Migrate multi_head_jagged_flash_attention SLL ops to OSS

Summary: - Migrate `multi_head_jagged_flash_attention` SLL ops to OSS

Differential Revision: D66972360

.github/scripts/fbgemm_gpu_test.bash

@@ -89,6 +89,7 @@ __configure_fbgemm_gpu_test_cpu () {
     ./sll/jagged_flash_attention_basic_test.py
     ./sll/jagged_jagged_bmm_jagged_out_test.py
     ./sll/jagged_dense_flash_attention_test.py
+    ./sll/multi_head_jagged_flash_attention_test.py
   )
 }
 

fbgemm_gpu/fbgemm_gpu/sll/__init__.py

@@ -48,6 +48,7 @@
     jagged_jagged_bmm,
     jagged_jagged_bmm_jagged_out,
     jagged_softmax,
+    multi_head_jagged_flash_attention,
     triton_jagged_self_substraction_jagged_out,
 )
 
@@ -263,6 +264,19 @@ def register_sll_op(op_name: str, functors: Dict[str, Callable]) -> None:
         """
     )
 
+if "fbgemm::sll_multi_head_jagged_flash_attention" not in torch.library._defs:
+    lib.define(
+        """sll_multi_head_jagged_flash_attention(
+            Tensor q_weights,
+            Tensor k_weights,
+            Tensor v_weights,
+            Tensor offsets,
+            int max_seq_len,
+            bool allow_tf32=True
+        ) -> Tensor
+        """
+    )
+
 # NOTE: here we register the op for AutogradCUDA/CPU and CUDA/CPU with the same function
 # however, this is not ideal because in the inference case, we don't need the autograd forward
 # to save the context because we don't need to do backward.
@@ -396,3 +410,11 @@ def register_sll_op(op_name: str, functors: Dict[str, Callable]) -> None:
         "AutogradCPU": cpu_jagged_dense_flash_attention,
     },
 )
+
+register_sll_op(
+    "sll_multi_head_jagged_flash_attention",
+    {
+        "CUDA": multi_head_jagged_flash_attention,
+        "AutogradCUDA": multi_head_jagged_flash_attention,
+    },
+)