[Bug]: Whisper (PyTorch backend) fails on any repeated request when KV block reuse is on: "Request requires multimodal_embed_mask_cumsum for chunked prefill or KV-cache reuse"

[ravo42-maker]

System Info

x86
5070ti

Who can help?

Follow-up to #18609 / #19450. With the #19450 guard applied (built from v1.3.0rc25) and
enable_block_reuse left at its default (True), the first Whisper request transcribes, but the
second request that shares a prompt prefix with an earlier one (e.g. a warm-up followed by real
traffic, or any two clips with the same decoder prompt) fails:

ValueError: Request requires multimodal_embed_mask_cumsum for chunked prefill or KV-cache reuse (begin_compute=3, end_compute=4, prompt_len=4) but py_multimodal_data has keys {'mm_bidirectional_blocks', 'audio'} with no cumsum. The input processor may be missing a discriminator (override get_mm_token_ids or ensure get_vocab_size resolves).
  File ".../tensorrt_llm/_torch/pyexecutor/model_engine.py", line 5300, in _prepare_tp_inputs
  File ".../tensorrt_llm/inputs/multimodal.py", line 1144, in check_mm_embed_cumsum_if_needed

Environment

RTX 5070 Ti (sm_120), nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc25 + #19450 (capacityScheduler.cpp hunk)

• the cross-KV sizing fix from the sibling issue, openai/whisper-large-v3, LLM API directly.

Reproduction

Same script as #18609 with max_input_len=1500, enable_block_reuse=True (default), and the
llm.generate(...) call executed twice in the same process on the same 30 s clip. Run 1 succeeds,
run 2 raises the error above.

Expected

Either the Whisper input processor provides the multimodal discriminator (get_mm_token_ids) so the
reuse path can compute the cumsum, or KV-cache reuse is disabled automatically for features-based
encoders. With KvCacheConfig(enable_block_reuse=False, ...) everything is stable across repeated
requests (0.34 s per 30 s clip warm, fp16).

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

.

Expected behavior

.

actual behavior

.

additional notes

.

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.