[Perf][Attention] Avoid redundant sparse attention metadata operations (#57885)

Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
Co-authored-by: OpenAI Codex <noreply@openai.com>

vllm/v1/attention/backends/mla/indexer.py

@@ -1261,7 +1261,6 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
         assert num_decode_tokens + num_prefill_tokens == num_tokens
 
         compressed_slot_mapping = slot_mapping
-        compressed_seq_lens = seq_lens
         indexer_block_table = block_table
         if self.compress_ratio > 1:
             kernel_block_size = self.kernel_block_size
@@ -1297,10 +1296,12 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
                     self.compressed_slot_mapping_buffer[:padded_num_tokens],
                     dim=0,
                 )
-            compressed_seq_lens = seq_lens // self.compress_ratio
 
         prefill_metadata = None
         if num_prefills > 0:
+            compressed_seq_lens = (
+                seq_lens // self.compress_ratio if self.compress_ratio > 1 else seq_lens
+            )
             # This CPU value is an upper bound for async-spec extend rows.  It
             # is safe for chunking/allocation because CUDA metadata below is
             # built from exact device seq_lens and gather ignores the tail.

vllm/v1/attention/backends/mla/sparse_swa.py

@@ -613,7 +613,7 @@ class DeepseekSparseSWAMetadataBuilder(AttentionMetadataBuilder):
         )
 
         is_valid_token = self.is_valid_token[: slot_mapping.shape[0]]
-        is_valid_token.copy_(slot_mapping >= 0)
+        torch.ge(slot_mapping, 0, out=is_valid_token)
         if replay_start is None:
             # Graph captures build without it; their dummy batches replay nothing.
             replay_start = self.no_replay_start