Merge branch 'v0.15.1-fix_custom_op' into 'v0.15.1-dev'

修改sparse_attn hip后端

See merge request dcutoolkit/deeplearing/vllm!498

vllm/model_executor/custom_op.py

@@ -184,7 +184,7 @@ class CustomOp(nn.Module):
             return self.maybe_compile(self.forward_native, enable=compile_native)
 
         if current_platform.is_rocm():
-            return self.forward_cuda
+            return self.forward_hip
         elif current_platform.is_cpu():
             return self.forward_cpu
         elif current_platform.is_tpu():

vllm/model_executor/layers/sparse_attn_indexer.py

@@ -296,8 +296,7 @@ class SparseAttnIndexer(CustomOp):
         if current_platform.is_cuda():
             return self.forward_cuda(hidden_states, q_fp8, k, weights)
         elif current_platform.is_rocm():
-            # return self.forward_hip(hidden_states, q_fp8, k, weights)
-            return self.forward_cuda(hidden_states, q_fp8, k, weights)
+            return self.forward_hip(hidden_states, q_fp8, k, weights)
         else:
             raise NotImplementedError(
                 "SparseAttnIndexer native forward is only implemented for "
@@ -349,9 +348,22 @@ class SparseAttnIndexer(CustomOp):
                 self.max_model_len,
                 self.max_total_seq_len,
                 self.topk_indices_buffer,
-            )
+            )        
         else:
-            raise RuntimeError(
-                "Sparse attention indexer ROCm custom op requires ROCm "
-                "Aiter ops to be enabled."
+            return torch.ops.vllm.sparse_attn_indexer(
+                hidden_states,
+                self.k_cache.prefix,
+                self.k_cache.kv_cache[0],
+                q_fp8,
+                k,
+                weights,
+                self.quant_block_size,
+                self.scale_fmt,
+                self.topk_tokens,
+                self.head_dim,
+                self.max_model_len,
+                self.max_total_seq_len,
+                self.topk_indices_buffer,
             )
+        
+