[Bugfix] Fix the remaining Python 3.10 breakages (#7985)

Signed-off-by: Armaan Amatya <armaanamatya2014@gmail.com>
Signed-off-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>
Co-authored-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>

pyproject.toml

@@ -56,7 +56,9 @@ dev = [
     # 0.3.12 includes acknowledged TCP writes (Mooncake #2850): completion
     # must mean the receive buffer is ready before zero-copy consumers read it.
     "mooncake-transfer-engine-cuda13==0.3.12.post1",
-    "av==18.0.0", # for ComfyUI tests
+    "av==18.0.0; python_version >= '3.11'", # for ComfyUI tests
+    "av==17.1.0; python_version < '3.11'", # last av release that supports Python 3.10
+    "tomli==2.4.1; python_version < '3.11'", # tomllib backport used by tests/helpers/mark.py
     "openpyxl==3.1.5", # for nightly CI
     "pyttsx3==2.99",
     "opencc==1.4.1",

requirements/common.txt

@@ -7,6 +7,11 @@
 # transformers_keys_to_ignore_compat in model_executor/models/utils.py.
 transformers >= 5.13.0, < 5.15
 av>=14.0.0
+# typing_extensions.Concatenate accepts a trailing `...` on 3.10 only from
+# 4.13 (older 4.12.x re-export typing.Concatenate, which rejects it). The
+# Nemotron VoiceChat codec annotates `Concatenate[int, ...]`, so pin the
+# floor instead of relying on the version vLLM/torch pull transitively.
+typing_extensions>=4.13
 omegaconf>=2.3.0
 diffusers==0.40.0
 # Hub kernels, including LTX-2.5 NATTEN. 0.16+ resolves torch-stable-abi*

tests/diffusion/streaming_interactive_generation/test_diffusion_streaming_output.py

@@ -470,9 +470,15 @@ class TestPipelineStreamingOutputToEntrypoint:
         fixture: OrchestratorFixture,
         inline_client: InlineStageDiffusionClient,
     ) -> None:
-        if omni.final_output_task is not None:
-            omni.final_output_task.cancel()
-            await asyncio.gather(omni.final_output_task, return_exceptions=True)
+        task = omni.final_output_task
+        omni.final_output_task = None
+        # The session ran inside ``TestClient``, whose loop is closed by the
+        # time the ``with`` block exits and the task with it already cancelled.
+        # Only a still-running task is worth cancelling and awaiting; gathering
+        # a finished one touches its closed loop on Python 3.10.
+        if task is not None and not task.done():
+            task.cancel()
+            await asyncio.gather(task, return_exceptions=True)
         inline_client.shutdown()
         fixture.request_sync_q.put_nowait(ShutdownRequestMessage())
         await asyncio.to_thread(fixture.thread.join, 5)

tests/engine/duplex/test_duplex_commands.py

@@ -451,3 +451,20 @@ def test_a_hint_survives_when_its_typed_field_is_unset():
     command = AppendAudio(audio=b"\x00\x00" * 8, hints={"is_speech": True})
 
     assert command.payload()["is_speech"] is True
+
+
+def test_duplex_command_mixin_keeps_empty_slots():
+    """The concrete commands combine ``DuplexCommand`` with a slotted wire command.
+
+    On Python 3.10 ``dataclass(slots=True)`` repeats the inherited fields in
+    ``__slots__`` (fixed in 3.11), so a slotted ``DuplexCommand`` gives every
+    concrete command two slotted bases and the module does not import (#7475).
+    The mixin spells out empty slots instead; instances stay ``__dict__``-free
+    on every version.
+    """
+    # ``__dataclass_params__`` only records ``slots`` from Python 3.12 on; an
+    # older interpreter proves the same thing by importing the module at all.
+    params = getattr(DuplexCommand, "__dataclass_params__")
+    assert getattr(params, "slots", False) is False, "slots=True on DuplexCommand breaks the import on Python 3.10"
+    assert DuplexCommand.__slots__ == ()
+    assert not hasattr(AppendAudio(audio=b"\x00\x00" * 8), "__dict__")

tests/model_executor/models/indextts2/test_optional_dependencies.py

@@ -9,9 +9,13 @@ import types
 from pathlib import Path
 
 import pytest
-import tomllib
 from packaging.requirements import Requirement
 
+try:
+    import tomllib
+except ModuleNotFoundError:  # Python < 3.11
+    import tomli as tomllib  # type: ignore[no-redef]
+
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
 ROOT = Path(__file__).parents[4]

vllm_omni/diffusion/models/lingbot_world/transformer.py

@@ -8,11 +8,12 @@ import math
 from collections.abc import Iterable
 from dataclasses import dataclass
 from types import SimpleNamespace
-from typing import TYPE_CHECKING, Any, Self, cast
+from typing import TYPE_CHECKING, Any, cast
 
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
+from typing_extensions import Self
 from vllm.distributed import (
     get_tensor_model_parallel_rank,
     get_tensor_model_parallel_world_size,

vllm_omni/engine/duplex/commands.py

@@ -47,7 +47,7 @@ class DuplexCommandError(RealtimeProtocolError):
     """
 
 
-@dataclass(frozen=True, slots=True, kw_only=True)
+@dataclass(frozen=True, kw_only=True)
 class DuplexCommand(RealtimeCommand):
     """A Realtime command as the duplex engine handles it.
 
@@ -55,6 +55,12 @@ class DuplexCommand(RealtimeCommand):
     command; every concrete class below pairs this with its protocol twin.
     """
 
+    # No instance fields of its own, so the slots are empty. Spelled out rather
+    # than ``slots=True`` because on Python 3.10 ``dataclass`` repeats the
+    # inherited fields in ``__slots__`` (fixed in 3.11); the concrete classes
+    # below would then combine two slotted layouts and fail to import.
+    __slots__ = ()
+
     #: Mailbox event type this command renders to (see ``payload()``).
     type: ClassVar[str] = ""
 

vllm_omni/model_executor/models/nemotron_voicechat/nemo_vendored/ear_tts_vae_codec.py

@@ -19,7 +19,9 @@ import functools
 import math
 from collections.abc import Callable
 from contextlib import contextmanager
-from typing import Any, Concatenate
+from typing import Any
+
+from typing_extensions import Concatenate
 
 import torch
 from omegaconf import DictConfig