Implement FA4 varlen based on Paddle.

Author: baoqiwen

flashmask/flash_mask/__init__.py

@@ -69,6 +69,15 @@
 except ImportError:
     pass  # cute module not installed or dependencies missing
 
+# ============================================================
+# FA4 varlen / standard interface (framework-routed)
+# ============================================================
+try:
+    from .interface import flash_attn_func, flash_attn_varlen_func, flash_attn_combine
+    __all__ += ["flash_attn_func", "flash_attn_varlen_func", "flash_attn_combine"]
+except ImportError:
+    pass
+
 if not _fa3_available and not _fa4_available:
     print("[WARNING] flash_mask: neither FA3 nor FA4 is available. "
           "Check your installation.")

flashmask/flash_mask/flash_attn_v4/__init__.py

@@ -9,12 +9,21 @@
 
 import cutlass.cute as cute
 
-from .interface import (
-    flash_attn_func,
-    flash_attn_varlen_func,
-)
-
-from flash_mask.flash_attn_v4.cute_dsl_utils import cute_compile_patched
+# Auto-detect framework: prefer torch, fall back to paddle.
+try:
+    import torch  # noqa: F401
+
+    from flash_mask.flash_attn_v4.torch.interface import (
+        flash_attn_func,
+        flash_attn_varlen_func,
+    )
+    from flash_mask.flash_attn_v4.torch.cute_dsl_utils import cute_compile_patched
+except ImportError:
+    from flash_mask.flash_attn_v4.paddle.interface import (
+        flash_attn_func,
+        flash_attn_varlen_func,
+    )
+    from flash_mask.flash_attn_v4.paddle.cute_dsl_utils import cute_compile_patched
 
 # Patch cute.compile to optionally dump SASS
 cute.compile = cute_compile_patched