[mm] Accept per-item embedding lists from DataEmbeddingFunc

[houseroad]

Motivation

DataEmbeddingFunc currently must return a single combined [tokens, hidden] tensor (or an EVSEmbeddingResult). Some encoders naturally produce per-item outputs — for example, an audio AutoEncoder that loops over clips. To satisfy the current contract they have to torch.cat their per-item results, which the per-item consumer (_get_chunked_embedding_by_item) then immediately splits back apart via torch.split. Those split views also pin the whole concatenated buffer for as long as any single item stays cached.

Modifications

• DataEmbeddingFunc may now also return List[torch.Tensor] (one tensor per input item).
• _get_chunked_embedding_by_item keeps the per-item tensors as separate cache entries — no concat, no split, and each cache entry owns its storage.
• _get_chunked_embedding_full flattens the per-item form back to one tensor (it caches the combined per-request embedding) via a new _flatten_embedding_result helper.
• The combined-tensor path is unchanged.

The per-item form is opt-in; existing encoders returning a single tensor or EVSEmbeddingResult are unaffected.

Tests

Adds CPU-only unit tests (test/registered/chunked_prefill/test_mm_chunked_embedding_unit.py) that exercise mm_utils internals directly (no engine/GPU) and assert:

• the per-item and combined forms produce bitwise-identical chunked-prefill embeddings for both _get_chunked_embedding_by_item and _get_chunked_embedding_full;
• per-item cache entries own their storage, while combined-tensor split views share the concatenated buffer.

These tests could not be executed in the porting environment (no torch/pytest available); they run identically in the source environment.

Original commits

• d4fd0560e

---

CI States

Latest PR Test (Base): ✅ Run #30053812808
Latest PR Test (Extra): ❌ Run #30053812690

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[houseroad]

/tag-and-rerun-ci

[houseroad]

/rerun-failed-ci

[houseroad]

/rerun-failed-ci

[houseroad]

/rerun-failed-ci

[houseroad]

CI recovery needed for the required pr-test.yml run: https://github.com/sgl-project/sglang/actions/runs/29940709888

The run was cancelled while jobs were queued/running; no executed test job failed. The PR-author /rerun-failed-ci command cannot recover it because the slash handler currently processes only workflow conclusions failure and skipped, not cancelled, and the PR author does not have Actions write permission.

Could a maintainer please use Re-run failed jobs on that run? This will retain the successful jobs and rerun only the cancelled/failed dependency chain.

[merrymercy]

/rerun-test test/registered/vlm/test_vision_chunked_prefill.py test/registered/vlm/test_vlm_input_format.py

[github-actions]

Results for /rerun-test test/registered/vlm/test_vision_chunked_prefill.py test/registered/vlm/test_vlm_input_format.py:

🚀 1-gpu-h100 (2 tests): ✅ View workflow run

cd test/ && python3 registered/vlm/test_vision_chunked_prefill.py
cd test/ && python3 registered/vlm/test_vlm_input_format.py

[mingfeima]

@houseroad this pr failed xeon ci: as you can see here: https://github.com/sgl-project/sglang/actions/runs/30229696257/job/89886278479?pr=24959 and every pr-test-xeon workflow.

Could you please fix this?

Two things need to be updated:

• change register_cpu_ci(est_time=10, suite="base-b-test-cpu") to register_cpu_ci(est_time=10, suite="base-a-test-cpu"). base-a right now is generic cpu test and base-b is intel xeon cpu test (test/registered/cpu/test_xxx.py)
• probably need to update _acknowledge_deferred_cuda_ipc_cache_hits, make an early-return when no item has a CudaIpcTensorTransportProxy feature, before touching attn_tp_rank. that depends on whether tp rank has been set on base-a-test-cpu host machine. (this ci machine is not managed by intel, base-b-test-cpu is and tp rank not set on the intel machine, that's the root cause of the fail)

CC @1pikachu @ZailiWang