fix(passthrough): charge remote high-detail images at the high-detail upper bound

litellm/litellm_core_utils/token_counter.py

@@ -172,6 +172,13 @@ def calculate_tiles_needed(
     return total_tiles
 
 
+def high_detail_image_token_upper_bound(base_tokens: int = 85) -> int:
+    largest_tile_count: Final = calculate_tiles_needed(
+        MAX_LONG_SIDE_FOR_IMAGE_HIGH_RES, MAX_SHORT_SIDE_FOR_IMAGE_HIGH_RES
+    )
+    return base_tokens + (base_tokens * 2) * largest_tile_count
+
+
 def _unpack_ints(fmt: str, buffer: bytes) -> tuple[int, ...]:
     return struct.unpack(fmt, buffer)
 

litellm/proxy/pass_through_endpoints/llm_provider_handlers/openai_passthrough_logging_handler.py

@@ -13,11 +13,11 @@ import httpx
 
 import litellm
 from litellm._logging import verbose_proxy_logger
-from litellm.constants import DEFAULT_IMAGE_TOKEN_COUNT
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
 from litellm.litellm_core_utils.litellm_logging import (
     get_standard_logging_object_payload,
 )
+from litellm.litellm_core_utils.token_counter import high_detail_image_token_upper_bound
 from litellm.llms.openai.openai import OpenAIConfig
 from litellm.llms.openai.openai import OpenAIConfig as OpenAIConfigType
 from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
@@ -133,7 +133,7 @@ def count_relayed_prompt_tokens(model: str, messages: Sequence[Mapping[str, obje
     ]
     return (
         litellm.token_counter(model=model, messages=local_messages)
-        + DEFAULT_IMAGE_TOKEN_COUNT * remote_high_detail_images
+        + high_detail_image_token_upper_bound() * remote_high_detail_images
     )
 
 

ruff-strict-budget.json

@@ -57,7 +57,7 @@
     "limit": 3
   },
   "BLE001": {
-    "limit": 2904
+    "limit": 2900
   },
   "C401": {
     "limit": 8
@@ -201,7 +201,7 @@
     "limit": 310
   },
   "SIM103": {
-    "limit": 110
+    "limit": 108
   },
   "SIM113": {
     "limit": 3

tests/test_litellm/litellm_core_utils/test_token_counter.py

@@ -1,5 +1,6 @@
 #### What this tests ####
 #    This tests litellm.token_counter.token_counter() function
+import base64
 import importlib
 import time
 import traceback
@@ -14,7 +15,11 @@ import litellm
 from litellm import create_pretrained_tokenizer, decode, encode, get_modified_max_tokens
 from litellm import token_counter as token_counter_old
 import litellm.constants
-from litellm.litellm_core_utils.token_counter import _get_tiktoken_count_function
+from litellm.litellm_core_utils.token_counter import (
+    _get_tiktoken_count_function,
+    calculate_img_tokens,
+    high_detail_image_token_upper_bound,
+)
 from litellm.litellm_core_utils.token_counter import token_counter as token_counter_new
 from tests.large_text import text
 from tests.test_litellm.litellm_core_utils.messages_with_counts import (
@@ -1412,3 +1417,18 @@ def test_openai_file_block_without_inline_bytes_counts_what_it_carries():
     assert _count_user_content([prompt, named]) == _count_user_content(
         [prompt, {"type": "text", "text": "report.pdf"}]
     )
+
+
+def _png_data_url(width: int, height: int) -> str:
+    ihdr = b"\x89PNG\r\n\x1a\n" + (13).to_bytes(4, "big") + b"IHDR" + width.to_bytes(4, "big") + height.to_bytes(4, "big")
+    return "data:image/png;base64," + base64.b64encode(ihdr + b"\x08\x06\x00\x00\x00").decode()
+
+
+@pytest.mark.parametrize(("width", "height"), [(1, 1), (768, 768), (2000, 768), (768, 2000), (4096, 4096), (8000, 3072)])
+def test_high_detail_image_token_upper_bound_covers_every_image_size(width: int, height: int) -> None:
+    assert calculate_img_tokens(_png_data_url(width, height), mode="high") <= high_detail_image_token_upper_bound()
+
+
+def test_high_detail_image_token_upper_bound_is_reached_by_the_largest_high_res_image() -> None:
+    assert calculate_img_tokens(_png_data_url(2000, 768), mode="high") == high_detail_image_token_upper_bound()
+    assert calculate_img_tokens(_png_data_url(1, 1), mode="high") < high_detail_image_token_upper_bound()

tests/test_litellm/llms/azure/passthrough/test_azure_passthrough_transformation.py

@@ -5,9 +5,9 @@ import httpx
 import pytest
 
 import litellm
-from litellm.constants import DEFAULT_IMAGE_TOKEN_COUNT
 
 
+from litellm.litellm_core_utils.token_counter import high_detail_image_token_upper_bound
 from litellm.llms.azure.passthrough.transformation import AzurePassthroughConfig
 from litellm.types.utils import ModelResponse
 
@@ -179,7 +179,7 @@ def test_azure_passthrough_streaming_chunks_count_remote_image_prompt_tokens_wit
     text_only_messages = [{"role": "user", "content": [{"type": "text", "text": "Describe this"}]}]
     assert isinstance(response, ModelResponse)
     assert response.usage.prompt_tokens == (
-        litellm.token_counter(model="gpt-4.1-mini", messages=text_only_messages) + DEFAULT_IMAGE_TOKEN_COUNT
+        litellm.token_counter(model="gpt-4.1-mini", messages=text_only_messages) + high_detail_image_token_upper_bound()
     )
 
 

tests/test_litellm/proxy/pass_through_endpoints/llm_provider_handlers/test_openai_passthrough_logging_handler.py

@@ -8,8 +8,8 @@ import pytest
 
 
 import litellm
-from litellm.constants import DEFAULT_IMAGE_TOKEN_COUNT
 from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj
+from litellm.litellm_core_utils.token_counter import high_detail_image_token_upper_bound
 from litellm.proxy.pass_through_endpoints.llm_provider_handlers.openai_passthrough_logging_handler import (
     OpenAIPassthroughLoggingHandler,
     count_relayed_prompt_tokens,
@@ -2074,12 +2074,12 @@ def test_count_relayed_prompt_tokens_keeps_a_low_detail_remote_image_at_the_base
     assert count_relayed_prompt_tokens("gpt-4.1-mini", messages) == litellm.token_counter(
         model="gpt-4.1-mini", messages=messages
     )
-    assert count_relayed_prompt_tokens("gpt-4.1-mini", messages) < DEFAULT_IMAGE_TOKEN_COUNT
+    assert count_relayed_prompt_tokens("gpt-4.1-mini", messages) < high_detail_image_token_upper_bound()
 
 
-def test_count_relayed_prompt_tokens_estimates_only_the_remote_high_detail_image():
+def test_count_relayed_prompt_tokens_charges_only_the_remote_high_detail_image_at_the_upper_bound():
     messages = _image_messages(UNREACHABLE_IMAGE_URL, "high")
 
     assert count_relayed_prompt_tokens("gpt-4.1-mini", messages) == (
-        litellm.token_counter(model="gpt-4.1-mini", messages=TEXT_ONLY_MESSAGES) + DEFAULT_IMAGE_TOKEN_COUNT
+        litellm.token_counter(model="gpt-4.1-mini", messages=TEXT_ONLY_MESSAGES) + high_detail_image_token_upper_bound()
     )

type-discipline-budget.json

@@ -1,6 +1,6 @@
 {
   "LIT001": {
-    "limit": 22156
+    "limit": 22146
   },
   "LIT002": {
     "limit": 26745
@@ -27,10 +27,10 @@
     "limit": 0
   },
   "LIT010": {
-    "limit": 16434
+    "limit": 16422
   },
   "LIT011": {
-    "limit": 5504
+    "limit": 5502
   },
   "LIT012": {
     "limit": 4486