[Attention] Clean up iRoPE in V1 (#21188)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>

vllm/v1/attention/backends/cpu_attn.py

@@ -446,17 +446,12 @@ class TorchSDPABackendImpl(AttentionImpl[TorchSDPAMetadata]):
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
         kv_sharing_target_layer_name: Optional[str] = None,
-        use_irope: bool = False,
     ) -> None:
         if kv_sharing_target_layer_name is not None:
             raise NotImplementedError("KV sharing is not supported in V0.")
         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.paged_attn_impl = _get_paged_attn_impl()
         self.num_heads = num_heads
         self.head_size = head_size

vllm/v1/attention/backends/flash_attn.py

@@ -352,7 +352,6 @@ class FlashAttentionImpl(AttentionImpl):
         logits_soft_cap: Optional[float] = None,
         attn_type: AttentionType = AttentionType.DECODER,
         kv_sharing_target_layer_name: Optional[str] = None,
-        use_irope: bool = False,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
@@ -381,7 +380,6 @@ class FlashAttentionImpl(AttentionImpl):
                                       "encoder/decoder cross-attention "
                                       "are not implemented for "
                                       "FlashAttentionImpl")
-        self.use_irope = use_irope
         self.vllm_flash_attn_version = get_flash_attn_version()
         if is_quantized_kv_cache(self.kv_cache_dtype) \
             and not flash_attn_supports_fp8():

vllm/v1/attention/backends/flashinfer.py

@@ -493,7 +493,6 @@ class FlashInferImpl(AttentionImpl):
         logits_soft_cap: Optional[float] = None,
         attn_type: AttentionType = AttentionType.DECODER,
         kv_sharing_target_layer_name: Optional[int] = None,
-        use_irope: bool = False,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
@@ -509,7 +508,6 @@ class FlashInferImpl(AttentionImpl):
         self.kv_cache_dtype = kv_cache_dtype
         self.logits_soft_cap = logits_soft_cap
         self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
-        self.use_irope = use_irope
 
         self.num_queries_per_kv = self.num_heads // self.num_kv_heads
 

vllm/v1/attention/backends/pallas.py

@@ -148,12 +148,7 @@ class PallasAttentionBackendImpl(AttentionImpl):
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
         kv_sharing_target_layer_name: Optional[int] = None,
-        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/v1/attention/backends/rocm_aiter_fa.py

@@ -337,7 +337,6 @@ class AiterFlashAttentionImpl(AttentionImpl):
         logits_soft_cap: Optional[float] = None,
         attn_type: AttentionType = AttentionType.DECODER,
         kv_sharing_target_layer_name: Optional[int] = None,
-        use_irope: bool = False,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
@@ -367,7 +366,6 @@ class AiterFlashAttentionImpl(AttentionImpl):
                                       "encoder/decoder cross-attention "
                                       "are not implemented for "
                                       "FlashAttentionImpl")
-        self.use_irope = use_irope
         if is_quantized_kv_cache(self.kv_cache_dtype):
             raise NotImplementedError(
                 "AiterFlashAttention does not support fp8 kv-cache on this "

vllm/v1/attention/backends/triton_attn.py

@@ -72,9 +72,6 @@ class TritonAttentionMetadataBuilder(
             vllm_config.parallel_config)
         self.headdim = model_config.get_head_size()
 
-        self.attention_chunk_size = getattr(vllm_config.scheduler_config,
-                                            'attention_chunk_size', None)
-
     def build_for_cudagraph_capture(
         self, common_attn_metadata: CommonAttentionMetadata
     ) -> TritonAttentionMetadata:
@@ -208,7 +205,6 @@ class TritonAttentionImpl(AttentionImpl):
         logits_soft_cap: Optional[float] = None,
         attn_type: AttentionType = AttentionType.DECODER,
         kv_sharing_target_layer_name: Optional[int] = None,
-        use_irope: bool = False,
     ) -> None:
         self.num_heads = num_heads
         self.head_size = head_size
@@ -228,8 +224,6 @@ class TritonAttentionImpl(AttentionImpl):
         self.logits_soft_cap = logits_soft_cap
         self.kv_sharing_target_layer_name = kv_sharing_target_layer_name
 
-        self.use_irope = use_irope
-
         self.num_queries_per_kv = self.num_heads // self.num_kv_heads
 
         TritonAttentionBackend.validate_head_size(head_size)