[Hardware][Intel] Add CPU inference backend (#3634)

Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Co-authored-by: Yuan Zhou <yuan.zhou@intel.com>

setup.py

@@ -15,6 +15,8 @@ from torch.utils.cpp_extension import CUDA_HOME
 
 ROOT_DIR = os.path.dirname(__file__)
 logger = logging.getLogger(__name__)
+# Target device of vLLM, supporting [cuda (by default), rocm, neuron, cpu]
+VLLM_TARGET_DEVICE = os.getenv("VLLM_TARGET_DEVICE", "cuda")
 
 # vLLM only supports Linux platform
 assert sys.platform.startswith(
@@ -112,6 +114,7 @@ class cmake_build_ext(build_ext):
             '-DCMAKE_BUILD_TYPE={}'.format(cfg),
             '-DCMAKE_LIBRARY_OUTPUT_DIRECTORY={}'.format(outdir),
             '-DCMAKE_ARCHIVE_OUTPUT_DIRECTORY={}'.format(self.build_temp),
+            '-DVLLM_TARGET_DEVICE={}'.format(VLLM_TARGET_DEVICE),
         ]
 
         verbose = bool(int(os.getenv('VERBOSE', '0')))
@@ -185,11 +188,14 @@ class cmake_build_ext(build_ext):
 
 
 def _is_cuda() -> bool:
-    return torch.version.cuda is not None and not _is_neuron()
+    return VLLM_TARGET_DEVICE == "cuda" \
+            and torch.version.cuda is not None \
+            and not _is_neuron()
 
 
 def _is_hip() -> bool:
-    return torch.version.hip is not None
+    return (VLLM_TARGET_DEVICE == "cuda"
+            or VLLM_TARGET_DEVICE == "rocm") and torch.version.hip is not None
 
 
 def _is_neuron() -> bool:
@@ -201,6 +207,10 @@ def _is_neuron() -> bool:
     return torch_neuronx_installed
 
 
+def _is_cpu() -> bool:
+    return VLLM_TARGET_DEVICE == "cpu"
+
+
 def _install_punica() -> bool:
     return bool(int(os.getenv("VLLM_INSTALL_PUNICA_KERNELS", "0")))
 
@@ -296,6 +306,8 @@ def get_vllm_version() -> str:
         if neuron_version != MAIN_CUDA_VERSION:
             neuron_version_str = neuron_version.replace(".", "")[:3]
             version += f"+neuron{neuron_version_str}"
+    elif _is_cpu():
+        version += "+cpu"
     else:
         raise RuntimeError("Unknown runtime environment")
 
@@ -322,6 +334,9 @@ def get_requirements() -> List[str]:
     elif _is_neuron():
         with open(get_path("requirements-neuron.txt")) as f:
             requirements = f.read().strip().split("\n")
+    elif _is_cpu():
+        with open(get_path("requirements-cpu.txt")) as f:
+            requirements = f.read().strip().split("\n")
     else:
         raise ValueError(
             "Unsupported platform, please use CUDA, ROCM or Neuron.")