[Hardware][AMD] integrate aiter chunked prefill into vllm

vllm/envs.py

@@ -87,6 +87,7 @@
     VLLM_ROCM_USE_AITER_MOE: bool = True
     VLLM_ROCM_USE_AITER_RMSNORM: bool = True
     VLLM_ROCM_USE_AITER_MLA: bool = True
+    VLLM_ROCM_USE_AITER_MHA: bool = True
     VLLM_ROCM_USE_SKINNY_GEMM: bool = True
     VLLM_ROCM_FP8_PADDING: bool = True
     VLLM_ROCM_MOE_PADDING: bool = True
@@ -653,6 +654,13 @@ def get_vllm_port() -> Optional[int]:
     "VLLM_ROCM_USE_AITER_MLA":
     lambda: (os.getenv("VLLM_ROCM_USE_AITER_MLA", "True").lower() in
              ("true", "1")),
+
+    # Whether to use aiter mha ops.
+    # By default is enabled.
+    "VLLM_ROCM_USE_AITER_MHA":
+    lambda: (os.getenv("VLLM_ROCM_USE_AITER_MHA", "True").lower() in
+             ("true", "1")),
+
     # use rocm skinny gemms
     "VLLM_ROCM_USE_SKINNY_GEMM":
     lambda: (os.getenv("VLLM_ROCM_USE_SKINNY_GEMM", "True").lower() in

vllm/platforms/rocm.py

@@ -215,9 +215,15 @@ def get_attn_backend_cls(cls, selected_backend, head_size, dtype,
             selected_backend = _Backend.ROCM_FLASH
 
         if envs.VLLM_USE_V1:
-            logger.info("Using Triton Attention backend on V1 engine.")
-            return ("vllm.v1.attention.backends."
-                    "triton_attn.TritonAttentionBackend")
+            if envs.VLLM_ROCM_USE_AITER and envs.VLLM_ROCM_USE_AITER_MHA \
+                and on_gfx9():
+                logger.info("Using Flash Attention backend on V1 engine.")
+                return ("vllm.v1.attention.backends."
+                        "rocm_aiter_fa.AiterFlashAttentionBackend")
+            else:
+                logger.info("Using Triton Attention backend on V1 engine.")
+                return ("vllm.v1.attention.backends."
+                        "triton_attn.TritonAttentionBackend")
         if selected_backend == _Backend.ROCM_FLASH:
             if not cls.has_device_capability(90):
                 # not Instinct series GPUs.