scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-29 01:42:19 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 79

Merge a303fcbb48 into 35225709ed

Browse source
This commit is contained in:
Jeremy Schoemaker 2026-09-23 20:26:55 +00:00 • committed by GitHub
commit b1ebfd46be
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 285 additions and 31 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

66
litellm/llms/gemini/realtime/transformation.py
View file

@ -132,7 +132,15 @@ def _grounding_metadata_from_frame(frame: Mapping[str, object]) -> tuple[Mapping
# 175 text tokens/min of output (ai.google.dev/gemini-api/docs/pricing).
GEMINI_LIVE_TRANSCRIBE_AUDIO_TOKENS_PER_SECOND: Final = 25
GEMINI_LIVE_TRANSCRIBE_OUTPUT_TEXT_TOKENS_PER_MINUTE: Final = 175
PCM16_INPUT_AUDIO_BYTES_PER_SECOND: Final = 48000
# The MIME rate controls server resampling and must describe the input bytes.
GEMINI_LIVE_INPUT_AUDIO_SAMPLE_RATE_HZ: Final = 16000
PCM16_BYTES_PER_SAMPLE: Final = 2
# Bound client-controlled rates because they affect the duration billed.
MIN_ACCEPTED_INPUT_AUDIO_SAMPLE_RATE_HZ: Final = 8000
MAX_ACCEPTED_INPUT_AUDIO_SAMPLE_RATE_HZ: Final = 48000
# The beta pcm16 codec implies 24kHz, matching the beta-to-GA conversion.
BETA_PCM16_INPUT_AUDIO_SAMPLE_RATE_HZ: Final = 24000
def _base64_decoded_byte_count(data: str) -> int:
@ -149,7 +157,9 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
# Gemini Live sometimes emits usageMetadata in a standalone frame between
# turns; buffer it here so the next response.done carries the token counts.
self._pending_usage_metadata: dict | None = None
self._unbilled_input_audio_bytes: int = 0
# Store per-chunk duration so later rate changes cannot reprice earlier audio.
self._unbilled_input_audio_seconds: float = 0.0
self._input_audio_sample_rate_hz: int = GEMINI_LIVE_INPUT_AUDIO_SAMPLE_RATE_HZ
def is_setup_message(self, msg_obj: dict) -> bool:
return "setup" in msg_obj
@ -243,7 +253,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
def get_audio_mime_type(self, input_audio_format: str = "pcm16"):
mime_types: Final = {
"pcm16": "audio/pcm;rate=24000",
"pcm16": f"audio/pcm;rate={self._input_audio_sample_rate_hz}",
"g711_ulaw": "audio/pcmu",
"g711_alaw": "audio/pcma",
}
@ -463,6 +473,45 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
)
return setup
@staticmethod
def _declared_rate_from_ga_shape(session_payload: Mapping[str, object]) -> object:
audio = session_payload.get("audio")
if not isinstance(audio, dict):
return None
audio_input = audio.get("input")
if not isinstance(audio_input, dict):
return None
audio_format = audio_input.get("format")
if not isinstance(audio_format, dict):
return None
return audio_format.get("rate")
@staticmethod
def _declared_rate_from_beta_shape(session_payload: Mapping[str, object]) -> object:
"""Only pcm16 implies a rate here because append frames are sent as pcm16."""
if session_payload.get("input_audio_format") == "pcm16":
return BETA_PCM16_INPUT_AUDIO_SAMPLE_RATE_HZ
return None
def _record_input_audio_sample_rate(self, session_payload: Mapping[str, object]) -> None:
"""Read both shapes so the OpenAI-Beta header cannot change the audio's declared rate."""
rate = self._declared_rate_from_ga_shape(session_payload)
if rate is None:
rate = self._declared_rate_from_beta_shape(session_payload)
if isinstance(rate, bool) or not isinstance(rate, int):
return
if not (MIN_ACCEPTED_INPUT_AUDIO_SAMPLE_RATE_HZ <= rate <= MAX_ACCEPTED_INPUT_AUDIO_SAMPLE_RATE_HZ):
verbose_logger.warning(
"Gemini Realtime: ignoring declared input audio rate %s, outside the accepted "
"%s-%s Hz range; keeping %s Hz",
rate,
MIN_ACCEPTED_INPUT_AUDIO_SAMPLE_RATE_HZ,
MAX_ACCEPTED_INPUT_AUDIO_SAMPLE_RATE_HZ,
self._input_audio_sample_rate_hz,
)
return
self._input_audio_sample_rate_hz = rate
def _handle_session_update(
self,
json_message: _OpenAIRealtimeClientEvent,
@ -488,6 +537,7 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
# explicit modality / transcription / turn-detection settings
# would be silently dropped because ``map_openai_params`` only
# recognises the flat OpenAI-beta key names.
self._record_input_audio_sample_rate(session_payload)
session_payload = self._normalize_session_payload_for_mapping(session_payload)
new_overrides: Final = self.map_openai_params(optional_params={}, non_default_params=session_payload)
@ -618,7 +668,9 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
if msg_type == "input_audio_buffer.append":
audio_b64: Final = json_message["audio"]
if isinstance(audio_b64, str):
self._unbilled_input_audio_bytes += _base64_decoded_byte_count(audio_b64)
self._unbilled_input_audio_seconds += _base64_decoded_byte_count(audio_b64) / (
self._input_audio_sample_rate_hz * PCM16_BYTES_PER_SAMPLE
)
realtime_input_dict["audio"] = HttpxBlobType(mimeType=self.get_audio_mime_type(), data=audio_b64)
realtime_input_dict = cast(
@ -1213,10 +1265,10 @@ class GeminiRealtimeConfig(BaseRealtimeConfig):
def _consume_input_transcription_usage_estimate(self, model: str) -> RealtimeInputAudioTranscriptionUsage | None:
"""Gemini Live sends no usageMetadata for transcribe sessions; estimate billing from streamed audio duration."""
if self._unbilled_input_audio_bytes <= 0 or not self._is_text_only_live_model(model):
if self._unbilled_input_audio_seconds <= 0 or not self._is_text_only_live_model(model):
return None
audio_seconds: Final = self._unbilled_input_audio_bytes / PCM16_INPUT_AUDIO_BYTES_PER_SECOND
self._unbilled_input_audio_bytes = 0
audio_seconds: Final = self._unbilled_input_audio_seconds
self._unbilled_input_audio_seconds = 0.0
audio_tokens: Final = round(audio_seconds * GEMINI_LIVE_TRANSCRIBE_AUDIO_TOKENS_PER_SECOND)
output_tokens: Final = round(audio_seconds * GEMINI_LIVE_TRANSCRIBE_OUTPUT_TEXT_TOKENS_PER_MINUTE / 60)
usage: Final[RealtimeInputAudioTranscriptionUsage] = {

17
litellm/llms/vertex_ai/realtime/transformation.py
View file

@ -93,19 +93,6 @@ class VertexAIRealtimeConfig(GeminiRealtimeConfig):
headers["x-goog-user-project"] = self._project
return headers
# ------------------------------------------------------------------
# Audio MIME type — Vertex AI needs the sample rate in the MIME string
# ------------------------------------------------------------------
def get_audio_mime_type(self, input_audio_format: str = "pcm16") -> str:
mime_types: Final = {
# Gemini Live native audio (OpenAI GA realtime default) is 24kHz PCM.
"pcm16": "audio/pcm;rate=24000",
"g711_ulaw": "audio/pcmu",
"g711_alaw": "audio/pcma",
}
return mime_types.get(input_audio_format, "application/octet-stream")
# ------------------------------------------------------------------
# Session setup message
# ------------------------------------------------------------------
@ -212,6 +199,10 @@ class VertexAIRealtimeConfig(GeminiRealtimeConfig):
msg_type: Final = json_message.get("type")
if msg_type == "session.update":
# Vertex handles session.update itself and never reaches the parent's handler, so the
# declared input audio rate has to be recorded here or the Vertex path silently keeps
# the default no matter what the client declares.
self._record_input_audio_sample_rate(json_message.get("session") or {})
if session_configuration_request is None:
setup_config: Final = self._build_vertex_ai_setup_config(model, json_message.get("session") or {})
gemini_setup_msg: Final = json.dumps({"setup": setup_config})

233
tests/unit/llms/gemini/realtime/test_gemini_realtime_transformation.py
View file

@ -1,9 +1,10 @@
import json
from collections.abc import Mapping
from typing import cast
from typing import Final, cast
from unittest.mock import MagicMock
import pytest
from pydantic import TypeAdapter
import litellm
@ -2007,10 +2008,219 @@ def _input_audio_append_message(raw_byte_count: int) -> str:
)
def _session_update_message(session: Mapping[str, object]) -> str:
return json.dumps({"type": "session.update", "session": dict(session)})
def _sent_audio_mime_type(config: GeminiRealtimeConfig, raw_byte_count: int = 32000) -> str:
sent: Final = config.transform_realtime_request(
_input_audio_append_message(raw_byte_count), "gemini-3.5-transcribe-live"
)
frame: Final = TypeAdapter(dict[str, dict[str, dict[str, str]]]).validate_json(sent[0])
return frame["realtimeInput"]["audio"]["mimeType"]
def test_input_audio_mime_type_declares_the_native_16khz_input_rate():
"""The Live API resamples against the MIME rate, and its documented native input rate is
16kHz; 24kHz is the output rate. Declaring the output rate on the input path mislabels
correctly-encoded audio (ai.google.dev/gemini-api/docs/live-api/capabilities)."""
from typing import Final
config: Final = GeminiRealtimeConfig()
assert config.get_audio_mime_type() == "audio/pcm;rate=16000"
assert _sent_audio_mime_type(config) == "audio/pcm;rate=16000"
def test_vertex_realtime_sends_the_native_16khz_input_rate() -> None:
"""Vertex's default input rate must match the serialized audio frame."""
from typing import Final
from litellm.llms.vertex_ai.realtime.transformation import VertexAIRealtimeConfig
config: Final = VertexAIRealtimeConfig(access_token="t", project="p", location="us-central1")
assert config.get_audio_mime_type() == "audio/pcm;rate=16000"
assert _sent_audio_mime_type(config) == "audio/pcm;rate=16000"
def test_client_declared_input_audio_rate_is_honored_on_the_wire():
"""A GA client that declares a non-native input rate must have that rate forwarded, not the
default; the whole point of the MIME rate is to describe the bytes actually sent."""
from typing import Final
config: Final = GeminiRealtimeConfig()
config.transform_realtime_request(
_session_update_message({"audio": {"input": {"format": {"type": "audio/pcm", "rate": 24000}}}}),
"gemini-3.5-transcribe-live",
)
assert _sent_audio_mime_type(config) == "audio/pcm;rate=24000"
@pytest.mark.parametrize(
"session",
[
{},
{"audio": {}},
{"audio": {"input": {}}},
{"audio": {"input": {"format": "audio/pcm"}}},
{"audio": {"input": {"format": {"type": "audio/pcm"}}}},
{"audio": {"input": {"format": {"type": "audio/pcm", "rate": 0}}}},
{"audio": {"input": {"format": {"type": "audio/pcm", "rate": -1}}}},
{"audio": {"input": {"format": {"type": "audio/pcm", "rate": 7999}}}},
{"audio": {"input": {"format": {"type": "audio/pcm", "rate": 48001}}}},
{"audio": {"input": {"format": {"type": "audio/pcm", "rate": 100_000_000}}}},
{"audio": {"input": {"format": {"type": "audio/pcm", "rate": "24000"}}}},
{"audio": {"input": {"format": {"type": "audio/pcm", "rate": True}}}},
{"input_audio_format": "g711_ulaw"},
],
)
def test_malformed_or_absent_declared_rate_keeps_the_native_default(session):
"""Anything that is not a plausible PCM rate, including a bool (which is an int subclass) and
rates outside 8000-48000, must leave the 16kHz default alone. The out-of-range cases matter
because the rate feeds the spend estimate: an unclamped 100MHz declaration would bill a long
session as a few milliseconds. A beta codec name other than pcm16 is also left alone, because
``get_audio_mime_type`` labels every append pcm16 regardless."""
from typing import Final
config: Final = GeminiRealtimeConfig()
config.transform_realtime_request(_session_update_message(session), "gemini-3.5-transcribe-live")
assert _sent_audio_mime_type(config) == "audio/pcm;rate=16000"
def test_beta_input_audio_format_declares_its_specified_24khz_rate():
"""The beta shape has no rate field, but pcm16 is specified as 24kHz, so a client that sends
the flat codec name has declared 24kHz audio and the MIME label has to say so."""
from typing import Final
config: Final = GeminiRealtimeConfig()
config.transform_realtime_request(
_session_update_message({"input_audio_format": "pcm16"}), "gemini-3.5-transcribe-live"
)
assert _sent_audio_mime_type(config) == "audio/pcm;rate=24000"
def test_beta_session_reaches_the_same_rate_through_the_ga_remap():
"""The proxy only forwards the flat beta shape untouched when the client sent the OpenAI-Beta
header. Without it, RealTimeStreaming rewrites the payload into the GA shape first. Driving the
real converter rather than hand-building the GA dict is what makes this able to fail: both
routes must land on the same rate, or an identical audio stream gets labelled 16kHz or 24kHz
depending on a header that says nothing about sample rates."""
from typing import Final
from litellm.litellm_core_utils.realtime_streaming import RealTimeStreaming
remapped: Final = RealTimeStreaming._remap_beta_session_to_ga({"input_audio_format": "pcm16"})
assert remapped["audio"]["input"]["format"] == {"type": "audio/pcm", "rate": 24000}
via_remap: Final = GeminiRealtimeConfig()
via_remap.transform_realtime_request(_session_update_message(remapped), "gemini-3.5-transcribe-live")
passthrough: Final = GeminiRealtimeConfig()
passthrough.transform_realtime_request(
_session_update_message({"input_audio_format": "pcm16"}), "gemini-3.5-transcribe-live"
)
assert _sent_audio_mime_type(via_remap) == _sent_audio_mime_type(passthrough) == "audio/pcm;rate=24000"
def test_ga_declared_rate_wins_over_the_beta_codec_name():
"""A session carrying both shapes has stated a rate outright; the name-implied one is a
fallback for when it has not."""
from typing import Final
config: Final = GeminiRealtimeConfig()
config.transform_realtime_request(
_session_update_message(
{
"input_audio_format": "pcm16",
"audio": {"input": {"format": {"type": "audio/pcm", "rate": 16000}}},
}
),
"gemini-3.5-transcribe-live",
)
assert _sent_audio_mime_type(config) == "audio/pcm;rate=16000"
def test_declared_rate_also_drives_the_billed_audio_duration(patch_gemini_transcribe_live_cost_map_entry):
"""The MIME label and the duration estimate read the same rate, so a client that declares
24kHz is billed for 24kHz audio: 96000 pcm16 bytes = 2s -> 50 in / 6 out."""
from typing import Final
config: Final = GeminiRealtimeConfig()
config.transform_realtime_request(
_session_update_message({"audio": {"input": {"format": {"type": "audio/pcm", "rate": 24000}}}}),
"gemini-3.5-transcribe-live",
)
config.transform_realtime_request(_input_audio_append_message(96000), "gemini-3.5-transcribe-live")
usage: Final = config.unbilled_usage_on_session_close("gemini-3.5-transcribe-live")
assert usage is not None
assert usage["input_tokens"] == 50
assert usage["output_tokens"] == 6
def test_vertex_records_the_declared_rate_from_its_own_session_update():
"""VertexAIRealtimeConfig handles session.update itself and never calls the parent's handler, so
without recording the rate on that path a Vertex client's declaration is silently discarded."""
from typing import Final
from litellm.llms.vertex_ai.realtime.transformation import VertexAIRealtimeConfig
config: Final = VertexAIRealtimeConfig(access_token="t", project="p", location="us-central1")
config.transform_realtime_request(
_session_update_message({"audio": {"input": {"format": {"type": "audio/pcm", "rate": 24000}}}}),
"gemini-3.5-transcribe-live",
)
assert _sent_audio_mime_type(config) == "audio/pcm;rate=24000"
def test_a_later_rate_declaration_cannot_reprice_already_buffered_audio(
patch_gemini_transcribe_live_cost_map_entry,
):
"""Audio is converted to seconds at the rate in force when it arrived. Otherwise a client could
stream at 16kHz and then declare 24kHz before the estimate is consumed, billing 2/3 of what it
actually sent while the backend still processed all of it."""
from typing import Final
config: Final = GeminiRealtimeConfig()
# 96000 bytes at the 16kHz default is 3s -> 75 in / 9 out.
config.transform_realtime_request(_input_audio_append_message(96000), "gemini-3.5-transcribe-live")
config.transform_realtime_request(
_session_update_message({"audio": {"input": {"format": {"type": "audio/pcm", "rate": 24000}}}}),
"gemini-3.5-transcribe-live",
)
usage: Final = config.unbilled_usage_on_session_close("gemini-3.5-transcribe-live")
assert usage is not None
assert usage["input_tokens"] == 75
assert usage["output_tokens"] == 9
def test_each_chunk_is_billed_at_the_rate_declared_when_it_arrived(
patch_gemini_transcribe_live_cost_map_entry,
):
"""Mixed-rate sessions bill per chunk: 96000 bytes at 16kHz (3s) then 96000 at 24kHz (2s) is 5s
total, not 5s at either single rate."""
from typing import Final
config: Final = GeminiRealtimeConfig()
config.transform_realtime_request(_input_audio_append_message(96000), "gemini-3.5-transcribe-live")
config.transform_realtime_request(
_session_update_message({"audio": {"input": {"format": {"type": "audio/pcm", "rate": 24000}}}}),
"gemini-3.5-transcribe-live",
)
config.transform_realtime_request(_input_audio_append_message(96000), "gemini-3.5-transcribe-live")
usage: Final = config.unbilled_usage_on_session_close("gemini-3.5-transcribe-live")
assert usage is not None
assert usage["input_tokens"] == 125 # 5s * 25 audio tokens/sec
assert usage["output_tokens"] == 15 # round(5 * 175 / 60)
def test_transcribe_live_completed_event_carries_estimated_usage(patch_gemini_transcribe_live_cost_map_entry):
"""Gemini Live sends no usageMetadata for transcribe sessions, so LiteLLM bills
from streamed audio duration at Google's published estimate (25 audio tok/sec in,
175 text tok/min out): 96000 pcm16 bytes = 2s at 24kHz -> 50 in / 6 out."""
175 text tok/min out): 96000 pcm16 bytes = 3s at the Live API's native 16kHz input
rate -> 75 in / 9 out."""
from typing import Final
from litellm.types.llms.gemini import BidiGenerateContentServerMessage
@ -2048,10 +2258,10 @@ def test_transcribe_live_completed_event_carries_estimated_usage(patch_gemini_tr
assert completed[0]["transcript"] == "ahoy there"
expected_usage: Final[RealtimeInputAudioTranscriptionUsage] = {
"type": "tokens",
"input_tokens": 50,
"output_tokens": 6,
"total_tokens": 56,
"input_token_details": {"text_tokens": 0, "audio_tokens": 50},
"input_tokens": 75,
"output_tokens": 9,
"total_tokens": 84,
"input_token_details": {"text_tokens": 0, "audio_tokens": 75},
}
assert completed[0]["usage"] == expected_usage
@ -2114,7 +2324,8 @@ def test_non_transcription_live_model_completed_event_has_no_usage(patch_gemini_
def test_unbilled_usage_on_session_close_flushes_trailing_audio(patch_gemini_transcribe_live_cost_map_entry):
"""Audio appended after the last transcript frame is still unbilled when the
session closes; the session-close hook must hand back the estimate exactly once
so the streaming layer can bill it (144000 pcm16 bytes = 3s -> 75 in / 9 out)."""
so the streaming layer can bill it (144000 pcm16 bytes = 4.5s at the native 16kHz
input rate -> 112 in / 13 out; 112.5 rounds to even)."""
from typing import Final
from litellm.types.realtime import RealtimeInputAudioTranscriptionUsage
@ -2126,10 +2337,10 @@ def test_unbilled_usage_on_session_close_flushes_trailing_audio(patch_gemini_tra
expected: Final[RealtimeInputAudioTranscriptionUsage] = {
"type": "tokens",
"input_tokens": 75,
"output_tokens": 9,
"total_tokens": 84,
"input_token_details": {"text_tokens": 0, "audio_tokens": 75},
"input_tokens": 112,
"output_tokens": 13,
"total_tokens": 125,
"input_token_details": {"text_tokens": 0, "audio_tokens": 112},
}
assert usage == expected
assert config.unbilled_usage_on_session_close("gemini-3.5-transcribe-live") is None