[ROCm] Split AITER unified attention into its own backend (#25507)

Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
2025-10-06 18:49:23 -04:00
parent 2161efe978
commit f231e5bc21
8 changed files with 325 additions and 301 deletions
--- a/vllm/attention/selector.py
+++ b/vllm/attention/selector.py
@@ -254,3 +254,4 @@ def global_force_attn_backend_context_manager(
    finally:
        # Revert the original global backend override, if any
        global_force_attn_backend(original_value)
+        _cached_get_attn_backend.cache_clear()