[Platform] platform agnostic for EngineArgs initialization (#11225)

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

vllm/engine/arg_utils.py

@@ -112,9 +112,7 @@ class EngineArgs:
     pipeline_parallel_size: int = 1
     tensor_parallel_size: int = 1
     max_parallel_loading_workers: Optional[int] = None
-    # NOTE(kzawora): default block size for Gaudi should be 128
+    block_size: Optional[int] = None
-    # smaller sizes still work, but very inefficiently
-    block_size: int = 16 if not current_platform.is_hpu() else 128
     enable_prefix_caching: Optional[bool] = None
     disable_sliding_window: bool = False
     use_v2_block_manager: bool = True
@@ -1036,9 +1034,7 @@ class EngineArgs:
             self.enable_prefix_caching = False
 
         cache_config = CacheConfig(
-            # neuron needs block_size = max_model_len
+            block_size=self.block_size,
-            block_size=self.block_size if self.device != "neuron" else
-            (self.max_model_len if self.max_model_len is not None else 0),
             gpu_memory_utilization=self.gpu_memory_utilization,
             swap_space=self.swap_space,
             cache_dtype=self.kv_cache_dtype,

vllm/platforms/cpu.py

@@ -60,6 +60,9 @@ class CpuPlatform(Platform):
 
         cache_config = vllm_config.cache_config
 
+        if cache_config and cache_config.block_size is None:
+            cache_config.block_size = 16
+
         kv_cache_space = envs.VLLM_CPU_KVCACHE_SPACE
 
         if kv_cache_space >= 0:

vllm/platforms/cuda.py

@@ -137,6 +137,10 @@ class CudaPlatformBase(Platform):
                 else:
                     parallel_config.worker_cls = "vllm.worker.worker.Worker"
 
+        cache_config = vllm_config.cache_config
+        if cache_config and cache_config.block_size is None:
+            cache_config.block_size = 16
+
 
 # NVML utils
 # Note that NVML is not affected by `CUDA_VISIBLE_DEVICES`,

vllm/platforms/hpu.py

@@ -48,6 +48,12 @@ class HpuPlatform(Platform):
         if parallel_config.worker_cls == "auto":
             parallel_config.worker_cls = "vllm.worker.hpu_worker.HPUWorker"
 
+        # NOTE(kzawora): default block size for Gaudi should be 128
+        # smaller sizes still work, but very inefficiently
+        cache_config = vllm_config.cache_config
+        if cache_config and cache_config.block_size is None:
+            cache_config.block_size = 128
+
     @classmethod
     def is_pin_memory_available(cls):
         logger.warning("Pin memory is not supported on HPU.")

vllm/platforms/neuron.py

@@ -33,6 +33,12 @@ class NeuronPlatform(Platform):
             parallel_config.worker_cls = \
                 "vllm.worker.neuron_worker.NeuronWorker"
 
+        cache_config = vllm_config.cache_config
+        if cache_config:
+            # neuron needs block_size = max_model_len
+            vllm_config.cache_config.block_size = \
+                vllm_config.model_config.max_model_len
+
     @classmethod
     def is_pin_memory_available(cls) -> bool:
         logger.warning("Pin memory is not supported on Neuron.")

vllm/platforms/openvino.py

@@ -87,6 +87,9 @@ class OpenVinoPlatform(Platform):
         # check and update cache config
         ov_core = ov.Core()
         cache_config = vllm_config.cache_config
+        if cache_config and cache_config.block_size is None:
+            cache_config.block_size = 16
+
         if envs.VLLM_OPENVINO_CPU_KV_CACHE_PRECISION == "u8":
             if not OpenVinoPlatform.is_openvino_cpu():
                 logger.info("VLLM_OPENVINO_CPU_KV_CACHE_PRECISION is"

vllm/platforms/rocm.py

@@ -84,6 +84,10 @@ class RocmPlatform(Platform):
 
     @classmethod
     def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
+        cache_config = vllm_config.cache_config
+        if cache_config and cache_config.block_size is None:
+            cache_config.block_size = 16
+
         parallel_config = vllm_config.parallel_config
         scheduler_config = vllm_config.scheduler_config
         if parallel_config.worker_cls == "auto":

vllm/platforms/tpu.py

@@ -46,6 +46,11 @@ class TpuPlatform(Platform):
     @classmethod
     def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         from vllm.config import CompilationLevel
+
+        cache_config = vllm_config.cache_config
+        if cache_config and cache_config.block_size is None:
+            cache_config.block_size = 16
+
         compilation_config = vllm_config.compilation_config
         if compilation_config.level == CompilationLevel.NO_COMPILATION:
             # TPU does not support NO_COMPILATION

vllm/platforms/xpu.py

@@ -51,6 +51,10 @@ class XPUPlatform(Platform):
 
     @classmethod
     def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
+        cache_config = vllm_config.cache_config
+        if cache_config and cache_config.block_size is None:
+            cache_config.block_size = 16
+
         # check and update model config
         model_config = vllm_config.model_config
         if model_config.dtype == torch.bfloat16: