[Build] Cython compilation support fix (#14296)

Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>

Dockerfile.rocm

@@ -40,7 +40,7 @@ ARG USE_CYTHON
 RUN cd vllm \
     && python3 -m pip install -r requirements/rocm.txt \
     && python3 setup.py clean --all  \
-    && if [ ${USE_CYTHON} -eq "1" ]; then python3 setup_cython.py build_ext --inplace; fi \
+    && if [ ${USE_CYTHON} -eq "1" ]; then python3 tests/build_cython.py build_ext --inplace; fi \
     && python3 setup.py bdist_wheel --dist-dir=dist
 FROM scratch AS export_vllm
 ARG COMMON_WORKDIR

pyproject.toml

@@ -86,6 +86,7 @@ exclude = [
 "vllm/triton_utils/**/*.py" = ["UP006", "UP035"]
 "vllm/vllm_flash_attn/**/*.py" = ["UP006", "UP035"]
 "vllm/worker/**/*.py" = ["UP006", "UP035"]
+"vllm/utils.py" = ["UP006", "UP035"]
 
 [tool.ruff.lint]
 select = [

tests/build_cython.py

@@ -0,0 +1,38 @@
+# SPDX-License-Identifier: Apache-2.0
+import Cython.Compiler.Options
+from Cython.Build import cythonize
+from setuptools import setup
+
+Cython.Compiler.Options.annotate = True
+
+infiles = []
+
+infiles += [
+    "vllm/engine/llm_engine.py",
+    "vllm/transformers_utils/detokenizer.py",
+    "vllm/engine/output_processor/single_step.py",
+    "vllm/outputs.py",
+    "vllm/engine/output_processor/stop_checker.py",
+]
+
+infiles += [
+    "vllm/core/scheduler.py",
+    "vllm/sequence.py",
+    "vllm/core/block_manager.py",
+]
+
+infiles += [
+    "vllm/model_executor/layers/sampler.py",
+    "vllm/sampling_params.py",
+    "vllm/utils.py",
+]
+
+setup(ext_modules=cythonize(infiles,
+                            annotate=False,
+                            force=True,
+                            compiler_directives={
+                                'language_level': "3",
+                                'infer_types': True
+                            }))
+
+# example usage: python3 build_cython.py build_ext --inplace

vllm/engine/llm_engine.py

@@ -1249,7 +1249,7 @@ class LLMEngine:
         return None
 
     def _advance_to_next_step(
-            self, output: List[SamplerOutput],
+            self, output: SamplerOutput,
             seq_group_metadata_list: List[SequenceGroupMetadata],
             scheduled_seq_groups: List[ScheduledSequenceGroup]) -> None:
         """Given model output from a single run, append the tokens to the

vllm/model_executor/layers/sampler.py

@@ -1187,7 +1187,8 @@ def _build_sampler_output(
         deferred_sample_results_args=deferred_sample_results_args)
 
 
-def _get_next_prompt_tokens(seq_group: SequenceGroupToSample) -> List[int]:
+def _get_next_prompt_tokens(
+        seq_group: SequenceGroupToSample) -> tuple[int, ...]:
     """Get a list of next prompt tokens to compute logprob from a
         given sequence group.
 

vllm/utils.py

@@ -37,7 +37,7 @@ from collections.abc import (AsyncGenerator, Awaitable, Generator, Hashable,
 from dataclasses import dataclass, field
 from functools import cache, lru_cache, partial, wraps
 from typing import (TYPE_CHECKING, Any, Callable, Generic, Literal, NamedTuple,
-                    Optional, TypeVar, Union)
+                    Optional, Type, TypeVar, Union)
 from uuid import uuid4
 
 import cloudpickle
@@ -1544,9 +1544,9 @@ class LazyDict(Mapping[str, T], Generic[T]):
         return len(self._factory)
 
 
-class ClassRegistry(UserDict[type[T], _V]):
+class ClassRegistry(UserDict[Type[T], _V]):
 
-    def __getitem__(self, key: type[T]) -> _V:
+    def __getitem__(self, key: Type[T]) -> _V:
         for cls in key.mro():
             if cls in self.data:
                 return self.data[cls]