chore: import upstream snapshot with attribution

python/requirements/llm/llm-requirements.txt

@@ -0,0 +1,18 @@
+# Keep this in sync with the definition in setup.py for ray[llm]
+vllm[audio]==0.24.0
+prometheus-fastapi-instrumentator>=8.0.0  # for starlette v1.0.1+
+# Keep NIXL in sync with vLLM's kv_connectors requirements.
+# https://github.com/vllm-project/vllm/blob/v0.24.0/requirements/kv_connectors.txt#L2
+nixl==1.2.0
+nixl-cu13==1.2.0
+anyio>=4.5.0
+# For json mode
+jsonref>=1.1.0
+jsonschema
+ninja
+# async-timeout is a backport of asyncio.timeout for python < 3.11
+async-timeout; python_version < '3.11'
+typer
+meson
+pybind11
+hf_transfer

python/requirements/llm/llm-test-requirements.txt

@@ -0,0 +1,8 @@
+pytest
+aiohttp
+pillow
+httpx>=0.27.2
+pynvml>=12.0.0
+jupytext>1.13.6
+backoff
+datasets

python/requirements/llm/patches/vllm-device-aware-compile-cache.patch

@@ -0,0 +1,30 @@
+diff --git a/vllm/compilation/backends.py b/vllm/compilation/backends.py
+index dee7cdde744d..472d1f1a7588 100644
+--- a/vllm/compilation/backends.py
++++ b/vllm/compilation/backends.py
+@@ -1031,7 +1031,9 @@ def __call__(self, graph: fx.GraphModule, example_inputs: Sequence[Any]) -> Any:
+         self.compilation_config.cache_dir = cache_dir
+         rank = vllm_config.parallel_config.rank
+         dp_rank = vllm_config.parallel_config.data_parallel_index
+-        local_cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}", self.prefix)
++        dev = torch.accelerator.current_device_index()
++        local_cache_dir = os.path.join(
++            cache_dir, f"rank_{rank}_{dp_rank}_dev{dev}", self.prefix)
+         os.makedirs(local_cache_dir, exist_ok=True)
+         self.compilation_config.local_cache_dir = local_cache_dir
+ 
+diff --git a/vllm/compilation/decorators.py b/vllm/compilation/decorators.py
+index 9c55a42a4924..753b429a041d 100644
+--- a/vllm/compilation/decorators.py
++++ b/vllm/compilation/decorators.py
+@@ -509,7 +509,9 @@ def __call__(self: type[_T], *args: Any, **kwargs: Any) -> Any:
+ 
+             rank = self.vllm_config.parallel_config.rank
+             dp_rank = self.vllm_config.parallel_config.data_parallel_index
+-            cache_dir = os.path.join(cache_dir, f"rank_{rank}_{dp_rank}")
++            dev = torch.accelerator.current_device_index()
++            cache_dir = os.path.join(cache_dir,
++                                     f"rank_{rank}_{dp_rank}_dev{dev}")
+             aot_compilation_path = os.path.join(cache_dir, "model")
+             if not envs.VLLM_DISABLE_COMPILE_CACHE:
+                 loaded_fn = _try_load_aot_compiled_fn(self, aot_compilation_path)