[Model]Support MiniCPM-o 4.5 (#3642)

Signed-off-by: tc-mb <tianchi_cai@icloud.com>
Co-authored-by: GKangaroo <1095103651@qq.com>
Co-authored-by: GKangaroo <gqx24@mails.tsinghua.edu.cn>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

pyproject.toml

@@ -76,6 +76,18 @@ local = [
     "FlagEmbedding",
 ]
 
+# Optional runtime dependencies for the MiniCPM-o 4.5 talker pipeline.
+# The talker loads MiniCPMTTS via `trust_remote_code` from the model repo,
+# and that module performs `from stepaudio2 import Token2wav` against the
+# MiniCPM-o-flavored Token2wav vocoder (PyPI: `stepaudio2-minicpmo`, NOT
+# upstream `stepfun-ai/Step-Audio2` — the upstream signature rejects
+# `n_timesteps` and crashes at construction). A clean `vllm-omni` install
+# does not pull this in by default to avoid forcing the dep onto users
+# who only need other models; opt in with `pip install vllm-omni[minicpmo]`.
+minicpmo = [
+    "stepaudio2-minicpmo",
+]
+
 docs = [
     "mkdocs>=1.5.0",
     "mkdocs-api-autonav",
@@ -101,6 +113,13 @@ Documentation = "https://vllm-omni.readthedocs.io"
 [project.scripts]
 vllm-omni = "vllm_omni.entrypoints.cli.main:main"
 
+# Register vllm-omni's model registrations on every vllm subprocess. vllm
+# auto-loads anything advertised under `vllm.general_plugins` in its own
+# `load_general_plugins()`, so worker subprocesses spawned by vllm (which
+# only import vllm, not vllm_omni) will still pick up the omni archs.
+[project.entry-points."vllm.general_plugins"]
+vllm_omni_register_models = "vllm_omni.engine.arg_utils:register_omni_models_to_vllm"
+
 
 [tool.setuptools.packages.find]
 where = ["."]
@@ -149,6 +168,10 @@ ignore = [
 [tool.ruff.lint.per-file-ignores]
 "examples/**" = ["E501"]  # Allow long lines in examples
 "tests/**" = ["E501"]  # Allow long lines in tests
+# MiniCPM-o reference modeling code vendored from upstream HF/transformers/MiniCPM-o.
+# These files keep upstream conventions (star imports, long lines, math-symbol
+# argument names) so they stay easy to diff against upstream releases.
+"vllm_omni/model_executor/models/minicpmo_4_5/minicpmo_4_5_omni_llm.py" = ["E402", "E501", "F403", "F405", "N803"]
 
 [tool.ruff.lint.flake8-tidy-imports.banned-api]
 "librosa".msg = "The librosa module is banned, use vllm.multimodal helpers instead"