fix compatibility

Author: NorthmanPKU

vllm/compilation/mirage_backend.py

@@ -331,7 +331,6 @@ def compile_or_call(*args):
 
                 self.compiled = True
 
-            logger.info(f"[Mirage] Calling the compiled result...")
             result_hidden_states = self.mpk()
 
             return (result_hidden_states,)

vllm/platforms/cuda.py

@@ -346,7 +346,7 @@ def get_attn_backend_cls(
         FLASH_ATTN_V1 = "vllm.v1.attention.backends.flash_attn.FlashAttentionBackend"  # noqa: E501
         TREE_ATTN_V1 = "vllm.v1.attention.backends.tree_attn.TreeAttentionBackend"  # noqa: E501
         XFORMERS_V1 = "vllm.v1.attention.backends.xformers.XFormersAttentionBackend"  # noqa: E501
-        MIRAGE_V1 = "vllm.v1.attention.backends.mirage.MirageBackend"  # noqa: E501
+        MIRAGE_V1 = "vllm.v1.attention.backends.mirage.MirageAttentionBackend"  # noqa: E501
 
         use_fp8_kv_cache = kv_cache_dtype is not None and kv_cache_dtype.startswith(
             "fp8"

vllm/v1/attention/backends/mirage.py

@@ -22,7 +22,8 @@
 
 from vllm.platforms import current_platform
 from vllm.triton_utils import tl, triton
-from vllm.utils import cdiv, is_pin_memory_available
+from vllm.utils.math_utils import cdiv
+from vllm.utils.platform_utils import is_pin_memory_available
 from vllm.v1.attention.backends.utils import (
     AttentionCGSupport,
     AttentionMetadataBuilder,