Add CUDA block size check

Signed-off-by: Konrad Zawora <[email protected]>

Author: kzawora-intel

vllm/config.py

@@ -917,6 +917,11 @@ def _verify_args(self) -> None:
             raise ValueError(
                 "GPU memory utilization must be less than 1.0. Got "
                 f"{self.gpu_memory_utilization}.")
+        if (current_platform.is_cuda_alike()
+                and self.block_size not in [8, 16, 32]):
+            raise ValueError(
+                "CUDA Paged Attention kernel only supports "
+                f"block sizes [8, 16, 32]. Got {self.block_size}.")
 
     def _verify_cache_dtype(self) -> None:
         if self.cache_dtype == "auto":

vllm/engine/arg_utils.py

@@ -426,7 +426,9 @@ def add_cli_args(parser: FlexibleArgumentParser) -> FlexibleArgumentParser:
                             choices=[8, 16, 32, 64, 128],
                             help='Token block size for contiguous chunks of '
                             'tokens. This is ignored on neuron devices and '
-                            'set to max-model-len')
+                            'set to max-model-len. On CUDA devices, '
+                            'only block sizes up to 32 are supported. '
+                            'On HPU devices, block size defaults to 128.')
 
         parser.add_argument(
             "--enable-prefix-caching",