[Fix] Update FlashAttention import handling for compatibility with newer releases

xtuner/v1/ops/flash_attn/__init__.py

@@ -19,7 +19,14 @@ def get_flash_attn_varlen() -> FlashAttnVarlenProtocol:
 
         if os.environ.get("XTUNER_USE_FA3", "0") == "1":
             try:
-                from flash_attn_interface import flash_attn_3_cuda
+                # flash_attn_interface renamed its v3 entrypoint in newer releases:
+                # - old: flash_attn_3_cuda
+                # - new: flash_attn_3_gpu
+                # We only import it here as an availability check; the actual impl lives in `.gpu`.
+                try:
+                    from flash_attn_interface import flash_attn_3_gpu  # noqa: F401
+                except ImportError:
+                    from flash_attn_interface import flash_attn_3_cuda  # noqa: F401
             except ImportError as e:
                 raise ImportError(f"Import FlashAttention 3 failed {e}, Please install it manually.")
             from .gpu import gpu_flash_varlen_attn_v3 as flash_attn_varlen_func

xtuner/v1/ops/flash_attn/gpu.py

@@ -5,7 +5,14 @@
 
 
 try:
-    from flash_attn_interface import flash_attn_3_cuda
+    # flash_attn_interface renamed its v3 entrypoint in newer releases:
+    # - old: flash_attn_3_cuda
+    # - new: flash_attn_3_gpu
+    # Keep the rest of this file stable by aliasing whichever exists to `flash_attn_3_cuda`.
+    try:
+        from flash_attn_interface import flash_attn_3_gpu as flash_attn_3_cuda
+    except ImportError:
+        from flash_attn_interface import flash_attn_3_cuda as flash_attn_3_cuda
     from flash_attn_interface import maybe_contiguous as maybe_contiguous_v3
 
     @torch.library.custom_op("flash_attn::_flash_attn_varlen_forward_v3", mutates_args=(), device_types="cuda")