review comments

Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>

vllm/attention/backends/abstract.py

@@ -276,3 +276,19 @@ class AttentionImpl(ABC, Generic[T]):
         output: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
         raise NotImplementedError
+
+
+class MLAAttentionImpl(AttentionImpl):
+
+    @abstractmethod
+    def forward(
+        self,
+        layer: AttentionLayer,
+        hidden_states_or_cq: torch.Tensor,
+        kv_c_normed: torch.Tensor,
+        k_pe: torch.Tensor,
+        kv_cache: torch.Tensor,
+        attn_metadata: T,
+        output: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        raise NotImplementedError

vllm/attention/backends/mla/utils.py

@@ -6,8 +6,9 @@ import torch
 
 from vllm import _custom_ops as ops
 from vllm import envs
-from vllm.attention.backends.abstract import (AttentionImpl, AttentionLayer,
-                                              AttentionMetadata)
+from vllm.attention.backends.abstract import (AttentionLayer,
+                                              AttentionMetadata,
+                                              MLAAttentionImpl)
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                RowParallelLinear)
@@ -18,11 +19,11 @@ from vllm.vllm_flash_attn import flash_attn_varlen_func
 @dataclass(kw_only=True)
 class MLAMetadataCommon(AttentionMetadata):
     # Input positions for rotrary embeddings since for MLA the rotarty
-    # position encoding
+    # position embeddings are applied inside the attention backend
     input_positions: torch.Tensor
 
 
-class MLAImplCommon(AttentionImpl):
+class MLACommonImpl(MLAAttentionImpl):
     """
     Common class for implementing repeated parts 
     

vllm/attention/backends/triton_mla.py

@@ -20,7 +20,7 @@ from vllm.attention.backends.abstract import (AttentionBackend,
                                               AttentionMetadata,
                                               AttentionMetadataBuilder,
                                               AttentionState, AttentionType)
-from vllm.attention.backends.mla.utils import MLAImplCommon, MLAMetadataCommon
+from vllm.attention.backends.mla.utils import MLACommonImpl, MLAMetadataCommon
 from vllm.attention.backends.utils import (PAD_SLOT_ID, compute_slot_mapping,
                                            compute_slot_mapping_start_idx,
                                            is_block_tables_empty)
@@ -585,7 +585,7 @@ class TritonMLAMetadataBuilder(AttentionMetadataBuilder[TritonMLAMetadata]):
         )
 
 
-class TritonMLAImpl(MLAImplCommon):
+class TritonMLAImpl(MLACommonImpl):
 
     def __init__(
             self,

vllm/model_executor/models/deepseek_v2.py

@@ -331,7 +331,7 @@ class DeepseekV2MLAAttention(nn.Module):
     Main reference: DeepseekV2 paper, and FlashInfer Implementation
     (https://arxiv.org/abs/2405.04434 and https://github.com/flashinfer-ai/flashinfer/pull/551).
     
-    For more info see MLAImplCommon in: vllm/attention/backends/mla/utils.py
+    For more info see MLACommonImpl in: vllm/attention/backends/mla/utils.py
     """
 
     def __init__(