The tap records block_hashes + output_token_ids but no per-request multimodal identity.
Problem: two requests with the same text prefix but different images are indistinguishable, so multimodal prefix-cache replay can't tell them apart.
Proposed: capture vLLM's per-input mm_hashes in the trace record.
The tap records
block_hashes+output_token_idsbut no per-request multimodal identity.Problem: two requests with the same text prefix but different images are indistinguishable, so multimodal prefix-cache replay can't tell them apart.
Proposed: capture vLLM's per-input
mm_hashesin the trace record.