Merge pull request #37377 from BerriAI/devin_ai_lit5757_dashscope_nested_cache_creation

fix(types): map nested prompt_tokens_details.cache_creation_input_tokens to cache_write_tokens

litellm/types/utils.py

@@ -1631,8 +1631,20 @@ class PromptTokensDetailsWrapper(
 
     def __init__(self, *args, **kwargs) -> None:
         super().__init__(*args, **kwargs)
+        extra_fields: Final = self.model_extra
+        nested_cache_creation_input_tokens: Final = (
+            extra_fields.get("cache_creation_input_tokens") if extra_fields is not None else None
+        )
         self.cache_write_tokens = (
-            self.cache_write_tokens if self.cache_write_tokens is not None else self.cache_creation_tokens
+            self.cache_write_tokens
+            if self.cache_write_tokens is not None
+            else (
+                self.cache_creation_tokens
+                if self.cache_creation_tokens is not None
+                else (
+                    nested_cache_creation_input_tokens if isinstance(nested_cache_creation_input_tokens, int) else None
+                )
+            )
         )
         if self.character_count is None:
             del self.character_count

tests/test_litellm/llms/dashscope/test_dashscope_cost_calculator.py

@@ -271,6 +271,47 @@ class TestDashscopeCostCalculator:
 
         assert math.isclose(prompt_cost, expected_prompt_cost, rel_tol=1e-10)
 
+    def test_dashscope_nested_cache_creation_input_tokens_bill_at_cache_write_rate(self):
+        """
+        Regression (LIT-5757): DashScope nests cache_creation_input_tokens inside
+        prompt_tokens_details; those tokens must bill at the tier's cache-creation
+        rate instead of being folded into text tokens at the input rate.
+        """
+        self._register_tiered_model(
+            "dashscope/qwen-nested-cache-write-test",
+            [
+                {
+                    "range": [0, 128000],
+                    "input_cost_per_token": 4e-07,
+                    "cache_read_input_token_cost": 1.6e-07,
+                    "cache_creation_input_token_cost": 5e-07,
+                    "output_cost_per_token": 1.6e-06,
+                }
+            ],
+        )
+
+        usage = Usage(
+            prompt_tokens=2059,
+            completion_tokens=201,
+            total_tokens=2260,
+            prompt_tokens_details={
+                "cached_tokens": 0,
+                "text_tokens": 2059,
+                "cache_type": "ephemeral",
+                "cache_creation_input_tokens": 2048,
+                "cache_creation": {"ephemeral_5m_input_tokens": 2048},
+            },
+            completion_tokens_details={"reasoning_tokens": 170},
+        )
+
+        prompt_cost, _ = dashscope_cost_per_token(
+            model="qwen-nested-cache-write-test", usage=usage
+        )
+
+        assert math.isclose(
+            prompt_cost, (2048 * 5e-07) + (11 * 4e-07), rel_tol=1e-10
+        )
+
     def test_dashscope_tiered_cache_creation_falls_back_to_tier_input_rate(self):
         """
         Tiers without a cache_creation_input_token_cost bill cache-creation tokens at

tests/test_litellm/proxy/spend_tracking/test_spend_tracking_utils.py

@@ -124,6 +124,30 @@ def test_get_logging_payload_maps_openai_cache_write_tokens_to_cache_creation_in
     assert additional_usage_values["prompt_tokens_details"]["cache_write_tokens"] == 800
 
 
+def test_get_logging_payload_maps_nested_cache_creation_input_tokens():
+    """
+    Regression (LIT-5757): DashScope nests cache_creation_input_tokens inside
+    prompt_tokens_details; SpendLogs must record it as cache_creation_input_tokens.
+    """
+    additional_usage_values: Final = _get_additional_usage_values_for_usage(
+        litellm.Usage(
+            prompt_tokens=2059,
+            completion_tokens=31,
+            total_tokens=2090,
+            prompt_tokens_details={
+                "cached_tokens": 0,
+                "text_tokens": 2059,
+                "cache_type": "ephemeral",
+                "cache_creation_input_tokens": 2048,
+                "cache_creation": {"ephemeral_5m_input_tokens": 2048},
+            },
+        )
+    )
+
+    assert additional_usage_values["cache_creation_input_tokens"] == 2048
+    assert additional_usage_values["prompt_tokens_details"]["cache_write_tokens"] == 2048
+
+
 def test_get_logging_payload_preserves_anthropic_cache_creation_input_tokens():
     additional_usage_values = _get_additional_usage_values_for_usage(
         litellm.Usage(

tests/test_litellm/test_responses_api_bridge_non_stream.py

@@ -1,6 +1,6 @@
 import os
 import sys
-from typing import Optional
+from typing import Final, Optional
 from unittest.mock import Mock
 
 import pytest
@@ -192,6 +192,32 @@ def test_transform_usage_with_cached_tokens_only():
     print("✓ Transformation works with cached_tokens only")
 
 
+def test_transform_usage_maps_nested_cache_creation_input_tokens():
+    """
+    Regression (LIT-5757): DashScope nests cache_creation_input_tokens inside
+    prompt_tokens_details; the bridge must surface it as cache_write_tokens.
+    """
+    usage: Final = Usage(
+        prompt_tokens=2059,
+        completion_tokens=31,
+        total_tokens=2090,
+        prompt_tokens_details={
+            "cached_tokens": 0,
+            "text_tokens": 2059,
+            "cache_type": "ephemeral",
+            "cache_creation_input_tokens": 2048,
+            "cache_creation": {"ephemeral_5m_input_tokens": 2048},
+        },
+    )
+
+    responses_usage: Final = LiteLLMCompletionResponsesConfig._transform_chat_completion_usage_to_responses_usage(
+        usage
+    )
+
+    assert responses_usage.input_tokens_details is not None
+    assert responses_usage.input_tokens_details.cache_write_tokens == 2048
+
+
 def test_transform_usage_with_reasoning_tokens_only():
     """
     Test transformation when only reasoning_tokens is provided (no cached_tokens).

tests/test_litellm/types/test_types_utils.py

@@ -1,5 +1,6 @@
 import os
 import sys
+from typing import Final
 
 import pytest
 
@@ -75,6 +76,29 @@ def test_usage_dump():
     assert new_usage.prompt_tokens_details.web_search_requests == 1
 
 
+def test_prompt_tokens_details_maps_nested_cache_creation_input_tokens():
+    """Regression (LIT-5757): DashScope nests the Anthropic-spelled
+    cache_creation_input_tokens inside prompt_tokens_details. It must populate
+    the canonical cache_write_tokens/cache_creation_tokens pair, without
+    overriding an explicitly provided canonical value."""
+    from litellm.types.utils import PromptTokensDetailsWrapper
+
+    nested: Final = PromptTokensDetailsWrapper(
+        cached_tokens=0, text_tokens=2059, cache_creation_input_tokens=2048
+    )
+    assert nested.cache_write_tokens == 2048
+    assert nested.cache_creation_tokens == 2048
+
+    explicit: Final = PromptTokensDetailsWrapper(
+        cache_write_tokens=100, cache_creation_input_tokens=2048
+    )
+    assert explicit.cache_write_tokens == 100
+    assert explicit.cache_creation_tokens == 100
+
+    non_int: Final = PromptTokensDetailsWrapper(cache_creation_input_tokens=None)
+    assert not hasattr(non_int, "cache_write_tokens")
+
+
 def test_usage_server_tool_use_dict_is_coerced_and_round_trips():
     from litellm.types.utils import ServerToolUse, Usage