[V1][BugFix] Detect interleaved sliding window attention (vllm-project#14896)

Signed-off-by: Woosuk Kwon <[email protected]>

Author: WoosukKwon

vllm/v1/worker/gpu_model_runner.py

@@ -82,8 +82,15 @@ def __init__(
             self.kv_cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
                 cache_config.cache_dtype]
 
-        self.is_multimodal_model = model_config.is_multimodal_model
+        # NOTE(woosuk): sliding_window is None for models with interleaved
+        # attention. Use interleaved_sliding_window instead.
         self.sliding_window = model_config.get_sliding_window()
+        self.interleaved_sliding_window = getattr(
+            model_config.hf_text_config, "interleaved_sliding_window", None)
+        self.window_size = (self.sliding_window
+                            or self.interleaved_sliding_window)
+
+        self.is_multimodal_model = model_config.is_multimodal_model
         self.block_size = cache_config.block_size
         self.max_model_len = model_config.max_model_len
         self.max_num_blocks_per_req = cdiv(self.max_model_len, self.block_size)
@@ -674,7 +681,7 @@ def _compute_cascade_attn_prefix_len(
             num_query_heads=self.num_query_heads,
             num_kv_heads=self.num_kv_heads,
             use_alibi=False,  # FIXME
-            use_sliding_window=self.sliding_window is not None,
+            use_sliding_window=self.window_size is not None,
             num_sms=self.num_sms,
         )
         return common_prefix_len if use_cascade else 0