Add vllm dependencies on nightly download.pytorch.org (#7177)

I want to be able to pip install vllm --pre --index-url
http://download.pytorch.org/whl/nightly/cu128 now that vLLM nightly
binaries are available. Also fix some lint issues from these scripts.

---------

Signed-off-by: Huy Do <[email protected]>

Author: huydhn

s3_management/manage.py

@@ -226,6 +226,99 @@
         "setuptools",
         "setuptools_scm",
         "wheel",
+        # vllm
+        "ninja",
+        "cuda_python",
+        "cuda_bindings",
+        "cuda_pathfinder",
+        "pynvml",
+        "nvidia_ml_py",
+        "einops",
+        "packaging",
+        "nvidia_cudnn_frontend",
+        "cachetools",
+        "blake3",
+        "py_cpuinfo",
+        "transformers",
+        "hf_xet",
+        "tokenizers",
+        "protobuf",
+        "fastapi",
+        "annotated_types",
+        "anyio",
+        "pydantic",
+        "pydantic_core",
+        "sniffio",
+        "starlette",
+        "typing_inspection",
+        "openai",
+        "distro",
+        "h11",
+        "httpcore",
+        "httpx",
+        "jiter",
+        "prometheus_client",
+        "prometheus_fastapi_instrumentator",
+        "lm_format_enforcer",
+        "interegular",
+        "llguidance",
+        "outlines_core",
+        "diskcache",
+        "lark",
+        "xgrammar",
+        "partial_json_parser",
+        "pyzmq",
+        "msgspec",
+        "gguf",
+        "mistral_common",
+        "rpds_py",
+        "pycountry",
+        "referencing",
+        "pydantic_extra_types",
+        "jsonschema_specifications",
+        "jsonschema",
+        "opencv_python_headless",
+        "compressed_tensors",
+        "frozendict",
+        "depyf",
+        "astor",
+        "cloudpickle",
+        "watchfiles",
+        "python_json_logger",
+        "scipy",
+        "pybase64",
+        "cbor2",
+        "setproctitle",
+        "openai_harmony",
+        "numba",
+        "llvmlite",
+        "ray",
+        "click",
+        "msgpack",
+        "fastapi_cli",
+        "fastapi_cloud_cli",
+        "httptools",
+        "markdown_it_py",
+        "pygments",
+        "python_dotenv",
+        "rich",
+        "rich_toolkit",
+        "shellingham",
+        "typer",
+        "uvicorn",
+        "uvloop",
+        "websockets",
+        "python_multipart",
+        "email_validator",
+        "dnspython",
+        "mdurl",
+        "rignore",
+        "sentry_sdk",
+        "cupy_cuda12x",
+        "fastrlock",
+        "soundfile",
+        "cffi",
+        "pycparser",
         "vllm",
         "flashinfer_python",
     ]

s3_management/update_dependencies.py

@@ -8,7 +8,7 @@
 CLIENT = boto3.client("s3")
 BUCKET = S3.Bucket("pytorch")
 
-PACKAGES_PER_PROJECT = {
+PACKAGES_PER_PROJECT: Dict[str, List[Dict[str, str]]] = {
     "sympy": [{"version": "latest", "project": "torch"}],
     "mpmath": [{"version": "latest", "project": "torch"}],
     "pillow": [{"version": "latest", "project": "torch"}],
@@ -18,7 +18,6 @@
     "filelock": [{"version": "latest", "project": "torch"}],
     "fsspec": [{"version": "latest", "project": "torch"}],
     "nvidia-cudnn-cu11": [{"version": "latest", "project": "torch"}],
-    "nvidia-cudnn-cu12": [{"version": "latest", "project": "torch"}],
     "typing-extensions": [{"version": "latest", "project": "torch"}],
     "nvidia-cuda-nvrtc-cu12": [
         {
@@ -378,6 +377,99 @@
     "onemkl-sycl-sparse": [{"version": "latest", "project": "torch_xpu"}],
     "onemkl-sycl-rng": [{"version": "latest", "project": "torch_xpu"}],
     "mkl": [{"version": "latest", "project": "torch_xpu"}],
+    # vLLM
+    "ninja": [{"version": "latest", "project": "vllm"}],
+    "cuda-python": [{"version": "12.9.0", "project": "vllm"}],
+    "cuda-bindings": [{"version": "12.9.2", "project": "vllm"}],
+    "cuda-pathfinder": [{"version": "latest", "project": "vllm"}],
+    "pynvml": [{"version": "latest", "project": "vllm"}],
+    "nvidia-ml-py": [{"version": "latest", "project": "vllm"}],
+    "einops": [{"version": "latest", "project": "vllm"}],
+    "packaging": [{"version": "latest", "project": "vllm"}],
+    "nvidia-cudnn-frontend": [{"version": "latest", "project": "vllm"}],
+    "cachetools": [{"version": "latest", "project": "vllm"}],
+    "blake3": [{"version": "latest", "project": "vllm"}],
+    "py-cpuinfo": [{"version": "latest", "project": "vllm"}],
+    "transformers": [{"version": "latest", "project": "vllm"}],
+    "hf-xet": [{"version": "latest", "project": "vllm"}],
+    "tokenizers": [{"version": "latest", "project": "vllm"}],
+    "protobuf": [{"version": "latest", "project": "vllm"}],
+    "fastapi": [{"version": "latest", "project": "vllm"}],
+    "annotated-types": [{"version": "latest", "project": "vllm"}],
+    "anyio": [{"version": "latest", "project": "vllm"}],
+    "pydantic": [{"version": "latest", "project": "vllm"}],
+    "pydantic-core": [{"version": "2.33.2", "project": "vllm"}],
+    "sniffio": [{"version": "latest", "project": "vllm"}],
+    "starlette": [{"version": "latest", "project": "vllm"}],
+    "typing-inspection": [{"version": "latest", "project": "vllm"}],
+    "openai": [{"version": "latest", "project": "vllm"}],
+    "distro": [{"version": "latest", "project": "vllm"}],
+    "h11": [{"version": "latest", "project": "vllm"}],
+    "httpcore": [{"version": "latest", "project": "vllm"}],
+    "httpx": [{"version": "latest", "project": "vllm"}],
+    "jiter": [{"version": "latest", "project": "vllm"}],
+    "prometheus-client": [{"version": "latest", "project": "vllm"}],
+    "prometheus-fastapi-instrumentator": [{"version": "latest", "project": "vllm"}],
+    "lm-format-enforcer": [{"version": "latest", "project": "vllm"}],
+    "interegular": [{"version": "latest", "project": "vllm"}],
+    "llguidance": [{"version": "0.7.11", "project": "vllm"}],
+    "outlines-core": [{"version": "0.2.10", "project": "vllm"}],
+    "diskcache": [{"version": "latest", "project": "vllm"}],
+    "lark": [{"version": "latest", "project": "vllm"}],
+    "xgrammar": [{"version": "0.1.23", "project": "vllm"}],
+    "partial-json-parser": [{"version": "latest", "project": "vllm"}],
+    "pyzmq": [{"version": "latest", "project": "vllm"}],
+    "msgspec": [{"version": "latest", "project": "vllm"}],
+    "gguf": [{"version": "latest", "project": "vllm"}],
+    "mistral-common": [{"version": "latest", "project": "vllm"}],
+    "rpds-py": [{"version": "latest", "project": "vllm"}],
+    "pycountry": [{"version": "latest", "project": "vllm"}],
+    "referencing": [{"version": "latest", "project": "vllm"}],
+    "pydantic-extra-types": [{"version": "latest", "project": "vllm"}],
+    "jsonschema-specifications": [{"version": "latest", "project": "vllm"}],
+    "jsonschema": [{"version": "latest", "project": "vllm"}],
+    "opencv-python-headless": [{"version": "latest", "project": "vllm"}],
+    "compressed-tensors": [{"version": "latest", "project": "vllm"}],
+    "frozendict": [{"version": "latest", "project": "vllm"}],
+    "depyf": [{"version": "latest", "project": "vllm"}],
+    "astor": [{"version": "latest", "project": "vllm"}],
+    "cloudpickle": [{"version": "latest", "project": "vllm"}],
+    "watchfiles": [{"version": "latest", "project": "vllm"}],
+    "python-json-logger": [{"version": "latest", "project": "vllm"}],
+    "scipy": [{"version": "latest", "project": "vllm"}],
+    "pybase64": [{"version": "latest", "project": "vllm"}],
+    "cbor2": [{"version": "latest", "project": "vllm"}],
+    "setproctitle": [{"version": "latest", "project": "vllm"}],
+    "openai-harmony": [{"version": "latest", "project": "vllm"}],
+    "numba": [{"version": "0.61.2", "project": "vllm"}],
+    "llvmlite": [{"version": "latest", "project": "vllm"}],
+    "ray": [{"version": "latest", "project": "vllm"}],
+    "click": [{"version": "latest", "project": "vllm"}],
+    "msgpack": [{"version": "latest", "project": "vllm"}],
+    "fastapi-cli": [{"version": "latest", "project": "vllm"}],
+    "httptools": [{"version": "latest", "project": "vllm"}],
+    "markdown-it-py": [{"version": "latest", "project": "vllm"}],
+    "pygments": [{"version": "latest", "project": "vllm"}],
+    "python-dotenv": [{"version": "latest", "project": "vllm"}],
+    "rich": [{"version": "latest", "project": "vllm"}],
+    "rich-toolkit": [{"version": "latest", "project": "vllm"}],
+    "shellingham": [{"version": "latest", "project": "vllm"}],
+    "typer": [{"version": "latest", "project": "vllm"}],
+    "uvicorn": [{"version": "latest", "project": "vllm"}],
+    "uvloop": [{"version": "latest", "project": "vllm"}],
+    "websockets": [{"version": "latest", "project": "vllm"}],
+    "python-multipart": [{"version": "latest", "project": "vllm"}],
+    "email-validator": [{"version": "latest", "project": "vllm"}],
+    "dnspython": [{"version": "2.7.0", "project": "vllm"}],
+    "fastapi-cloud-cli": [{"version": "latest", "project": "vllm"}],
+    "mdurl": [{"version": "latest", "project": "vllm"}],
+    "rignore": [{"version": "latest", "project": "vllm"}],
+    "sentry-sdk": [{"version": "latest", "project": "vllm"}],
+    "cupy-cuda12x": [{"version": "latest", "project": "vllm"}],
+    "fastrlock": [{"version": "latest", "project": "vllm"}],
+    "soundfile": [{"version": "latest", "project": "vllm"}],
+    "cffi": [{"version": "latest", "project": "vllm"}],
+    "pycparser": [{"version": "latest", "project": "vllm"}],
 }