = raw_response_json.get("model", model)
-        model_response.object = raw_response_json.get("object", "list")
+        response_data: Final[_PerplexityEmbeddingResponse] = _EMBEDDING_RESPONSE_ADAPTER.validate_python(
+            raw_response_json
+        )
 
-        raw_data: Final = raw_response_json.get("data", [])
+        model_response.model = response_data.get("model", model)
+        model_response.object = response_data.get("object", "list")
+
+        raw_data: Final = response_data.get("data", [])
         decoded_data: Final[list[dict[str, object]]] = []
         for item in raw_data:
             decoded_item = dict(_JSON_OBJECT.validate_python(item))
@@ -168,7 +185,7 @@ class PerplexityEmbeddingConfig(BaseEmbeddingConfig):
             decoded_data.append(decoded_item)
         model_response.data = decoded_data
 
-        usage_data: Final = _JSON_OBJECT.validate_python(raw_response_json.get("usage", {}))
+        usage_data: Final = _JSON_OBJECT.validate_python(response_data.get("usage", {}))
         usage: Final = Usage(
             prompt_tokens=_INT.validate_python(
                 usage_data.get("prompt_tokens", 0) or usage_data.get("total_tokens", 0) or 0

litellm/llms/predibase/chat/handler.py

@@ -14,6 +14,7 @@ from litellm.llms.custom_httpx.http_handler import (
     AsyncHTTPHandler,
     get_async_httpx_client,
 )
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines
 from litellm.utils import CustomStreamWrapper, ModelResponse
 
 from ..common_utils import PredibaseError
@@ -34,7 +35,7 @@ async def make_call(
     if response.status_code != 200:
         raise PredibaseError(status_code=response.status_code, message=response.text)
 
-    completion_stream: Final = response.aiter_lines()
+    completion_stream: Final = aiter_sse_lines(response.aiter_text())
     # LOGGING
     logging_obj.post_call(
         input=messages,
@@ -163,7 +164,7 @@ class PredibaseChatCompletion:
                 timeout=timeout,
             )
             _response: Final = CustomStreamWrapper(
-                response.iter_lines(),
+                iter_sse_lines(response.iter_text()),
                 model,
                 custom_llm_provider="predibase",
                 logging_obj=logging_obj,

litellm/llms/tencent/cost_calculator.py

@@ -1,6 +1,49 @@
-from litellm.litellm_core_utils.llm_cost_calc.utils import generic_cost_per_token
-from litellm.types.utils import Usage
+from typing import Final
+
+from litellm.litellm_core_utils.llm_cost_calc.utils import (
+    calculate_image_response_cost_from_usage,
+    flat_image_cost,
+    generic_cost_per_token,
+    resolve_image_model_info,
+)
+from litellm.types.utils import ImageResponse, ModelInfo, Usage
 
 
 def cost_per_token(model: str, usage: Usage) -> tuple[float, float]:
     return generic_cost_per_token(model=model, usage=usage, custom_llm_provider="tencent")
+
+
+def _deployment_image_token_pricing(model_info: ModelInfo | None) -> ModelInfo | None:
+    if model_info is None or model_info.get("output_cost_per_image_token") is not None:
+        return model_info
+    output_cost_per_token: Final = model_info.get("output_cost_per_token")
+    if output_cost_per_token is None:
+        return model_info
+    priced: Final[ModelInfo] = {**model_info, "output_cost_per_image_token": output_cost_per_token}
+    return priced
+
+
+def _deployment_prices_per_image_only(model_info: ModelInfo | None) -> bool:
+    return (
+        model_info is not None
+        and model_info.get("output_cost_per_image") is not None
+        and model_info.get("output_cost_per_image_token") is None
+        and model_info.get("output_cost_per_token") is None
+    )
+
+
+def image_cost_calculator(model: str, image_response: ImageResponse, model_info: ModelInfo | None = None) -> float:
+    resolved_model_info: Final = resolve_image_model_info(
+        model=model, custom_llm_provider="tencent", model_info=_deployment_image_token_pricing(model_info)
+    )
+    if _deployment_prices_per_image_only(model_info):
+        return flat_image_cost(resolved_model_info, image_response)
+    token_based_cost: Final = calculate_image_response_cost_from_usage(
+        model=model,
+        image_response=image_response,
+        custom_llm_provider="tencent",
+        model_info=resolved_model_info,
+    )
+    if token_based_cost is not None:
+        return token_based_cost
+    return flat_image_cost(resolved_model_info, image_response)

litellm/llms/tencent/image_edit/transformation.py

@@ -0,0 +1,147 @@
+from collections.abc import Mapping
+from types import MappingProxyType
+from typing import TYPE_CHECKING, Final, TypeAlias
+
+import httpx
+from httpx._types import RequestFiles
+from pydantic import TypeAdapter
+from typing_extensions import override
+
+from litellm.images.utils import image_to_data_url
+from litellm.llms.base_llm.chat.transformation import BaseLLMException
+from litellm.llms.base_llm.image_edit.transformation import BaseImageEditConfig
+from litellm.llms.tencent.image_generation.transformation import (
+    TOKENHUB_IMAGE_PARAMS,
+    tokenhub_image_headers,
+    tokenhub_image_params,
+    tokenhub_image_response,
+    tokenhub_image_url,
+)
+from litellm.types.images.main import ImageEditOptionalRequestParams
+from litellm.types.router import GenericLiteLLMParams
+from litellm.types.utils import ImageResponse
+
+if TYPE_CHECKING:
+    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+
+_IMAGES: Final = TypeAdapter(tuple[object, ...])
+_FORM_BOOLEANS: Final = MappingProxyType({"true": True, "false": False})
+_OPENAI_PARAMS: Final = frozenset(TOKENHUB_IMAGE_PARAMS)
+
+_FormEntries: TypeAlias = tuple[tuple[tuple[str, ...], object], ...]
+
+
+def _form_scalar(value: object) -> object:
+    match value:
+        case str() if value.lower() in _FORM_BOOLEANS:
+            return _FORM_BOOLEANS[value.lower()]
+        case str() if value.removeprefix("-").isdigit():
+            return int(value)
+        case _:
+            return value
+
+
+def _form_path(key: str) -> tuple[str, ...]:
+    root, bracket, rest = key.partition("[")
+    if not bracket or not rest.endswith("]"):
+        return (key,)
+    return (root, *rest.removesuffix("]").split("]["))
+
+
+def _nested_form_value(entries: _FormEntries) -> object:
+    match entries:
+        case (((), value),):
+            return _form_scalar(value)
+        case _:
+            return _nested_form_fields(tuple(entry for entry in entries if entry[0]))
+
+
+def _nested_form_fields(entries: _FormEntries) -> Mapping[str, object]:
+    roots: Final = tuple(dict.fromkeys(path[0] for path, _ in entries))
+    return {
+        root: _nested_form_value(tuple((path[1:], value) for path, value in entries if path[0] == root))
+        for root in roots
+    }
+
+
+def _tokenhub_form_params(params: Mapping[str, object]) -> Mapping[str, object]:
+    tokenhub_entries: Final = tuple(
+        (_form_path(key), value) for key, value in params.items() if key not in _OPENAI_PARAMS
+    )
+    return {
+        **{key: value for key, value in params.items() if key in _OPENAI_PARAMS},
+        **_nested_form_fields(tokenhub_entries),
+    }
+
+
+class TencentImageEditConfig(BaseImageEditConfig):
+    @override
+    def get_supported_openai_params(self, model: str) -> list[str]:  # mutable-ok: mirrors override contract
+        return list(TOKENHUB_IMAGE_PARAMS)
+
+    @override
+    def map_openai_params(
+        self,
+        image_edit_optional_params: ImageEditOptionalRequestParams,
+        model: str,
+        drop_params: bool,
+    ) -> dict[str, object]:  # mutable-ok: mirrors override contract
+        return dict(tokenhub_image_params(image_edit_optional_params, model, drop_params))
+
+    @override
+    def validate_environment(
+        self,
+        headers: Mapping[str, str],
+        model: str,
+        api_key: str | None = None,
+        litellm_params: Mapping[str, object] | None = None,
+        api_base: str | None = None,
+    ) -> dict[str, str]:  # mutable-ok: mirrors override contract
+        return dict(tokenhub_image_headers(headers, api_key, model))
+
+    @override
+    def get_complete_url(
+        self,
+        model: str,
+        api_base: str | None,
+        litellm_params: Mapping[str, object],
+    ) -> str:
+        return tokenhub_image_url(api_base)
+
+    @override
+    def use_multipart_form_data(self) -> bool:
+        return False
+
+    @override
+    def transform_image_edit_request(
+        self,
+        model: str,
+        prompt: str | None,
+        image: object,
+        image_edit_optional_request_params: Mapping[str, object],
+        litellm_params: GenericLiteLLMParams,
+        headers: Mapping[str, str],
+    ) -> tuple[dict[str, object], RequestFiles]:  # mutable-ok: mirrors override contract
+        images: Final = _IMAGES.validate_python(image) if isinstance(image, list) else (image,)
+        return {
+            "model": model,
+            "prompt": prompt,
+            "images": tuple(image_to_data_url(img) for img in images if img is not None),
+            **_tokenhub_form_params(image_edit_optional_request_params),
+        }, ()
+
+    @override
+    def transform_image_edit_response(
+        self,
+        model: str,
+        raw_response: httpx.Response,
+        logging_obj: "LiteLLMLoggingObj",
+    ) -> ImageResponse:
+        image_response: Final = tokenhub_image_response(raw_response)
+        if image_response is None:
+            raise BaseLLMException(
+                status_code=raw_response.status_code,
+                message=raw_response.text,
+                headers=raw_response.headers,
+            )
+        return image_response

litellm/llms/tencent/image_generation/transformation.py

@@ -0,0 +1,200 @@
+from collections.abc import Mapping, Sequence
+from typing import TYPE_CHECKING, Final
+
+import httpx
+from pydantic import BaseModel, ConfigDict, TypeAdapter, ValidationError
+from typing_extensions import override
+
+from litellm.exceptions import AuthenticationError, UnsupportedParamsError
+from litellm.litellm_core_utils.url_utils import ensure_api_route, join_api_base
+from litellm.llms.base_llm.chat.transformation import BaseLLMException
+from litellm.llms.base_llm.image_generation.transformation import (
+    BaseImageGenerationConfig,
+)
+from litellm.secret_managers.main import get_secret_str
+from litellm.types.llms.openai import (
+    AllMessageValues,
+    OpenAIImageGenerationOptionalParams,
+)
+from litellm.types.utils import (
+    ImageObject,
+    ImageResponse,
+    ImageUsage,
+    ImageUsageInputTokensDetails,
+    LlmProviders,
+)
+
+if TYPE_CHECKING:
+    from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+
+TOKENHUB_API_BASE: Final = "https://tokenhub-intl.tencentcloudmaas.com/v1"
+TOKENHUB_IMAGE_PATH: Final = "/wand/si-image/generation"
+TOKENHUB_IMAGE_MODEL_PREFIX: Final = "seedream-"
+TOKENHUB_IMAGE_PARAMS: Final[tuple[OpenAIImageGenerationOptionalParams, ...]] = (
+    "n",
+    "size",
+    "response_format",
+    "output_format",
+    "background",
+    "user",
+)
+
+_JSON_OBJECT: Final = TypeAdapter(Mapping[str, object], config=ConfigDict(hide_input_in_errors=True))
+
+
+class _TokenHubImage(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    url: str | None = None
+    b64_json: str | None = None
+    size: str | None = None
+    output_format: str | None = None
+
+
+class _TokenHubUsage(BaseModel):
+    model_config = ConfigDict(frozen=True)
+
+    total_tokens: int
+
+
+class _TokenHubImageResponse(BaseModel):
+    model_config = ConfigDict(frozen=True, hide_input_in_errors=True)
+
+    created: int
+    data: tuple[_TokenHubImage, ...]
+    tokenhub_usage: _TokenHubUsage | None = None
+
+
+def is_tokenhub_image_model(model: str) -> bool:
+    return model.removeprefix(f"{LlmProviders.TENCENT.value}/").startswith(TOKENHUB_IMAGE_MODEL_PREFIX)
+
+
+def tokenhub_image_url(api_base: str | None) -> str:
+    base: Final = api_base or get_secret_str("TENCENT_API_BASE") or TOKENHUB_API_BASE
+    return join_api_base(ensure_api_route(base, "v1"), TOKENHUB_IMAGE_PATH)
+
+
+def tokenhub_image_headers(headers: Mapping[str, str], api_key: str | None, model: str) -> Mapping[str, str]:
+    resolved_api_key: Final = api_key or get_secret_str("TENCENT_API_KEY")
+    if not resolved_api_key:
+        raise AuthenticationError(
+            message="TENCENT_API_KEY is not set. Set it in the environment or pass api_key.",
+            llm_provider="tencent",
+            model=model,
+        )
+    return {**headers, "Authorization": f"Bearer {resolved_api_key}", "Content-Type": "application/json"}
+
+
+def tokenhub_image_params(params: Mapping[str, object], model: str, drop_params: bool) -> Mapping[str, object]:
+    n: Final = params.get("n")
+    if n is not None and n != 1 and not drop_params:
+        raise UnsupportedParamsError(
+            status_code=400,
+            message=f"Tencent TokenHub returns one image per request, so n={n} is not supported for {model}. "
+            + "To send n=1 instead, set `litellm.drop_params = True`.",
+        )
+    return {
+        key: value
+        for key, value in params.items()
+        if key != "n" and value is not None and not (key == "background" and value == "auto")
+    }
+
+
+def tokenhub_image_response(raw_response: httpx.Response) -> ImageResponse | None:
+    try:
+        parsed: Final = _TokenHubImageResponse.model_validate_json(raw_response.content)
+    except ValidationError:
+        return None
+    total_tokens: Final = parsed.tokenhub_usage.total_tokens if parsed.tokenhub_usage is not None else 0
+    first_image: Final = parsed.data[0] if parsed.data else _TokenHubImage()
+    return ImageResponse(
+        created=parsed.created,
+        data=[ImageObject(url=image.url, b64_json=image.b64_json) for image in parsed.data],
+        size=first_image.size,
+        output_format=first_image.output_format,
+        usage=ImageUsage(
+            input_tokens=0,
+            input_tokens_details=ImageUsageInputTokensDetails(image_tokens=0, text_tokens=0),
+            output_tokens=total_tokens,
+            total_tokens=total_tokens,
+        ),
+    )
+
+
+class TencentImageGenerationConfig(BaseImageGenerationConfig):
+    @override
+    def get_supported_openai_params(
+        self, model: str
+    ) -> list[OpenAIImageGenerationOptionalParams]:  # mutable-ok: mirrors override contract
+        return list(TOKENHUB_IMAGE_PARAMS)
+
+    @override
+    def map_openai_params(
+        self,
+        non_default_params: Mapping[str, object],
+        optional_params: Mapping[str, object],
+        model: str,
+        drop_params: bool,
+    ) -> dict[str, object]:  # mutable-ok: mirrors override contract
+        return {**optional_params, **tokenhub_image_params(non_default_params, model, drop_params)}
+
+    @override
+    def get_complete_url(
+        self,
+        api_base: str | None,
+        api_key: str | None,
+        model: str,
+        optional_params: Mapping[str, object],
+        litellm_params: Mapping[str, object],
+        stream: bool | None = None,
+    ) -> str:
+        return tokenhub_image_url(api_base)
+
+    @override
+    def validate_environment(
+        self,
+        headers: Mapping[str, str],
+        model: str,
+        messages: Sequence[AllMessageValues],
+        optional_params: Mapping[str, object],
+        litellm_params: Mapping[str, object],
+        api_key: str | None = None,
+        api_base: str | None = None,
+    ) -> dict[str, str]:  # mutable-ok: mirrors override contract
+        return dict(tokenhub_image_headers(headers, api_key, model))
+
+    @override
+    def transform_image_generation_request(
+        self,
+        model: str,
+        prompt: str,
+        optional_params: Mapping[str, object],
+        litellm_params: Mapping[str, object],
+        headers: Mapping[str, str],
+    ) -> dict[str, object]:  # mutable-ok: mirrors override contract
+        extra_body: Final = _JSON_OBJECT.validate_python(optional_params.get("extra_body") or {})
+        params: Final = {key: value for key, value in optional_params.items() if key != "extra_body"}
+        return {"model": model, "prompt": prompt, **params, **extra_body}
+
+    @override
+    def transform_image_generation_response(
+        self,
+        model: str,
+        raw_response: httpx.Response,
+        model_response: ImageResponse,
+        logging_obj: "LiteLLMLoggingObj",
+        request_data: Mapping[str, object],
+        optional_params: Mapping[str, object],
+        litellm_params: Mapping[str, object],
+        encoding: object,
+        api_key: str | None = None,
+        json_mode: bool | None = None,
+    ) -> ImageResponse:
+        image_response: Final = tokenhub_image_response(raw_response)
+        if image_response is None:
+            raise BaseLLMException(
+                status_code=raw_response.status_code,
+                message=raw_response.text,
+                headers=raw_response.headers,
+            )
+        return image_response

litellm/llms/vertex_ai/agent_engine/transformation.py

@@ -22,6 +22,7 @@ from litellm.litellm_core_utils.prompt_templates.common_utils import (
     convert_content_list_to_str,
 )
 from litellm.llms.base_llm.chat.transformation import BaseConfig, BaseLLMException
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines
 from litellm.llms.vertex_ai.agent_engine.sse_iterator import (
     VertexAgentEngineResponseIterator,
 )
@@ -357,7 +358,7 @@ class VertexAgentEngineConfig(BaseConfig, VertexBase):
     ) -> VertexAgentEngineResponseIterator:
         """Return a streaming iterator for SSE responses."""
         return VertexAgentEngineResponseIterator(
-            streaming_response=raw_response.iter_lines(),
+            streaming_response=iter_sse_lines(raw_response.iter_text()),
             sync_stream=True,
         )
 
@@ -465,7 +466,7 @@ class VertexAgentEngineConfig(BaseConfig, VertexBase):
 
         # Create iterator for SSE stream (async)
         completion_stream: Final = VertexAgentEngineResponseIterator(
-            streaming_response=response.aiter_lines(),
+            streaming_response=aiter_sse_lines(response.aiter_text()),
             sync_stream=False,
         )
 

litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py

@@ -41,6 +41,7 @@ from litellm.llms.custom_httpx.http_handler import (
     get_async_httpx_client,
     get_httpx_client,
 )
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines
 from litellm.types.llms.anthropic import AnthropicThinkingParam
 from litellm.types.llms.base import CachedTokensDetails
 from litellm.types.llms.gemini import BidiGenerateContentServerMessage
@@ -2743,10 +2744,10 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
                 chunk_indices = support.get("groundingChunkIndices", [])
 
                 if start_index is not None and end_index is not None and chunk_indices:
-                    # Use the first chunk's URL for the annotation
-                    first_chunk_idx = chunk_indices[0]
-                    if first_chunk_idx in chunk_to_uri_map:
-                        uri_info = chunk_to_uri_map[first_chunk_idx]
+                    for chunk_index in chunk_indices:
+                        if chunk_index not in chunk_to_uri_map:
+                            continue
+                        uri_info = chunk_to_uri_map[chunk_index]
 
                         url_citation: ChatCompletionAnnotationURLCitation = {
                             "start_index": start_index,
@@ -3241,7 +3242,7 @@ async def make_call(
             raise
 
     completion_stream: Final = ModelResponseIterator(
-        streaming_response=response.aiter_lines(),
+        streaming_response=aiter_sse_lines(response.aiter_text()),
         sync_stream=False,
         logging_obj=logging_obj,
         response_headers=response.headers,
@@ -3312,7 +3313,7 @@ def make_sync_call(
             raise
 
     completion_stream: Final = ModelResponseIterator(
-        streaming_response=response.iter_lines(),
+        streaming_response=iter_sse_lines(response.iter_text()),
         sync_stream=True,
         logging_obj=logging_obj,
         response_headers=response.headers,
@@ -4190,7 +4191,7 @@ class ModelResponseIterator:
             log_failed_hop(e)
             return self._truncated_stream_end(token)
         self._start_hop(hop_response, token)
-        self.streaming_response = hop_response.aiter_lines()
+        self.streaming_response = aiter_sse_lines(hop_response.aiter_text())
         self.async_response_iterator = self.streaming_response
         return None
 
@@ -4204,7 +4205,7 @@ class ModelResponseIterator:
             log_failed_hop(e)
             return self._truncated_stream_end(token)
         self._start_hop(hop_response, token)
-        self.streaming_response = hop_response.iter_lines()
+        self.streaming_response = iter_sse_lines(hop_response.iter_text())
         self.response_iterator = self.streaming_response
         return None
 

litellm/llms/vertex_ai/vertex_gemma_models/transformation.py

@@ -9,10 +9,11 @@ The actual message transformation reuses OpenAIGPTConfig since Gemma uses OpenAI
 """
 
 from collections.abc import Callable
-from typing import TYPE_CHECKING, Any, Final, cast
+from typing import TYPE_CHECKING, Final, cast
 
 import httpx
-from pydantic import ValidationError
+from pydantic import ConfigDict, TypeAdapter, ValidationError
+from typing_extensions import NotRequired, ReadOnly, TypeAliasType, TypedDict
 
 from litellm.llms.base_llm.chat.transformation import BaseLLMException
 from litellm.llms.custom_httpx.http_handler import (
@@ -32,6 +33,17 @@ if TYPE_CHECKING:
     from litellm.litellm_core_utils.tokenizer import Encoding as Tokenizer
 
 
+class _PredictionsEnvelope(TypedDict):
+    predictions: NotRequired[ReadOnly[dict[str, object]]]
+
+
+_PredictionsEnvelopeType = TypeAliasType("_PredictionsEnvelopeType", _PredictionsEnvelope)
+_PREDICTIONS_ENVELOPE: Final[TypeAdapter[_PredictionsEnvelope]] = TypeAdapter(
+    _PredictionsEnvelopeType,
+    config=ConfigDict(strict=True, extra="allow"),
+)
+
+
 def parse_vertex_gemma_container_error(predictions: object) -> VertexGemmaContainerError | None:
     try:
         return VertexGemmaContainerError.model_validate(predictions)
@@ -142,7 +154,7 @@ class VertexGemmaConfig(OpenAIGPTConfig):
 
     def _unwrap_predictions_response(
         self,
-        response_json: dict[str, Any],
+        response_json: _PredictionsEnvelope,
     ) -> dict[str, object]:
         """
         Unwrap the Vertex Gemma predictions format to OpenAI format.
@@ -354,7 +366,7 @@ class VertexGemmaConfig(OpenAIGPTConfig):
                 message=f"Request failed: {response.text}",
             )
 
-        response_json: Final = response.json()
+        response_json: Final = _PREDICTIONS_ENVELOPE.validate_python(response.json())
 
         # Unwrap predictions to get OpenAI-compatible response
         openai_response: Final = self._unwrap_predictions_response(response_json)
@@ -449,7 +461,7 @@ class VertexGemmaConfig(OpenAIGPTConfig):
                 message=f"Request failed: {response.text}",
             )
 
-        response_json: Final = response.json()
+        response_json: Final = _PREDICTIONS_ENVELOPE.validate_python(response.json())
 
         # Unwrap predictions to get OpenAI-compatible response
         openai_response: Final = self._unwrap_predictions_response(response_json)

litellm/llms/volcengine/responses/transformation.py

@@ -3,6 +3,7 @@ from types import UnionType
 from typing import TYPE_CHECKING, Any, Final, Literal, Protocol, Union, get_args, get_origin
 
 import httpx
+from pydantic import ConfigDict, TypeAdapter
 from pydantic import fields as pyd_fields
 
 import litellm
@@ -37,6 +38,8 @@ if TYPE_CHECKING:
 else:
     LiteLLMLoggingObj = Any
 
+_RESPONSE_BODY: Final = TypeAdapter(dict[str, object], config=ConfigDict(strict=True))
+
 
 class _EventModelClass(Protocol):
     @property
@@ -420,7 +423,7 @@ class VolcEngineResponsesAPIConfig(OpenAIResponsesAPIConfig):
 
     @staticmethod
     def _parsed_response_body(raw_response: httpx.Response) -> dict[str, object]:
-        return raw_response.json()
+        return _RESPONSE_BODY.validate_python(raw_response.json())
 
     @staticmethod
     def _annotation_origin(annotation: object) -> object:

litellm/model_prices_and_context_window_backup.json

@@ -21589,7 +21589,13 @@
         ],
         "supports_vision": true,
         "supports_audio_input": true,
-        "supports_video_input": true
+        "supports_video_input": true,
+        "input_cost_per_token_priority": 5.6252e-07,
+        "output_cost_per_token_priority": 3.37498e-06,
+        "cache_read_input_token_cost_priority": 5.628e-08,
+        "cache_creation_input_token_cost_priority": 5.6252e-07,
+        "input_cost_per_audio_token_priority": 1.12497e-06,
+        "cache_read_input_audio_token_cost_priority": 1.1249e-07
     },
     "databricks/databricks-gemini-3-1-flash-image": {
         "cache_creation_input_token_cost": 6.2503e-07,
@@ -21676,7 +21682,14 @@
         "supports_reasoning": true,
         "supports_tool_choice": true,
         "supports_video_input": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 4.50002e-06,
+        "output_cost_per_token_priority": 2.699998e-05,
+        "cache_read_input_token_cost_priority": 4.5003e-07,
+        "cache_creation_input_token_cost_priority": 4.50002e-06,
+        "input_cost_per_token_above_200k_tokens_priority": 8.99997e-06,
+        "output_cost_per_token_above_200k_tokens_priority": 4.049997e-05,
+        "cache_read_input_token_cost_above_200k_tokens_priority": 8.9999e-07
     },
     "databricks/databricks-gemini-3-flash": {
         "cache_creation_input_token_cost": 6.2503e-07,
@@ -21710,7 +21723,13 @@
         ],
         "supports_vision": true,
         "supports_audio_input": true,
-        "supports_video_input": true
+        "supports_video_input": true,
+        "input_cost_per_token_priority": 1.12497e-06,
+        "output_cost_per_token_priority": 6.75003e-06,
+        "cache_read_input_token_cost_priority": 1.1249e-07,
+        "cache_creation_input_token_cost_priority": 1.12497e-06,
+        "input_cost_per_audio_token_priority": 2.25001e-06,
+        "cache_read_input_audio_token_cost_priority": 2.2498e-07
     },
     "databricks/databricks-gemini-3-pro": {
         "cache_creation_input_token_cost": 2.49998e-06,
@@ -21757,7 +21776,11 @@
         "supports_reasoning": true,
         "supports_sampling_params": false,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 1.35002e-06,
+        "output_cost_per_token_priority": 6.75003e-06,
+        "cache_read_input_token_cost_priority": 1.3503e-07,
+        "cache_creation_input_token_cost_priority": 1.35002e-06
     },
     "databricks/databricks-gemini-3-7-flash": {
         "cache_creation_input_token_cost": 7.4998e-07,
@@ -21785,10 +21808,14 @@
         "supports_reasoning": true,
         "supports_sampling_params": false,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 1.35002e-06,
+        "output_cost_per_token_priority": 6.75003e-06,
+        "cache_read_input_token_cost_priority": 1.3503e-07,
+        "cache_creation_input_token_cost_priority": 1.35002e-06
     },
     "databricks/databricks-gemini-3-6-flash": {
-        "cache_creation_input_token_cost": 1.87502e-06,
+        "cache_creation_input_token_cost": 7.4998e-07,
         "cache_read_input_token_cost": 7.497e-08,
         "input_cost_per_token": 7.4998e-07,
         "input_dbu_cost_per_token": 1.0714e-05,
@@ -21816,7 +21843,11 @@
         "supports_reasoning": true,
         "supports_sampling_params": false,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 1.35002e-06,
+        "output_cost_per_token_priority": 6.75003e-06,
+        "cache_read_input_token_cost_priority": 1.3503e-07,
+        "cache_creation_input_token_cost_priority": 1.35002e-06
     },
     "databricks/databricks-gemini-3-5-flash": {
         "cache_creation_input_token_cost": 1.87502e-06,
@@ -21846,7 +21877,11 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 3.37498e-06,
+        "output_cost_per_token_priority": 2.025002e-05,
+        "cache_read_input_token_cost_priority": 3.3747e-07,
+        "cache_creation_input_token_cost_priority": 3.37498e-06
     },
     "databricks/databricks-gemini-3-5-flash-lite": {
         "cache_creation_input_token_cost": 3.7499e-07,
@@ -21878,7 +21913,11 @@
         "supports_response_schema": true,
         "supports_sampling_params": false,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 6.7501e-07,
+        "output_cost_per_token_priority": 5.62499e-06,
+        "cache_read_input_token_cost_priority": 6.748e-08,
+        "cache_creation_input_token_cost_priority": 6.7501e-07
     },
     "databricks/databricks-gemma-3-12b": {
         "cache_creation_input_token_cost": 1.5001e-07,
@@ -21921,7 +21960,11 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": false
+        "supports_vision": false,
+        "input_cost_per_token_priority": 2.45e-06,
+        "output_cost_per_token_priority": 7.7e-06,
+        "cache_read_input_token_cost_priority": 4.55e-07,
+        "cache_creation_input_token_cost_priority": 2.45e-06
     },
     "databricks/databricks-glm-5-3": {
         "cache_creation_input_token_cost": 1.4e-06,
@@ -21951,7 +21994,11 @@
         "supports_response_schema": true,
         "supports_tool_choice": true,
         "supports_vision": false,
-        "thinking_always_on": true
+        "thinking_always_on": true,
+        "input_cost_per_token_priority": 2.45e-06,
+        "output_cost_per_token_priority": 7.7e-06,
+        "cache_read_input_token_cost_priority": 4.55e-07,
+        "cache_creation_input_token_cost_priority": 2.45e-06
     },
     "databricks/databricks-glm-5-3-flash": {
         "cache_creation_input_token_cost": 1.5001e-07,
@@ -21980,7 +22027,11 @@
         "supports_reasoning": true,
         "supports_tool_choice": true,
         "supports_vision": true,
-        "thinking_always_on": true
+        "thinking_always_on": true,
+        "input_cost_per_token_priority": 2.625e-07,
+        "output_cost_per_token_priority": 8.75e-07,
+        "cache_read_input_token_cost_priority": 5.25e-08,
+        "cache_creation_input_token_cost_priority": 2.625e-07
     },
     "databricks/databricks-gpt-5": {
         "cache_creation_input_token_cost": 1.24999e-06,
@@ -21999,7 +22050,11 @@
         "source": "https://www.databricks.com/product/pricing/proprietary-foundation-model-serving",
         "supports_function_calling": true,
         "supports_prompt_caching": true,
-        "supports_tool_choice": true
+        "supports_tool_choice": true,
+        "input_cost_per_token_priority": 2.49998e-06,
+        "output_cost_per_token_priority": 1.999998e-05,
+        "cache_read_input_token_cost_priority": 2.4997e-07,
+        "cache_creation_input_token_cost_priority": 2.49998e-06
     },
     "databricks/databricks-gpt-5-1": {
         "cache_creation_input_token_cost": 1.24999e-06,
@@ -22018,7 +22073,11 @@
         "source": "https://www.databricks.com/product/pricing/proprietary-foundation-model-serving",
         "supports_function_calling": true,
         "supports_prompt_caching": true,
-        "supports_tool_choice": true
+        "supports_tool_choice": true,
+        "input_cost_per_token_priority": 2.49998e-06,
+        "output_cost_per_token_priority": 1.999998e-05,
+        "cache_read_input_token_cost_priority": 2.4997e-07,
+        "cache_creation_input_token_cost_priority": 2.49998e-06
     },
     "databricks/databricks-gpt-5-2": {
         "cache_creation_input_token_cost": 1.75e-06,
@@ -22037,7 +22096,11 @@
         "source": "https://www.databricks.com/product/pricing/proprietary-foundation-model-serving",
         "supports_function_calling": true,
         "supports_prompt_caching": true,
-        "supports_tool_choice": true
+        "supports_tool_choice": true,
+        "input_cost_per_token_priority": 3.5e-06,
+        "output_cost_per_token_priority": 2.8e-05,
+        "cache_read_input_token_cost_priority": 3.5e-07,
+        "cache_creation_input_token_cost_priority": 3.5e-06
     },
     "databricks/databricks-gpt-5-3-codex": {
         "cache_creation_input_token_cost": 1.75e-06,
@@ -22056,7 +22119,11 @@
         "source": "https://www.databricks.com/product/pricing/proprietary-foundation-model-serving",
         "supports_function_calling": true,
         "supports_prompt_caching": true,
-        "supports_tool_choice": true
+        "supports_tool_choice": true,
+        "input_cost_per_token_priority": 3.5e-06,
+        "output_cost_per_token_priority": 2.8e-05,
+        "cache_read_input_token_cost_priority": 3.5e-07,
+        "cache_creation_input_token_cost_priority": 3.5e-06
     },
     "databricks/databricks-gpt-5-4": {
         "cache_creation_input_token_cost": 2.49998e-06,
@@ -22087,7 +22154,14 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 5.00003e-06,
+        "output_cost_per_token_priority": 2.999997e-05,
+        "cache_read_input_token_cost_priority": 5.0001e-07,
+        "cache_creation_input_token_cost_priority": 5.00003e-06,
+        "input_cost_per_token_above_272k_tokens_priority": 5.00003e-06,
+        "output_cost_per_token_above_272k_tokens_priority": 2.999997e-05,
+        "cache_read_input_token_cost_above_272k_tokens_priority": 5.0001e-07
     },
     "databricks/databricks-gpt-5-4-mini": {
         "cache_creation_input_token_cost": 7.4998e-07,
@@ -22115,7 +22189,11 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 1.50003e-06,
+        "output_cost_per_token_priority": 8.99997e-06,
+        "cache_read_input_token_cost_priority": 1.5001e-07,
+        "cache_creation_input_token_cost_priority": 1.50003e-06
     },
     "databricks/databricks-gpt-5-4-nano": {
         "cache_creation_input_token_cost": 1.9999e-07,
@@ -22175,7 +22253,15 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 8.00002e-06,
+        "output_cost_per_token_priority": 4.000003e-05,
+        "cache_read_input_token_cost_priority": 8.0003e-07,
+        "cache_creation_input_token_cost_priority": 9.99999e-06,
+        "input_cost_per_token_above_272k_tokens_priority": 1.599997e-05,
+        "output_cost_per_token_above_272k_tokens_priority": 6.000001e-05,
+        "cache_read_input_token_cost_above_272k_tokens_priority": 1.59999e-06,
+        "cache_creation_input_token_cost_above_272k_tokens_priority": 1.999998e-05
     },
     "databricks/databricks-gpt-5-6-terra": {
         "cache_creation_input_token_cost": 2.49998e-06,
@@ -22207,7 +22293,15 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 4.00001e-06,
+        "output_cost_per_token_priority": 2.399999e-05,
+        "cache_read_input_token_cost_priority": 3.9998e-07,
+        "cache_creation_input_token_cost_priority": 5.00003e-06,
+        "input_cost_per_token_above_272k_tokens_priority": 8.00002e-06,
+        "output_cost_per_token_above_272k_tokens_priority": 3.600002e-05,
+        "cache_read_input_token_cost_above_272k_tokens_priority": 8.0003e-07,
+        "cache_creation_input_token_cost_above_272k_tokens_priority": 9.99999e-06
     },
     "databricks/databricks-gpt-5-6-luna": {
         "cache_creation_input_token_cost": 2.4997e-07,
@@ -22239,7 +22333,15 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 3.9998e-07,
+        "output_cost_per_token_priority": 2.40002e-06,
+        "cache_read_input_token_cost_priority": 3.997e-08,
+        "cache_creation_input_token_cost_priority": 5.0001e-07,
+        "input_cost_per_token_above_272k_tokens_priority": 8.0003e-07,
+        "output_cost_per_token_above_272k_tokens_priority": 3.60003e-06,
+        "cache_read_input_token_cost_above_272k_tokens_priority": 8.001e-08,
+        "cache_creation_input_token_cost_above_272k_tokens_priority": 1.00002e-06
     },
     "databricks/databricks-gpt-5-5": {
         "cache_creation_input_token_cost": 5.00003e-06,
@@ -22270,7 +22372,14 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 1.249997e-05,
+        "output_cost_per_token_priority": 7.50001e-05,
+        "cache_read_input_token_cost_priority": 1.24999e-06,
+        "cache_creation_input_token_cost_priority": 1.249997e-05,
+        "input_cost_per_token_above_272k_tokens_priority": 1.249997e-05,
+        "output_cost_per_token_above_272k_tokens_priority": 7.50001e-05,
+        "cache_read_input_token_cost_above_272k_tokens_priority": 1.24999e-06
     },
     "databricks/databricks-gpt-5-5-pro": {
         "cache_creation_input_token_cost": 2.999997e-05,
@@ -22319,7 +22428,11 @@
         "source": "https://www.databricks.com/product/pricing/proprietary-foundation-model-serving",
         "supports_function_calling": true,
         "supports_prompt_caching": true,
-        "supports_tool_choice": true
+        "supports_tool_choice": true,
+        "input_cost_per_token_priority": 4.5003e-07,
+        "output_cost_per_token_priority": 3.60003e-06,
+        "cache_read_input_token_cost_priority": 4.501e-08,
+        "cache_creation_input_token_cost_priority": 4.5003e-07
     },
     "databricks/databricks-gpt-5-nano": {
         "cache_creation_input_token_cost": 4.998e-08,
@@ -22469,7 +22582,11 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 5.25e-06,
+        "output_cost_per_token_priority": 2.625e-05,
+        "cache_read_input_token_cost_priority": 5.25e-07,
+        "cache_creation_input_token_cost_priority": 5.25e-06
     },
     "databricks/system.ai.kimi-k3": {
         "litellm_provider": "databricks",
@@ -22551,7 +22668,11 @@
         "supports_reasoning": true,
         "supports_tool_choice": true,
         "supports_vision": false,
-        "thinking_always_on": true
+        "thinking_always_on": true,
+        "input_cost_per_token_priority": 3.85e-07,
+        "output_cost_per_token_priority": 3.85e-06,
+        "cache_read_input_token_cost_priority": 3.85e-07,
+        "cache_creation_input_token_cost_priority": 3.85e-07
     },
     "databricks/databricks-qwen3-next-80b-a3b-instruct": {
         "cache_creation_input_token_cost": 1.5001e-07,
@@ -28732,6 +28853,7 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.039,
         "output_cost_per_image_token": 3e-05,
+        "output_cost_per_image_token_batches": 1.5e-05,
         "output_cost_per_reasoning_token": 2.5e-06,
         "output_cost_per_token": 2.5e-06,
         "output_cost_per_token_batches": 1.25e-06,
@@ -28789,6 +28911,7 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.134,
         "output_cost_per_image_token": 0.00012,
+        "output_cost_per_image_token_batches": 6e-05,
         "output_cost_per_token": 1.2e-05,
         "output_cost_per_token_above_200k_tokens": 1.8e-05,
         "output_cost_per_token_above_200k_tokens_priority": 3.24e-05,
@@ -28842,12 +28965,13 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.134,
         "output_cost_per_image_token": 0.00012,
+        "output_cost_per_image_token_batches": 6e-05,
         "output_cost_per_token": 1.2e-05,
         "output_cost_per_token_priority": 2.16e-05,
         "output_cost_per_token_above_200k_tokens": 1.8e-05,
         "output_cost_per_token_above_200k_tokens_priority": 3.24e-05,
         "output_cost_per_token_batches": 6e-06,
-        "source": "https://ai.google.dev/gemini-api/docs/pricing",
+        "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/images/generations",
@@ -28893,6 +29017,7 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.0672,
         "output_cost_per_image_token": 6e-05,
+        "output_cost_per_image_token_batches": 3e-05,
         "output_cost_per_token": 3e-06,
         "output_cost_per_token_batches": 1.5e-06,
         "output_cost_per_token_flex": 1.5e-06,
@@ -28938,9 +29063,10 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.0672,
         "output_cost_per_image_token": 6e-05,
+        "output_cost_per_image_token_batches": 3e-05,
         "output_cost_per_token": 3e-06,
         "output_cost_per_token_batches": 1.5e-06,
-        "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models",
+        "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/images/generations",
@@ -28985,6 +29111,7 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.0336,
         "output_cost_per_image_token": 3e-05,
+        "output_cost_per_image_token_batches": 1.5e-05,
         "output_cost_per_token": 1.5e-06,
         "output_cost_per_token_batches": 7.5e-07,
         "output_cost_per_token_flex": 7.5e-07,
@@ -29551,10 +29678,12 @@
         },
         "search_free_requests": {"count": 10000, "period": "day", "shared_group": "gemini-2.5-pro"},
         "google_maps_grounding_cost_per_query": 0.025,
+        "input_cost_per_token_above_200k_tokens_batches": 1.25e-06,
         "input_cost_per_token_above_200k_tokens_priority": 4.5e-06,
         "input_cost_per_token_batches": 6.25e-07,
         "input_cost_per_token_flex": 6.25e-07,
         "input_cost_per_token_priority": 2.25e-06,
+        "output_cost_per_token_above_200k_tokens_batches": 7.5e-06,
         "output_cost_per_token_above_200k_tokens_priority": 2.7e-05,
         "output_cost_per_token_batches": 5e-06,
         "output_cost_per_token_flex": 5e-06,
@@ -29569,6 +29698,7 @@
         "google_maps_grounding_cost_per_query": 0.014,
         "input_cost_per_token": 2e-06,
         "input_cost_per_token_above_200k_tokens": 4e-06,
+        "input_cost_per_token_above_200k_tokens_batches": 2e-06,
         "input_cost_per_token_above_200k_tokens_priority": 7.2e-06,
         "input_cost_per_token_batches": 1e-06,
         "input_cost_per_token_flex": 1e-06,
@@ -29580,6 +29710,7 @@
         "output_cost_per_image": 0.00012,
         "output_cost_per_token": 1.2e-05,
         "output_cost_per_token_above_200k_tokens": 1.8e-05,
+        "output_cost_per_token_above_200k_tokens_batches": 9e-06,
         "output_cost_per_token_above_200k_tokens_priority": 3.24e-05,
         "output_cost_per_token_batches": 6e-06,
         "output_cost_per_token_flex": 6e-06,
@@ -30108,6 +30239,7 @@
         "google_maps_grounding_cost_per_query": 0.014,
         "input_cost_per_token": 2e-06,
         "input_cost_per_token_above_200k_tokens": 4e-06,
+        "input_cost_per_token_above_200k_tokens_batches": 2e-06,
         "input_cost_per_token_above_200k_tokens_priority": 7.2e-06,
         "input_cost_per_token_batches": 1e-06,
         "input_cost_per_token_flex": 1e-06,
@@ -30119,6 +30251,7 @@
         "output_cost_per_image": 0.00012,
         "output_cost_per_token": 1.2e-05,
         "output_cost_per_token_above_200k_tokens": 1.8e-05,
+        "output_cost_per_token_above_200k_tokens_batches": 9e-06,
         "output_cost_per_token_above_200k_tokens_priority": 3.24e-05,
         "output_cost_per_token_batches": 6e-06,
         "output_cost_per_token_flex": 6e-06,
@@ -51789,6 +51922,7 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.039,
         "output_cost_per_image_token": 3e-05,
+        "output_cost_per_image_token_batches": 1.5e-05,
         "output_cost_per_reasoning_token": 2.5e-06,
         "output_cost_per_token": 2.5e-06,
         "output_cost_per_token_batches": 1.25e-06,
@@ -51847,6 +51981,7 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.134,
         "output_cost_per_image_token": 0.00012,
+        "output_cost_per_image_token_batches": 6e-05,
         "output_cost_per_token": 1.2e-05,
         "output_cost_per_token_above_200k_tokens": 1.8e-05,
         "output_cost_per_token_above_200k_tokens_priority": 3.24e-05,
@@ -51881,13 +52016,14 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.134,
         "output_cost_per_image_token": 0.00012,
+        "output_cost_per_image_token_batches": 6e-05,
         "output_cost_per_token": 1.2e-05,
         "output_cost_per_token_priority": 2.16e-05,
         "output_cost_per_token_above_200k_tokens": 1.8e-05,
         "output_cost_per_token_above_200k_tokens_priority": 3.24e-05,
         "output_cost_per_token_batches": 6e-06,
         "supports_reasoning": false,
-        "source": "https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-pro-image",
+        "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/images/generations",
@@ -51913,6 +52049,7 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.0672,
         "output_cost_per_image_token": 6e-05,
+        "output_cost_per_image_token_batches": 3e-05,
         "output_cost_per_token": 3e-06,
         "output_cost_per_token_batches": 1.5e-06,
         "output_cost_per_token_flex": 1.5e-06,
@@ -51940,10 +52077,11 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.0672,
         "output_cost_per_image_token": 6e-05,
+        "output_cost_per_image_token_batches": 3e-05,
         "output_cost_per_token": 3e-06,
         "output_cost_per_token_batches": 1.5e-06,
         "supports_reasoning": false,
-        "source": "https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models",
+        "source": "https://cloud.google.com/gemini-enterprise-agent-platform/generative-ai/pricing",
         "supported_endpoints": [
             "/v1/chat/completions",
             "/v1/images/generations",
@@ -51968,6 +52106,7 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.0336,
         "output_cost_per_image_token": 3e-05,
+        "output_cost_per_image_token_batches": 1.5e-05,
         "output_cost_per_token": 1.5e-06,
         "output_cost_per_token_batches": 7.5e-07,
         "output_cost_per_token_flex": 7.5e-07,
@@ -61865,6 +62004,39 @@
         "supports_reasoning": true,
         "supports_vision": false
     },
+    "tencent/seedream-image-v5.0-flash": {
+        "litellm_provider": "tencent",
+        "mode": "image_generation",
+        "output_cost_per_image": 0.018,
+        "output_cost_per_image_token": 1.6e-06,
+        "source": "https://cloud.tencent.com/document/product/1823/130055",
+        "supported_endpoints": [
+            "/v1/images/generations",
+            "/v1/images/edits"
+        ]
+    },
+    "tencent/seedream-image-v5.0-lite": {
+        "litellm_provider": "tencent",
+        "mode": "image_generation",
+        "output_cost_per_image": 0.035,
+        "output_cost_per_image_token": 1.6e-06,
+        "source": "https://www.tencentcloud.com/document/product/1300/78937",
+        "supported_endpoints": [
+            "/v1/images/generations",
+            "/v1/images/edits"
+        ]
+    },
+    "tencent/seedream-image-v5.0-pro": {
+        "litellm_provider": "tencent",
+        "mode": "image_generation",
+        "output_cost_per_image": 0.09,
+        "output_cost_per_image_token": 1.6e-06,
+        "source": "https://www.tencentcloud.com/document/product/1300/78937",
+        "supported_endpoints": [
+            "/v1/images/generations",
+            "/v1/images/edits"
+        ]
+    },
     "cognition/swe-1.6": {
         "input_cost_per_token": 5e-07,
         "output_cost_per_token": 2.5e-06,
@@ -74246,6 +74418,7 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.0336,
         "output_cost_per_image_token": 3e-05,
+        "output_cost_per_image_token_batches": 1.5e-05,
         "output_cost_per_token": 7.5e-06,
         "output_cost_per_token_batches": 3.75e-06,
         "output_cost_per_token_priority": 1.35e-05,
@@ -74291,6 +74464,7 @@
         "mode": "image_generation",
         "output_cost_per_image": 0.0336,
         "output_cost_per_image_token": 3e-05,
+        "output_cost_per_image_token_batches": 1.5e-05,
         "output_cost_per_token": 7.5e-06,
         "output_cost_per_token_batches": 3.75e-06,
         "output_cost_per_token_priority": 1.35e-05,
@@ -77774,7 +77948,11 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 5.25e-07,
+        "output_cost_per_token_priority": 2.1e-06,
+        "cache_read_input_token_cost_priority": 5.25e-08,
+        "cache_creation_input_token_cost_priority": 5.25e-07
     },
     "databricks/databricks-deepseek-v4-1-flash-1017": {
         "cache_creation_input_token_cost": 3.0002e-07,
@@ -77897,7 +78075,11 @@
         "supports_reasoning": true,
         "supports_tool_choice": true,
         "supports_video_input": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 4.50002e-06,
+        "output_cost_per_token_priority": 2.250003e-05,
+        "cache_read_input_token_cost_priority": 2.2498e-07,
+        "cache_creation_input_token_cost_priority": 4.50002e-06
     },
     "databricks/databricks-gpt-5-4-pro": {
         "cache_creation_input_token_cost": 2.999997e-05,
@@ -77959,7 +78141,15 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 4.00001e-06,
+        "output_cost_per_token_priority": 1.99997e-05,
+        "cache_read_input_token_cost_priority": 1.9999e-07,
+        "cache_creation_input_token_cost_priority": 5.00003e-06,
+        "input_cost_per_token_above_272k_tokens_priority": 8.00002e-06,
+        "output_cost_per_token_above_272k_tokens_priority": 2.99999e-05,
+        "cache_read_input_token_cost_above_272k_tokens_priority": 3.9998e-07,
+        "cache_creation_input_token_cost_above_272k_tokens_priority": 9.99999e-06
     },
     "databricks/databricks-gpt-6-1-terra": {
         "cache_creation_input_token_cost": 1.24999e-06,
@@ -77991,7 +78181,15 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 1.99997e-06,
+        "output_cost_per_token_priority": 9.99999e-06,
+        "cache_read_input_token_cost_priority": 1.0003e-07,
+        "cache_creation_input_token_cost_priority": 2.49998e-06,
+        "input_cost_per_token_above_272k_tokens_priority": 4.00001e-06,
+        "output_cost_per_token_above_272k_tokens_priority": 1.500002e-05,
+        "cache_read_input_token_cost_above_272k_tokens_priority": 1.9999e-07,
+        "cache_creation_input_token_cost_above_272k_tokens_priority": 5.00003e-06
     },
     "databricks/databricks-gpt-6-astra": {
         "cache_creation_input_token_cost": 1.249997e-05,
@@ -78023,7 +78221,15 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 1.999998e-05,
+        "output_cost_per_token_priority": 9.99999e-05,
+        "cache_read_input_token_cost_priority": 1.99997e-06,
+        "cache_creation_input_token_cost_priority": 2.500001e-05,
+        "input_cost_per_token_above_272k_tokens_priority": 4.000003e-05,
+        "output_cost_per_token_above_272k_tokens_priority": 0.0001500002,
+        "cache_read_input_token_cost_above_272k_tokens_priority": 4.00001e-06,
+        "cache_creation_input_token_cost_above_272k_tokens_priority": 5.000002e-05
     },
     "databricks/databricks-gpt-6-luna": {
         "cache_creation_input_token_cost": 1.2502e-07,
@@ -78055,7 +78261,15 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 1.9999e-07,
+        "output_cost_per_token_priority": 1.0003e-06,
+        "cache_read_input_token_cost_priority": 2.002e-08,
+        "cache_creation_input_token_cost_priority": 2.4997e-07,
+        "input_cost_per_token_above_272k_tokens_priority": 3.9998e-07,
+        "output_cost_per_token_above_272k_tokens_priority": 1.5001e-06,
+        "cache_read_input_token_cost_above_272k_tokens_priority": 3.997e-08,
+        "cache_creation_input_token_cost_above_272k_tokens_priority": 5.0001e-07
     },
     "databricks/databricks-gpt-6-sol": {
         "cache_creation_input_token_cost": 2.49998e-06,
@@ -78087,7 +78301,15 @@
         "supports_prompt_caching": true,
         "supports_reasoning": true,
         "supports_tool_choice": true,
-        "supports_vision": true
+        "supports_vision": true,
+        "input_cost_per_token_priority": 4.00001e-06,
+        "output_cost_per_token_priority": 1.99997e-05,
+        "cache_read_input_token_cost_priority": 3.9998e-07,
+        "cache_creation_input_token_cost_priority": 5.00003e-06,
+        "input_cost_per_token_above_272k_tokens_priority": 8.00002e-06,
+        "output_cost_per_token_above_272k_tokens_priority": 2.99999e-05,
+        "cache_read_input_token_cost_above_272k_tokens_priority": 8.0003e-07,
+        "cache_creation_input_token_cost_above_272k_tokens_priority": 9.99999e-06
     },
     "databricks/databricks-gpt-image-1": {
         "cache_creation_input_token_cost": 5.00003e-06,

litellm/provider_endpoints_support_backup.json

@@ -3793,8 +3793,8 @@
         "responses": true,
         "text_completion": true,
         "embeddings": false,
-        "image_generations": false,
-        "image_edits": false,
+        "image_generations": true,
+        "image_edits": true,
         "audio_transcriptions": false,
         "audio_speech": false,
         "rerank": false,

litellm/proxy/_lazy_openapi_snapshot.json

@@ -3487,6 +3487,13 @@
               "title": "Entities",
               "type": "object"
             },
+            "guardrails": {
+              "additionalProperties": {
+                "$ref": "#/components/schemas/GuardrailMetric"
+              },
+              "title": "Guardrails",
+              "type": "object"
+            },
             "mcp_servers": {
               "additionalProperties": {
                 "$ref": "#/components/schemas/MetricWithMetadata"
@@ -3802,6 +3809,74 @@
           "title": "ExportType",
           "type": "string"
         },
+        "GuardrailKeyMetric": {
+          "properties": {
+            "metadata": {
+              "$ref": "#/components/schemas/KeyMetadata"
+            },
+            "metrics": {
+              "$ref": "#/components/schemas/GuardrailSpendMetrics"
+            }
+          },
+          "required": [
+            "metrics",
+            "metadata"
+          ],
+          "title": "GuardrailKeyMetric",
+          "type": "object"
+        },
+        "GuardrailMetadata": {
+          "properties": {
+            "guardrail_name": {
+              "title": "Guardrail Name",
+              "type": "string"
+            }
+          },
+          "required": [
+            "guardrail_name"
+          ],
+          "title": "GuardrailMetadata",
+          "type": "object"
+        },
+        "GuardrailMetric": {
+          "properties": {
+            "api_key_breakdown": {
+              "additionalProperties": {
+                "$ref": "#/components/schemas/GuardrailKeyMetric"
+              },
+              "title": "Api Key Breakdown",
+              "type": "object"
+            },
+            "metadata": {
+              "$ref": "#/components/schemas/GuardrailMetadata"
+            },
+            "metrics": {
+              "$ref": "#/components/schemas/GuardrailSpendMetrics"
+            }
+          },
+          "required": [
+            "metrics",
+            "metadata"
+          ],
+          "title": "GuardrailMetric",
+          "type": "object"
+        },
+        "GuardrailSpendMetrics": {
+          "properties": {
+            "spend": {
+              "default": 0.0,
+              "title": "Spend",
+              "type": "number"
+            },
+            "untracked_units": {
+              "default": 0,
+              "title": "Untracked Units",
+              "type": "integer"
+            }
+          },
+          "title": "GuardrailSpendMetrics",
+          "type": "object"
+        },
         "HTTPAuthSecurityScheme": {
           "description": "Defines a security scheme using HTTP authentication.",
           "properties": {

litellm/proxy/agent_endpoints/a2a_endpoints.py

@@ -26,6 +26,7 @@ from litellm._logging import verbose_proxy_logger
 from litellm.litellm_core_utils.core_helpers import as_str_mapping
 from litellm.litellm_core_utils.url_utils import SSRFError, validate_url
 from litellm.llms.a2a.common_utils import resolve_a2a_hop_auth_header
+from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines
 from litellm.proxy._types import UserAPIKeyAuth
 from litellm.proxy.a2a.version_convert import (
     A2AVersion,
@@ -320,7 +321,7 @@ async def _a2a_sse_event_source(
                 "error": {"code": -32603, "message": resp.reason_phrase},
             }
             return
-        async for line in resp.aiter_lines():
+        async for line in aiter_sse_lines(resp.aiter_text()):
             stripped = line.strip()
             if not stripped.startswith("data:"):
                 continue

litellm/proxy/anthropic_endpoints/streaming_model_restamp.py

@@ -14,6 +14,8 @@ from typing import Final
 
 from pydantic import TypeAdapter, ValidationError
 
+from litellm.llms.custom_httpx.sse_lines import split_sse_lines_keepends
+
 _MESSAGE_START_EVENT: Final = "message_start"
 _MESSAGE_START_MARKER: Final = b"message_start"
 _SSE_DATA_FIELD: Final = "data:"
@@ -52,7 +54,7 @@ def _restamped_data_line(line: str, requested_model: str) -> str | None:
 
 
 def _restamped_frame(frame: str, requested_model: str) -> str | None:
-    lines: Final = frame.splitlines(keepends=True)
+    lines: Final = split_sse_lines_keepends(frame)
     restamped: Final = tuple(_restamped_data_line(line, requested_model) for line in lines)
     if all(line is None for line in restamped):
         return None

litellm/proxy/auth/user_api_key_auth.py

@@ -370,7 +370,7 @@ class _KeyModelBudgetLimiter(Protocol):
 
 class _UserModelBudgetLimiter(Protocol):
     async def is_user_within_model_budget(
-        self, user_id: str, user_model_max_budget: Mapping[str, object], model: str
+        self, user_id: str, user_model_max_budget: Mapping[str, object], model: str, team_id: str | None = None
     ) -> bool: ...
 
 
@@ -388,11 +388,15 @@ class _EveryModelBudgetLimiter(Protocol):
     async def is_key_within_model_budget(self, user_api_key_dict: UserAPIKeyAuth, model: str) -> bool: ...
 
     async def is_user_within_model_budget(
-        self, user_id: str, user_model_max_budget: Mapping[str, object], model: str
+        self, user_id: str, user_model_max_budget: Mapping[str, object], model: str, team_id: str | None = None
     ) -> bool: ...
 
     async def is_end_user_within_model_budget(
-        self, end_user_id: str, end_user_model_max_budget: Mapping[str, object], model: str
+        self,
+        end_user_id: str,
+        end_user_model_max_budget: Mapping[str, object],
+        model: str,
+        team_id: str | None = None,
     ) -> bool: ...
 
     async def is_team_within_model_budget(
@@ -503,6 +507,7 @@ async def _check_user_model_budget(
             user_id=valid_token.user_id,
             user_model_max_budget=user_model_max_budget,
             model=model_name,
+            team_id=valid_token.team_id,
         )
 
 
@@ -547,6 +552,7 @@ async def _is_within_every_model_budget(
                 user_id=valid_token.user_id,
                 user_model_max_budget=user_model_max_budget,
                 model=model,
+                team_id=valid_token.team_id,
             )
         if (
             valid_token.end_user_id is not None
@@ -557,6 +563,7 @@ async def _is_within_every_model_budget(
                 end_user_id=valid_token.end_user_id,
                 end_user_model_max_budget=end_user_model_max_budget,
                 model=model,
+                team_id=valid_token.team_id,
             )
         if valid_token.team_id is not None and isinstance(team_model_max_budget, Mapping) and team_model_max_budget:
             await model_max_budget_limiter.is_team_within_model_budget(
@@ -3154,6 +3161,7 @@ async def validate_resolved_virtual_key(  # noqa: C901  # Preserve ordering of e
                             end_user_id=valid_token.end_user_id,
                             end_user_model_max_budget=end_user_mmb,
                             model=model_name,
+                            team_id=valid_token.team_id,
                         )
 
         # Check 6: Additional Common Checks across jwt + key auth
@@ -4733,6 +4741,7 @@ async def _run_post_custom_auth_checks(
                 end_user_id=valid_token.end_user_id,
                 end_user_model_max_budget=end_user_mmb,
                 model=model_name,
+                team_id=valid_token.team_id,
             )
 
     # team / user / end_user / project context objects are fetched by

litellm/proxy/client/models.py

@@ -3,9 +3,12 @@ from collections.abc import Mapping
 from typing import Any, Final
 
 import requests
+from pydantic import ConfigDict, TypeAdapter
 
 from .exceptions import NotFoundError, UnauthorizedError
 
+_JSON_OBJECT: Final = TypeAdapter(dict[str, object], config=ConfigDict(strict=True))
+
 
 class ModelsManagementClient:
     def __init__(self, base_url: str, api_key: str | None = None, timeout: int = 30):
@@ -109,7 +112,7 @@ class ModelsManagementClient:
         try:
             response: Final = session.send(request.prepare(), timeout=self._timeout)
             response.raise_for_status()
-            return response.json()
+            return _JSON_OBJECT.validate_python(response.json())
         except requests.exceptions.HTTPError as e:
             if e.response.status_code == 401:
                 raise UnauthorizedError(e)
@@ -145,7 +148,7 @@ class ModelsManagementClient:
         try:
             response: Final = session.send(request.prepare(), timeout=self._timeout)
             response.raise_for_status()
-            return response.json()
+            return _JSON_OBJECT.validate_python(response.json())
         except requests.exceptions.HTTPError as e:
             if e.response.status_code == 401:
                 raise UnauthorizedError(e)
@@ -287,7 +290,7 @@ class ModelsManagementClient:
         try:
             response: Final = session.send(request.prepare(), timeout=self._timeout)
             response.raise_for_status()
-            return response.json()
+            return _JSON_OBJECT.validate_python(response.json())
         except requests.exceptions.HTTPError as e:
             if e.response.status_code == 401:
                 raise UnauthorizedError(e)

litellm/proxy/common_request_processing.py

@@ -139,8 +139,8 @@ from litellm.proxy.utils import (  # noqa: F401  # legacy module exports
 )
 from litellm.responses.streaming_iterator import partial_responses_incomplete_event_on_disconnect, spawn_logging_task
 from litellm.router import Router
+from litellm.router_strategy.auto_router.auto_router_model_naming import AUTO_ROUTER_MODEL_PREFIX
 from litellm.router_utils.add_retry_fallback_headers import get_hidden_params_dict
-from litellm.router_utils.auto_router_model_naming import AUTO_ROUTER_MODEL_PREFIX
 from litellm.router_utils.common_utils import resolve_model_group_alias, resolve_served_model
 from litellm.types.extraction import ExtractionResponse
 from litellm.types.guardrails import GuardrailEventHooks
@@ -266,6 +266,8 @@ from litellm.proxy.anthropic_endpoints.streaming_model_restamp import (
     AnthropicStreamModelRestamper,
 )
 from litellm.proxy.litellm_pre_call_utils import (
+    ROUTER_SCHEDULED_ROUTE_TYPES,
+    SCHEDULER_PRIORITY_HEADER,
     LiteLLMProxyRequestSetup,
     add_litellm_data_to_request,
     get_metadata_variable_name,
@@ -2478,14 +2480,20 @@ class ProxyBaseLLMRequestProcessing:
                     self.data["model"] = alias_target
 
         aliased_model: Final = self.data.get("model")
+        if llm_router is not None and route_type in ROUTER_SCHEDULED_ROUTE_TYPES and isinstance(aliased_model, str):
+            self.data.update(
+                llm_router.priority_header_update(
+                    model=aliased_model, kwargs=self.data, header=request.headers.get(SCHEDULER_PRIORITY_HEADER)
+                )
+            )
 
         litellm_call_id: Final = resolve_litellm_call_id(request.headers.get("x-litellm-call-id"))
         self.data["litellm_call_id"] = litellm_call_id
         bind_telemetry_call_id(litellm_call_id)
-        DDSpanTagger.tag_call_id(self.data.get("litellm_call_id"))
+        DDSpanTagger.tag_call_id(litellm_call_id)
         DDSpanTagger.tag_request(
             user_api_key_dict=user_api_key_dict,
-            requested_model=self.data.get("model"),
+            requested_model=aliased_model,
         )
 
         ### AUTO STREAM USAGE TRACKING ###

litellm/proxy/db/accounting_ledger.py

@@ -0,0 +1,269 @@
+from __future__ import annotations
+
+import hashlib
+import json
+from collections.abc import AsyncGenerator, Callable
+from contextlib import AbstractAsyncContextManager, asynccontextmanager
+from dataclasses import dataclass
+from datetime import datetime, timedelta, timezone
+from typing import TYPE_CHECKING, Final, Literal, TypeAlias
+
+from pydantic import AwareDatetime, ConfigDict, JsonValue, TypeAdapter
+from typing_extensions import assert_never
+
+from litellm.accounting import CostEvent, DeliveryTarget, EventKind, InboundUsageEvent, OutboundUsageEvent
+from litellm.proxy.db.db_span import db_span
+from litellm.proxy.db.routing_prisma_wrapper import writer_wrapper
+
+if TYPE_CHECKING:
+    from prisma import Prisma
+
+    from litellm.proxy.utils import PrismaClient
+
+AccountingEvent: TypeAlias = InboundUsageEvent | OutboundUsageEvent | CostEvent
+_JSON: Final = TypeAdapter(dict[str, JsonValue])
+_INSTANT: Final[TypeAdapter[datetime]] = TypeAdapter(AwareDatetime)
+_STRINGS: Final = TypeAdapter(tuple[dict[str, str], ...])
+_TIMEOUT: Final = timedelta(seconds=10)
+
+
+class AccountingConflictError(ValueError):
+    pass
+
+
+_Transaction: TypeAlias = AbstractAsyncContextManager["Prisma"]
+
+
+@dataclass(frozen=True, slots=True)
+class _AcceptedEvent:
+    event_id: str
+    kind: EventKind
+    source_event_id: str | None
+    occurred_at: datetime
+    payload: str
+    payload_hash: str
+    targets: tuple[str, ...]
+    schema_version: int
+    version: int
+
+
+def _json(value: object) -> str:
+    return json.dumps(value, sort_keys=True, separators=(",", ":"), allow_nan=False)
+
+
+def _prepare(
+    event: AccountingEvent,
+    targets: tuple[DeliveryTarget, ...],
+    envelope: str,
+    schema_version: int,
+    version: int,
+) -> _AcceptedEvent:
+    if type(schema_version) is not int or type(version) is not int or min(schema_version, version) < 1:
+        raise ValueError("Accounting versions must be positive integers")
+    validated: Final = type(event).model_validate_json(event.model_dump_json())
+    payload: Final = _JSON.validate_json(envelope)
+    if payload.get("event") != validated.model_dump(mode="json"):
+        raise ValueError("Accounting envelope must contain the matching public event")
+    destinations: Final = tuple(sorted(frozenset(DeliveryTarget(target).value for target in targets)))
+    serialized: Final = _json(payload)
+    identity: Final = _json((schema_version, version, destinations, payload))
+    kind: Final = (
+        EventKind.COST
+        if isinstance(validated, CostEvent)
+        else EventKind.INBOUND_USAGE
+        if isinstance(validated, InboundUsageEvent)
+        else EventKind.OUTBOUND_USAGE
+    )
+    return _AcceptedEvent(
+        event_id=validated.event_id,
+        kind=kind,
+        source_event_id=validated.usage_event_id if isinstance(validated, CostEvent) else None,
+        occurred_at=validated.occurred_at if isinstance(validated, CostEvent) else validated.ended_at,
+        payload=serialized,
+        payload_hash=hashlib.sha256(identity.encode()).hexdigest(),
+        targets=destinations,
+        schema_version=schema_version,
+        version=version,
+    )
+
+
+_INSERT_EVENT: Final = """
+INSERT INTO "LiteLLM_AccountingEvent"
+    (event_id, kind, source_event_id, schema_version, payload_hash, occurred_at, payload)
+VALUES ($1, $2::"AccountingEventKind", $3, $4::int, $5, $6::timestamp, $7::jsonb)
+ON CONFLICT (event_id) DO NOTHING RETURNING event_id
+"""
+_INSERT_DELIVERIES: Final = """
+INSERT INTO "LiteLLM_AccountingDelivery" (event_id, target, version)
+SELECT $1, value::"AccountingDeliveryTarget", $3::int FROM jsonb_array_elements_text($2::jsonb)
+"""
+_LOCK_SOURCE: Final = """
+SELECT event_id FROM "LiteLLM_AccountingEvent"
+WHERE event_id = $1 AND kind = 'outbound_usage' FOR UPDATE
+"""
+_EXISTING_COST: Final = """
+SELECT event_id FROM "LiteLLM_AccountingEvent" WHERE kind = 'cost' AND source_event_id = $1
+"""
+_LOCK_EVENT: Final = """
+SELECT payload_hash FROM "LiteLLM_AccountingEvent" WHERE event_id = $1 FOR UPDATE
+"""
+_COMPLETED_FAMILY: Final = """
+root.source_event_id IS NULL AND root.kind IN ('inbound_usage', 'outbound_usage')
+AND root.recorded_at < $1::timestamp
+AND (root.kind = 'inbound_usage' OR EXISTS (
+    SELECT 1 FROM "LiteLLM_AccountingEvent" AS cost WHERE cost.source_event_id = root.event_id AND cost.kind = 'cost'
+))
+AND NOT EXISTS (
+    SELECT 1 FROM "LiteLLM_AccountingEvent" AS member
+    WHERE (member.event_id = root.event_id OR (member.kind = 'cost' AND member.source_event_id = root.event_id))
+    AND (member.recorded_at >= $1::timestamp
+        OR EXISTS (SELECT 1 FROM "LiteLLM_AccountingDelivery" AS delivery
+            WHERE delivery.event_id = member.event_id AND delivery.state != 'applied'))
+)
+"""
+_RETENTION_CANDIDATES: Final = (
+    'SELECT root.event_id FROM "LiteLLM_AccountingEvent" AS root WHERE '
+    + _COMPLETED_FAMILY
+    + " ORDER BY root.recorded_at, root.event_id LIMIT $2::int FOR UPDATE OF root SKIP LOCKED"
+)
+_ELIGIBLE_GROUP: Final = (
+    'SELECT root.event_id FROM "LiteLLM_AccountingEvent" AS root WHERE root.event_id = $2 AND ' + _COMPLETED_FAMILY
+)
+_DELETE_DELIVERIES: Final = """
+DELETE FROM "LiteLLM_AccountingDelivery" WHERE event_id IN (
+    SELECT event_id FROM "LiteLLM_AccountingEvent" WHERE event_id = $1 OR (kind = 'cost' AND source_event_id = $1)
+)
+"""
+_DELETE_EVENTS: Final = """
+DELETE FROM "LiteLLM_AccountingEvent" WHERE event_id = $1 OR (kind = 'cost' AND source_event_id = $1)
+"""
+
+
+_Conflict: TypeAlias = Literal["missing_source", "competing_cost", "identity_conflict"]
+
+
+async def _accept(db: Prisma, record: _AcceptedEvent) -> Literal["inserted", "duplicate"] | _Conflict:
+    if record.source_event_id is not None:
+        source: Final = _STRINGS.validate_python(await db.query_raw(_LOCK_SOURCE, record.source_event_id))
+        if not source:
+            return "missing_source"
+        costs: Final = _STRINGS.validate_python(tuple(await db.query_raw(_EXISTING_COST, record.source_event_id)))
+        if any(cost["event_id"] != record.event_id for cost in costs):
+            return "competing_cost"
+    inserted: Final = _STRINGS.validate_python(
+        await db.query_raw(
+            _INSERT_EVENT,
+            record.event_id,
+            record.kind.value,
+            record.source_event_id,
+            record.schema_version,
+            record.payload_hash,
+            record.occurred_at.astimezone(timezone.utc).isoformat(),
+            record.payload,
+        )
+    )
+    if not inserted:
+        existing: Final = _STRINGS.validate_python(tuple(await db.query_raw(_LOCK_EVENT, record.event_id)))
+        if not existing or existing[0]["payload_hash"] != record.payload_hash:
+            return "identity_conflict"
+        return "duplicate"
+    await db.execute_raw(_INSERT_DELIVERIES, record.event_id, _json(record.targets), record.version)
+    return "inserted"
+
+
+async def _delete_completed(db: Prisma, event_id: str, cutoff: str) -> int:
+    if not _STRINGS.validate_python(await db.query_raw(_ELIGIBLE_GROUP, cutoff, event_id)):
+        return 0
+    await db.execute_raw(_DELETE_DELIVERIES, event_id)
+    return await db.execute_raw(_DELETE_EVENTS, event_id)
+
+
+class AccountingLedger:
+    def __init__(self, transaction: Callable[[], _Transaction]) -> None:
+        self._transaction: Final = transaction
+
+    @classmethod
+    def for_client(cls, client: PrismaClient) -> AccountingLedger:
+        from prisma.client import TransactionManager
+
+        managers: Final = TypeAdapter(TransactionManager, config=ConfigDict(arbitrary_types_allowed=True))
+
+        @asynccontextmanager
+        async def transaction() -> AsyncGenerator[Prisma]:
+            primary: Final = writer_wrapper(client.db)
+            manager: Final = managers.validate_python(primary.tx(timeout=_TIMEOUT))
+            async with db_span("accounting_ledger", "LiteLLM_AccountingEvent"), manager as db:
+                await db.execute_raw("SET LOCAL statement_timeout = 5000")
+                await db.execute_raw("SET LOCAL lock_timeout = 1000")
+                yield db
+
+        return cls(transaction)
+
+    async def insert_event(
+        self,
+        event: AccountingEvent,
+        targets: tuple[DeliveryTarget, ...],
+        *,
+        envelope: str,
+        schema_version: int = 1,
+        version: int = 1,
+    ) -> Literal["inserted", "duplicate"]:
+        """Accept T5/T6's serialized envelope, containing its public event under `event`.
+
+        The immutable acceptance hash includes the target set and delivery version.
+        Changing either for an accepted event conflicts. T7 must provide a separate
+        add-target or new-version workflow that preserves receipts and prevents
+        duplicate contributions; redelivery cannot silently create or reset work.
+
+        The Prisma adapter opens one interactive transaction per event, including
+        two SET LOCAL statements and a source-row lock for costs. I/O is async,
+        but awaiting this method inline adds commit latency per event, potentially
+        several times per request. T5 owns bounded asynchronous capture and an
+        explicit batching strategy; looping over this method does not batch commits.
+        Success acknowledges commit, independently of HTTP response acknowledgement.
+        """
+        record: Final = _prepare(event, targets, envelope, schema_version, version)
+        async with self._transaction() as db:
+            result: Final = await _accept(db, record)
+        match result:
+            case "inserted" | "duplicate":
+                return result
+            case "missing_source":
+                raise AccountingConflictError("Cost source must be a durably accepted outbound usage event")
+            case "competing_cost":
+                raise AccountingConflictError("Usage event already has an authoritative cost")
+            case "identity_conflict":
+                raise AccountingConflictError("Event identity conflicts with its accepted content or delivery set")
+            case _:
+                assert_never(result)
+
+    async def purge_completed(
+        self,
+        *,
+        redelivery_before: datetime | None = None,
+        recovery_before: datetime | None = None,
+        limit: int = 100,
+    ) -> int:
+        """Delete at most `limit` completed families once both retention owners permit it.
+
+        T8 must supply a coherent recovery watermark including outstanding holds.
+        Omitting either watermark disables deletion. Each usage row and its cost
+        remain together so a retained charge never loses its source or vice versa.
+        Applied receipts are deleted with the family, ending its retained dedup
+        protection. The caller must guarantee events recorded before
+        redelivery_before cannot be redelivered through supported paths.
+        T8's recovery_before must also protect checkpoints,
+        outstanding holds and unfinished settlements; age alone is insufficient.
+        """
+        if type(limit) is not int or not 1 <= limit <= 1000:
+            raise ValueError("Accounting retention batch size must be between 1 and 1000")
+        if redelivery_before is None or recovery_before is None:
+            return 0
+        cutoff: Final = min(
+            _INSTANT.validate_python(redelivery_before).astimezone(timezone.utc),
+            _INSTANT.validate_python(recovery_before).astimezone(timezone.utc),
+        ).isoformat()
+        async with self._transaction() as db:
+            roots: Final = _STRINGS.validate_python(tuple(await db.query_raw(_RETENTION_CANDIDATES, cutoff, limit)))
+            deleted: Final = tuple([await _delete_completed(db, root["event_id"], cutoff) for root in roots])
+        return sum(deleted)

litellm/proxy/db/budget_window_spend_writer.py

@@ -64,6 +64,16 @@ _ROLL_WINDOW_SPEND_SQL: Final = (
     "AND window_start < ($4::timestamptz AT TIME ZONE 'UTC')"
 )
 
+_OPEN_WINDOW_SPEND_SQL: Final = (
+    'INSERT INTO "LiteLLM_BudgetWindowSpend" '
+    "(entity_type, entity_id, window_duration, window_start, spend, created_at, updated_at) "
+    "VALUES ($1, $2, $3, ($4::timestamptz AT TIME ZONE 'UTC'), 0, "
+    "($4::timestamptz AT TIME ZONE 'UTC'), ($4::timestamptz AT TIME ZONE 'UTC')) "
+    "ON CONFLICT (entity_type, entity_id, window_duration) DO UPDATE SET "
+    "window_start = EXCLUDED.window_start, spend = 0, updated_at = EXCLUDED.updated_at "
+    "WHERE \"LiteLLM_BudgetWindowSpend\".window_start <= ($5::timestamptz AT TIME ZONE 'UTC')"
+)
+
 _SEED_FROM_SPEND_LOGS_KEY_SQL: Final = (
     "SELECT COALESCE(SUM(spend), 0.0) AS total, "
     "COALESCE(SUM(spend) FILTER (WHERE \"startTime\" < ($3::timestamptz AT TIME ZONE 'UTC')), 0.0) AS before_batch "
@@ -389,3 +399,29 @@ async def roll_window_spend_row(
             to_naive_utc(new_window_start),
             to_naive_utc(datetime.now(timezone.utc)),
         )
+
+
+async def open_window_spend_row(
+    prisma_client: "PrismaClient",
+    entity_type: str,
+    entity_id: str,
+    window_duration: str,
+    window_start: datetime,
+    replaces_window_start: datetime | None,
+) -> None:
+    """Start a window at `window_start` unless another writer already started one.
+
+    The row is created when it does not exist, and moved onto the new window only
+    while it still holds `replaces_window_start`, the ended window the caller saw.
+    A writer that lost the race leaves the winner's window in place, so reading
+    the row back gives every writer the same window start.
+    """
+    async with db_span("open_window_spend_row", "LiteLLM_BudgetWindowSpend"):
+        await prisma_client.db.execute_raw(
+            _OPEN_WINDOW_SPEND_SQL,
+            entity_type,
+            entity_id,
+            window_duration,
+            to_naive_utc(window_start),
+            None if replaces_window_start is None else to_naive_utc(replaces_window_start),
+        )

litellm/proxy/guardrails/auto_router_compression.py

@@ -14,7 +14,7 @@ from typing import TYPE_CHECKING, Final
 
 from litellm._logging import verbose_proxy_logger
 from litellm.litellm_core_utils.core_helpers import get_or_create_metadata_bucket
-from litellm.router_utils.auto_router_model_naming import AUTO_ROUTER_MODEL_PREFIX
+from litellm.router_strategy.auto_router.auto_router_model_naming import AUTO_ROUTER_MODEL_PREFIX
 from litellm.types.utils import GenericGuardrailAPIInputs
 
 if TYPE_CHECKING:

litellm/proxy/guardrails/guardrail_hooks/openai/moderations.py

@@ -169,7 +169,7 @@ class OpenAIModerationGuardrail(OpenAIGuardrailBase, CustomGuardrail):
 
         from litellm.types.llms.openai import OpenAIModerationResponse
 
-        return OpenAIModerationResponse(**response.json())
+        return OpenAIModerationResponse.model_validate(response.json())
 
     def _check_moderation_result(self, moderation_response: "OpenAIModerationResponse") -> None:
         """

litellm/proxy/health_check.py

@@ -27,7 +27,7 @@ from litellm.constants import (
     HEALTH_CHECK_TIMEOUT_SECONDS,
 )
 from litellm.litellm_core_utils.health_check_helpers import native_health_check_mode
-from litellm.router_utils.auto_router_model_naming import (
+from litellm.router_strategy.auto_router.auto_router_model_naming import (
     StrategyRouterDependency,
     classify_strategy_router_model,
     strategy_router_dependencies,

litellm/proxy/hooks/autorouter_baseline_cache.py

@@ -37,8 +37,8 @@ from litellm.proxy.spend_tracking.savings import (
     _effective_model_info,  # pyright: ignore[reportPrivateUsage]  # existing deployment-price owner
     _proxy_llm_router,  # pyright: ignore[reportPrivateUsage]  # existing optional proxy-router owner
 )
+from litellm.router_strategy.auto_router.baseline_request import baseline_request
 from litellm.router_strategy.complexity_router.context_compaction import compaction_applied
-from litellm.router_utils.baseline_request import baseline_request
 from litellm.types.llms.base import LiteLLMBaseModel
 from litellm.types.router import BaselineRouteStamp
 from litellm.types.utils import CallTypes, ModelInfo, Usage

litellm/proxy/hooks/model_budget_shared_windows.py

@@ -0,0 +1,176 @@
+from collections.abc import Awaitable, Callable
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import TYPE_CHECKING, Final, Protocol, TypeAlias
+
+from litellm._logging import verbose_proxy_logger
+from litellm.caching.in_memory_cache import InMemoryCache
+from litellm.proxy.db.budget_window_spend_writer import open_window_spend_row
+from litellm.proxy.db.db_transaction_queue.window_spend_update_queue import (
+    WindowSpendTransaction,
+    build_window_spend_transaction,
+)
+from litellm.repositories.table_repositories import BudgetWindowSpendRepository
+from litellm.router_strategy.budget_limiter import budget_window_seconds
+
+if TYPE_CHECKING:
+    from litellm.proxy.utils import PrismaClient
+
+MODEL_BUDGET_WINDOW_ENTITY_TYPE: Final = "model_max_budget"
+_READ_TTL_SECONDS: Final = 5
+
+
+@dataclass(frozen=True, slots=True)
+class SharedWindow:
+    started_at: float
+    spend: float
+
+    def is_open(self, budget_duration: str, now: float) -> bool:
+        return now < self.started_at + budget_window_seconds(self.started_at, budget_duration)
+
+
+class LocalTTLCache(Protocol):
+    def get_cache(self, key: str) -> object: ...
+
+    def set_cache(self, key: str, value: object, *, ttl: float) -> None: ...
+
+    def delete_cache(self, key: str) -> None: ...
+
+
+class SharedWindows(Protocol):
+    async def read(self, spend_key: str, budget_duration: str) -> SharedWindow | None: ...
+
+    async def open(self, spend_key: str, budget_duration: str, now: float) -> float | None: ...
+
+    async def add(self, spend_key: str, budget_duration: str, started_at: float, spend: float) -> None: ...
+
+
+PrismaClientSource: TypeAlias = Callable[[], "PrismaClient | None"]
+WindowSpendSink: TypeAlias = Callable[[WindowSpendTransaction], Awaitable[None]]
+
+
+def _proxy_prisma_client() -> "PrismaClient | None":
+    try:
+        from litellm.proxy.proxy_server import prisma_client
+    except ImportError:
+        return None
+    return prisma_client
+
+
+async def _queue_on_proxy_spend_writer(transaction: WindowSpendTransaction) -> None:
+    from litellm.proxy.proxy_server import proxy_logging_obj
+
+    await proxy_logging_obj.db_spend_update_writer.window_spend_update_queue.add_update(transaction)
+
+
+@dataclass(frozen=True, slots=True)
+class _Reading:
+    window: SharedWindow | None
+
+
+class DatabaseSharedWindows:
+    def __init__(
+        self,
+        prisma_client_source: PrismaClientSource = _proxy_prisma_client,
+        sink: WindowSpendSink = _queue_on_proxy_spend_writer,
+        read_ttl_seconds: int = _READ_TTL_SECONDS,
+    ) -> None:
+        self._prisma_client_source = prisma_client_source
+        self._sink = sink
+        self._read_ttl_seconds = read_ttl_seconds
+        self._readings: LocalTTLCache = InMemoryCache(max_size_in_memory=10_000, default_ttl=read_ttl_seconds)
+
+    async def read(self, spend_key: str, budget_duration: str) -> SharedWindow | None:
+        cached: Final = self._readings.get_cache(key=spend_key)
+        if isinstance(cached, _Reading):
+            return cached.window
+        window: Final = await self._read_row(spend_key=spend_key, budget_duration=budget_duration)
+        self._readings.set_cache(key=spend_key, value=_Reading(window=window), ttl=self._read_ttl_seconds)
+        return window
+
+    async def open(self, spend_key: str, budget_duration: str, now: float) -> float | None:
+        """When the window every replica shares started, opening one at `now` if none is open.
+
+        Without a database every window is this replica's own, so it starts at `now`. None means the database did
+        not answer, so this replica cannot know which window the others are in yet.
+        """
+        prisma_client: Final = self._prisma_client_source()
+        if prisma_client is None:
+            return now
+        try:
+            seen: Final = await _fetch_row(
+                prisma_client=prisma_client, spend_key=spend_key, budget_duration=budget_duration, use_writer=True
+            )
+            if seen is not None and seen.is_open(budget_duration=budget_duration, now=now):
+                return seen.started_at
+            await open_window_spend_row(
+                prisma_client=prisma_client,
+                entity_type=MODEL_BUDGET_WINDOW_ENTITY_TYPE,
+                entity_id=spend_key,
+                window_duration=budget_duration,
+                window_start=datetime.fromtimestamp(now, tz=timezone.utc),
+                replaces_window_start=None
+                if seen is None
+                else datetime.fromtimestamp(seen.started_at, tz=timezone.utc),
+            )
+            opened: Final = await _fetch_row(
+                prisma_client=prisma_client, spend_key=spend_key, budget_duration=budget_duration, use_writer=True
+            )
+        except Exception:  # noqa: BLE001  # an unreachable database leaves this replica on its own window
+            verbose_proxy_logger.exception("Unable to open the shared model budget window")
+            return None
+        if opened is None or not opened.is_open(budget_duration=budget_duration, now=now):
+            return None
+        return opened.started_at
+
+    async def add(self, spend_key: str, budget_duration: str, started_at: float, spend: float) -> None:
+        if self._prisma_client_source() is None:
+            return
+        try:
+            await self._sink(
+                build_window_spend_transaction(
+                    entity_type=MODEL_BUDGET_WINDOW_ENTITY_TYPE,
+                    entity_id=spend_key,
+                    window_duration=budget_duration,
+                    window_start=datetime.fromtimestamp(started_at, tz=timezone.utc),
+                    spend=spend,
+                )
+            )
+        except Exception:  # noqa: BLE001  # spend tracking must never fail the success callback
+            verbose_proxy_logger.debug("Unable to queue shared model budget spend", exc_info=True)
+
+    async def _read_row(self, spend_key: str, budget_duration: str) -> SharedWindow | None:
+        prisma_client: Final = self._prisma_client_source()
+        if prisma_client is None:
+            return None
+        try:
+            return await _fetch_row(
+                prisma_client=prisma_client, spend_key=spend_key, budget_duration=budget_duration, use_writer=False
+            )
+        except Exception:  # noqa: BLE001  # an unreadable row leaves enforcement on this replica's own counter
+            verbose_proxy_logger.exception("Unable to read shared model budget spend")
+            return None
+
+
+async def _fetch_row(
+    prisma_client: "PrismaClient", spend_key: str, budget_duration: str, use_writer: bool
+) -> SharedWindow | None:
+    row: Final = await BudgetWindowSpendRepository(prisma_client, use_writer=use_writer).table.find_unique(
+        where={
+            "entity_type_entity_id_window_duration": {
+                "entity_type": MODEL_BUDGET_WINDOW_ENTITY_TYPE,
+                "entity_id": spend_key,
+                "window_duration": budget_duration,
+            }
+        }
+    )
+    if row is None:
+        return None
+    window_start: Final = row.window_start
+    started_at: Final = (
+        window_start if window_start.tzinfo is not None else window_start.replace(tzinfo=timezone.utc)
+    ).timestamp()
+    return SharedWindow(started_at=started_at, spend=float(row.spend or 0.0))
+
+
+PROXY_SHARED_WINDOWS: Final = DatabaseSharedWindows()

litellm/proxy/litellm_pre_call_utils.py

@@ -118,6 +118,9 @@ _SESSION_ID_VALUE_RE: Final = re.compile(r"^[a-zA-Z0-9_\-]{8,}$")
 
 _SHA256_HEX_RE: Final = re.compile(r"^[0-9a-f]{64}$")
 
+SCHEDULER_PRIORITY_HEADER: Final = "x-litellm-priority"
+ROUTER_SCHEDULED_ROUTE_TYPES: Final = frozenset({"acompletion", "atext_completion"})
+
 # W3C Trace Context traceparent header: https://www.w3.org/TR/trace-context/
 # e.g. "00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-01"
 _TRACEPARENT_RE: Final = re.compile(r"^[0-9a-f]{2}-([0-9a-f]{32})-[0-9a-f]{16}-[0-9a-f]{2}$", re.IGNORECASE)
@@ -688,7 +691,7 @@ def _metadata_string_tags(metadata_objects: Iterable[Mapping[str, object]]) -> t
 
 
 def _strip_untrusted_request_header_controls(
-    headers: Any,
+    headers: object,
     *,
     allow_client_message_redaction_opt_out: bool = False,
 ) -> None:
@@ -1692,7 +1695,7 @@ class LiteLLMProxyRequestSetup:
 
     @staticmethod
     def add_internal_user_from_user_mapping(
-        general_settings: dict | None,
+        general_settings: Mapping[str, object] | None,
         user_api_key_dict: UserAPIKeyAuth,
         headers: dict,
     ) -> UserAPIKeyAuth:
@@ -1711,7 +1714,7 @@ class LiteLLMProxyRequestSetup:
         return user_api_key_dict
 
     @staticmethod
-    def get_user_from_headers(headers: dict, general_settings: dict | None = None) -> str | None:
+    def get_user_from_headers(headers: dict, general_settings: Mapping[str, object] | None = None) -> str | None:
         """
         Get the user from the specified header if `general_settings.user_header_name` is set.
         """
@@ -1732,7 +1735,9 @@ class LiteLLMProxyRequestSetup:
         return user
 
     @staticmethod
-    def get_openai_org_id_from_headers(headers: dict, general_settings: dict | None = None) -> str | None:
+    def get_openai_org_id_from_headers(
+        headers: dict, general_settings: Mapping[str, object] | None = None
+    ) -> str | None:
         """
         Get the OpenAI Org ID from the headers.
         """
@@ -1823,7 +1828,7 @@ class LiteLLMProxyRequestSetup:
         headers: dict,
         request_data: Mapping[str, object],
         user_api_key_dict: UserAPIKeyAuth,
-        general_settings: dict[str, object] | None = None,
+        general_settings: Mapping[str, object] | None = None,
     ) -> LitellmDataForBackendLLMCall:
         """
         - Adds user from headers
@@ -3285,7 +3290,7 @@ def _per_user_credential_names_for_groups(
         GITHUB_COPILOT_AUTH_TYPE_KEY,
         GITHUB_COPILOT_PER_USER_AUTH_TYPE,
     )
-    from litellm.router_utils.auto_router_model_naming import classify_strategy_router_model
+    from litellm.router_strategy.auto_router.auto_router_model_naming import classify_strategy_router_model
 
     names: Final[list[str]] = []  # mutable-ok: accumulates one name per per-user deployment
     for group in model_groups:

litellm/proxy/management_endpoints/auto_router_endpoints.py

@@ -60,16 +60,16 @@ from litellm.repositories.table_repositories import PrismaTableRepository
 from litellm.repositories.team_repository import TeamRepository
 from litellm.repositories.user_repository import UserRepository
 from litellm.repositories.verification_token_repository import VerificationTokenRepository
+from litellm.router_strategy.auto_router.auto_router_model_naming import (
+    StrategyRouterDependencyRole,
+    classify_strategy_router_model,
+    strategy_router_dependencies,
+)
 from litellm.router_strategy.complexity_router import ComplexityRouter
 from litellm.router_strategy.complexity_router.request_models import (
     AutoRouterRoutingTestRequest,
     RequestComplexityRouterConfig,
 )
-from litellm.router_utils.auto_router_model_naming import (
-    StrategyRouterDependencyRole,
-    classify_strategy_router_model,
-    strategy_router_dependencies,
-)
 from litellm.types.llms.base import LiteLLMBaseModel
 from litellm.types.management_endpoints.auto_router_endpoints import (
     SHADOW_EVAL_TURN_VALVE,
@@ -231,9 +231,12 @@ def _shadow_eval_attempts(prisma_client: "PrismaClient") -> _ShadowEvalAttemptTa
     return _ShadowEvalAttemptRepository(prisma_client).table
 
 
+_RAW_ROWS: Final = TypeAdapter(Sequence[dict[str, object]], config=ConfigDict(strict=True))
+
+
 async def _query_raw(prisma_client: "PrismaClient", query: str, *args: object) -> Sequence[Mapping[str, object]]:
     async with db_span("auto_router_report_query", sql_relation(query)):
-        return await prisma_client.db.query_raw(query, *args)
+        return _RAW_ROWS.validate_python(await prisma_client.db.query_raw(query, *args))
 
 
 async def _execute_raw(prisma_client: "PrismaClient", query: str, *args: object) -> int:
@@ -408,7 +411,7 @@ async def validate_complexity_router_config(
         auto_router_environment_reference_error(data.complexity_router_config, user_api_key_dict)
     )
 
-    from litellm.router_utils.auto_router_model_naming import (
+    from litellm.router_strategy.auto_router.auto_router_model_naming import (
         validate_complexity_router_config_write,
     )
 

litellm/proxy/management_endpoints/common_daily_activity.py

@@ -3,6 +3,7 @@ from collections.abc import Awaitable, Callable, Mapping, Sequence
 from collections.abc import Set as AbstractSet
 from dataclasses import dataclass, replace
 from datetime import date, datetime, timedelta
+from itertools import groupby
 from types import MappingProxyType
 from typing import Final, Literal, NoReturn, Protocol
 
@@ -28,6 +29,10 @@ from litellm.types.proxy.management_endpoints.common_daily_activity import (
     DailySpendData,
     DailySpendMetadata,
     GroupedData,
+    GuardrailKeyMetric,
+    GuardrailMetadata,
+    GuardrailMetric,
+    GuardrailSpendMetrics,
     KeyMetadata,
     KeyMetricWithMetadata,
     MetricWithMetadata,
@@ -39,6 +44,7 @@ from litellm.types.repositories.daily_activity import (
     DailyActivityTable,
     EntityRollupRow,
     GroupingSetsRow,
+    GuardrailSpendRow,
     KeyMetadataRow,
     RollupMetricsRow,
     SpendLogsWindow,
@@ -1013,17 +1019,68 @@ def _fold_entity_rollups_sync(
             )
 
 
+def _guardrail_metric(day: DailySpendData, rows: Sequence[GuardrailSpendRow]) -> GuardrailMetric:
+    return GuardrailMetric(
+        metrics=GuardrailSpendMetrics(
+            spend=sum(row.spend for row in rows),
+            untracked_units=sum(row.untracked_units for row in rows),
+        ),
+        metadata=GuardrailMetadata(guardrail_name=rows[0].guardrail_name),
+        api_key_breakdown={
+            row.api_key: GuardrailKeyMetric(
+                metrics=GuardrailSpendMetrics(spend=row.spend, untracked_units=row.untracked_units),
+                metadata=day.breakdown.api_keys[row.api_key].metadata,
+            )
+            for row in rows
+            if row.api_key in day.breakdown.api_keys
+        },
+    )
+
+
+def _day_with_guardrails(day: DailySpendData, rows: Sequence[GuardrailSpendRow]) -> DailySpendData:
+    by_guardrail: Final = groupby(sorted(rows, key=lambda row: row.guardrail_id), key=lambda row: row.guardrail_id)
+    guardrails: Final = {guardrail_id: _guardrail_metric(day, tuple(group)) for guardrail_id, group in by_guardrail}
+    return day.model_copy(update={"breakdown": day.breakdown.model_copy(update={"guardrails": guardrails})})
+
+
+def _with_guardrail_breakdown(
+    results: Sequence[DailySpendData], rows: Sequence[GuardrailSpendRow]
+) -> tuple[DailySpendData, ...]:
+    rows_by_date: Final = MappingProxyType(
+        {day: tuple(group) for day, group in groupby(sorted(rows, key=lambda row: row.date), key=lambda row: row.date)}
+    )
+    return tuple(_day_with_guardrails(day, rows_by_date.get(day.date.strftime("%Y-%m-%d"), ())) for day in results)
+
+
+async def _guardrail_spend_or_empty(
+    repository: DailyActivityRepository, scope: DailyActivityScope
+) -> tuple[GuardrailSpendRow, ...]:
+    from prisma.errors import PrismaError
+
+    try:
+        return await repository.guardrail_spend(scope)
+    except PrismaError as exc:
+        verbose_proxy_logger.warning("Returning daily activity without its guardrail breakdown: %s", exc)
+        return ()
+
+
 async def get_daily_activity_aggregated(
     repository: DailyActivityRepository,
     scope: DailyActivityScope,
     *,
     entity_metadata_field: Mapping[str, dict[str, object]] | None = None,
     include_entity_breakdown: bool = False,
+    include_guardrail_breakdown: bool = False,
     api_key_limit: int = constants.USAGE_TOP_API_KEYS_DEFAULT,
 ) -> SpendAnalyticsPaginatedResponse:
     try:
-        aggregated_rows: Final = await repository.aggregated(
-            scope, include_entity_breakdown=include_entity_breakdown, api_key_limit=api_key_limit
+        aggregated_rows, guardrail_rows = await asyncio.gather(
+            repository.aggregated(
+                scope, include_entity_breakdown=include_entity_breakdown, api_key_limit=api_key_limit
+            ),
+            _guardrail_spend_or_empty(repository, scope)
+            if include_guardrail_breakdown and not scope.model
+            else asyncio.sleep(0, result=()),
         )
         records: Final = aggregated_rows.grouping_rows
         aggregated: Final = await _aggregate_grouping_sets_records(
@@ -1058,8 +1115,13 @@ async def get_daily_activity_aggregated(
                 api_key_metadata=entity_key_metadata,
                 entity_metadata_field=entity_metadata_field,
             )
+        results: Final = (
+            list(await asyncio.to_thread(_with_guardrail_breakdown, aggregated["results"], guardrail_rows))
+            if guardrail_rows
+            else aggregated["results"]
+        )
         return SpendAnalyticsPaginatedResponse(
-            results=aggregated["results"],
+            results=results,
             metadata=DailySpendMetadata(
                 total_spend=aggregated["totals"].spend,
                 total_flat_cost=aggregated["totals"].flat_cost,

litellm/proxy/management_endpoints/daily_activity_routes.py

@@ -300,6 +300,7 @@ def _register_aggregated_route(router: APIRouter, resolver: EntityScopeResolver,
                 resolved.scope,
                 entity_metadata_field=resolved.entity_metadata,
                 include_entity_breakdown=resolver.include_entity_breakdown,
+                include_guardrail_breakdown=resolver.include_guardrail_breakdown,
                 api_key_limit=api_key_limit,
             )
             if resolver.entity != "team":

litellm/proxy/management_endpoints/daily_activity_scopes.py

@@ -56,6 +56,7 @@ class EntityScopeResolver:
     resolve: EntityScopeResolve
     include_entity_breakdown: bool
     operation_names: OperationNames
+    include_guardrail_breakdown: bool = False
 
 
 def _query_ids(value: str | None) -> tuple[str, ...] | None:
@@ -364,6 +365,7 @@ USER_RESOLVER = EntityScopeResolver(
     query=_user_query,
     resolve=_resolve_user,
     include_entity_breakdown=False,
+    include_guardrail_breakdown=True,
     operation_names=MappingProxyType(
         {
             "aggregated": "get_user_daily_activity_aggregated",
@@ -385,6 +387,7 @@ TEAM_RESOLVER = EntityScopeResolver(
     query=_team_query,
     resolve=_resolve_team,
     include_entity_breakdown=True,
+    include_guardrail_breakdown=True,
     operation_names=MappingProxyType(
         {
             "aggregated": "get_team_daily_activity_aggregated",
@@ -423,6 +426,7 @@ ORGANIZATION_RESOLVER = EntityScopeResolver(
     query=_organization_query,
     resolve=_resolve_organization,
     include_entity_breakdown=True,
+    include_guardrail_breakdown=True,
     operation_names=MappingProxyType(
         {
             "aggregated": "get_organization_daily_activity_aggregated",

litellm/proxy/management_endpoints/internal_user_endpoints.py

@@ -499,7 +499,7 @@ async def new_user(
     - permissions: Optional[dict] - [Not Implemented Yet] User-specific permissions, eg. turning off pii masking.
     - metadata: Optional[dict] - Metadata for user, store information for user. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" }
     - max_parallel_requests: Optional[int] - Rate limit a user based on the number of parallel requests. Raises 429 error, if user's parallel requests > x.
-    - model_max_budget: Optional[dict] - Model-specific max budget for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys)
+    - model_max_budget: Optional[dict] - Model-specific max budget for user, keyed by model name, `*` pattern, or model access group. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys)
     - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
     - model_rpm_limit: Optional[float] - Model-specific rpm limit for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-limits-to-keys)
     - mcp_rpm_limit: Optional[dict] - Per-MCP-server rpm limit, keyed by MCP server name {"github": 100, "slack": 200}. Enforced for keys and teams only; values set on a user are stored but not enforced per user.
@@ -1702,7 +1702,7 @@ async def user_update(
         - permissions: Optional[dict] - [Not Implemented Yet] User-specific permissions, eg. turning off pii masking.
         - metadata: Optional[dict] - Metadata for user, store information for user. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" }
         - max_parallel_requests: Optional[int] - Rate limit a user based on the number of parallel requests. Raises 429 error, if user's parallel requests > x.
-        - model_max_budget: Optional[dict] - Model-specific max budget for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys)
+        - model_max_budget: Optional[dict] - Model-specific max budget for user, keyed by model name, `*` pattern, or model access group. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-budgets-to-keys)
         - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
         - model_rpm_limit: Optional[float] - Model-specific rpm limit for user. [Docs](https://docs.litellm.ai/docs/proxy/users#add-model-specific-limits-to-keys)
         - mcp_rpm_limit: Optional[dict] - Per-MCP-server rpm limit, keyed by MCP server name {"github": 100, "slack": 200}. Enforced for keys and teams only; values set on a user are stored but not enforced per user.

litellm/proxy/management_endpoints/key_management_endpoints.py

@@ -2328,7 +2328,7 @@ async def generate_key_fn(
     - throttle_on_budget_exceeded: Optional[bool] - When the key exceeds its max_budget, throttle its tpm/rpm to the global budget_exceeded_throttle_percentage instead of blocking the key entirely.
     - enable_prompt_caching: Optional[bool] - Auto-inject prompt caching breakpoints (Anthropic cache_control markers) on requests made with this key. Supported Claude models on Anthropic, Bedrock, Vertex AI, and Azure AI only.
     - permissions: Optional[dict] - key-specific permissions. Currently just used for turning off pii masking (if connected). Example - {"pii": false}
-    - model_max_budget: Optional[Dict[str, BudgetConfig]] - Model-specific budgets {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}}. IF null or {} then no model specific budget.
+    - model_max_budget: Optional[Dict[str, BudgetConfig]] - Model-specific budgets keyed by model name, `*` pattern like "claude-opus*", or model access group, e.g. {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}}. IF null or {} then no model specific budget.
     - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
     - model_rpm_limit: Optional[dict] - key-specific model rpm limit. Example - {"text-davinci-002": 1000, "gpt-3.5-turbo": 1000}. IF null or {} then no model specific rpm limit.
     - model_tpm_limit: Optional[dict] - key-specific model tpm limit. Example - {"text-davinci-002": 1000, "gpt-3.5-turbo": 1000}. IF null or {} then no model specific tpm limit.
@@ -2538,7 +2538,7 @@ async def generate_service_account_key_fn(
     - metadata: Optional[dict] - Metadata for key, store information for key. Example metadata = {"team": "core-infra", "app": "app2", "email": "ishaan@berri.ai" }
     - guardrails: Optional[List[str]] - List of active guardrails for the key
     - permissions: Optional[dict] - key-specific permissions. Currently just used for turning off pii masking (if connected). Example - {"pii": false}
-    - model_max_budget: Optional[Dict[str, BudgetConfig]] - Model-specific budgets {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}}. IF null or {} then no model specific budget.
+    - model_max_budget: Optional[Dict[str, BudgetConfig]] - Model-specific budgets keyed by model name, `*` pattern like "claude-opus*", or model access group, e.g. {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}}. IF null or {} then no model specific budget.
     - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
     - model_rpm_limit: Optional[dict] - key-specific model rpm limit. Example - {"text-davinci-002": 1000, "gpt-3.5-turbo": 1000}. IF null or {} then no model specific rpm limit.
     - model_tpm_limit: Optional[dict] - key-specific model tpm limit. Example - {"text-davinci-002": 1000, "gpt-3.5-turbo": 1000}. IF null or {} then no model specific tpm limit.
@@ -3855,7 +3855,7 @@ async def update_key_fn(
     - enforced_params: Optional[List[str]] - List of enforced params for the key (Enterprise only). [Docs](https://docs.litellm.ai/docs/proxy/enterprise#enforce-required-params-for-llm-requests)
     - spend: Optional[float] - Amount spent by key
     - max_budget: Optional[float] - Max budget for key
-    - model_max_budget: Optional[Dict[str, BudgetConfig]] - Model-specific budgets {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}
+    - model_max_budget: Optional[Dict[str, BudgetConfig]] - Model-specific budgets keyed by model name, `*` pattern like "claude-opus*", or model access group, e.g. {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}
     - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
     - budget_duration: Optional[str] - Budget reset period ("30d", "1h", etc.)
     - soft_budget: Optional[float] - Soft budget limit (warning vs. hard stop). Will trigger a slack alert when this soft budget is reached. Set to null to remove the soft budget.
@@ -6303,7 +6303,7 @@ async def regenerate_key_fn(
         - tags: Optional[List[str]] - Tags for organizing keys (Enterprise only)
         - spend: Optional[float] - Amount spent by key
         - max_budget: Optional[float] - Max budget for key
-        - model_max_budget: Optional[Dict[str, BudgetConfig]] - Model-specific budgets {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}
+        - model_max_budget: Optional[Dict[str, BudgetConfig]] - Model-specific budgets keyed by model name, `*` pattern like "claude-opus*", or model access group, e.g. {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}
         - budget_fallbacks: Optional[Dict[str, List[str]]] - Per-model fallback chain tried in order when that model's own `model_max_budget` is exceeded, e.g. {"gpt-4o": ["gpt-4o-mini"]}.
         - budget_duration: Optional[str] - Budget reset period ("30d", "1h", etc.)
         - soft_budget: Optional[float] - Soft budget limit (warning vs. hard stop). Will trigger a slack alert when this soft budget is reached.

litellm/proxy/management_endpoints/model_management_endpoints.py

@@ -131,6 +131,23 @@ from litellm.repositories.project_repository import ProjectRepository
 from litellm.repositories.table_repositories import ModelTableRepository
 from litellm.repositories.team_repository import TeamRepository
 from litellm.router import Router
+from litellm.router_strategy.auto_router.auto_router_model_naming import (
+    GATED_AUTO_ROUTER_CAPABILITIES,
+    STRATEGY_ROUTER_PARAM_FIELDS,
+    GatedAutoRouterCapability,
+    capability_limit_violation,
+    carries_complexity_router_settings,
+    count_capability_routers,
+    gated_capabilities_of,
+    is_complexity_router_model,
+    validate_complexity_router_config_placement,
+    validate_complexity_router_config_write,
+    validate_strategy_router_model_write,
+)
+from litellm.router_strategy.auto_router.auto_router_tuning_baseline import (
+    is_mutable_tuned_candidate,
+    tuning_quota_violation,
+)
 from litellm.router_strategy.complexity_router import (
     DEFAULT_CLASSIFIER_CONTEXT_WINDOW_SIZE,
     ClassificationRubric,
@@ -144,20 +161,6 @@ from litellm.router_strategy.complexity_router import (
     normalize_classification_prompt,
 )
 from litellm.router_strategy.complexity_router.config import resolve_complexity_router_config_write
-from litellm.router_utils.auto_router_model_naming import (
-    GATED_AUTO_ROUTER_CAPABILITIES,
-    STRATEGY_ROUTER_PARAM_FIELDS,
-    GatedAutoRouterCapability,
-    capability_limit_violation,
-    carries_complexity_router_settings,
-    count_capability_routers,
-    gated_capabilities_of,
-    is_complexity_router_model,
-    validate_complexity_router_config_placement,
-    validate_complexity_router_config_write,
-    validate_strategy_router_model_write,
-)
-from litellm.router_utils.auto_router_tuning_baseline import is_mutable_tuned_candidate, tuning_quota_violation
 from litellm.types.llms.base import LiteLLMBaseModel
 from litellm.types.llms.bedrock import AwsSessionTag
 from litellm.types.proxy.management_endpoints.model_management_endpoints import (

litellm/proxy/management_endpoints/router_settings_endpoints.py

@@ -10,7 +10,7 @@ GET /router/fields - Get router settings field definitions without values (for U
 import inspect
 from collections.abc import Mapping
 from types import MappingProxyType
-from typing import Any, Final, get_args
+from typing import Final, get_args
 
 from fastapi import APIRouter, Depends
 from pydantic import Field
@@ -32,7 +32,7 @@ router: Final = APIRouter()
 
 class RouterSettingsResponse(LiteLLMBaseModel):
     fields: list[RouterSettingsField] = Field(description="List of all configurable router settings with metadata")
-    current_values: dict[str, Any] = Field(description="Current values of router settings")
+    current_values: dict[str, object] = Field(description="Current values of router settings")
     routing_strategy_descriptions: dict[str, str] = Field(description="Descriptions for each routing strategy option")
     routing_group_strategies: tuple[str, ...] = Field(
         description="Strategies supported when constructing a routing group"

litellm/proxy/management_endpoints/team_endpoints.py

@@ -11,6 +11,7 @@ All /team management endpoints
 
 import asyncio
 import copy
+import itertools
 import json
 import math
 import traceback
@@ -139,7 +140,12 @@ from litellm.proxy.common_utils.user_api_key_cache import (
 from litellm.proxy.db.db_span import db_span
 from litellm.proxy.hooks.key_management_event_hooks import KeyManagementEventHooks
 from litellm.proxy.hooks.model_max_budget_limiter import (
+    RoutedModels,
     build_model_max_budget_usage,
+    entries_outranked_by,
+    entry_names_model,
+    is_enforced_budget_entry,
+    proxy_routed_models,
     resolve_model_budget,
 )
 from litellm.proxy.management.teams.authz import TEAM_OR_ORG_ADMIN, TeamRole, is_team_admin, team_access_denied
@@ -1467,8 +1473,10 @@ def _check_team_model_budget_update_authority(
     data: UpdateTeamRequest,
     user_api_key_dict: UserAPIKeyAuth,
     existing_model_max_budget: Mapping[str, object] | None,
+    routed_models_source: Callable[[str | None], RoutedModels] = proxy_routed_models,
 ) -> None:
-    """Like `_check_team_budget_update_authority`: only a proxy admin may raise, re-window or drop a per-model cap."""
+    """Like `_check_team_budget_update_authority`: only a proxy admin may raise, re-window or drop a per-model cap, or
+    add an entry that takes precedence over one."""
     if user_api_key_dict.user_role == LitellmUserRoles.PROXY_ADMIN:
         return
     if "model_max_budget" not in data.model_fields_set or not existing_model_max_budget:
@@ -1487,29 +1495,97 @@ def _check_team_model_budget_update_authority(
                 )
             },
         )
-    for model_name, proposed in requested.items():
-        governing = resolve_model_budget(model=model_name, model_max_budget=existing_model_max_budget)
-        if governing is None:
-            continue
-        cap = governing.budget_config
-        if cap.max_budget is None:
-            continue
-        if (
-            proposed.max_budget is None
-            or proposed.max_budget > cap.max_budget
-            or proposed.budget_duration != cap.budget_duration
-        ):
-            raise HTTPException(
-                status_code=403,
-                detail={
-                    "error": (
-                        f"Only a proxy admin can raise a team's model_max_budget for {model_name!r} or change its "
-                        f"budget_duration. Current max_budget={cap.max_budget} per {cap.budget_duration} "
-                        f"(entry {governing.budget_model!r}), requested={proposed.max_budget} per "
-                        f"{proposed.budget_duration}."
-                    )
-                },
+    routed_models: Final = routed_models_source(data.team_id)
+    for entry, proposed in requested.items():
+        existing = _existing_model_cap(existing_model_max_budget.get(entry))
+        if existing is not None:
+            _check_existing_model_cap_kept(entry=entry, cap_entry=entry, proposed=proposed, existing=existing)
+            if is_enforced_budget_entry(existing) or not is_enforced_budget_entry(proposed):
+                continue
+        governing = resolve_model_budget(model=entry, model_max_budget=existing_model_max_budget, team_id=data.team_id)
+        if governing is not None and entry_names_model(entry=governing.budget_model, model=entry):
+            _check_existing_model_cap_kept(
+                entry=entry, cap_entry=governing.budget_model, proposed=proposed, existing=governing.budget_config
             )
+        _check_new_entry_outranks_no_cap(
+            entry=entry,
+            existing_model_max_budget=existing_model_max_budget,
+            routed_models=routed_models,
+            team_id=data.team_id,
+        )
+
+
+def _check_existing_model_cap_kept(entry: str, cap_entry: str, proposed: BudgetConfig, existing: BudgetConfig) -> None:
+    if existing.max_budget is None or not _looser_model_cap(proposed=proposed, cap=existing):
+        return
+    raise HTTPException(
+        status_code=403,
+        detail={
+            "error": (
+                f"Only a proxy admin can raise a team's model_max_budget for {entry!r} or change its "
+                f"budget_duration. Current max_budget={existing.max_budget} per {existing.budget_duration} "
+                f"(entry {cap_entry!r}), requested={proposed.max_budget} per {proposed.budget_duration}."
+            )
+        },
+    )
+
+
+def _check_new_entry_outranks_no_cap(
+    entry: str,
+    existing_model_max_budget: Mapping[str, object],
+    routed_models: RoutedModels,
+    team_id: str | None,
+) -> None:
+    """A new entry gets its own counter, so the models it takes from a capped entry stop drawing from that cap."""
+    outranked: Final = next(
+        _capped_entries_outranked_by(
+            entry=entry,
+            existing_model_max_budget=existing_model_max_budget,
+            routed_models=routed_models,
+            team_id=team_id,
+        ),
+        None,
+    )
+    if outranked is None:
+        return
+    outranked_name, cap = outranked
+    raise HTTPException(
+        status_code=403,
+        detail={
+            "error": (
+                f"Only a proxy admin can add a team model_max_budget entry {entry!r} that takes precedence over "
+                f"{outranked_name!r}, since the models they share would stop drawing from its max_budget="
+                f"{cap.max_budget} per {cap.budget_duration}."
+            )
+        },
+    )
+
+
+def _capped_entries_outranked_by(
+    entry: str,
+    existing_model_max_budget: Mapping[str, object],
+    routed_models: RoutedModels,
+    team_id: str | None,
+) -> Iterator[tuple[str, BudgetConfig]]:
+    governing: Final = resolve_model_budget(model=entry, model_max_budget=existing_model_max_budget, team_id=team_id)
+    outranked_names: Final = itertools.chain(
+        ()
+        if governing is None or entry_names_model(entry=governing.budget_model, model=entry)
+        else (governing.budget_model,),
+        entries_outranked_by(entry=entry, model_max_budget=existing_model_max_budget, routed_models=routed_models),
+    )
+    outranked_caps: Final = ((name, _existing_model_cap(existing_model_max_budget[name])) for name in outranked_names)
+    yield from (
+        (name, cap) for name, cap in outranked_caps if name != entry and cap is not None and cap.max_budget is not None
+    )
+
+
+def _looser_model_cap(proposed: BudgetConfig, cap: BudgetConfig) -> bool:
+    return (
+        proposed.max_budget is None
+        or (cap.max_budget is not None and proposed.max_budget > cap.max_budget)
+        or (is_enforced_budget_entry(cap) and proposed.budget_duration != cap.budget_duration)
+    )
 
 
 def _should_auto_add_team_creator(
@@ -1572,7 +1648,7 @@ async def new_team(
     - prompts: Optional[List[str]] - List of prompts that the team is allowed to use.
     - organization_id: Optional[str] - The organization id of the team. Default is None. Create via `/organization/new`.
     - model_aliases: Optional[dict] - Model aliases for the team. [Docs](https://docs.litellm.ai/docs/proxy/team_based_routing#create-team-with-model-alias)
-    - model_max_budget: Optional[dict] - Per-model max budget every key on the team inherits unless the key sets its own for that model. Example: {"gpt-4o": {"max_budget": 10, "budget_duration": "1d"}}
+    - model_max_budget: Optional[dict] - Per-model max budget every key on the team inherits unless the key sets its own for that model, keyed by model name, `*` pattern, or model access group. Example: {"gpt-4o": {"max_budget": 10, "budget_duration": "1d"}}
     - guardrails: Optional[List[str]] - Guardrails for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails)
     - policies: Optional[List[str]] - Policies for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails/guardrail_policies)
     - disable_global_guardrails: Optional[bool] - Whether to disable global guardrails for the team. Proxy admin only.
@@ -2350,7 +2426,7 @@ async def update_team(
     - tags: Optional[List[str]] - Tags for [tracking spend](https://litellm.vercel.app/docs/proxy/enterprise#tracking-spend-for-custom-tags) and/or doing [tag-based routing](https://litellm.vercel.app/docs/proxy/tag_routing).
     - organization_id: Optional[str] - The organization id of the team. Default is None. Create via `/organization/new`.
     - model_aliases: Optional[dict] - Model aliases for the team. [Docs](https://docs.litellm.ai/docs/proxy/team_based_routing#create-team-with-model-alias)
-    - model_max_budget: Optional[dict] - Per-model max budget every key on the team inherits unless the key sets its own for that model. Example: {"gpt-4o": {"max_budget": 10, "budget_duration": "1d"}}
+    - model_max_budget: Optional[dict] - Per-model max budget every key on the team inherits unless the key sets its own for that model, keyed by model name, `*` pattern, or model access group. Example: {"gpt-4o": {"max_budget": 10, "budget_duration": "1d"}}
     - guardrails: Optional[List[str]] - Guardrails for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails)
     - policies: Optional[List[str]] - Policies for the team. [Docs](https://docs.litellm.ai/docs/proxy/guardrails/guardrail_policies)
     - disable_global_guardrails: Optional[bool] - Whether to disable global guardrails for the team. Proxy admin only.

litellm/proxy/management_helpers/auto_router_availability.py

@@ -7,14 +7,14 @@ from typing import Final
 from pydantic import Json, TypeAdapter, ValidationError
 
 from litellm.proxy.common_utils.encrypt_decrypt_utils import decrypt_value_helper
-from litellm.router_utils.auto_router_model_naming import (
+from litellm.router_strategy.auto_router.auto_router_model_naming import (
     GATED_AUTO_ROUTER_CAPABILITIES,
     capability_limit_violation,
     classify_strategy_router_model,
     count_capability_routers,
     gated_capabilities_of,
 )
-from litellm.router_utils.auto_router_tuning_baseline import (
+from litellm.router_strategy.auto_router.auto_router_tuning_baseline import (
     is_mutable_tuned_candidate,
     mutable_tuned_identities,
     tuning_quota_violation,