Adapt OpenVINO CPU plugin implementation

vllm/executor/openvino_executor.py

@@ -18,6 +18,7 @@
 from vllm.sampling_params import SamplingParams
 
 import openvino as ov
+import openvino.properties.hint as hints
 
 logger = init_logger(__name__)
 
@@ -59,6 +60,11 @@ def __init__(
         self.num_layers = model_config.get_num_layers(parallel_config)
         self.num_heads = model_config.get_num_kv_heads(parallel_config)
 
+        if device_config.device.type == "cpu":
+            if cache_config.block_size != 1:
+                cache_config.num_cpu_blocks *= cache_config.block_size
+                cache_config.block_size = 1
+                print(f"Warning: CPU only support block_size = 1, it's forced to 1, num_cpu_blocks is set to {cache_config.num_cpu_blocks}.")
         self.block_size = cache_config.block_size
         self.num_gpu_blocks = cache_config.num_gpu_blocks
         self.num_cpu_blocks = cache_config.num_cpu_blocks
@@ -146,7 +152,15 @@ def get_cache_dtype(
         # probably, we need to force OpenVINO kv cache data types per device and assert
         # if user specified a different value
         if cache_dtype == "auto":
-            cache_dtype = model_config.dtype
+            if device_config.device.type == "cpu":
+                core = ov.Core()
+                inference_precision = core.get_property("CPU", hints.inference_precision)
+                if inference_precision == ov.Type.bf16:
+                    cache_dtype = torch.bfloat16
+                else:
+                    cache_dtype = torch.float16
+            else:
+                cache_dtype = model_config.dtype
         else:
             cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_dtype]
         ov_cache_dtype = TORCH_DTYPE_TO_OPENVINO_DTYPE[cache_dtype]
@@ -469,7 +483,9 @@ def _init_worker(self):
             self.scheduler_config,
             self.device_config,
             lora_config=self.lora_config,
-            kv_cache_dtype=self.cache_config.cache_dtype,
+            kv_cache_dtype=OpenVINOCacheEngine.get_cache_dtype(self.cache_config.cache_dtype,
+                                                               self.model_config,
+                                                               self.device_config)
         )
         self.driver_worker.init_model()
         self.driver_worker.load_model()

vllm/model_executor/openvino_model_loader.py

@@ -12,6 +12,7 @@
 from vllm.utils import is_openvino_optimum_intel
 
 import openvino as ov
+from openvino import Type
 
 
 def _flattenize_inputs(inputs):
@@ -56,7 +57,8 @@ def ov_wrapper(self, *args, **kwargs) -> torch.Tensor:
 
 def patch_stateful_model(
     model: ov.Model,
-    factory):
+    factory,
+    kv_cache_dtype: Type):
     print('TRANSFORMING OPTIMUM-INTEL MODEL TO vLLM COMPATIBLE FORM')
     from openvino.runtime.passes import Manager, MatcherPass, WrapType, Matcher, AnyInput, Or
     from openvino.runtime import opset13
@@ -128,8 +130,8 @@ def callback(m: Matcher) -> bool:
                 real_v = mapping[v_current]
                 hidden_shape = real_q.get_partial_shape()
                 hidden_dim = hidden_shape[hidden_shape.rank.get_length() - 1].get_length()  # TODO: What if it is a dynamic? Need to insert a ShapeOf sub-graph instead
-                k_parameter = opset13.parameter(shape=[-1, -1, -1, -1, -1], dtype=np.float32)
-                v_parameter = opset13.parameter(shape=[-1, -1, -1, -1], dtype=np.float32)
+                k_parameter = opset13.parameter(shape=[-1, -1, -1, -1, -1], dtype=kv_cache_dtype)
+                v_parameter = opset13.parameter(shape=[-1, -1, -1, -1], dtype=kv_cache_dtype)
                 kv_parameters.append(k_parameter)
                 kv_parameters.append(v_parameter)
                 # TODO: The rank 4 is used in the following code, but it is not guaranteed for all models, adopt to other ranks.
@@ -274,7 +276,8 @@ def has_parameter(model, name):
 
 def _patch_model_with_openvino(
         pt_model: torch.nn.Module,
-        model_config: ModelConfig):
+        model_config: ModelConfig,
+        kv_cache_dtype: Type):
     print(' ============= PATCHING MODEL =============')
     from vllm.model_executor.layers.attention.attention import Attention
     from openvino.frontend.pytorch import ModuleExtension
@@ -294,7 +297,15 @@ def _patch_model_with_openvino(
 
     # Prepare example inputs
 
-    kv_cache_dtype = torch.float32
+    torch_dtype_maping = {
+        Type.boolean: torch.bool,
+        Type.f32: torch.float32,
+        Type.f16: torch.float16,
+        Type.bf16: torch.bfloat16,
+        Type.i32: torch.int32,
+        Type.i64: torch.int64
+    }
+    kv_cache_dtype = torch_dtype_maping[kv_cache_dtype]
     num_heads = pt_model.config.num_attention_heads
     num_kv_heads = num_heads
     head_size = pt_model.config.hidden_size // num_kv_heads
@@ -423,6 +434,7 @@ def ov_sample(
 
 def get_model(model_config: ModelConfig,
               device_config: DeviceConfig,
+              kv_cache_dtype: Type,
               **kwargs) -> torch.nn.Module:
     lora_config = kwargs.get("lora_config", None)
     if lora_config:
@@ -443,7 +455,7 @@ def get_model(model_config: ModelConfig,
             # Keep factory to destroy it in a particular moment when all other objects referencing custom nodes are destoyed
             pt_model.ov_node_factory = NodeFactory()
             pt_model.ov_node_factory.add_extension('libuser_ov_extensions.so')
-        patch_stateful_model(pt_model.model, pt_model.ov_node_factory)
+        patch_stateful_model(pt_model.model, pt_model.ov_node_factory, kv_cache_dtype)
         core = ov.Core()
         ov_compiled = core.compile_model(pt_model.model, "CPU")
         pt_model._ov_request = ov_compiled.create_infer_request()
@@ -457,6 +469,6 @@ def get_model(model_config: ModelConfig,
     else:
         from vllm.model_executor.model_loader import get_model
         pt_model = get_model(model_config, device_config, **kwargs)
-        _patch_model_with_openvino(pt_model, model_config)
+        _patch_model_with_openvino(pt_model, model_config, kv_cache_dtype)
 
     return pt_model

vllm/worker/model_runner.py

@@ -88,6 +88,7 @@ def load_model(self) -> None:
         with measure_cuda_memory(self.device) as m:
             self.model = get_model(self.model_config,
                                    self.device_config,
+                                   kv_cache_dtype=self.kv_cache_dtype,
                                    lora_config=self.lora_config,
                                    parallel_config=self.parallel_config,
                                    scheduler_config=self.scheduler_config)