[model] make llama4 compatible with pure dense layers (vllm-project#17315)

Signed-off-by: Lucia Fang <[email protected]>
Signed-off-by: Yuqi Zhang <[email protected]>

Author: luccafong

vllm/model_executor/models/llama4.py

@@ -273,8 +273,8 @@ def __init__(
             cache_config=cache_config,
             prefix=f"{prefix}.self_attn",
         )
-        is_moe_layer = (self.layer_idx +
-                        1) % config.interleave_moe_layer_step == 0
+        is_moe_layer = config.interleave_moe_layer_step > 0 and (
+            self.layer_idx + 1) % config.interleave_moe_layer_step == 0
         if is_moe_layer:
             self.feed_forward = Llama4MoE(
                 config=config,