Merge remote-tracking branch 'origin/main' into litellm_fireworks_minimax_m3_vision_pin

This commit is contained in:
mateo-berri 2026-10-28 18:24:20 -07:00
commit 28c3a27265
313 changed files with 13028 additions and 2031 deletions

View file

@ -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

1
.github/CODEOWNERS vendored
View file

@ -3,3 +3,4 @@
/ci_cd/cost_map_pins.json @mateo-berri @ryan-crabbe-berri @kerry-berri @joshua-berri
/helm/ @yassin-berriai
/terraform/ @yassin-berriai
/.github/workflows/zizmor.yml @yuneng-berri

View file

@ -1,7 +1,7 @@
name: Cost map guard
on: # zizmor: ignore[dangerous-triggers] runs the base branch's code only; the PR's cost map files are read as data and never executed
pull_request_target:
on:
pull_request:
branches:
- main
- "litellm_**"
@ -19,6 +19,7 @@ jobs:
steps:
- uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
ref: ${{ github.event.pull_request.base.sha }}
persist-credentials: false
- name: Fetch the pull request head, its merge base, and the base branch tip
id: revisions

View file

@ -1,13 +1,9 @@
name: Spend logs guard
on: # zizmor: ignore[dangerous-triggers] runs the base branch's code only; the PR's diff is read as data and never executed
pull_request_target:
on:
pull_request:
branches:
- main
pull_request_review:
types:
- submitted
- dismissed
permissions:
contents: read
@ -23,7 +19,7 @@ jobs:
steps:
- uses: actions/checkout@08eba0b27e820071cde6df949e0beb9ba4906955 # v4.3.0
with:
ref: ${{ github.event.pull_request.base.ref }}
ref: ${{ github.event.pull_request.base.sha }}
persist-credentials: false
- name: Fetch the pull request head, its merge base and its reviews
id: revisions

View file

@ -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

View file

@ -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

View file

@ -27,10 +27,15 @@ jobs:
with:
persist-credentials: false
- name: Install zizmor
run: |
curl -fsSL --retry 3 -o "$RUNNER_TEMP/zizmor.tar.gz" \
https://github.com/zizmorcore/zizmor/releases/download/v1.31.1/zizmor-x86_64-unknown-linux-gnu.tar.gz
echo "d185f726bdaf92df78273217dfdd24ebcd6b0ef48cf3ce776ff5762b6e138c99 $RUNNER_TEMP/zizmor.tar.gz" | sha256sum -c -
tar xzf "$RUNNER_TEMP/zizmor.tar.gz" -C "$RUNNER_TEMP" zizmor
- name: Run zizmor
uses: zizmorcore/zizmor-action@5f14fd08f7cf1cb1609c1e344975f152c7ee938d # v0.5.6
with:
version: "1.24.1"
min-severity: medium
advanced-security: false
annotations: true
env:
GH_TOKEN: ${{ github.token }}
run: |
"$RUNNER_TEMP/zizmor" --no-ignores --no-config --min-severity=medium --format=github .

View file

@ -2,9 +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. 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
@ -13,8 +16,10 @@ import argparse
import re
import subprocess
import sys
from collections.abc import 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
@ -25,7 +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):
@ -37,31 +45,77 @@ class Review(BaseModel):
model_config = ConfigDict(frozen=True, extra="ignore")
state: str
user: Reviewer
commit_id: str | None = None
_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:
@ -70,11 +124,33 @@ 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"
def latest_approval(reviews: Sequence[Review], login: str) -> Review | None:
"""`login`'s most recent APPROVED review on the pull request, the one GitHub keeps in force after pushes."""
own: Final = tuple(review for review in reviews if review.user.login == login and review.state == "APPROVED")
return own[-1] if own else None
def approval_covers(
approval: Review, head: str, hits: Sequence[str], hits_at: Callable[[str], Sequence[str] | None]
) -> bool:
"""True when the approval is on the head, or the head adds no spend-log line the approved commit did not."""
if approval.commit_id is None:
return False
if approval.commit_id == head:
return True
reviewed: Final = hits_at(approval.commit_id)
return reviewed is not None and set(hits) <= set(reviewed)
def added_diff(base: str, head: str) -> str:
result: Final = subprocess.run(
("git", "diff", "--unified=0", "--no-color", f"{base}..{head}", "--", *WATCHED, MIGRATIONS_PREFIX),
@ -85,6 +161,27 @@ 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), schema_blocks(base, commit)) if fetched(commit) else None
class Arguments(BaseModel):
model_config = ConfigDict(frozen=True)
base: str
@ -100,18 +197,28 @@ 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
if approved_by(_REVIEWS.validate_json(args.reviews.read_bytes()), args.approver):
reviews: Final = _REVIEWS.validate_json(args.reviews.read_bytes())
approval: Final = latest_approval(reviews, args.approver) if approved_by(reviews, args.approver) else None
if approval is not None and approval_covers(
approval, args.head, hits, lambda commit: reviewed_hits(args.base, commit)
):
sys.stdout.write(f"LiteLLM_SpendLogs change approved by {args.approver}: {', '.join(hits)}\n")
return 0
listed: Final = "".join(f" {hit}\n" for hit in hits)
sys.stdout.write(
reason: Final = (
f"This pull request changes LiteLLM_SpendLogs and needs an approving review from @{args.approver}:\n"
f"{listed}The check re-runs when a review is submitted.\n"
if approval is None
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}After approving, re-run this job.\n")
return 1

View file

@ -32,6 +32,7 @@ from litellm.integrations.custom_guardrail import (
CustomGuardrail,
log_guardrail_information,
)
from litellm.llms.base_llm.guardrail_translation.utils import effective_skip_tool_message_for_guardrail
from litellm.proxy._types import UserAPIKeyAuth
from litellm.proxy.guardrails._content_utils import iter_tool_json_fields, map_tool_json_text, walk_user_text
from litellm.types.guardrails import GuardrailEventHooks
@ -753,7 +754,12 @@ class _ENTERPRISE_SecretDetection(CustomGuardrail):
_masked_entity_count.set({})
# Covers multimodal list content + Responses-API input.
inspected = walk_user_text(data, self.redact_text, self.redact_tool_payload) + self._redact_prompt(data)
inspected = walk_user_text(
data,
self.redact_text,
self.redact_tool_payload,
skip_tool_message=effective_skip_tool_message_for_guardrail(self),
) + self._redact_prompt(data)
if inspected == 0:
# Image-only, empty-text, and unsupported payloads inspected

View file

@ -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":

View file

@ -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

View file

@ -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

View file

@ -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

View file

@ -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

View file

@ -4387,12 +4387,17 @@ name = "litellm-router"
version = "0.1.0"
dependencies = [
"litellm-config",
"litellm-core-utils",
"litellm-inference",
"litellm-inference-messages",
"litellm-router-types",
"rstest",
]
[[package]]
name = "litellm-router-core"
version = "0.1.0"
[[package]]
name = "litellm-router-types"
version = "0.1.0"

View file

@ -1,13 +1,25 @@
pub fn parse_str_bool(value: &str) -> Option<bool> {
let token = value.trim_matches(|character: char| {
fn python_strip(value: &str) -> &str {
value.trim_matches(|character: char| {
character.is_whitespace() || matches!(character, '\u{1c}'..='\u{1f}')
});
})
}
pub fn parse_str_bool(value: &str) -> Option<bool> {
let token = python_strip(value);
if token.eq_ignore_ascii_case("true") {
return Some(true);
}
token.eq_ignore_ascii_case("false").then_some(false)
}
pub fn parse_pydantic_bool(value: &str) -> Option<bool> {
match python_strip(value).to_ascii_lowercase().as_str() {
"1" | "on" | "t" | "true" | "y" | "yes" => Some(true),
"0" | "off" | "f" | "false" | "n" | "no" => Some(false),
_ => None,
}
}
/// `redis-py` string Booleans: only `1`, `true`, and `yes` (case-insensitive) are true.
pub fn parse_redis_bool(value: &str) -> bool {
value == "1" || value.eq_ignore_ascii_case("true") || value.eq_ignore_ascii_case("yes")
@ -34,4 +46,27 @@ mod tests {
) {
assert_eq!(parse_str_bool(input), expected, "{input:?}");
}
#[rstest]
#[case::yes(" YES ", Some(true))]
#[case::on("On", Some(true))]
#[case::one("1", Some(true))]
#[case::t("t", Some(true))]
#[case::y("y", Some(true))]
#[case::control_whitespace_true("\u{1c}TRUE\u{1f}", Some(true))]
#[case::no("No", Some(false))]
#[case::off("OFF", Some(false))]
#[case::zero("0", Some(false))]
#[case::f("F", Some(false))]
#[case::n("n", Some(false))]
#[case::false_word("false", Some(false))]
#[case::two("2", None)]
#[case::empty("", None)]
#[case::unknown("maybe", None)]
fn boolean_words_follow_pydantic_lax_bool_parsing(
#[case] input: &str,
#[case] expected: Option<bool>,
) {
assert_eq!(parse_pydantic_bool(input), expected, "{input:?}");
}
}

View file

@ -37,6 +37,7 @@ pub fn build_inference(config: &Config) -> Result<Arc<Gateway>, Error> {
.map(|value| value.expose().to_owned())
.or_else(|| std::env::var(name).ok())
};
let router = ModelRouter::from_config(config, &lookup);
let settings = &config.litellm_settings;
let http_settings = HttpSettings::from_layers([
HttpSettingsLayer::from_environment(&lookup),
@ -64,12 +65,7 @@ pub fn build_inference(config: &Config) -> Result<Arc<Gateway>, Error> {
Arc::new(EnvironmentSecrets::python_compatible(client)),
));
let resources = CoreResources::new(pool);
Ok(Arc::new(Gateway::new(
resources,
http,
secrets,
ModelRouter::from_model_list(&config.model_list),
)?))
Ok(Arc::new(Gateway::new(resources, http, secrets, router)?))
}
pub async fn build_redis(config: &Config) -> Result<RedisClients, Error> {

View file

@ -141,9 +141,13 @@ async fn provider_call(
&env_lookup,
)?;
let body = config.transform_anthropic_messages_request(
trimmed,
&MessagesTransformContext::new(shaping.capabilities, shaping.settings.drop_params),
let body = config.request_body(
config.transform_anthropic_messages_request(
trimmed,
&MessagesTransformContext::new(shaping.capabilities, shaping.settings.drop_params),
)?,
&validated.headers,
shaping.settings.drop_params,
)?;
let environment = ValidatedEnvironment {
headers: config.request_headers(
@ -514,6 +518,84 @@ mod tests {
);
}
fn vertex_call_requesting(
shaping: MessagesShaping,
requested: &str,
drop_params: bool,
) -> MessagesCall {
MessagesCall {
body: body(
json!({"model": "claude-sonnet-4-5", "messages": [{"role": "user", "content": "hi"}], "max_tokens": 16}),
),
api_key: None,
api_base: Some("https://vertex.test".into()),
custom_llm_provider: Some("vertex_ai".into()),
litellm_params: Default::default(),
extra_headers: Some(Map::from_iter([
("authorization".to_string(), json!("Bearer vertex-token")),
("anthropic-version".to_string(), json!(requested)),
])),
provider_specific_header: None,
timeout: None,
shaping: MessagesShaping {
settings: MessagesSettings {
drop_params,
..shaping.settings
},
..shaping
},
}
}
fn vertex_project(name: &str) -> Option<String> {
(name == "VERTEXAI_PROJECT").then(|| "version-project".to_string())
}
#[rstest]
#[case::sdk_2("2026-10-21", false, "2026-10-21")]
#[case::sdk_1("2023-06-01", false, "vertex-2023-10-16")]
#[case::unknown_under_drop_params("latest", true, "vertex-2023-10-16")]
#[tokio::test]
async fn vertex_gets_the_requested_version_in_the_body_alone(
shaping: MessagesShaping,
#[case] requested: &str,
#[case] drop_params: bool,
#[case] body_version: &str,
) {
let call = provider_call_with_secrets(
vertex_call_requesting(shaping, requested, drop_params),
&vertex_project,
)
.await
.unwrap();
let version_headers: Vec<&str> = call
.wire
.headers
.iter()
.filter(|(name, _)| name.eq_ignore_ascii_case("anthropic-version"))
.map(|(_, value)| value.as_str())
.collect();
assert_eq!(
(version_headers, &call.wire.body["anthropic_version"]),
(Vec::<&str>::new(), &json!(body_version))
);
}
#[rstest]
#[tokio::test]
async fn vertex_refuses_an_unsupported_version_without_drop_params(shaping: MessagesShaping) {
assert!(matches!(
provider_call_with_secrets(
vertex_call_requesting(shaping, "latest", false),
&vertex_project
)
.await,
Err(Error::Llm(litellm_llms::Error::InvalidRequest(
litellm_llms::ErrorDetail::UnsupportedParameter { .. }
)))
));
}
#[rstest]
#[tokio::test]
async fn missing_credentials_are_reported_before_the_request_is_transformed(

View file

@ -77,6 +77,15 @@ pub trait BaseMessagesConfig: Sync {
headers
}
fn request_body(
&self,
request: MessagesRequest,
_headers: &Headers,
_drop_params: bool,
) -> Result<MessagesRequest, Error> {
Ok(request)
}
/// The JSON that goes on the wire, for a host whose body differs from the typed request:
/// Python's configs `pop("model")` when the model is addressed by the URL.
fn wire_body(&self, body: Value) -> Value {

View file

@ -1,6 +1,7 @@
# rules
- Claude on Vertex AI: reuses `anthropic/messages` shaping, moves the model into the publisher URL, puts the Vertex `anthropic_version` in the body, and authenticates with a Google access token
- Claude on Vertex AI: reuses `anthropic/messages` shaping, moves the model into the publisher URL, and authenticates with a Google access token
- Vertex reads the API version only from the body's `anthropic_version` and answers 400 when an `anthropic-version` header names a different one, so the caller's header never reaches Vertex. A version Vertex accepts in the body is carried there, `2023-06-01`, `2023-01-01`, or no header get `vertex-2023-10-16`, and any other version drops under `drop_params` and is refused without it
- Applies Vertex's beta policy and strips `output_config.effort` for models that do not accept it
# references

View file

@ -3,10 +3,10 @@ use litellm_auth_gcp::{
VertexConfig, get_vertex_ai_location, get_vertex_ai_project,
get_vertex_ai_project_from_credentials, secret_names,
};
use litellm_http::request::has_header;
use litellm_http::request::{has_header, header_value, without_headers};
use litellm_llms_types::{
formats::messages::{MessagesOptionalParams, MessagesRequest, MessagesTool, OutputConfig},
providers::anthropic::{AnthropicBeta, BetaProvider, BetaSet},
providers::anthropic::{AnthropicBeta, BetaProvider, BetaSet, VERSION_HEADER},
recognized::Recognized,
};
use litellm_router_types::LitellmParams;
@ -36,6 +36,8 @@ use litellm_llms_types::providers::vertex_ai::{
};
const WEB_SEARCH_TOOL_PREFIX: &str = "web_search";
const VERSIONS_SERVED_AS_VERTEX_DEFAULT: [&str; 3] = ["", "2023-01-01", "2023-06-01"];
const VERSIONS_VERTEX_ACCEPTS_IN_BODY: [&str; 2] = [ANTHROPIC_VERSION, "2026-10-21"];
/// Claude on Vertex AI, Python's `VertexAIPartnerModelsAnthropicMessagesConfig`: the Anthropic
/// payload addressed to a project and location, authenticated with a Google access token.
@ -90,12 +92,6 @@ impl BaseMessagesConfig for VertexAiPartnerModelsAnthropicMessagesConfig {
.params
.output_config
.and_then(|config| sanitize_output_config(config, accepts_effort)),
extra: request
.params
.extra
.into_iter()
.chain([("anthropic_version".to_string(), json!(ANTHROPIC_VERSION))])
.collect(),
..request.params
},
..request
@ -133,7 +129,35 @@ impl BaseMessagesConfig for VertexAiPartnerModelsAnthropicMessagesConfig {
}
fn request_headers(&self, headers: Headers, request: &MessagesRequest) -> Headers {
merge_beta_headers(headers, vertex_feature_betas(request))
merge_beta_headers(
without_headers(headers, &[VERSION_HEADER]),
vertex_feature_betas(request),
)
}
fn request_body(
&self,
request: MessagesRequest,
headers: &Headers,
drop_params: bool,
) -> Result<MessagesRequest, Error> {
let version = vertex_version(
&request.model,
header_value(headers, VERSION_HEADER),
drop_params,
)?;
Ok(MessagesRequest {
params: MessagesOptionalParams {
extra: request
.params
.extra
.into_iter()
.chain([("anthropic_version".to_string(), json!(version))])
.collect(),
..request.params
},
..request
})
}
/// Vertex addresses the model in the URL and rejects it in the body.
@ -225,6 +249,26 @@ fn vertex_feature_betas(request: &MessagesRequest) -> BetaSet {
)
}
fn vertex_version<'a>(
model: &str,
requested: Option<&'a str>,
drop_params: bool,
) -> Result<&'a str, Error> {
match requested.map(str::trim).unwrap_or_default() {
version if VERSIONS_SERVED_AS_VERTEX_DEFAULT.contains(&version) => Ok(ANTHROPIC_VERSION),
version if VERSIONS_VERTEX_ACCEPTS_IN_BODY.contains(&version) => Ok(version),
_ if drop_params => Ok(ANTHROPIC_VERSION),
version => Err(Error::InvalidRequest(
crate::ErrorDetail::UnsupportedParameter {
model: model.into(),
param: VERSION_HEADER.into(),
value: version.into(),
hint: String::new(),
},
)),
}
}
#[cfg(test)]
mod tests {
use rstest::rstest;
@ -471,6 +515,28 @@ mod tests {
);
}
#[rstest]
#[case::sdk_spelling("anthropic-version")]
#[case::any_case("Anthropic-Version")]
fn a_forwarded_anthropic_version_never_reaches_vertex(#[case] name: &str) {
let request: MessagesRequest = serde_json::from_value(json!({
"model": "claude-sonnet-4-5",
"max_tokens": 16,
"messages": [{"role": "user", "content": "hi"}],
}))
.unwrap();
assert_eq!(
VERTEX_ANTHROPIC_MESSAGES_CONFIG.request_headers(
vec![
(name.to_string(), "2026-10-21".to_string()),
("x-request-id".to_string(), "req-1".to_string()),
],
&request,
),
vec![("x-request-id".to_string(), "req-1".to_string())]
);
}
fn transformed(value: Value, context: &MessagesTransformContext) -> Value {
let request: MessagesRequest = serde_json::from_value(value).unwrap();
serde_json::to_value(
@ -482,16 +548,87 @@ mod tests {
}
#[rstest]
fn the_body_carries_the_vertex_version_and_the_wire_drops_the_model() {
let body = transformed(
json!({
"model": "claude-sonnet-4-5@20250929",
"max_tokens": 16,
"messages": [{"role": "user", "content": "hi"}]
}),
&MessagesTransformContext::default(),
#[case::sdk_2(&[("anthropic-version", "2026-10-21")], false, "2026-10-21")]
#[case::any_case(&[("Anthropic-Version", "2026-10-21")], false, "2026-10-21")]
#[case::vertex_name(&[("anthropic-version", "vertex-2023-10-16")], false, "vertex-2023-10-16")]
#[case::sdk_1(&[("anthropic-version", "2023-06-01")], false, "vertex-2023-10-16")]
#[case::first_version(&[("anthropic-version", "2023-01-01")], false, "vertex-2023-10-16")]
#[case::blank(&[("anthropic-version", " ")], false, "vertex-2023-10-16")]
#[case::no_header(&[], false, "vertex-2023-10-16")]
#[case::unknown_under_drop_params(&[("anthropic-version", "latest")], true, "vertex-2023-10-16")]
fn the_requested_version_goes_in_the_body(
#[case] forwarded: &[(&str, &str)],
#[case] drop_params: bool,
#[case] body_version: &str,
) {
let request: MessagesRequest = serde_json::from_value(json!({
"model": "claude-sonnet-4-5",
"max_tokens": 16,
"messages": [{"role": "user", "content": "hi"}],
}))
.unwrap();
let headers: Headers = forwarded
.iter()
.map(|(name, value)| (name.to_string(), value.to_string()))
.collect();
let body = serde_json::to_value(
VERTEX_ANTHROPIC_MESSAGES_CONFIG
.request_body(request, &headers, drop_params)
.unwrap(),
)
.unwrap();
assert_eq!(body["anthropic_version"], json!(body_version));
}
#[rstest]
#[case::unreleased("2025-01-01")]
#[case::alias("latest")]
fn an_unsupported_version_is_refused_without_drop_params(#[case] requested: &str) {
let request: MessagesRequest = serde_json::from_value(json!({
"model": "claude-sonnet-4-5",
"max_tokens": 16,
"messages": [{"role": "user", "content": "hi"}],
}))
.unwrap();
let headers: Headers = vec![("anthropic-version".to_string(), requested.to_string())];
let Error::InvalidRequest(detail) = VERTEX_ANTHROPIC_MESSAGES_CONFIG
.request_body(request, &headers, false)
.unwrap_err()
else {
panic!("expected an invalid request");
};
assert_eq!(
detail.to_string(),
format!(
"claude-sonnet-4-5 does not support anthropic-version={requested}. \
To drop unsupported params, set `litellm.drop_params = True`."
)
);
assert_eq!(body["anthropic_version"], json!("vertex-2023-10-16"));
}
#[rstest]
fn the_body_carries_the_vertex_version_and_the_wire_drops_the_model() {
let request: MessagesRequest = serde_json::from_value(json!({
"model": "claude-sonnet-4-5@20250929",
"max_tokens": 16,
"messages": [{"role": "user", "content": "hi"}]
}))
.unwrap();
let body = serde_json::to_value(
VERTEX_ANTHROPIC_MESSAGES_CONFIG
.request_body(
VERTEX_ANTHROPIC_MESSAGES_CONFIG
.transform_anthropic_messages_request(
request,
&MessagesTransformContext::default(),
)
.unwrap(),
&Vec::new(),
false,
)
.unwrap(),
)
.unwrap();
assert_eq!(body["model"], json!("claude-sonnet-4-5@20250929"));
let wire = VERTEX_ANTHROPIC_MESSAGES_CONFIG.wire_body(body);
assert_eq!(wire.get("model"), None);

View file

@ -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;

View file

@ -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]

View file

@ -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.

View file

@ -7,6 +7,7 @@ repository.workspace = true
[dependencies]
litellm-config.workspace = true
litellm-core-utils.workspace = true
litellm-inference.workspace = true
litellm-inference-messages.workspace = true
litellm-router-types.workspace = true

View file

@ -1,4 +1,4 @@
`litellm-router` scaffolds the model-list setup and deployment lookup portion of Python's `litellm.Router`. `Router::from_model_list(&config.model_list)` maps configured public names to provider deployments. Programmatic callers can collect `(String, Deployment)` entries into a `Router`
`litellm-router` scaffolds the model-list setup and deployment lookup portion of Python's `litellm.Router`. `Router::from_config(&config, &env)` maps configured public names to provider deployments and reads `drop_params` from `litellm_settings` and each deployment, resolving an `os.environ/NAME` value through `env`. Programmatic callers can collect `(String, Deployment)` entries into a `Router`
Lookup is exact and returns `None` for an unknown name. This extraction preserves the gateway's existing behavior: the last entry wins when public names repeat. Multiple deployments per model group, routing strategies, retries, cooldowns, and fallbacks are not implemented yet

View file

@ -2,16 +2,33 @@ mod deployment;
use std::collections::HashMap;
use litellm_config::Model;
use litellm_config::{Config, Spelled};
use litellm_core_utils::{serde_compat::parse_pydantic_bool, settings::Lookup};
use litellm_inference_messages::{MessagesSettings, MessagesShaping};
pub use deployment::Deployment;
#[derive(Clone, Debug, Default)]
pub struct Router(HashMap<String, Deployment>);
fn is_on(flag: Option<&Spelled<bool>>, env: &impl Lookup) -> bool {
match flag {
Some(Spelled::Value(on)) => *on,
Some(Spelled::Text(text)) => match text.strip_prefix("os.environ/") {
Some(name) => env
.get(name)
.is_some_and(|value| parse_pydantic_bool(&value) == Some(true)),
None => parse_pydantic_bool(text) == Some(true),
},
None => false,
}
}
impl Router {
pub fn from_model_list(model_list: &[Model]) -> Self {
model_list
pub fn from_config(config: &Config, env: &impl Lookup) -> Self {
let drops_params_globally = is_on(config.litellm_settings.drop_params.as_ref(), env);
config
.model_list
.iter()
.map(|model| {
(
@ -26,6 +43,14 @@ impl Router {
api_base: model.litellm_params.api_base.clone(),
custom_llm_provider: model.litellm_params.custom_llm_provider.clone(),
litellm_params: model.litellm_params.clone(),
shaping: MessagesShaping {
settings: MessagesSettings {
drop_params: drops_params_globally
|| is_on(model.litellm_params.drop_params.as_ref(), env),
..MessagesSettings::default()
},
..MessagesShaping::default()
},
..Deployment::default()
},
)
@ -43,3 +68,60 @@ impl FromIterator<(String, Deployment)> for Router {
Self(entries.into_iter().collect())
}
}
#[cfg(test)]
mod tests {
use litellm_config::Config;
use rstest::rstest;
use super::Router;
#[rstest]
#[case::unset("", "", false)]
#[case::deployment_flag("", "\n drop_params: true", true)]
#[case::deployment_word("", "\n drop_params: \" True \"", true)]
#[case::deployment_off("", "\n drop_params: false", false)]
#[case::deployment_unknown_word("", "\n drop_params: maybe", false)]
#[case::deployment_yes("", "\n drop_params: yes", true)]
#[case::deployment_no("", "\n drop_params: \"no\"", false)]
#[case::deployment_env_reference("", "\n drop_params: os.environ/DROP_PARAMS_FLAG", true)]
#[case::deployment_unset_env_reference(
"",
"\n drop_params: os.environ/MISSING_FLAG",
false
)]
#[case::global_env_reference(
"litellm_settings:\n drop_params: os.environ/DROP_PARAMS_FLAG\n",
"",
true
)]
#[case::global_flag("litellm_settings:\n drop_params: true\n", "", true)]
#[case::global_word("litellm_settings:\n drop_params: \"true\"\n", "", true)]
#[case::global_off("litellm_settings:\n drop_params: false\n", "", false)]
#[case::global_over_deployment_off(
"litellm_settings:\n drop_params: true\n",
"\n drop_params: false",
true
)]
fn configured_drop_params_reach_the_deployment(
#[case] settings: &str,
#[case] deployment: &str,
#[case] expected: bool,
) {
let config = Config::from_yaml(&format!(
"{settings}model_list:\n - model_name: public-model\n litellm_params:\n model: provider/model{deployment}"
))
.unwrap();
assert_eq!(
Router::from_config(&config, &|name: &str| (name == "DROP_PARAMS_FLAG")
.then(|| " Yes ".to_owned()))
.get("public-model")
.unwrap()
.shaping
.settings
.drop_params,
expected
);
}
}

View file

@ -16,7 +16,7 @@ fn configuration_preserves_deployment_parameters(#[case] parameters: &str) {
"model_list:\n - model_name: public-model\n litellm_params:\n model: provider/model\n {parameters}"
))
.unwrap();
let router = Router::from_model_list(&config.model_list);
let router = Router::from_config(&config, &|_: &str| None);
let deployment = router.get(&config.model_list[0].model_name).unwrap();
let params = &config.model_list[0].litellm_params;
@ -44,7 +44,7 @@ fn configuration_types_the_provider_params_of_a_deployment() {
azure_ad_token: ignored-by-every-group",
)
.unwrap();
let router = Router::from_model_list(&config.model_list);
let router = Router::from_config(&config, &|_: &str| None);
let params = &router.get("public-model").unwrap().litellm_params;
assert_eq!(
@ -74,7 +74,7 @@ fn lookup_uses_public_names(#[case] name: &str, #[case] expected: Option<&str>)
model: provider/b",
)
.unwrap();
let router = Router::from_model_list(&config.model_list);
let router = Router::from_config(&config, &|_: &str| None);
assert_eq!(router.get(name).map(|entry| entry.model.as_str()), expected);
}
@ -84,7 +84,7 @@ fn empty_configuration_has_no_deployment() {
let config = Config::from_yaml("model_list: []").unwrap();
assert!(
Router::from_model_list(&config.model_list)
Router::from_config(&config, &|_: &str| None)
.get("")
.is_none()
);

View file

@ -1336,7 +1336,6 @@ all_embedding_models = (
)
####### IMAGE GENERATION MODELS ###################
openai_image_generation_models = ["dall-e-2", "dall-e-3"]
####### VIDEO GENERATION MODELS ###################
openai_video_generation_models = ["sora-2"]

View file

@ -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

View file

@ -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()

View file

@ -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

View file

@ -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

View file

@ -764,7 +764,6 @@ ANTHROPIC_WEB_SEARCH_TOOL_MAX_USES: Final = {
# Used for web search interception across providers
LITELLM_WEB_SEARCH_TOOL_NAME: Final = "litellm_web_search"
DEFAULT_IMAGE_ENDPOINT_MODEL: Final = "dall-e-2"
DEFAULT_VIDEO_ENDPOINT_MODEL: Final = "sora-2"
DEFAULT_GOOGLE_VIDEO_DURATION_SECONDS: Final = get_env_int("DEFAULT_GOOGLE_VIDEO_DURATION_SECONDS", 8)

View file

@ -1592,14 +1592,6 @@ def completion_cost(
for part in ws_tier_parts
)
if (
(call_type == "aimage_generation" or call_type == "image_generation")
and model is not None
and isinstance(model, str)
and len(model) == 0
and custom_llm_provider == "azure"
):
model = "dall-e-2" # for dall-e-2, azure expects an empty model name
# Handle Inputs to completion_cost
prompt_tokens = 0
prompt_characters: int | None = None
@ -1674,7 +1666,8 @@ def completion_cost(
raw_candidates if custom_pricing is True else _priced_candidates_before_unpriced_rules(raw_candidates)
)
for idx, (model, candidate_pricing_provider) in enumerate(potential_model_names):
for idx, (candidate_model, candidate_pricing_provider) in enumerate(potential_model_names):
model = candidate_model
try:
if verbose_logger.isEnabledFor(logging.DEBUG):
verbose_logger.debug("selected model name for cost calculation: %s", model)

View file

@ -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

View file

@ -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
@ -16,7 +16,6 @@ import litellm
# client is imported from litellm as it's a decorator
from litellm import client
from litellm.constants import DEFAULT_IMAGE_ENDPOINT_MODEL
from litellm.constants import request_timeout as DEFAULT_REQUEST_TIMEOUT
from litellm.exceptions import LiteLLMUnknownProvider
from litellm.litellm_core_utils.get_llm_provider_logic import (
@ -61,6 +60,7 @@ from litellm.types.utils import (
ProviderRoutingHints,
)
from litellm.utils import (
EXTRA_BODY_ROUTING_KEYS,
ImageResponse,
ProviderConfigManager,
get_llm_provider,
@ -298,17 +298,14 @@ def image_generation(
if custom_llm_provider == "chatgpt" or (model is not None and model.startswith("chatgpt/"))
else None
)
dynamic_api_key: str | None = None
if model is not None or custom_llm_provider is not None:
model, custom_llm_provider, dynamic_api_key, api_base = get_llm_provider(
model=model,
custom_llm_provider=custom_llm_provider,
api_base=api_base,
litellm_params=chatgpt_params,
)
else:
model = "dall-e-2"
custom_llm_provider = "openai" # default to dall-e-2 on openai
if model is None and custom_llm_provider is None:
raise litellm.BadRequestError(message="model is required", model="", llm_provider="")
model, custom_llm_provider, dynamic_api_key, api_base = get_llm_provider(
model=model,
custom_llm_provider=custom_llm_provider,
api_base=api_base,
litellm_params=chatgpt_params,
)
model_response.hidden_params["model"] = model
openai_params: Final = [
"user",
@ -466,7 +463,7 @@ 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}")
@ -711,7 +708,7 @@ async def aimage_variation(*args, **kwargs) -> ImageResponse:
@client
def image_variation(
image: FileTypes,
model: str = "dall-e-2", # set to dall-e-2 by default - like OpenAI.
model: str | None = None,
n: int = 1,
response_format: Literal["url", "b64_json"] = "url",
size: str | None = None,
@ -727,7 +724,7 @@ def image_variation(
litellm_params: Final = get_litellm_params(**kwargs)
# get the custom llm provider
model, custom_llm_provider, dynamic_api_key, api_base = get_llm_provider(
model=model,
model=model or "",
custom_llm_provider=litellm_params.get("custom_llm_provider", None),
api_base=litellm_params.get("api_base", None),
api_key=litellm_params.get("api_key", None),
@ -746,7 +743,7 @@ def image_variation(
response: ImageResponse | None = None
provider_config: Final = ProviderConfigManager.get_provider_model_info(
model=model or "", # openai defaults to dall-e-2
model=model,
provider=llm_provider,
)
@ -771,7 +768,7 @@ def image_variation(
model_response=model_response,
api_key=api_key,
api_base=api_base,
model=model,
model=model or None,
image=image,
timeout=litellm_params.get("timeout", None),
custom_llm_provider=custom_llm_provider,
@ -789,7 +786,7 @@ def image_variation(
model_response=model_response,
api_key=api_key,
api_base=api_base,
model=model,
model=model or None,
image=image,
timeout=litellm_params.get("timeout", None) or DEFAULT_REQUEST_TIMEOUT,
custom_llm_provider=custom_llm_provider,
@ -807,6 +804,32 @@ 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,
@ -874,7 +897,7 @@ def image_edit(
# get llm provider logic
litellm_params: Final = GenericLiteLLMParams(**kwargs)
model, custom_llm_provider, _, _ = get_llm_provider(
model=model or DEFAULT_IMAGE_ENDPOINT_MODEL,
model=model or "",
custom_llm_provider=custom_llm_provider,
)
@ -935,22 +958,24 @@ 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 +1064,19 @@ 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,

View file

@ -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')}"

View file

@ -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),

View file

@ -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

View file

@ -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

View file

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

View file

@ -437,7 +437,6 @@ def get_llm_provider(
model in litellm.open_ai_chat_completion_models
or "ft:gpt-3.5-turbo" in model
or "ft:gpt-4" in model # catches ft:gpt-4-0613, ft:gpt-4o
or model in litellm.openai_image_generation_models
or model.startswith("gpt-image")
or model in litellm.openai_video_generation_models
):

View file

@ -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,
)

View file

@ -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

View file

@ -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,

View file

@ -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:

View file

@ -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

View file

@ -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,

View file

@ -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"

View file

@ -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()

View file

@ -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,
)

View file

@ -13,13 +13,12 @@ from openai import (
BadRequestError,
OpenAI,
)
from pydantic import BaseModel, TypeAdapter
from pydantic import BaseModel
import litellm
from litellm.constants import AZURE_OPERATION_POLLING_TIMEOUT, DEFAULT_MAX_RETRIES
from litellm.constants import DEFAULT_MAX_RETRIES
from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
from litellm.litellm_core_utils.logging_utils import speech_request_body, track_llm_api_timing
from litellm.litellm_core_utils.url_utils import SSRFError, assert_same_origin
from litellm.llms.custom_httpx.http_handler import (
AsyncHTTPHandler,
HTTPHandler,
@ -34,7 +33,6 @@ from litellm.types.utils import (
from litellm.utils import (
CustomStreamWrapper,
convert_to_model_response_object,
modify_url,
)
from ...types.llms.openai import HttpxBinaryResponseContent
@ -59,8 +57,6 @@ from .image_generation import (
)
from .image_generation.http_utils import azure_deployment_image_generation_json_body
_DECODED_JSON: Final = TypeAdapter(object)
class AzureOpenAIAssistantsAPIConfig:
"""
@ -860,11 +856,6 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM):
headers: dict,
deployment_name: str | None = None,
) -> httpx.Response:
"""
Implemented for azure dall-e-2 image gen calls
Alternative to needing a custom transport implementation
"""
if client is None:
_params: Final = {}
if timeout is not None:
@ -881,95 +872,6 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM):
else:
async_handler = client
if (
"images/generations" in api_base
and api_version
in [ # dall-e-3 starts from `2023-12-01-preview` so we should be able to avoid conflict
"2023-06-01-preview",
"2023-07-01-preview",
"2023-08-01-preview",
"2023-09-01-preview",
"2023-10-01-preview",
]
): # CREATE + POLL for azure dall-e-2 calls
api_base = modify_url(original_url=api_base, new_path="/openai/images/generations:submit")
data.pop(
"model", None
) # REMOVE 'model' from dall-e-2 arg https://learn.microsoft.com/en-us/azure/ai-services/openai/reference#request-a-generated-image-dall-e-2-preview
response = await async_handler.post(
url=api_base,
data=json.dumps(data),
headers=headers,
)
if "operation-location" in response.headers:
operation_location_url: Final = response.headers["operation-location"]
else:
raise AzureOpenAIError(status_code=500, message=response.text)
# Reject polling URLs that don't share an origin with ``api_base``.
# Without this an upstream-controlled or attacker-controlled
# value would receive the operator's Azure API key in the
# request headers below. VERIA-51.
try:
assert_same_origin(operation_location_url, api_base)
except SSRFError as ssrf_err:
raise AzureOpenAIError(
status_code=502,
message=f"Rejected polling URL: {ssrf_err}",
)
response = await async_handler.get(
url=operation_location_url,
headers=headers,
)
await response.aread()
timeout_secs: Final[int] = AZURE_OPERATION_POLLING_TIMEOUT
start_time: Final = time.time()
if "status" not in response.json():
# Don't reflect the raw response body — when the polling
# URL points at an internal JSON API (cloud metadata
# service etc.) reflecting it here turns Blind SSRF into
# Full-Read SSRF. VERIA-51.
raise AzureOpenAIError(
status_code=502,
message="Polling response missing 'status' field",
)
while response.json()["status"] not in ["succeeded", "failed"]:
if time.time() - start_time > timeout_secs:
raise AzureOpenAIError(status_code=408, message="Operation polling timed out.")
await asyncio.sleep(int(response.headers.get("retry-after") or 10))
response = await async_handler.get(
url=operation_location_url,
headers=headers,
)
await response.aread()
if response.json()["status"] == "failed":
error_data: Final = response.json()
# Preserve Azure error details (e.g. content_policy_violation,
# inner_error, content_filter_results) as structured body so
# exception_type() can route them correctly.
_error_body: Final = error_data.get("error", error_data)
_error_msg: Final = (
_error_body.get("message", "Image generation failed")
if isinstance(_error_body, dict)
else json.dumps(error_data)
)
raise AzureOpenAIError(
status_code=400,
message=_error_msg,
body=error_data,
)
result: Final = _DECODED_JSON.validate_python(response.json()["result"])
return httpx.Response(
status_code=200,
headers=response.headers,
content=json.dumps(result).encode("utf-8"),
request=httpx.Request(method="POST", url="https://api.openai.com/v1"),
)
request_json: Final = azure_deployment_image_generation_json_body(api_base, data, deployment_name)
return await async_handler.post(
url=api_base,
@ -988,11 +890,6 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM):
headers: dict,
deployment_name: str | None = None,
) -> httpx.Response:
"""
Implemented for azure dall-e-2 image gen calls
Alternative to needing a custom transport implementation
"""
if client is None:
_params: Final = {}
if timeout is not None:
@ -1006,87 +903,6 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM):
else:
sync_handler = client
if (
"images/generations" in api_base
and api_version
in [ # dall-e-3 starts from `2023-12-01-preview` so we should be able to avoid conflict
"2023-06-01-preview",
"2023-07-01-preview",
"2023-08-01-preview",
"2023-09-01-preview",
"2023-10-01-preview",
]
): # CREATE + POLL for azure dall-e-2 calls
api_base = modify_url(original_url=api_base, new_path="/openai/images/generations:submit")
data.pop(
"model", None
) # REMOVE 'model' from dall-e-2 arg https://learn.microsoft.com/en-us/azure/ai-services/openai/reference#request-a-generated-image-dall-e-2-preview
response = sync_handler.post(
url=api_base,
data=json.dumps(data),
headers=headers,
)
if "operation-location" in response.headers:
operation_location_url: Final = response.headers["operation-location"]
else:
raise AzureOpenAIError(status_code=500, message=response.text)
try:
assert_same_origin(operation_location_url, api_base)
except SSRFError as ssrf_err:
raise AzureOpenAIError(
status_code=502,
message=f"Rejected polling URL: {ssrf_err}",
)
response = sync_handler.get(
url=operation_location_url,
headers=headers,
)
response.read()
timeout_secs: Final[int] = AZURE_OPERATION_POLLING_TIMEOUT
start_time: Final = time.time()
if "status" not in response.json():
raise AzureOpenAIError(
status_code=502,
message="Polling response missing 'status' field",
)
while response.json()["status"] not in ["succeeded", "failed"]:
if time.time() - start_time > timeout_secs:
raise AzureOpenAIError(status_code=408, message="Operation polling timed out.")
time.sleep(int(response.headers.get("retry-after") or 10))
response = sync_handler.get(
url=operation_location_url,
headers=headers,
)
response.read()
if response.json()["status"] == "failed":
error_data: Final = response.json()
# Preserve Azure error details (e.g. content_policy_violation,
# inner_error, content_filter_results) as structured body so
# exception_type() can route them correctly.
_error_body: Final = error_data.get("error", error_data)
_error_msg: Final = (
_error_body.get("message", "Image generation failed")
if isinstance(_error_body, dict)
else json.dumps(error_data)
)
raise AzureOpenAIError(
status_code=400,
message=_error_msg,
body=error_data,
)
result: Final = _DECODED_JSON.validate_python(response.json()["result"])
return httpx.Response(
status_code=200,
headers=response.headers,
content=json.dumps(result).encode("utf-8"),
request=httpx.Request(method="POST", url="https://api.openai.com/v1"),
)
request_json: Final = azure_deployment_image_generation_json_body(api_base, data, deployment_name)
return sync_handler.post(
url=api_base,
@ -1193,10 +1009,10 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM):
deployment_name=model,
)
provider_config: Final = get_azure_image_generation_config(data.get("model", "dall-e-2"))
provider_config: Final = get_azure_image_generation_config(data.get("model", ""))
if provider_config is not None:
return provider_config.transform_image_generation_response(
model=data.get("model", "dall-e-2"),
model=data.get("model", ""),
raw_response=httpx_response,
model_response=model_response or ImageResponse(),
logging_obj=logging_obj,
@ -1329,10 +1145,10 @@ class AzureChatCompletion(BaseAzureLLM, BaseLLM):
headers=request_headers,
deployment_name=model,
)
provider_config: Final = get_azure_image_generation_config(data.get("model", "dall-e-2"))
provider_config: Final = get_azure_image_generation_config(data.get("model", ""))
if isinstance(provider_config, AzureFoundryMAIImageGenerationConfig):
return provider_config.transform_image_generation_response(
model=data.get("model", "dall-e-2"),
model=data.get("model", ""),
raw_response=httpx_response,
model_response=model_response or ImageResponse(),
logging_obj=logging_obj,

View file

@ -4,14 +4,10 @@ from litellm.llms.base_llm.image_generation.transformation import (
BaseImageGenerationConfig,
)
from .dall_e_2_transformation import AzureDallE2ImageGenerationConfig
from .dall_e_3_transformation import AzureDallE3ImageGenerationConfig
from .gpt_transformation import AzureGPTImageGenerationConfig
from .http_utils import azure_deployment_image_generation_json_body
__all__ = [
"AzureDallE2ImageGenerationConfig",
"AzureDallE3ImageGenerationConfig",
"AzureGPTImageGenerationConfig",
"azure_deployment_image_generation_json_body",
]
@ -21,11 +17,7 @@ def get_azure_image_generation_config(model: str) -> BaseImageGenerationConfig:
model = model.lower()
model = model.replace("-", "")
model = model.replace("_", "")
if model == "" or "dalle2" in model: # empty model is dall-e-2
return AzureDallE2ImageGenerationConfig()
elif "dalle3" in model:
return AzureDallE3ImageGenerationConfig()
elif AzureFoundryMAIImageGenerationConfig.is_mai_model(model):
if AzureFoundryMAIImageGenerationConfig.is_mai_model(model):
return AzureFoundryMAIImageGenerationConfig()
else:
verbose_logger.debug(

View file

@ -1,7 +0,0 @@
from litellm.llms.openai.image_generation import DallE2ImageGenerationConfig
class AzureDallE2ImageGenerationConfig(DallE2ImageGenerationConfig):
"""
Azure dall-e-2 image generation config
"""

View file

@ -1,7 +0,0 @@
from litellm.llms.openai.image_generation import DallE3ImageGenerationConfig
class AzureDallE3ImageGenerationConfig(DallE3ImageGenerationConfig):
"""
Azure dall-e-3 image generation config
"""

View file

@ -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:"):

View file

@ -3,15 +3,11 @@ from litellm.llms.base_llm.image_generation.transformation import (
BaseImageGenerationConfig,
)
from .dall_e_2_transformation import AzureFoundryDallE2ImageGenerationConfig
from .dall_e_3_transformation import AzureFoundryDallE3ImageGenerationConfig
from .flux_transformation import AzureFoundryFluxImageGenerationConfig
from .gpt_transformation import AzureFoundryGPTImageGenerationConfig
from .mai_transformation import AzureFoundryMAIImageGenerationConfig
__all__ = [
"AzureFoundryDallE2ImageGenerationConfig",
"AzureFoundryDallE3ImageGenerationConfig",
"AzureFoundryFluxImageGenerationConfig",
"AzureFoundryGPTImageGenerationConfig",
"AzureFoundryMAIImageGenerationConfig",
@ -22,11 +18,7 @@ def get_azure_ai_image_generation_config(model: str) -> BaseImageGenerationConfi
model = model.lower()
model = model.replace("-", "")
model = model.replace("_", "")
if model == "" or "dalle2" in model: # empty model is dall-e-2
return AzureFoundryDallE2ImageGenerationConfig()
elif "dalle3" in model:
return AzureFoundryDallE3ImageGenerationConfig()
elif AzureFoundryMAIImageGenerationConfig.is_mai_model(model):
if AzureFoundryMAIImageGenerationConfig.is_mai_model(model):
return AzureFoundryMAIImageGenerationConfig()
elif "flux" in model:
return AzureFoundryFluxImageGenerationConfig()

View file

@ -1,7 +0,0 @@
from litellm.llms.openai.image_generation import DallE2ImageGenerationConfig
class AzureFoundryDallE2ImageGenerationConfig(DallE2ImageGenerationConfig):
"""
Azure dall-e-2 image generation config
"""

View file

@ -1,7 +0,0 @@
from litellm.llms.openai.image_generation import DallE3ImageGenerationConfig
class AzureFoundryDallE3ImageGenerationConfig(DallE3ImageGenerationConfig):
"""
Azure dall-e-3 image generation config
"""

View file

@ -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:

View file

@ -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

View file

@ -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(

View file

@ -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

View file

@ -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

View file

@ -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())

View file

@ -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:")
)

View file

@ -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 []

View file

@ -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:

View file

@ -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

View file

@ -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

View file

@ -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
@ -560,11 +561,8 @@ class BaseLLMAIOHTTPHandler:
organization: str | None = None,
headers: dict | None = None,
) -> ImageResponse:
if model is None:
raise ValueError("model is required for non-openai image variations")
provider_config: Final = ProviderConfigManager.get_provider_image_variation_config(
model=model, # openai defaults to dall-e-2
model=model or "",
provider=LlmProviders(custom_llm_provider),
)

View file

@ -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,
)

View file

@ -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),

View file

@ -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

View file

@ -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

View file

@ -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(

View file

@ -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, ()

View file

@ -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,
}

View file

@ -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]"

View file

@ -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")

View file

@ -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,

View file

@ -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:

View file

@ -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,

View file

@ -1,31 +1,5 @@
from typing import Final
from litellm.llms.base_llm.image_edit.transformation import BaseImageEditConfig
from .dalle2_transformation import DallE2ImageEditConfig
from .transformation import OpenAIImageEditConfig
__all__ = [
"DallE2ImageEditConfig",
"OpenAIImageEditConfig",
"get_openai_image_edit_config",
]
def get_openai_image_edit_config(model: str) -> BaseImageEditConfig:
"""
Get the appropriate OpenAI image edit config based on the model.
Args:
model: The model name (e.g., "dall-e-2", "gpt-image-1")
Returns:
The appropriate config instance for the model
"""
model_normalized: Final = model.lower().replace("-", "").replace("_", "")
if model_normalized == "dalle2":
return DallE2ImageEditConfig()
else:
# Default to standard OpenAI config for gpt-image-1 and other models
return OpenAIImageEditConfig()

View file

@ -1,98 +0,0 @@
from io import BufferedReader
from typing import TYPE_CHECKING, Any, Final, cast
from httpx._types import RequestFiles
import litellm
from litellm.images.utils import ImageEditRequestUtils
from litellm.types.images.main import ImageEditRequestParams
from litellm.types.llms.openai import FileTypes
from litellm.types.router import GenericLiteLLMParams
from .transformation import OpenAIImageEditConfig
if TYPE_CHECKING:
from litellm.litellm_core_utils.litellm_logging import Logging as _LiteLLMLoggingObj
LiteLLMLoggingObj = _LiteLLMLoggingObj
else:
LiteLLMLoggingObj = Any
class DallE2ImageEditConfig(OpenAIImageEditConfig):
"""
DALL-E-2 specific configuration for image edit API.
DALL-E-2 only supports editing a single image (not an array).
Uses "image" field name instead of "image[]".
"""
def transform_image_edit_request(
self,
model: str,
prompt: str | None,
image: FileTypes | None,
image_edit_optional_request_params: dict,
litellm_params: GenericLiteLLMParams,
headers: dict,
) -> tuple[dict, RequestFiles]:
"""
Transform image edit request for DALL-E-2.
DALL-E-2 only accepts a single image with field name "image" (not "image[]").
"""
request_params: Final = {
"model": model,
**image_edit_optional_request_params,
}
if image is not None:
request_params["image"] = image
if prompt is not None:
request_params["prompt"] = prompt
request: Final = ImageEditRequestParams(**request_params)
request_dict: Final = cast(dict, request)
#########################################################
# Separate images and masks as `files` and send other parameters as `data`
#########################################################
_image_list: Final = request_dict.get("image")
_mask = request_dict.get("mask")
data_without_files: Final = {k: v for k, v in request_dict.items() if k not in ["image", "mask"]}
files_list: Final[list[tuple[str, Any]]] = []
# Handle image parameter - DALL-E-2 only supports single image
if _image_list is not None:
image_list: Final = [_image_list] if not isinstance(_image_list, list) else _image_list
# Validate only one image is provided
if len(image_list) > 1:
raise litellm.BadRequestError(
message="DALL-E-2 only supports editing a single image. Please provide one image.",
model=model,
llm_provider="openai",
)
# Use "image" field name (singular) for DALL-E-2
for _image in image_list:
if _image is not None:
self._add_image_to_files(
files_list=files_list,
image=_image,
field_name="image",
)
# Handle mask parameter if provided
if _mask is not None:
# Handle case where mask can be a list (extract first mask)
if isinstance(_mask, list):
_mask = _mask[0] if _mask else None
if _mask is not None:
mask_content_type: Final[str] = ImageEditRequestUtils.get_image_content_type(_mask)
if isinstance(_mask, BufferedReader):
files_list.append(("mask", (_mask.name, _mask, mask_content_type)))
else:
files_list.append(("mask", ("mask.png", _mask, mask_content_type)))
return data_without_files, files_list

View file

@ -96,7 +96,7 @@ class OpenAIImageEditConfig(BaseImageEditConfig):
"""
# Build request params, only including non-None values
request_params: Final = {
"model": model,
**({"model": model} if model else {}),
**image_edit_optional_request_params,
}
if image is not None:

View file

@ -1,9 +1,3 @@
from litellm.llms.base_llm.image_generation.transformation import (
BaseImageGenerationConfig,
)
from .dall_e_2_transformation import DallE2ImageGenerationConfig
from .dall_e_3_transformation import DallE3ImageGenerationConfig
from .gpt_transformation import GPTImageGenerationConfig
from .guardrail_translation import (
OpenAIImageGenerationHandler,
@ -11,18 +5,7 @@ from .guardrail_translation import (
)
__all__ = [
"DallE2ImageGenerationConfig",
"DallE3ImageGenerationConfig",
"GPTImageGenerationConfig",
"OpenAIImageGenerationHandler",
"guardrail_translation_mappings",
]
def get_openai_image_generation_config(model: str) -> BaseImageGenerationConfig:
if model.startswith("dall-e-2") or model == "": # empty model is dall-e-2
return DallE2ImageGenerationConfig()
elif model.startswith("dall-e-3"):
return DallE3ImageGenerationConfig()
else:
return GPTImageGenerationConfig()

View file

@ -1,81 +0,0 @@
from typing import TYPE_CHECKING, Final
import httpx
from litellm.llms.base_llm.image_generation.transformation import (
BaseImageGenerationConfig,
)
from litellm.types.llms.openai import OpenAIImageGenerationOptionalParams
from litellm.types.utils import ImageResponse
from litellm.utils import convert_to_model_response_object
if TYPE_CHECKING:
from litellm.litellm_core_utils.logging import Logging as LiteLLMLoggingObj
from litellm.litellm_core_utils.tokenizer import Encoding as Tokenizer
class DallE2ImageGenerationConfig(BaseImageGenerationConfig):
"""
OpenAI dall-e-2 image generation config
"""
def get_supported_openai_params(self, model: str) -> list[OpenAIImageGenerationOptionalParams]:
return ["n", "response_format", "quality", "size", "user"]
def map_openai_params(
self,
non_default_params: dict,
optional_params: dict,
model: str,
drop_params: bool,
) -> dict:
supported_params: Final = self.get_supported_openai_params(model)
for k in non_default_params:
if k not in optional_params:
if k in supported_params:
optional_params[k] = non_default_params[k]
elif drop_params:
pass
else:
raise ValueError(
f"Parameter {k} is not supported for model {model}. Supported parameters are {supported_params}. Set drop_params=True to drop unsupported parameters."
)
return optional_params
def transform_image_generation_response(
self,
model: str,
raw_response: httpx.Response,
model_response: ImageResponse,
logging_obj: "LiteLLMLoggingObj",
request_data: dict,
optional_params: dict,
litellm_params: dict,
encoding: "Tokenizer | None",
api_key: str | None = None,
json_mode: bool | None = None,
) -> ImageResponse:
response: Final = raw_response.json()
stringified_response: Final = response
## LOGGING
logging_obj.post_call(
input=request_data.get("prompt", ""),
api_key=api_key,
additional_args={"complete_input_dict": request_data},
original_response=stringified_response,
)
image_response: Final[ImageResponse] = convert_to_model_response_object(
response_object=stringified_response,
model_response_object=model_response,
response_type="image_generation",
)
# set optional params
image_response.size = optional_params.get("size", "1024x1024") # default is always 1024x1024
image_response.quality = optional_params.get("quality", "standard") # always standard for dall-e-2
image_response.output_format = optional_params.get("output_format", "png") # always png for dall-e-2
return image_response

View file

@ -1,81 +0,0 @@
from typing import TYPE_CHECKING, Final
import httpx
from litellm.llms.base_llm.image_generation.transformation import (
BaseImageGenerationConfig,
)
from litellm.types.llms.openai import OpenAIImageGenerationOptionalParams
from litellm.types.utils import ImageResponse
from litellm.utils import convert_to_model_response_object
if TYPE_CHECKING:
from litellm.litellm_core_utils.logging import Logging as LiteLLMLoggingObj
from litellm.litellm_core_utils.tokenizer import Encoding as Tokenizer
class DallE3ImageGenerationConfig(BaseImageGenerationConfig):
"""
OpenAI dall-e-3 image generation config
"""
def get_supported_openai_params(self, model: str) -> list[OpenAIImageGenerationOptionalParams]:
return ["n", "response_format", "quality", "size", "user", "style"]
def map_openai_params(
self,
non_default_params: dict,
optional_params: dict,
model: str,
drop_params: bool,
) -> dict:
supported_params: Final = self.get_supported_openai_params(model)
for k in non_default_params:
if k not in optional_params:
if k in supported_params:
optional_params[k] = non_default_params[k]
elif drop_params:
pass
else:
raise ValueError(
f"Parameter {k} is not supported for model {model}. Supported parameters are {supported_params}. Set drop_params=True to drop unsupported parameters."
)
return optional_params
def transform_image_generation_response(
self,
model: str,
raw_response: httpx.Response,
model_response: ImageResponse,
logging_obj: "LiteLLMLoggingObj",
request_data: dict,
optional_params: dict,
litellm_params: dict,
encoding: "Tokenizer | None",
api_key: str | None = None,
json_mode: bool | None = None,
) -> ImageResponse:
response: Final = raw_response.json()
stringified_response: Final = response
## LOGGING
logging_obj.post_call(
input=request_data.get("prompt", ""),
api_key=api_key,
additional_args={"complete_input_dict": request_data},
original_response=stringified_response,
)
image_response: Final[ImageResponse] = convert_to_model_response_object(
response_object=stringified_response,
model_response_object=model_response,
response_type="image_generation",
)
# set optional params
image_response.size = optional_params.get("size", "1024x1024") # default is always 1024x1024
image_response.quality = optional_params.get("quality", "hd") # always hd for dall-e-3
image_response.output_format = optional_params.get("output_format", "png") # always png for dall-e-3
return image_response

View file

@ -151,7 +151,7 @@ class OpenAIImageVariationsHandler:
) -> ImageResponse:
try:
provider_config: Final = ProviderConfigManager.get_provider_image_variation_config(
model=model or "", # openai defaults to dall-e-2
model=model or "",
provider=LlmProviders.OPENAI,
)

View file

@ -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,

View file

@ -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(

View file

@ -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

View file

@ -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,

View file

@ -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)

View file

@ -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

Some files were not shown because too many files have changed in this diff Show more