mirror of
https://github.com/BerriAI/litellm.git
synced 2026-09-30 01:52:18 +00:00
feat(cost): bill azure_ai image edits by measured reference megapixels
This commit is contained in:
shrey kharbanda
parent
dd08db6f79
commit
1ceae6fc43
7 changed files with 259 additions and 55 deletions
|
|
@ -3,22 +3,45 @@ from typing import Any, Final
|
|||
|
||||
import litellm
|
||||
from litellm.litellm_core_utils.llm_cost_calc.utils import (
|
||||
_get_cost_per_unit,
|
||||
_get_cost_per_unit, # pyright: ignore[reportPrivateUsage] # shared rate-resolution helper used by sibling calculators
|
||||
calculate_image_response_cost_from_usage,
|
||||
resolve_image_model_info,
|
||||
)
|
||||
from litellm.types.utils import ImageResponse, ModelInfo
|
||||
|
||||
|
||||
def _input_cost_per_pixel(resolved: ModelInfo) -> float:
|
||||
deployment_price: Final = _get_cost_per_unit(resolved, "input_cost_per_pixel", default_value=None)
|
||||
if deployment_price is not None:
|
||||
return deployment_price
|
||||
model_cost_key: Final = resolved.get("key")
|
||||
shared_entry: Final = litellm.model_cost.get(model_cost_key) if model_cost_key is not None else None
|
||||
if shared_entry is None:
|
||||
return 0.0
|
||||
return shared_entry.get("input_cost_per_pixel") or 0.0
|
||||
def _rate(table: ModelInfo, key: str) -> float | None:
|
||||
return _get_cost_per_unit(table, key, default_value=None)
|
||||
|
||||
|
||||
def _size_pixels(size: str | None) -> int:
|
||||
if size is None:
|
||||
return 0
|
||||
for separator in ("x", "-x-"):
|
||||
if separator in size:
|
||||
parts = size.split(separator)
|
||||
if len(parts) != 2:
|
||||
continue
|
||||
try:
|
||||
width = int(parts[0])
|
||||
height = int(parts[1])
|
||||
except ValueError:
|
||||
continue
|
||||
if width > 0 and height > 0:
|
||||
return width * height
|
||||
continue
|
||||
return 0
|
||||
|
||||
|
||||
def _generated_pixels(
|
||||
optional_params: Mapping[str, object] | None, size: str | None, image_response: ImageResponse
|
||||
) -> int:
|
||||
width: Final = optional_params.get("width") if optional_params else None
|
||||
height: Final = optional_params.get("height") if optional_params else None
|
||||
if type(width) is int and type(height) is int and width > 0 and height > 0:
|
||||
return width * height
|
||||
raw_size: Final = size or image_response.size
|
||||
return _size_pixels(raw_size if isinstance(raw_size, str) else None)
|
||||
|
||||
|
||||
def cost_calculator(
|
||||
|
|
@ -32,44 +55,36 @@ def cost_calculator(
|
|||
"""
|
||||
Azure AI image generation cost calculator
|
||||
"""
|
||||
_model_info: Final = resolve_image_model_info(
|
||||
if not isinstance(image_response, ImageResponse):
|
||||
raise ValueError(f"image_response must be of type ImageResponse got type={type(image_response)}")
|
||||
|
||||
resolved: Final = resolve_image_model_info(
|
||||
model=model,
|
||||
custom_llm_provider=litellm.LlmProviders.AZURE_AI.value,
|
||||
model_info=model_info,
|
||||
)
|
||||
|
||||
if isinstance(image_response, ImageResponse):
|
||||
token_based_cost: Final = calculate_image_response_cost_from_usage(
|
||||
model=model,
|
||||
image_response=image_response,
|
||||
custom_llm_provider=litellm.LlmProviders.AZURE_AI.value,
|
||||
model_info=_model_info,
|
||||
)
|
||||
if token_based_cost is not None:
|
||||
return token_based_cost
|
||||
usage_cost: Final = calculate_image_response_cost_from_usage(
|
||||
model=model,
|
||||
image_response=image_response,
|
||||
custom_llm_provider=litellm.LlmProviders.AZURE_AI.value,
|
||||
model_info=resolved,
|
||||
)
|
||||
if usage_cost is not None:
|
||||
return usage_cost
|
||||
|
||||
num_images: Final = n if n is not None else len(image_response.data or ())
|
||||
output_cost_per_image: Final[float] = _model_info.get("output_cost_per_image") or 0.0
|
||||
if output_cost_per_image:
|
||||
return output_cost_per_image * num_images
|
||||
|
||||
if _input_cost_per_pixel(_model_info):
|
||||
from litellm.cost_calculator import default_image_cost_calculator
|
||||
|
||||
width: Final = optional_params.get("width") if optional_params else None
|
||||
height: Final = optional_params.get("height") if optional_params else None
|
||||
pixel_size: Final = (
|
||||
f"{width}x{height}"
|
||||
if type(width) is int and type(height) is int and width > 0 and height > 0
|
||||
else size or image_response.size
|
||||
)
|
||||
return default_image_cost_calculator(
|
||||
model=_model_info.get("key", model),
|
||||
custom_llm_provider=litellm.LlmProviders.AZURE_AI.value,
|
||||
size=pixel_size,
|
||||
n=num_images,
|
||||
model_info=model_info,
|
||||
)
|
||||
return 0.0
|
||||
|
||||
raise ValueError(f"image_response must be of type ImageResponse got type={type(image_response)}")
|
||||
data: Final = image_response.data
|
||||
num_images: Final = n if n is not None else (len(data) if isinstance(data, list) else 0)
|
||||
generated_meters: Final = (
|
||||
("output_cost_per_image", num_images),
|
||||
("input_cost_per_pixel", _generated_pixels(optional_params, size, image_response) * num_images),
|
||||
)
|
||||
generated: Final = next(
|
||||
(rate * units for key, units in generated_meters if (rate := _rate(resolved, key)) is not None),
|
||||
0.0,
|
||||
)
|
||||
reference_pixels: Final = getattr(image_response, "_reference_pixels", None)
|
||||
reference: Final = (_rate(resolved, "input_cost_per_reference_pixel") or 0.0) * (
|
||||
reference_pixels if isinstance(reference_pixels, int) and not isinstance(reference_pixels, bool) else 0
|
||||
)
|
||||
return generated + reference
|
||||
|
|
|
|||
|
|
@ -257,9 +257,9 @@ class VertexAIVideoConfig(BaseVideoConfig, VertexBase):
|
|||
|
||||
@staticmethod
|
||||
def _supports_resolution_inference(model: str) -> bool:
|
||||
model_key: Final = model if model.startswith("vertex_ai/") else f"vertex_ai/{model}"
|
||||
model_info: Final = litellm.model_cost.get(model_key)
|
||||
return model_info is not None and model_info.get("output_cost_per_second_1080p") is not None
|
||||
model_name: Final = model.removeprefix("vertex_ai/")
|
||||
model_info: Final = litellm.get_model_info(model=model_name, custom_llm_provider="vertex_ai")
|
||||
return model_info.get("output_cost_per_second_1080p") is not None
|
||||
|
||||
def validate_environment(
|
||||
self,
|
||||
|
|
|
|||
|
|
@ -791,6 +791,7 @@ class ModelGroupInfo(BaseModel):
|
|||
input_cost_per_token: float | None = None
|
||||
output_cost_per_token: float | None = None
|
||||
input_cost_per_pixel: float | None = None
|
||||
input_cost_per_reference_pixel: float | None = None
|
||||
mode: (
|
||||
str
|
||||
| Literal["chat", "embedding", "completion", "image_generation", "audio_transcription", "rerank", "moderations"]
|
||||
|
|
|
|||
|
|
@ -6201,6 +6201,8 @@ def _get_model_info_helper(
|
|||
output_cost_per_second_4k=_model_info.get("output_cost_per_second_4k", None),
|
||||
output_cost_per_video_per_second=_model_info.get("output_cost_per_video_per_second", None),
|
||||
output_cost_per_image=_model_info.get("output_cost_per_image", None),
|
||||
input_cost_per_pixel=_model_info.get("input_cost_per_pixel", None),
|
||||
input_cost_per_reference_pixel=_model_info.get("input_cost_per_reference_pixel", None),
|
||||
output_cost_per_pixel=_model_info.get("output_cost_per_pixel", None),
|
||||
output_cost_per_image_token=_model_info.get("output_cost_per_image_token", None),
|
||||
output_cost_per_video_token=_model_info.get("output_cost_per_video_token", None),
|
||||
|
|
|
|||
|
|
@ -127,7 +127,12 @@ def test_flux2_flex_model_info():
|
|||
assert model_info["max_input_tokens"] == 32000
|
||||
assert model_info["max_tokens"] == 32000
|
||||
assert model_info["supported_endpoints"] == ["/v1/images/generations", "/v1/images/edits"]
|
||||
assert catalog_info["input_cost_per_pixel"] == 5e-08
|
||||
assert catalog_info["input_cost_per_reference_pixel"] == catalog_info["input_cost_per_pixel"]
|
||||
assert catalog_info["input_cost_per_pixel"] * 1024 * 1024 == pytest.approx(0.05), (
|
||||
"Azure Retail Prices API, product 'Azure BFL Flux Models', meters 'Flex Megapixel' and "
|
||||
"'Flex Ref Megapixel' are $0.05 per MP where 1 MP = 1024x1024 pixels; confirmed against "
|
||||
"Azure Cost Management usage on 2026-09-22 (1024x1024 image metered as 1.0 MP)"
|
||||
)
|
||||
assert catalog_info["supported_modalities"] == ["text", "image"]
|
||||
assert catalog_info["supported_output_modalities"] == ["image"]
|
||||
|
||||
|
|
@ -148,7 +153,7 @@ def test_flux2_flex_cost_uses_generated_megapixels():
|
|||
call_type="image_generation",
|
||||
)
|
||||
|
||||
assert cost == pytest.approx(5e-08 * 2048 * 1024 * 2)
|
||||
assert cost == pytest.approx(litellm.model_cost["azure_ai/FLUX.2-flex"]["input_cost_per_pixel"] * 2048 * 1024 * 2)
|
||||
|
||||
|
||||
@pytest.mark.parametrize("model", ("FLUX-1.1-pro", "FLUX.1-Kontext-pro"))
|
||||
|
|
@ -189,7 +194,7 @@ def test_flux2_cost_uses_mapped_dimensions_after_response_transformation(dimensi
|
|||
completion_response=response,
|
||||
optional_params=params,
|
||||
call_type="image_generation",
|
||||
) == pytest.approx(5e-08 * 2048 * 1024 * 2)
|
||||
) == pytest.approx(litellm.model_cost["azure_ai/FLUX.2-flex"]["input_cost_per_pixel"] * 2048 * 1024 * 2)
|
||||
|
||||
|
||||
def test_flux2_flex_cost_accepts_lowercase_model_spelling():
|
||||
|
|
@ -202,7 +207,7 @@ def test_flux2_flex_cost_accepts_lowercase_model_spelling():
|
|||
call_type="image_generation",
|
||||
)
|
||||
|
||||
assert cost == pytest.approx(5e-08 * 1536 * 1024 * 2)
|
||||
assert cost == pytest.approx(litellm.model_cost["azure_ai/FLUX.2-flex"]["input_cost_per_pixel"] * 1536 * 1024 * 2)
|
||||
|
||||
|
||||
def test_flux2_flex_cost_prefers_deployment_input_cost_per_pixel() -> None:
|
||||
|
|
|
|||
|
|
@ -0,0 +1,163 @@
|
|||
from typing import Final
|
||||
|
||||
import pytest
|
||||
|
||||
import litellm
|
||||
from litellm.litellm_core_utils.get_model_cost_map import GetModelCostMap
|
||||
from litellm.litellm_core_utils.llm_cost_calc.utils import CostCalculatorUtils
|
||||
from litellm.llms.azure_ai.image_generation.cost_calculator import cost_calculator
|
||||
from litellm.types.utils import ImageObject, ImageResponse, ImageUsage, ImageUsageInputTokensDetails
|
||||
from litellm.utils import _invalidate_model_cost_lowercase_map
|
||||
|
||||
REFERENCE_PIXELS: Final = 2 * 1024 * 1024
|
||||
|
||||
|
||||
@pytest.fixture(autouse=True)
|
||||
def use_local_model_cost_map(monkeypatch: pytest.MonkeyPatch):
|
||||
monkeypatch.setattr(litellm, "model_cost", GetModelCostMap.load_local_model_cost_map())
|
||||
litellm.get_model_info.cache_clear()
|
||||
_invalidate_model_cost_lowercase_map()
|
||||
yield
|
||||
litellm.get_model_info.cache_clear()
|
||||
_invalidate_model_cost_lowercase_map()
|
||||
|
||||
|
||||
def _edit_response(reference_pixels: int | None = REFERENCE_PIXELS, **kwargs: object) -> ImageResponse:
|
||||
response: Final = ImageResponse(data=[ImageObject(b64_json="aW1n")], size="1024x1024", **kwargs)
|
||||
response._reference_pixels = reference_pixels
|
||||
return response
|
||||
|
||||
|
||||
def _edit_cost(response: ImageResponse, **kwargs: object) -> float:
|
||||
return CostCalculatorUtils.route_image_generation_cost_calculator(
|
||||
model="FLUX.2-flex",
|
||||
completion_response=response,
|
||||
custom_llm_provider="azure_ai",
|
||||
size="1024x1024",
|
||||
call_type="image_edit",
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
|
||||
def test_get_model_info_surfaces_flux2_flex_pixel_rates() -> None:
|
||||
model_info = litellm.get_model_info(model="FLUX.2-flex", custom_llm_provider="azure_ai")
|
||||
catalog_info = litellm.model_cost["azure_ai/FLUX.2-flex"]
|
||||
|
||||
assert model_info["input_cost_per_pixel"] == catalog_info["input_cost_per_pixel"]
|
||||
assert model_info["input_cost_per_reference_pixel"] == catalog_info["input_cost_per_reference_pixel"]
|
||||
|
||||
|
||||
def test_flux2_flex_catalog_pixel_rates_match_reference_rates() -> None:
|
||||
catalog_info = litellm.model_cost["azure_ai/FLUX.2-flex"]
|
||||
|
||||
assert catalog_info["input_cost_per_reference_pixel"] == catalog_info["input_cost_per_pixel"]
|
||||
assert catalog_info["input_cost_per_pixel"] * 1024 * 1024 == pytest.approx(0.05), (
|
||||
"Azure Retail Prices API, product 'Azure BFL Flux Models', meters 'Flex Megapixel' and "
|
||||
"'Flex Ref Megapixel' are $0.05 per MP where 1 MP = 1024x1024 pixels; confirmed against "
|
||||
"Azure Cost Management usage on 2026-09-22 (1024x1024 image metered as 1.0 MP)"
|
||||
)
|
||||
|
||||
|
||||
def test_edit_cost_adds_reference_pixels_to_generated_pixels() -> None:
|
||||
catalog_rate: Final = litellm.model_cost["azure_ai/FLUX.2-flex"]["input_cost_per_pixel"]
|
||||
|
||||
cost: Final = _edit_cost(_edit_response())
|
||||
|
||||
assert cost == pytest.approx(catalog_rate * 1024 * 1024 + catalog_rate * REFERENCE_PIXELS)
|
||||
|
||||
|
||||
def test_edit_cost_prefers_deployment_rates_over_catalog() -> None:
|
||||
cost: Final = _edit_cost(
|
||||
_edit_response(), model_info={"input_cost_per_pixel": 2e-07, "input_cost_per_reference_pixel": 3e-07}
|
||||
)
|
||||
|
||||
assert cost == pytest.approx(2e-07 * 1024 * 1024 + 3e-07 * REFERENCE_PIXELS)
|
||||
|
||||
|
||||
def test_edit_cost_honors_explicit_zero_reference_rate() -> None:
|
||||
cost: Final = _edit_cost(
|
||||
_edit_response(), model_info={"input_cost_per_pixel": 2e-07, "input_cost_per_reference_pixel": 0.0}
|
||||
)
|
||||
|
||||
assert cost == pytest.approx(2e-07 * 1024 * 1024)
|
||||
|
||||
|
||||
def test_edit_cost_honors_explicit_zero_generated_rate() -> None:
|
||||
cost: Final = _edit_cost(
|
||||
_edit_response(), model_info={"input_cost_per_pixel": 0.0, "input_cost_per_reference_pixel": 1e-07}
|
||||
)
|
||||
|
||||
assert cost == pytest.approx(1e-07 * REFERENCE_PIXELS)
|
||||
|
||||
|
||||
def test_per_image_rate_beats_per_pixel_rate(monkeypatch: pytest.MonkeyPatch) -> None:
|
||||
cost: Final = _edit_cost(
|
||||
_edit_response(),
|
||||
model_info={
|
||||
"output_cost_per_image": 0.04,
|
||||
"input_cost_per_pixel": 1e-07,
|
||||
"input_cost_per_reference_pixel": 1.5e-08,
|
||||
},
|
||||
)
|
||||
|
||||
assert cost == pytest.approx(0.04 + 1.5e-08 * REFERENCE_PIXELS)
|
||||
|
||||
|
||||
def test_usage_short_circuits_pixel_billing() -> None:
|
||||
response: Final = _edit_response(
|
||||
usage=ImageUsage(
|
||||
input_tokens=150,
|
||||
input_tokens_details=ImageUsageInputTokensDetails(image_tokens=100, text_tokens=50),
|
||||
output_tokens=1000,
|
||||
total_tokens=1150,
|
||||
)
|
||||
)
|
||||
|
||||
cost: Final = _edit_cost(
|
||||
response,
|
||||
model_info={
|
||||
"input_cost_per_pixel": 5e-08,
|
||||
"input_cost_per_reference_pixel": 5e-08,
|
||||
"input_cost_per_token": 1e-05,
|
||||
"input_cost_per_image_token": 2e-05,
|
||||
"output_cost_per_image_token": 4e-05,
|
||||
"output_cost_per_token": 4e-05,
|
||||
},
|
||||
)
|
||||
|
||||
assert cost == pytest.approx(50 * 1e-05 + 100 * 2e-05 + 1000 * 4e-05)
|
||||
|
||||
|
||||
def test_edit_without_measurement_bills_generated_pixels_only() -> None:
|
||||
catalog_rate: Final = litellm.model_cost["azure_ai/FLUX.2-flex"]["input_cost_per_pixel"]
|
||||
|
||||
cost: Final = _edit_cost(_edit_response(reference_pixels=None))
|
||||
|
||||
assert cost == pytest.approx(catalog_rate * 1024 * 1024)
|
||||
|
||||
|
||||
def test_optional_params_dimensions_beat_response_size() -> None:
|
||||
cost: Final = _edit_cost(
|
||||
_edit_response(reference_pixels=None),
|
||||
optional_params={"width": 2048, "height": 1024},
|
||||
)
|
||||
|
||||
assert cost == pytest.approx(litellm.model_cost["azure_ai/FLUX.2-flex"]["input_cost_per_pixel"] * 2048 * 1024)
|
||||
|
||||
|
||||
def test_unlisted_model_bills_deployment_rates() -> None:
|
||||
cost: Final = CostCalculatorUtils.route_image_generation_cost_calculator(
|
||||
model="unlisted-flux-deployment",
|
||||
completion_response=_edit_response(),
|
||||
custom_llm_provider="azure_ai",
|
||||
size="1024x1024",
|
||||
call_type="image_edit",
|
||||
model_info={"input_cost_per_pixel": 1e-07, "input_cost_per_reference_pixel": 1e-07},
|
||||
)
|
||||
|
||||
assert cost == pytest.approx(1e-07 * 1024 * 1024 + 1e-07 * REFERENCE_PIXELS)
|
||||
|
||||
|
||||
def test_non_image_response_raises() -> None:
|
||||
with pytest.raises(ValueError):
|
||||
cost_calculator(model="FLUX.2-flex", image_response=object())
|
||||
|
|
@ -229,7 +229,25 @@ def test_model_info_rejects_offset_aware_access_window_times():
|
|||
with pytest.raises(ValidationError):
|
||||
ModelInfo(
|
||||
id="x",
|
||||
access_windows=[
|
||||
{"start": "22:00+05:00", "end": "06:00", "timezone": "UTC", "team_ids": ["t"]}
|
||||
],
|
||||
access_windows=[{"start": "22:00+05:00", "end": "06:00", "timezone": "UTC", "team_ids": ["t"]}],
|
||||
)
|
||||
|
||||
|
||||
def test_model_group_info_surfaces_input_cost_per_reference_pixel():
|
||||
from typing import Final
|
||||
|
||||
from litellm import Router
|
||||
|
||||
configured_rate: Final = 3e-7
|
||||
router: Final = Router(
|
||||
model_list=[
|
||||
{
|
||||
"model_name": "flux-edit",
|
||||
"litellm_params": {"model": "azure_ai/FLUX.2-flex"},
|
||||
"model_info": {"id": "dep-flux-1", "input_cost_per_reference_pixel": configured_rate},
|
||||
}
|
||||
]
|
||||
)
|
||||
group: Final = router.get_model_group_info("flux-edit")
|
||||
assert group is not None
|
||||
assert group.input_cost_per_reference_pixel == configured_rate
|
||||
|
|
|
|||
Loading…
Add table
Reference in a new issue