diff --git a/vllm/engine/arg_utils.py b/vllm/engine/arg_utils.py
index 06b8542779dc0..f6d276fe7c0c8 100644
--- a/vllm/engine/arg_utils.py
+++ b/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
-    # smaller sizes still work, but very inefficiently
-    block_size: int = 16 if not current_platform.is_hpu() else 128
+    block_size: Optional[int] = None
     enable_prefix_caching: Optional[bool] = None
     disable_sliding_window: bool = False
     use_v2_block_manager: bool = True
@@ -1036,9 +1034,7 @@ def create_engine_config(self,
             self.enable_prefix_caching = False
 
         cache_config = CacheConfig(
-            # neuron needs block_size = max_model_len
-            block_size=self.block_size if self.device != "neuron" else
-            (self.max_model_len if self.max_model_len is not None else 0),
+            block_size=self.block_size,
             gpu_memory_utilization=self.gpu_memory_utilization,
             swap_space=self.swap_space,
             cache_dtype=self.kv_cache_dtype,
diff --git a/vllm/platforms/cpu.py b/vllm/platforms/cpu.py
index aad8755d9fcd8..d95a2b4cd5565 100644
--- a/vllm/platforms/cpu.py
+++ b/vllm/platforms/cpu.py
@@ -60,6 +60,9 @@ 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
+
         kv_cache_space = envs.VLLM_CPU_KVCACHE_SPACE
 
         if kv_cache_space >= 0:
diff --git a/vllm/platforms/cuda.py b/vllm/platforms/cuda.py
index ae1fd6d5ce068..3c5350b778345 100644
--- a/vllm/platforms/cuda.py
+++ b/vllm/platforms/cuda.py
@@ -137,6 +137,10 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
                 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`,
diff --git a/vllm/platforms/hpu.py b/vllm/platforms/hpu.py
index 2b947d280f9f8..0a44f2b74163a 100644
--- a/vllm/platforms/hpu.py
+++ b/vllm/platforms/hpu.py
@@ -48,6 +48,12 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         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.")
diff --git a/vllm/platforms/neuron.py b/vllm/platforms/neuron.py
index 86113523385f6..a4bbbd27c8a89 100644
--- a/vllm/platforms/neuron.py
+++ b/vllm/platforms/neuron.py
@@ -33,6 +33,12 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
             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.")
diff --git a/vllm/platforms/openvino.py b/vllm/platforms/openvino.py
index ccd94e8adb3b1..16eb8dc81efc2 100644
--- a/vllm/platforms/openvino.py
+++ b/vllm/platforms/openvino.py
@@ -87,6 +87,9 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
         # 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"
diff --git a/vllm/platforms/rocm.py b/vllm/platforms/rocm.py
index 0133f26a0b1bc..7778b565372cb 100644
--- a/vllm/platforms/rocm.py
+++ b/vllm/platforms/rocm.py
@@ -84,6 +84,10 @@ def is_async_output_supported(cls, enforce_eager: Optional[bool]) -> bool:
 
     @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":
diff --git a/vllm/platforms/tpu.py b/vllm/platforms/tpu.py
index 10d874349f36b..77f5c8401424b 100644
--- a/vllm/platforms/tpu.py
+++ b/vllm/platforms/tpu.py
@@ -46,6 +46,11 @@ def inference_mode(cls):
     @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
diff --git a/vllm/platforms/xpu.py b/vllm/platforms/xpu.py
index c20190e789d7e..78e17c2afec65 100644
--- a/vllm/platforms/xpu.py
+++ b/vllm/platforms/xpu.py
@@ -51,6 +51,10 @@ def inference_mode():
 
     @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: