- api_router: stop silently overriding --enable-prompt-tokens-details for
AnthropicServingMessages; pass through the user's CLI setting like the
other serving objects.
- _compute_cache_usage: rewrite docstring to document where
prompt_tokens_details is attached in vLLM's OpenAI streaming path
(terminal include_usage chunk only), why message_start cannot populate
cache fields today, and why cache_creation_input_tokens defaults to 0
rather than None when cache info is present.
- AnthropicUsage construction: omit cache fields entirely when the
underlying cache info is unknown (cache_read is None), rather than
emitting null. Applied uniformly to non-streaming responses,
message_start, and message_delta so "unknown" is signaled by key
absence rather than null, distinguishing it from a real zero.
- Tests: add TestStreamingCacheUsageSemantics covering the three usage
states (cache hit, cache miss with details, no details at all) for
both message_start and message_delta.
Signed-off-by: mistral0105 <zhangshuoming17@mails.ucas.ac.cn>
mistral0105
fd3e0cac12