[Bugfix][Minor] Fix potential NameError in mamba backend selector and misc typos (#35886)

Signed-off-by: Li <chuali@amd.com>

vllm/model_executor/models/kimi_k25.py

@@ -369,7 +369,7 @@ class KimiK25ForConditionalGeneration(
         target_dtype = next(self.vision_tower.parameters()).dtype
         pixel_values = pixel_values.to(target_dtype)
         assert isinstance(grid_thws, torch.Tensor), (
-            f"expect grid_thws to be a tensor, get {type(grid_thws)}"
+            f"expect grid_thws to be a tensor, got {type(grid_thws)}"
         )
         # In some cases (e.g. with merger), grid_thws has an extra middle dimension
         grid_thws = grid_thws.reshape(-1, grid_thws.shape[-1])

vllm/v1/attention/backends/flex_attention.py

@@ -749,7 +749,10 @@ class FlexAttentionMetadataBuilder(AttentionMetadataBuilder[FlexAttentionMetadat
         prefix_kv_lens = None
         suffix_kv_lens = None
         if use_cascade:
-            raise NotImplementedError("Not yet my friend")
+            raise NotImplementedError(
+                "Cascade prefix attention is not yet implemented "
+                "for FlexAttention backend"
+            )
 
         block_size = self.kv_cache_spec.block_size
         max_possible_seq_len = self.model_config.max_model_len

vllm/v1/attention/backends/utils.py

@@ -253,7 +253,7 @@ def make_local_attention_virtual_batches(
     #   seqlens_q_local = [2, 2, 1, 4, 4, 1, 4, 1]
     #
     # First Get batched arange. (E.g., [2, 4, 2] -> [0, 1, 0, 1, 2, 3, 0, 1])
-    #   (TODO: max a utility to share this code with _prepare_inputs)
+    #   (TODO: make a utility to share this code with _prepare_inputs)
     # arange step 1. [2, 4, 2] -> [2, 6, 8]
     cu_num_blocks = np.cumsum(local_blocks)
     virtual_batches = cu_num_blocks[-1]

vllm/v1/attention/selector.py

@@ -149,8 +149,8 @@ def _cached_get_mamba_attn_backend(
         selected_backend = MambaAttentionBackendEnum[backend_name]
     except KeyError as e:
         raise ValueError(
-            f"Invalid mamba attention backend type: '{backend_name}'. Valid "
-            f"backends are: {list(MambaAttentionBackendEnum.__members__.keys())}"
+            f"Invalid mamba attention backend type: '{mamba_type}'. Valid "
+            f"types are: {list(MAMBA_TYPE_TO_BACKEND_MAP.keys())}"
         ) from e
 
     mamba_attn_backend = selected_backend.get_class()