OmniRenderer wrapped the upstream renderer, forwarded unknown attributes via
__getattr__, and re-bound eight BaseRenderer methods (render_cmpl, render_chat,
process_for_engine, _process_enc_dec and their async twins) onto the wrapper
so that they would reach Omni's _process_singleton. That allow-list hid any
concrete-renderer override of those methods and would silently skip the
Omni extras for any future BaseRenderer entry point not on the list.
Make OmniRenderer a mixin and create the renderer as a real subclass of the
upstream class instead:
- omni_renderer_cls(base_cls) builds (and caches) Omni<Base> = (OmniRenderer,
base_cls); only _process_singleton / _process_singleton_async are extended,
everything else is the upstream renderer unchanged.
- build_omni_renderer(vllm_config) resolves the class the way upstream
renderer_from_config does (cached_tokenizer_from_config,
tokenizer_args_from_config, RENDERER_REGISTRY.load_renderer_cls) and
instantiates the Omni subclass; renderer_cls/tokenizer can be passed to
skip the resolution.
- build_stage0_input_processor passes the renderer to InputProcessor up
front instead of replacing it afterwards; the skip-tokenizer branch keeps
_build_token_only_renderer and never resolves a tokenizer.
Rebinding an existing instance's __class__ was rejected: CPython refuses it
for HfRenderer ("object layout differs"), so the subclass is constructed.
Tests: tests/inputs/test_preprocess.py builds the renderer with
omni_renderer_cls and adds checks that the subclass is created once, that a
concrete renderer's render_cmpl override is no longer suppressed, and that
build_omni_renderer resolves like upstream or honours an explicit class;
tests/engine/test_async_omni_engine_stage_init.py asserts the stage-0
processor receives the built subclass instance and that the skip-tokenizer
branch never calls build_omni_renderer. Two pre-existing mypy complaints in
test_preprocess.py are fixed with annotations.
Validation: ruff, pre-commit (incl. mypy) on the four files; the modules
import and omni_renderer_cls(HfRenderer) builds; tests/inputs (32),
tests/engine/test_async_omni_engine_stage_init.py (55) and tests/entrypoints
(548) pass on CPU.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MpXe2LeRrUQGjZfKFaocMZ
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Reuse upstream weight mapping and multimodal processing, consolidate Omni processor bridges, and trim unused Hunyuan internals. Route no-media kwargs through the active renderer and preserve GLM generation prompt formatting.
Validation: 453 CPU tests passed, 7 skipped; processor imports, Ruff, and diff checks passed. GPU and end-to-end CI not rerun.
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
Five further removals/renames, each traced from a failing 3009 job:
- StageEngineCoreProcManager mirrors the parent __init__ rather than calling
super(), and 0.29 added _request_shutdown_timeout there. The inherited
shutdown() read it, raised AttributeError, and left the engine core
subprocesses running, so the test process hung at interpreter exit until
the 15 minute job timeout. That is the whole of the Engine&Model Executor
exit-124: its pytest reported 567 passed, then sat idle for 663s. The
failure surfaced only as a StagePool warning, never as a test failure.
- PromptUpdateDetails.select_text is gone; prompt updates are token oriented
now. MiniCPM-o 4.5 kept three text-based image/video/audio replacements,
which crashed stage-0 startup and account for the whole MiniCPM-o cluster.
They encode the text and select the unk placeholder by token id instead.
- Qwen3-Omni: 0.29 marks audio_feature_lengths keep_on_cpu, while the audio
tower derives device placement from feature_lens, so the mask comparison
mixed cuda and cpu tensors. Moved explicitly, matching upstream
Qwen3OmniMoeThinker._process_audio_input.
- AttentionBackend.get_kv_cache_shape is gone; the per-layer shape comes from
the spec via compute_layer_kv_cache_shape_bytes, whose content dim is in
bytes.
- WeightsMapper.get_unstacked_mapper is renamed to get_rename_mapper.
tests/engine/test_async_omni_engine_abort.py::test_abort fails here both with
and without these changes, so it is environmental rather than a regression.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0146Xu5Sgj4vXxLxvLjV5YPj
Signed-off-by: tzhouam <tzhouam@connect.ust.hk>
