Populate cache_read_input_tokens and cache_creation_input_tokens in
the Anthropic Messages API response, which were previously always None.
Key changes:
- Add _get_cached_tokens() and _compute_cache_usage() helpers to map
vLLM's prefix cache hits to Anthropic's usage format
- Fix input_tokens semantics: Anthropic defines total_input =
input_tokens + cache_read + cache_creation, so input_tokens must
exclude cached tokens (previously it included them)
- Set cache_creation_input_tokens to 0 when cache info is available
(vLLM's prefix caching only tracks cache reads, not writes)
- Force enable_prompt_tokens_details=True for AnthropicServingMessages
so cache fields are always populated regardless of CLI flag
- Cover all three AnthropicUsage construction sites: non-streaming
full response, streaming message_start, and streaming message_delta
Fixes#33923
Co-authored-by: Claude
Signed-off-by: mistral0105 <zhangshuoming17@mails.ucas.ac.cn>
