DashScope inherits OpenAIGPTConfig which strips cache_control from
messages and tools by default. Override remove_cache_control_flag_from_messages_and_tools()
to preserve cache_control, following the same pattern used by ZAI, MiniMax, and Databricks.
Verified through 10-round multi-turn conversation tests:
- Explicit caching works correctly: cached_tokens grows each round from R4 onwards,
with cache_creation_tokens reported on first cache build.
- Implicit caching is not affected: models that rely on implicit prefix-matching caching
produce identical cached_tokens with and without this change, confirmed by comparing
results against both the reverted codebase and direct API calls bypassing litellm.
- No errors or regressions observed on any model, including those that do not support
explicit caching — the DashScope API silently ignores unrecognized cache_control fields.
Fixes#25330
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* add helper functions
* update generic_cost_per_token function
* add test
* formatting
* add examples in docstring for _calculate_tiered_cost
* Restore files to upstream/main version
* dashscope specific calculation
* improve for different costs
* remove _calculate_flat_cost function
* Added dashscope as a provider
* Fix some leftover references on nebius
* Porting the dashscope api endpoit international version
* explicit tool_choice = True in config
Ishaan Jaffer