Add warning for Attention backends that do not support irope yet (#16212)

Author: sarckk

vllm/attention/backends/flashinfer.py

@@ -38,9 +38,12 @@
 from vllm.attention.layer import Attention
 from vllm.attention.ops.paged_attn import PagedAttention
 from vllm.config import VllmConfig, get_current_vllm_config
+from vllm.logger import init_logger
 from vllm.utils import (async_tensor_h2d, get_kv_cache_torch_dtype,
                         make_tensor_with_pad)
 
+logger = init_logger(__name__)
+
 if TYPE_CHECKING:
     from vllm.worker.model_runner import (ModelInputForGPUBuilder,
                                           ModelInputForGPUWithSamplingMetadata)
@@ -907,7 +910,12 @@ def __init__(
         blocksparse_params: Optional[Dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
+        use_irope: bool = False,
     ) -> None:
+        if use_irope:
+            logger.warning_once(
+                "Using irope in FlashInfer is not supported yet, it will fall"
+                " back to global attention for long context.")
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)

vllm/attention/backends/hpu_attn.py

@@ -108,8 +108,13 @@ def __init__(
         blocksparse_params: Optional[Dict[str, Any]] = None,
         max_seq_len: int = 4096,
         attn_type: str = AttentionType.DECODER,
+        use_irope: bool = False,
     ) -> None:
         super(AttentionImpl, self).__init__()
+        if use_irope:
+            logger.warning_once(
+                "Using irope in HPU is not supported yet, it will fall back "
+                "to global attention for long context.")
         self.kv_cache_dtype = kv_cache_dtype
         self.num_heads = num_heads
         self.head_size = head_size

vllm/attention/backends/ipex_attn.py

@@ -14,6 +14,9 @@
 from vllm.attention.backends.utils import CommonAttentionState
 from vllm.attention.ops.paged_attn import (PagedAttention,
                                            PagedAttentionMetadata)
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
 
 _PARTITION_SIZE = 512
 
@@ -119,7 +122,12 @@ def __init__(
         blocksparse_params: Optional[Dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
+        use_irope: bool = False,
     ) -> None:
+        if use_irope:
+            logger.warning_once(
+                "Using irope in Ipex is not supported yet, it will fall"
+                " back to global attention for long context.")
         if blocksparse_params is not None:
             raise ValueError(
                 "IPEX backend does not support block-sparse attention.")

vllm/attention/backends/pallas.py

@@ -11,6 +11,9 @@
                                               AttentionMetadata, AttentionType,
                                               is_quantized_kv_cache)
 from vllm.attention.backends.utils import CommonAttentionState
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
 
 
 class PallasAttentionBackend(AttentionBackend):
@@ -105,7 +108,12 @@ def __init__(
         blocksparse_params: Optional[Dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
+        use_irope: bool = False,
     ) -> None:
+        if use_irope:
+            logger.warning_once(
+                "Using irope in Pallas is not supported yet, it will fall back "
+                "to global attention for long context.")
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)

vllm/attention/backends/rocm_flash_attn.py

@@ -462,7 +462,12 @@ def __init__(
         blocksparse_params: Optional[Dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
+        use_irope: bool = False,
     ) -> None:
+        if use_irope:
+            logger.warning_once(
+                "Using irope in ROCm Flash Attention is not supported yet, it "
+                "will fail back to global attention for long context.")
         if blocksparse_params is not None:
             raise ValueError(
                 "ROCmFlashAttention does not support blocksparse attention.")

vllm/attention/backends/torch_sdpa.py

@@ -404,13 +404,18 @@ def __init__(
         blocksparse_params: Optional[Dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
+        use_irope: bool = False,
     ) -> None:
         if blocksparse_params is not None:
             raise ValueError(
                 "Torch SPDA does not support block-sparse attention.")
         if logits_soft_cap is not None:
             logger.warning_once("Torch SPDA does not support logits soft cap. "
                                 "Outputs may be slightly off.")
+        if use_irope:
+            logger.warning_once(
+                "Using irope in Torch SPDA is not supported yet, it will fall"
+                " back to global attention for long context.")
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)

vllm/attention/backends/xformers.py

@@ -389,13 +389,18 @@ def __init__(
         blocksparse_params: Optional[Dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
+        use_irope: bool = False,
     ) -> None:
         if blocksparse_params is not None:
             raise ValueError(
                 "XFormers does not support block-sparse attention.")
         if logits_soft_cap is not None:
             logger.warning_once("XFormers does not support logits soft cap. "
                                 "Outputs may be slightly off.")
+        if use_irope:
+            logger.warning_once(
+                "Using irope in XFormers is not supported yet, it will fall"
+                " back to global attention for long context.")
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)

vllm/v1/attention/backends/pallas.py

@@ -10,6 +10,9 @@
 from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
                                               AttentionLayer, AttentionType)
 from vllm.attention.backends.utils import CommonAttentionState
+from vllm.logger import init_logger
+
+logger = init_logger(__name__)
 
 
 class PallasAttentionBackend(AttentionBackend):
@@ -80,7 +83,12 @@ def __init__(
         blocksparse_params: Optional[dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
+        use_irope: bool = False,
     ) -> None:
+        if use_irope:
+            logger.warning_once(
+                "Using irope in Pallas is not supported yet, it will fall back "
+                "to global attention for long context.")
         if blocksparse_params is not None:
             raise ValueError("Paged attention Pallas kernel does "
                              "not support block-sparse attention.")