[Bug] Fix 3 failing + 2 erroring tests in CI on `main`

[zhaochenyang20]

CI Run: https://github.com/sgl-project/sglang-omni/actions/runs/23229772248/job/67520568093
Branch: main (commit cd9aaf3)
Result: 3 failed, 2 errors, 27 passed, 2 skipped

---

Summary

The PR Test CI workflow (pytest tests/ -v) has 5 broken test cases on main, caused by two independent issues:

1. Encoder cache (cache_key) is not working — 3 FAILED tests
2. Missing preprocessor_config.json in CI model cache — 2 ERROR tests

---

Issue 1: Encoder cache tests failing (3 tests)

Failing tests

Test	Expected forward_count	Actual
test_cache_key.py::TestCacheHit::test_same_key_hits_cache	1	2
test_cache_key.py::TestCacheHit::test_multiple_hits	1	5
test_cache_key.py::TestCacheEviction::test_lru_eviction	3 (after hitting key-1 from cache)	4

Root cause

The encoder engine's result cache (use_cache=True) appears to never actually cache results. Every request with the same cache_key still triggers a model forward call, as if the cache lookup always misses.

Relevant code

• Test: tests/test_cache_key.py
• Engine factory: sglang_omni.engines.omni.create_encoder_engine
• Likely bug location: the cache get/set logic in the encoder engine runtime

---

Issue 2: Video integration tests erroring at setup (2 tests)

Erroring tests

Test	Error
test_video_integration.py::test_two_round_conversation	ERROR at setup (server_process fixture)
test_video_integration.py::test_server_stability_after_request	ERROR at setup (server_process fixture)

Root cause

Both tests share a server_process module-scoped fixture that launches sglang_omni.cli.cli serve --model-path Qwen/Qwen3-Omni-30B-A3B-Instruct. The server exits immediately with code 1 because:

1. Qwen3OmniMoeProcessor.from_pretrained() fails with an IndexError inside transformers image_processing_base.py:354 — the HuggingFace cache snapshot is likely incomplete or corrupt.
2. The fallback path _build_processor_local() then tries to open preprocessor_config.json directly from the model directory, which does not exist in the CI cache:

FileNotFoundError: [Errno 2] No such file or directory:
'/github/home/.cache/huggingface/hub/models--Qwen--Qwen3-Omni-30B-A3B-Instruct/
snapshots/26291f793822fb6be9555850f06dfe95f2d7e695/preprocessor_config.json'

Error traceback (abbreviated)

sglang_omni/models/qwen3_omni/components/preprocessor.py:108  # from_pretrained fails
  → preprocessor.py:116  # fallback to _build_processor_local
    → preprocessor.py:66  # calls _load_preprocessor_config
      → preprocessor.py:45  # FileNotFoundError on preprocessor_config.json

Relevant code

• Test fixture: tests/test_model/test_video_integration.py::server_process
• Preprocessor init: sglang_omni/models/qwen3_omni/components/preprocessor.py (lines 45, 66, 108, 116)

Possible fixes

• Ensure the model cache on the CI runner includes preprocessor_config.json (re-download or update the cached snapshot).
• Or: make the preprocessor more resilient when the file is missing (e.g., provide sensible defaults or download the missing config file on the fly).

---

Additional notes

• There is also a warning about fishaudio_s2_pro failing to import AssistantStreamEvent from openai.types.beta.assistant_stream_event — this doesn't cause test failures currently but indicates a dependency version mismatch with the openai package.
• The CI uses a pinned Docker image frankleeeee/sglang-omni:dev with Python 3.11.14.

[Ratish1]

Hey @zhaochenyang20 , I have opened the PR to fix this issue. Lmk what you think. Thanks